Let the new tests run on a Windows box and on a small machine

Two ways the tests were standing on this machine rather than on the one
they meant to describe.

Windows has no os.sysconf at all, and mock.patch.object insists the
attribute exists before it will replace it, so four tests failed at the
patch rather than in the body. `create=True` is what lets them stand
somewhere that has no such function, which is the case the code under test
already handles two lines down and which is checked on its own.

And the publisher order follows the memory on purpose: a runner with 7 GB
in it puts the two Gemma 4 rows last and is right to. The three tests that
read an order now say which machine they are standing on instead of
assuming the one that ran them has room for everything.

.github/workflows/whisper-vulkan.yml

@@ -0,0 +1,222 @@
+name: whisper.cpp Vulkan bundle
+
+# Only what the bundle is built from. Compiling the Vulkan shaders takes
+# tens of minutes, and a README typo is not worth one: what ties ggml.py to
+# this release is a handful of assertions in tests/test_packaging.py, and
+# those run on every pull request in milliseconds.
+on:
+  pull_request:
+    paths:
+      - packaging/whisper-vulkan/**
+      - .github/workflows/whisper-vulkan.yml
+  workflow_dispatch:
+    inputs:
+      whisper_version:
+        description: Upstream whisper.cpp version (without v)
+        required: true
+        default: "1.9.3"
+        type: string
+      whisper_commit:
+        description: Peeled commit SHA for that reviewed upstream tag
+        required: true
+        default: "371b5a7561823ab2bb32142d2751e35e7534727b"
+        type: string
+      expected_sha256:
+        description: >-
+          Reviewed archive SHA-256. Leave empty for a version this file has
+          not reviewed: the digest of what was built is reported instead of
+          being checked, and publishing is refused.
+        required: false
+        default: ""
+        type: string
+      publish:
+        description: Publish a Dikte dependency release
+        required: true
+        default: false
+        type: boolean
+
+permissions:
+  contents: read
+
+concurrency:
+  group: whisper-vulkan-${{ github.event.pull_request.number || github.ref }}
+  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+
+env:
+  WHISPER_VERSION: ${{ inputs.whisper_version || '1.9.3' }}
+  WHISPER_COMMIT: ${{ inputs.whisper_commit || '371b5a7561823ab2bb32142d2751e35e7534727b' }}
+  REVIEWED_WHISPER_VERSION: "1.9.3"
+  REVIEWED_WHISPER_SHA256: c25ca76504144da488eb74441390a7b9aa7ce547e5f2f391cbd831253c9b54d8
+
+jobs:
+  build:
+    runs-on: ubuntu-22.04
+    timeout-minutes: 45
+    steps:
+      - name: Check out Dikte
+        uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4
+        with:
+          persist-credentials: false
+
+      - name: Validate source coordinates
+        shell: bash
+        run: |
+          set -euo pipefail
+          [[ "$WHISPER_VERSION" =~ ^[0-9]+\.[0-9]+\.[0-9]+$ ]]
+          [[ "$WHISPER_COMMIT" =~ ^[0-9a-f]{40}$ ]]
+
+      - name: Check out pinned whisper.cpp source
+        uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 # v4
+        with:
+          repository: ggml-org/whisper.cpp
+          ref: ${{ env.WHISPER_COMMIT }}
+          path: vendor/whisper.cpp
+          fetch-depth: 0
+          persist-credentials: false
+
+      - name: Verify source version and commit
+        shell: bash
+        run: |
+          set -euo pipefail
+          test "$(git -C vendor/whisper.cpp rev-parse HEAD)" = "$WHISPER_COMMIT"
+          git -C vendor/whisper.cpp fetch --depth=1 origin \
+            "refs/tags/v$WHISPER_VERSION:refs/tags/v$WHISPER_VERSION"
+          test "$(git -C vendor/whisper.cpp rev-list -n1 "v$WHISPER_VERSION")" = "$WHISPER_COMMIT"
+          echo "SOURCE_DATE_EPOCH=$(git -C vendor/whisper.cpp show -s --format=%ct HEAD)" >> "$GITHUB_ENV"
+
+      - name: Build pinned build environment
+        run: docker build --pull=false -f packaging/whisper-vulkan/Dockerfile.build -t dikte-whisper-builder packaging/whisper-vulkan
+
+      - name: Build deterministic archive
+        run: |
+          docker run --rm \
+            -e WHISPER_VERSION -e WHISPER_COMMIT -e SOURCE_DATE_EPOCH \
+            -v "$PWD/vendor/whisper.cpp:/src:ro" \
+            -v "$PWD/packaging/whisper-vulkan:/packaging:ro" \
+            -v "$PWD/work:/work" \
+            dikte-whisper-builder \
+            bash /packaging/build-package.sh
+          mkdir -p dist
+          cp work/out/whisper-bin-ubuntu-vulkan-x64.* dist/
+
+      - name: Verify reviewed archive digest
+        shell: bash
+        env:
+          EXPECTED_SHA256: ${{ inputs.expected_sha256 }}
+          PUBLISH: ${{ inputs.publish }}
+        run: |
+          set -euo pipefail
+          read -r actual _ < dist/whisper-bin-ubuntu-vulkan-x64.tar.gz.sha256
+          echo "built archive sha256: $actual"
+          expected="$EXPECTED_SHA256"
+          if [ -z "$expected" ] \
+              && [ "$WHISPER_VERSION" = "$REVIEWED_WHISPER_VERSION" ]; then
+            expected="$REVIEWED_WHISPER_SHA256"
+          fi
+          if [ -z "$expected" ]; then
+            # The digest of a version nobody has reviewed yet cannot be known
+            # before it is built. Reporting it is the whole point of the run;
+            # a release out of it is not.
+            if [ "${PUBLISH:-false}" = true ]; then
+              echo "refusing to publish an archive whose digest has not been reviewed" >&2
+              exit 1
+            fi
+            echo "::notice::no reviewed digest for $WHISPER_VERSION." \
+                 "Review the one above, then dispatch again with expected_sha256."
+            exit 0
+          fi
+          test "$actual" = "$expected"
+
+      - name: Validate archive and ELF contract
+        run: OUT_DIR=dist packaging/whisper-vulkan/validate-package.sh
+
+      - name: Schema-validate CycloneDX 1.6 SBOM
+        run: |
+          docker run --rm \
+            -v "$PWD/dist/whisper-bin-ubuntu-vulkan-x64.cdx.json:/sbom.json:ro" \
+            cyclonedx/cyclonedx-cli@sha256:252c2e26f468c25fea1e63ecde1bc3198ad6e9dbb57f5ed3236bddcb2281b3a7 \
+            validate --input-file /sbom.json --input-format json \
+            --input-version v1_6 --fail-on-errors
+
+      - name: CPU fallback smoke test (no Vulkan loader)
+        run: OUT_DIR=dist packaging/whisper-vulkan/smoke-runtime.sh cpu
+
+      - name: Vulkan loader present, no device smoke test
+        run: OUT_DIR=dist packaging/whisper-vulkan/smoke-runtime.sh noicd
+
+      - name: Vulkan plugin-load smoke test (Mesa llvmpipe)
+        run: OUT_DIR=dist packaging/whisper-vulkan/smoke-runtime.sh vulkan
+
+      - name: Upload reviewed outputs
+        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4
+        with:
+          name: whisper-bin-ubuntu-vulkan-x64
+          path: dist/*
+          if-no-files-found: error
+          retention-days: 14
+
+  publish:
+    if: >-
+      github.event_name == 'workflow_dispatch' && inputs.publish &&
+      github.ref == 'refs/heads/master'
+    needs: build
+    runs-on: ubuntu-22.04
+    environment: dependency-release
+    permissions:
+      contents: write
+      id-token: write
+      attestations: write
+      artifact-metadata: write
+    steps:
+      - name: Download the exact tested outputs
+        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
+        with:
+          name: whisper-bin-ubuntu-vulkan-x64
+          path: dist
+
+      - name: Verify digest sidecar
+        run: (cd dist && sha256sum --check whisper-bin-ubuntu-vulkan-x64.tar.gz.sha256)
+
+      - name: Attest build provenance
+        uses: actions/attest@1e69f48acb82d1966a394da916b4c1698aa569d6 # v4
+        with:
+          subject-path: dist/whisper-bin-ubuntu-vulkan-x64.tar.gz
+
+      - name: Attest SBOM to archive
+        uses: actions/attest@1e69f48acb82d1966a394da916b4c1698aa569d6 # v4
+        with:
+          subject-path: dist/whisper-bin-ubuntu-vulkan-x64.tar.gz
+          sbom-path: dist/whisper-bin-ubuntu-vulkan-x64.cdx.json
+
+      - name: Publish dependency release
+        env:
+          GH_TOKEN: ${{ github.token }}
+          RELEASE_TAG: whisper.cpp-v${{ inputs.whisper_version }}
+          RELEASE_TITLE: whisper.cpp v${{ inputs.whisper_version }} Vulkan bundle
+          RELEASE_NOTES: >-
+            Pinned source: ggml-org/whisper.cpp@${{ inputs.whisper_commit }}.
+            Verify with: gh attestation verify
+            whisper-bin-ubuntu-vulkan-x64.tar.gz
+            --repo ${{ github.repository }}
+        run: |
+          set -euo pipefail
+          if gh release view "$RELEASE_TAG" --repo "$GITHUB_REPOSITORY" >/dev/null 2>&1; then
+            echo "refusing to replace existing release $RELEASE_TAG" >&2
+            exit 1
+          fi
+          if gh api "repos/$GITHUB_REPOSITORY/git/ref/tags/$RELEASE_TAG" >/dev/null 2>&1; then
+            echo "refusing to replace existing tag $RELEASE_TAG" >&2
+            exit 1
+          fi
+          gh api --method POST "repos/$GITHUB_REPOSITORY/git/refs" \
+            -f ref="refs/tags/$RELEASE_TAG" \
+            -f sha="$GITHUB_SHA" >/dev/null
+          test "$(gh api "repos/$GITHUB_REPOSITORY/git/ref/tags/$RELEASE_TAG" \
+            --jq .object.sha)" = "$GITHUB_SHA"
+          gh release create "$RELEASE_TAG" dist/* \
+            --repo "$GITHUB_REPOSITORY" \
+            --verify-tag \
+            --prerelease \
+            --latest=false \
+            --title "$RELEASE_TITLE" \
+            --notes "$RELEASE_NOTES"

README.md

@@ -115,9 +115,13 @@ or runs it on the spot.
 Speech to text and cleanup each pick a provider in the settings window, and both
 run here by default, on models of your own. The cloud is the other option:
 speech to text on **OpenAI**, **Groq** or **OpenRouter** (`gpt-4o-transcribe`),
-cleanup on OpenRouter (`google/gemini-3.5-flash-lite`) or, when either is
-installed, on Claude Code or Codex. The keys fall back to `OPENAI_API_KEY`,
-`GROQ_API_KEY` and `OPENROUTER_API_KEY`, and are stored in
+cleanup on OpenRouter (`google/gemini-3.5-flash-lite`), on **Google AI Studio**
+(`gemini-3.5-flash-lite`), on **OpenCode Go** (`deepseek-v4-flash`) or, when one
+of them is installed, on Claude Code, Codex or Antigravity. The first three are
+a single HTTP request; the three CLIs each open a whole session to do it, which
+is where their few extra seconds go. The keys fall back to `OPENAI_API_KEY`,
+`GROQ_API_KEY`, `OPENROUTER_API_KEY`, `GEMINI_API_KEY` and `OPENCODE_API_KEY`,
+and are stored in
 `~/.config/dikte/config.json`, mode 600, or in
 `~/Library/Application Support/Dikte` on a Mac. Cleanup can be switched off, in
 which case the raw transcript is pasted, and a thinking model's effort can be
@@ -139,7 +143,8 @@ set next to it.
 
 An indicator in the screen corner shows a red dot, a live waveform and the
 elapsed time, then the stage it is on. It never takes focus. Pressing
-`Ctrl+Space` again while Dikte is still working does nothing; nothing queues up.
+`Ctrl+Space` again while the last dictation is still being cleaned up starts
+the next one; it is transcribed and pasted in turn, behind the one still going.
 A dictation and a command to the agent do wait on each other for the microphone,
 which is one device, but for nothing else: each has its own indicator, and the
 second one stacks above the first while both are up.
@@ -156,9 +161,13 @@ running.
 - **It all runs on this machine by default.** Speech to text on whisper.cpp and
   cleanup on llama.cpp, neither installed beforehand: the settings window fetches
   the program and the model, verifies the sha256 and refuses a download published
-  without one, then keeps a server alive while you dictate. The graphics card is
+  without one, then keeps a server alive while you dictate. The model list is
+  grouped by model rather than by file size, and the row this machine's memory
+  and graphics can take is marked. The graphics card is
   reached through CUDA, ROCm or Vulkan where the build allows. No key, no
-  account, nothing leaving the machine.
+  account, nothing leaving the machine. On x86_64 Linux the same button fetches
+  a Vulkan build of whisper-server that Dikte publishes itself, because
+  upstream's Linux archive is processor-only.
 - **Silence never reaches the API.** Handed near-silence, a transcription model
   invents a sentence instead of returning nothing ("Thanks for watching", or in
   Turkish "Altyazı M.K."). A recording is dropped when nothing rose 10 dB above
@@ -188,11 +197,13 @@ running.
   what comes of it: the answer, or a sentence saying what was done. It is the
   session you would have opened yourself, so your skills and connected services
   are there, which is what makes "put that in my calendar on Thursday at three"
-  a thing you can say to a window that is not Claude. Codex (`codex exec`) runs
-  the same way, and OpenRouter is there as a plain question-and-answer fallback
-  for a machine with neither CLI on it. Provider, model, permissions and working
-  directory are under Settings → Agent, and commands close together stay in one
-  conversation.
+  a thing you can say to a window that is not Claude. Codex (`codex exec`) and
+  Antigravity (`agy -p`) run the same way, though Antigravity takes neither a
+  permission mode nor a sandbox from Dikte: what it may do without asking is
+  whatever its own allow-rules say. OpenRouter or OpenCode Go is there as a
+  plain question-and-answer fallback for a machine with no CLI on it. Provider,
+  model, permissions and working directory are under Settings → Agent, and
+  commands close together stay in one conversation.
 - **Meetings** are recorded from the microphone and the speaker output at the
   same time, which settles who said what by the channel a voice arrived on
   instead of guessing at it. The two sides are transcribed separately and
@@ -239,9 +250,9 @@ cli.py            the command line: every verb, and what it answers with
 ipc.py            one request and one reply over the local socket
 audio.py          PCM capture: pw-record for dictation, ffmpeg for a meeting
 meeting.py        channel split, speaker labelling, cleanup, minutes
-assistant.py      running a dictation through Claude Code, Codex or OpenRouter
+assistant.py      handing a dictation to Claude Code, Codex, agy or a chat model
 api.py            transcription and cleanup requests (stdlib only)
-cleanup.py        who rewrites the transcript: OpenRouter, here, Claude or Codex
+cleanup.py        who rewrites the transcript: a hosted model, one here, a CLI
 ggml.py           whisper.cpp and llama.cpp here: fetch, verify, keep serving
 hub.py            what GitHub and Hugging Face have on offer today
 update.py         whether a newer release is out, and the page it is on

README.tr.md

@@ -112,10 +112,14 @@ Sesi yazıya çevirme ve temizleme, ayarlar penceresinde ayrı ayrı sağlayıc
 seçer; ikisi de varsayılan olarak burada, kendi modellerinle çalışır. Bulutu
 seçersen sesi yazıya çevirme **OpenAI**, **Groq** ya da **OpenRouter**'da
 (varsayılan `gpt-4o-transcribe`), temizleme OpenRouter'da
-(`google/gemini-3.5-flash-lite`) ya da kuruluysa Claude Code veya Codex'te
-çalışır. Anahtarları boş bırakırsan `OPENAI_API_KEY`, `GROQ_API_KEY` ve
-`OPENROUTER_API_KEY` kullanılır; anahtarlar `~/.config/dikte/config.json`
-içinde, izinler 600, Mac'te ise `~/Library/Application Support/Dikte` altında.
+(`google/gemini-3.5-flash-lite`), **Google AI Studio**'da
+(`gemini-3.5-flash-lite`), **OpenCode Go**'da (`deepseek-v4-flash`) ya da
+kuruluysa Claude Code, Codex veya Antigravity'de çalışır. İlk üçü tek bir HTTP
+isteği; üç CLI ise bunun için birer oturum açar, fazladan giden birkaç saniye de
+oradan gelir. Anahtarları boş bırakırsan `OPENAI_API_KEY`, `GROQ_API_KEY`,
+`OPENROUTER_API_KEY`, `GEMINI_API_KEY` ve `OPENCODE_API_KEY` kullanılır;
+anahtarlar `~/.config/dikte/config.json` içinde, izinler 600, Mac'te ise
+`~/Library/Application Support/Dikte` altında.
 Temizlemeyi tamamen kapatabilirsin, o zaman ham transkript yapıştırılır; modelin
 yanındaki kutudan düşünme seviyesini de seçebilirsin.
 
@@ -134,8 +138,9 @@ yanındaki kutudan düşünme seviyesini de seçebilirsin.
 | Çık | Tepsi menüsü → *Çık*, ya da `dikte quit` |
 
 Ekranın köşesindeki gösterge kırmızı kayıt noktasını, canlı ses dalgasını ve
-süreyi, ardından hangi aşamada olduğunu gösterir. Odak almaz. Dikte çalışırken
-`Ctrl+Space`'e tekrar basmak bir şey yapmaz, sıraya da girmez. Dikte ile ajana
+süreyi, ardından hangi aşamada olduğunu gösterir. Odak almaz. Önceki dikte daha
+temizlenirken `Ctrl+Space`'e tekrar basmak yenisini başlatır; o da sırası
+gelince, öndekinin ardından yazılıp yapıştırılır. Dikte ile ajana
 verilen komut yalnızca mikrofon için birbirini bekler, o da tek aygıt olduğu
 için; başka hiçbir şeyde beklemezler. Her birinin kendi göstergesi var, ikisi
 birden ekrandayken ikincisi birincinin üstüne yerleşir.
@@ -153,8 +158,12 @@ olmasını ister.
   whisper.cpp, temizleme llama.cpp üzerinde; ikisini de önceden kurman gerekmez:
   ayarlar penceresi programı ve modeli indirir, sha256'sını doğrular,
   checksum'suz yayınlanmış bir indirmeyi reddeder, sen dikte ettikçe sunucuyu
-  ayakta tutar. Derleme destekliyorsa ekran kartına CUDA, ROCm ya da Vulkan
+  ayakta tutar. Model listesi dosya boyutuna değil modele göre gruplanır ve bu
+  makinenin belleğine ve ekran kartına uyan satır işaretlenir. Derleme
+  destekliyorsa ekran kartına CUDA, ROCm ya da Vulkan
   üzerinden ulaşılır. Anahtar yok, hesap yok, makineden çıkan bir şey yok.
+  x86_64 Linux'ta aynı düğme, whisper-server'ın Dikte'nin kendi yayınladığı
+  Vulkan derlemesini indirir; upstream'in Linux arşivi yalnızca işlemci için.
 - **Sessizlik API'ye gitmez.** Sessize yakın bir ses verildiğinde model boş dize
   döndürmez, bir cümle uydurur ("Altyazı M.K.", "Thanks for watching"). *O
   kaydın kendi* gürültü tabanının 10 dB üstüne en az 0,3 saniye çıkan bir şey
@@ -183,9 +192,11 @@ olmasını ister.
   yapıştırır: cevabı ya da ne yapıldığını söyleyen bir cümle. Kendi açacağın
   oturumun aynısıdır, yani skill'lerin ve bağlı servislerin oradadır; "bunu
   perşembe üçe takvime koy" cümlesini Claude olmayan bir pencerede söyleyebilir
-  olmanı sağlayan da budur. Codex (`codex exec`) da aynı şekilde çalışır;
-  OpenRouter ise ikisi de kurulu olmayan bir makinede düz soru cevap için
-  duruyor. Sağlayıcı, model, izinler ve çalışma dizini Ayarlar → Ajan
+  olmanı sağlayan da budur. Codex (`codex exec`) ile Antigravity (`agy -p`) da
+  aynı şekilde çalışır; ama Antigravity'ye Dikte bir izin kipi ya da sandbox
+  veremiyor, sormadan ne yapabileceğini kendi allow-rule'ları belirliyor.
+  OpenRouter ya da OpenCode Go ise hiçbiri kurulu olmayan bir makinede düz soru
+  cevap için duruyor. Sağlayıcı, model, izinler ve çalışma dizini Ayarlar → Ajan
   sekmesinde; arka arkaya verilen komutlar tek bir konuşmada kalır.
 - **Toplantılar** mikrofonla hoparlör çıkışından aynı anda kaydedilir; kimin ne
   dediği tahmin edilmez, sesin hangi kanaldan geldiğiyle belli olur. İki taraf
@@ -232,9 +243,9 @@ cli.py            komut satırı: bütün fiiller ve verdikleri cevap
 ipc.py            yerel sokette bir istek, bir cevap
 audio.py          PCM kaydı: diktede pw-record, toplantıda ffmpeg
 meeting.py        kanal ayırma, konuşmacı etiketi, temizleme, tutanak
-assistant.py      dikteyi Claude Code, Codex ya da OpenRouter'dan geçirme
+assistant.py      dikteyi Claude Code, Codex, agy ya da sohbet modelinden geçirme
 api.py            transkript ve temizleme istekleri (yalnız stdlib)
-cleanup.py        transkripti kim temizler: OpenRouter, burası, Claude ya da Codex
+cleanup.py        transkripti kim temizler: bulutta bir model, burası, bir CLI
 ggml.py           whisper.cpp ve llama.cpp'yi indirip burada çalıştırma
 hub.py            GitHub ve Hugging Face'te bugün ne olduğu
 update.py         yeni sürüm çıkmış mı, çıkmışsa hangi sayfada

dikte/__init__.py

@@ -10,4 +10,4 @@ business loading Qt to answer one question.
 # both the .dmg's Info.plist and the AppImage's file name are built from it. A
 # build off master rather than off a tag appends the commit to it, so that a
 # bug report from someone running "latest" names a commit.
-__version__ = "1.0.2"
+__version__ = "1.1.0"

dikte/api.py

@@ -1,10 +1,17 @@
-"""OpenAI, Groq, OpenRouter and this machine, stdlib only.
+"""OpenAI, Groq, OpenRouter, Google AI Studio and this machine, stdlib only.
 
-Transcription runs on any of the four: Groq and OpenRouter both mirror OpenAI's
-/audio/transcriptions endpoint field for field, and ggml.py starts whisper.cpp
-on that same path, so one multipart request serves all of them and only the key,
-the base URL and the model id change. llama.cpp answers /chat/completions the way
-OpenRouter does, so cleanup here is the same request too.
+Transcription runs on the first three and on this machine: Groq and OpenRouter
+both mirror OpenAI's /audio/transcriptions endpoint field for field, and ggml.py
+starts whisper.cpp on that same path, so one multipart request serves all of
+them and only the key, the base URL and the model id change. llama.cpp answers
+/chat/completions the way OpenRouter does, so cleanup here is the same request
+too.
+
+Google AI Studio is here for cleanup and nothing else. Its OpenAI-compatible
+endpoint answers /chat/completions and /models, but there is no
+/audio/transcriptions behind it: audio only goes in as base64 inside a chat
+message, and what comes back has none of the segment times a subtitle file or a
+meeting transcript is built out of.
 
 What is on this machine has no key, and its base URL is not known until a server
 is up, which is the one thing this module has to fill in for it.
@@ -31,6 +38,7 @@ USER_AGENT = f"dikte/1.0 (+{APP_URL})"
 OPENAI_URL = "https://api.openai.com/v1"
 GROQ_URL = "https://api.groq.com/openai/v1"
 OPENROUTER_URL = "https://openrouter.ai/api/v1"
+GEMINI_URL = "https://generativelanguage.googleapis.com/v1beta/openai"
 
 # The floor for a local request. The timeouts elsewhere are sized for a hosted
 # API, where a slow answer is a bill running; here the only thing being spent is
@@ -40,22 +48,33 @@ LOCAL_TIMEOUT = 3600
 
 # Where a transcription request goes; built by config.Config.transcribe_target().
 # `service` is the name the user sees in an error, `provider` the one the code
-# branches on.
-Target = collections.namedtuple("Target", "provider service api_key base_url model")
+# branches on. `file_model` is what a timestamped run asks for instead of
+# `model`, where the two differ; empty means the provider's own whisper.
+Target = collections.namedtuple(
+    "Target", "provider service api_key base_url model file_model",
+    defaults=[""])
+
+# What answers with segment times on OpenRouter when nothing else was chosen.
+OPENROUTER_FILE_MODEL = "openai/whisper-1"
 
 
-def timestamp_model(provider, selected=""):
+def timestamp_model(provider, selected="", file_model=""):
     """Which model answers with segment times.
 
-    OpenAI keeps them to whisper-1 and OpenRouter namespaces that id. Everything
-    Groq transcribes with is a whisper, so the model already chosen does it and
-    the fallback is only for a provider left on its default. So is everything the
-    local server runs, whatever the file is called, and there asking for another
-    model would name one it has never heard of.
+    OpenAI keeps them to whisper-1. Everything Groq transcribes with is a
+    whisper, so the model already chosen does it and the fallback is only for a
+    provider left on its default. So is everything the local server runs,
+    whatever the file is called, and there asking for another model would name
+    one it has never heard of. OpenRouter fronts several models that do times
+    and several that do not, and a request to the wrong one gets a transcript
+    with no segments in it, so the one to use is a setting of its own
+    (`file_model`) and whisper-1 is only where that setting is left empty.
     """
     if provider in ("groq", "local"):
         return selected or "whisper-large-v3-turbo"
-    return "openai/whisper-1" if provider == "openrouter" else "whisper-1"
+    if provider == "openrouter":
+        return file_model or OPENROUTER_FILE_MODEL
+    return "whisper-1"
 
 
 # What a gateway in front of the model answers of its own accord: the request
@@ -254,10 +273,23 @@ def _request(url, data, headers, timeout=120, aborter=None):
 
 
 def _extract_error(body):
+    """The line worth showing out of a failed request's body.
+
+    Whatever comes back, this has to end in a string: it is called while an
+    ApiError is being raised, and an exception thrown here would escape the
+    `except ApiError` every caller is holding and lose the dictation the raw
+    transcript would otherwise have been pasted from.
+    """
     try:
         payload = json.loads(body)
     except json.JSONDecodeError:
         return body[:300]
+    if isinstance(payload, list):
+        # Google answers some failures with an array holding the object the
+        # other providers send on its own.
+        payload = next((item for item in payload if isinstance(item, dict)), None)
+    if not isinstance(payload, dict):
+        return body[:300]
     err = payload.get("error")
     if isinstance(err, dict):
         return err.get("message") or json.dumps(err)[:300]
@@ -423,7 +455,8 @@ def transcribe_segments(target, audio_path, language="", prompt="", timeout=300,
                         aborter=None):
     """[(start_seconds, end_seconds, text)] using whisper-1's verbose response."""
     data = _transcribe_request(
-        target._replace(model=timestamp_model(target.provider, target.model)),
+        target._replace(model=timestamp_model(target.provider, target.model,
+                                              target.file_model)),
         audio_path, language, prompt, "verbose_json",
         granularity="segment", timeout=timeout, aborter=aborter,
     )
@@ -448,14 +481,24 @@ def transcribe_segments(target, audio_path, language="", prompt="", timeout=300,
     return out
 
 
+# The settings window offers OpenRouter's ladder, and Google has neither end of
+# it: "none" is refused outright with a 400, and there is nothing above "high".
+# Both ends land on the nearest rung that does exist, which costs the cleanup
+# rather than the dictation when it is wrong. "minimal" is where "off" goes, and
+# it is the quickest of them by a wide margin, which is what cleanup wants
+# anyway.
+GEMINI_EFFORT = {"none": "minimal", "xhigh": "high", "max": "high"}
+
+
 def _thinking(payload, provider, reasoning):
     """Ask for as much thinking as this provider understands, or for none.
 
     An empty level means "whatever the model does on its own", so nothing is
-    sent. The two mean opposite things by that, which is why the setting is kept
-    per provider: OpenRouter's cleanup models answer straight away, while a local
-    model that was trained to think will think, and cleanup is punctuation rather
-    than a job worth thinking about.
+    sent. The three mean opposite things by that, which is why the setting is
+    kept per provider: OpenRouter's cleanup models answer straight away, while a
+    local model that was trained to think will think, and a Gemini Flash left to
+    itself thinks too. Cleanup is punctuation rather than a job worth thinking
+    about.
     """
     if not reasoning:
         return
@@ -463,6 +506,13 @@ def _thinking(payload, provider, reasoning):
         # What llama.cpp passes to the chat template. The models that think read
         # it; the ones that do not ignore it.
         payload["chat_template_kwargs"] = {"enable_thinking": reasoning != "none"}
+    elif provider == "gemini":
+        # Google's compatibility layer takes OpenAI's flat field rather than
+        # OpenRouter's object, and it has no word for off, so "none" is asked
+        # for as the lowest rung it has rather than skipped: a Flash model left
+        # to decide for itself thinks, and thinking about a comma is the second
+        # this provider was chosen to save.
+        payload["reasoning_effort"] = GEMINI_EFFORT.get(reasoning, reasoning)
     elif reasoning != "none":
         # The thinking itself is never shown, so ask for it to be left out.
         payload["reasoning"] = {"effort": reasoning, "exclude": True}
@@ -524,15 +574,16 @@ def cleanup(text, api_key, model, system_prompt, reasoning="",
 
 
 def chat(messages, api_key, model, system_prompt, reasoning="",
-         base_url=OPENROUTER_URL, timeout=180):
+         base_url=OPENROUTER_URL, timeout=180, provider="openrouter",
+         service="OpenRouter"):
     """A conversation, rather than one transcript rewritten.
 
     The messages are the whole history and come back unchanged; the caller keeps
-    them, because there is no session on OpenRouter's side to resume.
+    them, because there is no session on the provider's side to resume.
     """
     if not api_key:
         raise ApiError(t("{service} API key is empty. Add it in Settings.",
-                         service="OpenRouter"))
+                         service=service))
     payload = {
         "model": model,
         "messages": [{"role": "system", "content": system_prompt}] + list(messages),
@@ -543,11 +594,11 @@ def chat(messages, api_key, model, system_prompt, reasoning="",
         data = _request(
             f"{base_url.rstrip('/')}/chat/completions",
             json.dumps(payload).encode("utf-8"),
-            _headers("openrouter", api_key, "application/json"),
+            _headers(provider, api_key, "application/json"),
             timeout=timeout,
         )
     except ApiError as exc:
-        raise explain(exc, "OpenRouter") from None
+        raise explain(exc, service) from None
     choices = data.get("choices") or []
     if not choices:
         raise ApiError(_extract_error(json.dumps(data)))
@@ -611,6 +662,37 @@ def openrouter_models(api_key="", transcription=False):
     return sorted(m["id"] for m in models if m.get("id"))
 
 
+# What a `gemini` id can be besides a model that answers a chat request: an
+# embedding, a picture, or a voice. None of them is any use to cleanup.
+NOT_CHAT = ("embedding", "-image", "-tts", "-audio")
+
+
+def gemini_models(api_key, base_url=GEMINI_URL):
+    """The Gemini models Google AI Studio will answer a chat request with.
+
+    Google serves its embedding, image and speech models out of the same list
+    and names them all `gemini` too, so the prefix alone is not the question;
+    none of those can clean up a sentence. The listing has also been known to
+    hand the ids back in their long form, `models/gemini-3.5-flash`, while a
+    request wants the short one; taking the prefix off costs nothing and is
+    right whichever form arrives.
+    """
+    service = "Google AI Studio"
+    if not api_key:
+        raise ApiError(t("{service} API key is empty. Add it in Settings.",
+                         service=service))
+    try:
+        data = _get_json(
+            f"{base_url.rstrip('/')}/models",
+            {"Authorization": f"Bearer {api_key}", "User-Agent": USER_AGENT},
+        )
+    except ApiError as exc:
+        raise explain(exc, service) from None
+    ids = [(m.get("id") or "").removeprefix("models/") for m in data.get("data", [])]
+    return sorted(i for i in ids
+                  if i.startswith("gemini") and not any(w in i for w in NOT_CHAT))
+
+
 def openai_models(api_key, base_url=OPENAI_URL, service="OpenAI"):
     """The audio models of anything that speaks OpenAI's /models, Groq included.
 

dikte/app.py

@@ -18,6 +18,7 @@ import socket
 import subprocess
 import sys
 import threading
+import time
 
 # A Wayland client cannot place a window in a screen corner, so the indicator
 # is drawn through XWayland.
@@ -49,6 +50,7 @@ from . import hub  # noqa: E402
 from . import i18n  # noqa: E402
 from . import integrate  # noqa: E402
 from . import ipc  # noqa: E402
+from . import mac_window  # noqa: E402
 from . import meeting  # noqa: E402
 from . import trayicon  # noqa: E402
 from . import update  # noqa: E402
@@ -116,6 +118,10 @@ class Dikte:
     def __init__(self, app):
         self.app = app
         self.conf = cfg.Config()
+        # run_app hands the one-instance lock over after construction; a Dikte
+        # built without one (the tests) restarts without touching a lock.
+        self.instance_lock = None
+        self._registry_shortcuts = True   # settled by _apply_settings below
         self.state = IDLE
         self.ask_state = IDLE
         # Which of the two the microphone is currently serving, or None.
@@ -145,12 +151,25 @@ class Dikte:
         # Which recording is the current one, so a timer set for the run that
         # started it cannot stop the one that came after.
         self._run_id = 0
+        # Dictations handed to the pipeline and not yet out of it. More than
+        # one is normal: the microphone is free while a transcript is being
+        # cleaned up, so the next dictation can already be spoken, and it then
+        # queues up behind the one still going.
+        self._transcripts_pending = 0
+        # The application that was in front when the recording started, which
+        # is where the transcript is meant to go, and the timer watching for
+        # the moment it has to be put back there. macOS only; see
+        # _give_the_front_back.
+        self.front_before = None
+        self._front_watch = None
 
-        self.overlay = Overlay(self.conf["overlay_corner"])
+        self.overlay = Overlay(self.conf["overlay_corner"],
+                               screen_name=self.conf["overlay_screen"])
         # The agent's indicator sits on top of the dictation one when both are
         # up, and drops into the corner when it is alone there.
         self.ask_overlay = Overlay(self.conf["overlay_corner"], below=self.overlay,
-                                   dismissable=True)
+                                   dismissable=True,
+                                   screen_name=self.conf["overlay_screen"])
         self.recorder = audio.Recorder()
         self.pipeline = Pipeline(self.conf)
         self.ask_pipeline = Pipeline(self.conf)
@@ -160,13 +179,20 @@ class Dikte:
         # Before anything of ours is started: a server from a Dikte that was
         # killed outright is still holding a model in memory.
         ggml.sweep()
+        # The first dictation with no microphone picked would otherwise pay for
+        # the ffmpeg device listing on the key press itself, seconds of nothing
+        # happening at the least explicable moment. Warmed the way the models
+        # are, off the main thread; the listing caches itself.
+        if sys.platform == "win32" and not self.conf["mic_target"]:
+            threading.Thread(target=audio.list_sources, daemon=True).start()
 
         self.recorder.level.connect(self._on_level)
         self.recorder.stopped.connect(self._on_recorded)
+        self.recorder.died.connect(self._on_recorder_died)
         self.recorder.failed.connect(self._on_recorder_error)
-        self.pipeline.stage.connect(self.overlay.show_busy)
+        self.pipeline.stage.connect(self._on_stage)
         self.pipeline.finished.connect(self._on_finished)
-        self.pipeline.failed.connect(self._on_error)
+        self.pipeline.failed.connect(self._on_pipeline_failed)
         self.ask_pipeline.stage.connect(self.ask_overlay.show_busy)
         self.ask_pipeline.finished.connect(self._on_ask_finished)
         self.ask_pipeline.failed.connect(self._on_ask_error)
@@ -184,7 +210,7 @@ class Dikte:
 
         self.elapsed = QElapsedTimer()
         self.meeting_elapsed = QElapsedTimer()
-        self.last_toggle = QElapsedTimer()
+        self.last_toggle = {}   # action name -> QElapsedTimer, see _repeated
         self.last_evdev = {}
         self.ticker = QTimer()
         self.ticker.setInterval(100)
@@ -336,13 +362,18 @@ class Dikte:
             BUSY: ("Working…", "view-refresh", "Dikte: working"),
         }
         label, icon, tip = labels[self.state]
+        if self.state == BUSY:
+            # Still working, but the microphone is free again: the menu offers
+            # the next dictation rather than a wait.
+            label = "Start recording"
         agent = assistant.display_name(self.conf)
 
         self.toggle_action.setText(t(label))
-        # Free while the other one is thinking, blocked only while it is holding
-        # the microphone.
+        # Blocked only while something is holding the microphone: a transcript
+        # still being cleaned up queues the next dictation behind it, and the
+        # agent thinking never blocked it at all.
         self.toggle_action.setEnabled(
-            self.state == RECORDING or (self.state == IDLE and not self.recording)
+            self.state == RECORDING or not self.recording
         )
         asked = i18n.name(agent, "dative")
         self.ask_action.setText(
@@ -429,7 +460,10 @@ class Dikte:
         # Where nothing was installed there is no shortcut to catch up, and
         # retiring the listener would leave the keys with nowhere to arrive.
         timer = self.last_evdev.get(name)
-        if (hotkey.installs_shortcuts() and self.evdev.running
+        # The snapshot from _apply_settings: which desktop this is cannot
+        # change under a running process, and asking hotkey again here costs a
+        # PATH scan on every key press.
+        if (self._registry_shortcuts and self.evdev.running
                 and timer is not None and timer.elapsed() < ECHO_MS):
             self._retire_listener()
             return
@@ -498,10 +532,6 @@ class Dikte:
 
     def _dictation_request(self, cmd, request, reply):
         before = self.state
-        # Only a request that said something about pasting changes it, so that
-        # the stop half of a `start --paste` does not undo the start half.
-        if "paste" in request:
-            self.paste_override[DICTATION] = request["paste"]
         if cmd == "toggle":
             self.toggle()
         elif cmd == "stop":
@@ -512,13 +542,20 @@ class Dikte:
             if seconds > 0 and self.state == RECORDING:
                 run = self._run_id
                 QTimer.singleShot(int(seconds * 1000), lambda: self._auto_stop(run))
+        # Armed only when the request actually moved this run along: a request
+        # that no-opped (the microphone held by the other one, or nothing to
+        # stop) must not leave a preference behind for some later, unrelated
+        # run to pick up. A stop that lands keeps changing the run it ends,
+        # so the stop half of a `start --paste` still does not undo the start.
+        if "paste" in request and self.state != before:
+            self.paste_override[DICTATION] = request["paste"]
         self._answer(DICTATION, before, self.state, request, reply)
 
     def _ask_request(self, request, reply):
         before = self.ask_state
-        if "paste" in request:
-            self.paste_override[ASK] = request["paste"]
         self.toggle_ask()
+        if "paste" in request and self.ask_state != before:
+            self.paste_override[ASK] = request["paste"]
         self._answer(ASK, before, self.ask_state, request, reply)
 
     def _meeting_request(self, cmd, request, reply):
@@ -583,40 +620,70 @@ class Dikte:
     def _toggle(self):
         # Two /dev/input nodes can carry the same keyboard, and a menu click can
         # land on top of a key press; swallow the immediate repeat.
-        if self._repeated():
+        if self._repeated("toggle"):
             return
         if self.state == RECORDING:
             self.stop()
-        elif self.state == IDLE:
+        else:
+            # BUSY does not block: the microphone is free while the last
+            # dictation is being cleaned up, and the next one starts now and
+            # waits its turn in the pipeline.
             self.start()
-        # a request during its own BUSY is ignored; nothing queues up
 
     def _toggle_ask(self):
-        if self._repeated():
+        if self._repeated("ask"):
             return
         if self.ask_state == RECORDING:
             self.stop_ask()
         elif self.ask_state == IDLE:
             self.start_ask()
 
-    def _repeated(self):
-        if self.last_toggle.isValid() and self.last_toggle.elapsed() < 400:
+    def _repeated(self, name):
+        # Per action, the way last_evdev already is: the window is meant to
+        # swallow a duplicate delivery of the same press, not a pause landing
+        # right after the toggle that started the recording.
+        timer = self.last_toggle.get(name)
+        if timer is None:
+            timer = self.last_toggle[name] = QElapsedTimer()
+        if timer.isValid() and timer.elapsed() < 400:
             return True
-        self.last_toggle.restart()
+        timer.restart()
         return False
 
+    def _the_front(self):
+        """The application a recording is about to start from, or None.
+
+        Asked before the indicator goes up rather than alongside the
+        microphone: putting a window on screen can take the front as well, and
+        once it has, the only answer left to the question is Dikte.
+        """
+        return mac_window.frontmost_pid() if sys.platform == "darwin" else None
+
     def start(self):
-        if self.state != IDLE or self.recording:
+        # Only a held microphone blocks: a previous dictation still being
+        # transcribed or cleaned up is the pipeline's business, not the
+        # recorder's.
+        if self.state == RECORDING or self.recording:
             return
+        self.front_before = self._the_front()
         self.overlay.show_recording()
         self._begin_recording(DICTATION)
+        # A recorder that could not start has already said so, synchronously,
+        # and the error handler put everything back; setting RECORDING on top
+        # of that would strand the state machine with no signal ever coming.
+        # The same guard start_meeting has always had.
+        if not self.recorder.active:
+            return
         self._set_state(RECORDING)
 
     def start_ask(self):
         if self.ask_state != IDLE or self.recording:
             return
+        self.front_before = self._the_front()
         self.ask_overlay.show_recording(asking=True)
         self._begin_recording(ASK)
+        if not self.recorder.active:
+            return
         self._set_ask_state(RECORDING)
 
     def _begin_recording(self, owner):
@@ -627,6 +694,80 @@ class Dikte:
         self.elapsed.restart()
         self.ticker.start()
         self.recorder.start(self.conf["mic_target"], self.conf["max_seconds"])
+        self._give_the_front_back(self.front_before)
+
+    def _give_the_front_back(self, was_in_front):
+        """Hand the front back to whoever had it when the recording started.
+
+        On macOS a recording goes through ffmpeg's avfoundation input, and
+        opening a capture session there brings the process that did it to the
+        front. ffmpeg is a child of Dikte with no bundle of its own, so the
+        system credits the move to Dikte: the window the user was typing in
+        loses the front, its caret stops, its title bar greys out, and the
+        Cmd+V at the end of the dictation has nowhere to land. Measured with a
+        TextEdit document in front:
+
+            press the shortcut          front = TextEdit
+            recorder.start returns      front = TextEdit
+            89 ms later                 front = Dikte
+
+        Nothing about the capture session can be asked not to do this. It is
+        not a window of ours and no flag reaches it. Starting ffmpeg in its own
+        session, and clearing __CFBundleIdentifier from its environment, were
+        both tried and both measured to make no difference. So it is undone
+        instead. The move lands a moment after the process starts rather than
+        during the call, hence the short watch rather than one attempt: it
+        gives up as soon as it has put the front back, and in any case after a
+        second and a half, which is longer than the microphone has ever taken
+        to open.
+
+        Silent off macOS, and silent when the recording started from Dikte
+        itself: there is nothing to give back.
+        """
+        # One watch at a time. A second recording started before the first
+        # watch had finished would otherwise leave two of them running, and the
+        # older one would put the front back where the older recording
+        # started, which by then is the wrong window.
+        if self._front_watch is not None:
+            self._front_watch.stop()
+            self._front_watch = None
+        if not was_in_front or was_in_front == os.getpid():
+            return
+        deadline = time.monotonic() + 1.5
+        watch = QTimer(self.app)
+        # Ten milliseconds because the front is already gone by the time this
+        # notices, and every tick it waits is a tick of the user's window drawn
+        # inactive: at forty the title bar visibly blinks, at ten it does not.
+        # Two messages to AppKit per tick, for at most a second and a half.
+        watch.setInterval(10)
+        # activateWithOptions: answers whether macOS accepted the request, not
+        # whether the other application is already back in front.  Keep the
+        # watch alive until that asynchronous handoff is observable; on Intel
+        # Macs it can take hundreds of milliseconds after the call returned.
+        restore_requested = False
+
+        def look():
+            nonlocal restore_requested
+            if time.monotonic() > deadline:
+                self._stop_watching_the_front()
+                return
+            if mac_window.is_frontmost():
+                if not restore_requested:
+                    restore_requested = mac_window.activate(was_in_front)
+            elif restore_requested:
+                # The request has landed.  Stop only now, rather than as soon
+                # as AppKit accepted it, so a delayed or failed handoff stays
+                # under observation until the deadline guard above.
+                self._stop_watching_the_front()
+
+        watch.timeout.connect(look)
+        self._front_watch = watch
+        watch.start()
+
+    def _stop_watching_the_front(self):
+        if self._front_watch is not None:
+            self._front_watch.stop()
+            self._front_watch = None
 
     def stop(self):
         if self.state != RECORDING:
@@ -634,7 +775,9 @@ class Dikte:
         self.ticker.stop()
         self._clear_pause()
         self._set_state(BUSY)
-        self.overlay.show_busy(t("Transcribing…"))
+        self.overlay.show_busy(t("Waiting for the one before it…")
+                               if self._transcripts_pending
+                               else t("Transcribing…"))
         self.recorder.stop()
 
     def stop_ask(self):
@@ -654,7 +797,7 @@ class Dikte:
         the phone call in the middle of a dictation never reaches the model and
         the sentence around it is still one sentence.
         """
-        if not self.recording or self._repeated():
+        if not self.recording or self._repeated("pause"):
             return
         self.paused = not self.paused
         if self.paused:
@@ -686,6 +829,8 @@ class Dikte:
         self.ticker.stop()
         self._clear_pause()
         self.recorder.cancel()
+        # The preference dies with the run it was given for.
+        self.paste_override.pop(ASK if asking else DICTATION, None)
         self.recorder_owner = None
         # What goes over the socket is read by a program as often as by a
         # person, so it stays in one language; only what a run itself said
@@ -697,7 +842,9 @@ class Dikte:
             self._settle(ASK, dropped)
         else:
             self.overlay.dismiss()
-            self._set_state(IDLE)
+            # An earlier dictation may still be in the pipeline; only the
+            # recording was thrown away.
+            self._set_state(BUSY if self._transcripts_pending else IDLE)
             self._settle(DICTATION, dropped)
 
     def cancel_ask(self):
@@ -742,6 +889,12 @@ class Dikte:
             return
         base = meeting.new_base()
         _, wav_path = cfg.meeting_paths(base)
+        # A meeting opens the same capture as a dictation does, and takes the
+        # front the same way: whoever is being recorded is in a call, and
+        # having their window go inactive mid-sentence is worse here than
+        # anywhere else. Kept as a local rather than on self: a dictation may
+        # already be waiting on its own note for where to paste.
+        was_in_front = self._the_front()
         self.meeting_recorder.start(
             str(wav_path),
             self.conf["meeting_mic_target"] or self.conf["mic_target"],
@@ -750,6 +903,7 @@ class Dikte:
         )
         if not self.meeting_recorder.active:
             return  # start() has already said what went wrong
+        self._give_the_front_back(was_in_front)
         self.meeting_base = base
         self.meeting_elapsed.restart()
         self.meeting_ticker.start()
@@ -876,31 +1030,56 @@ class Dikte:
     def _on_recorded(self, wav_path, duration, rms_values):
         owner, self.recorder_owner = self.recorder_owner, None
         wants_paste = self.paste_override.pop(owner, None)
+        focus, self.front_before = self.front_before, None
         if owner == ASK:
             self.ask_pipeline.run(wav_path, duration, rms_values, ask=True,
-                                  paste=wants_paste)
+                                  paste=wants_paste, focus=focus)
         else:
-            self.pipeline.run(wav_path, duration, rms_values, paste=wants_paste)
+            self._transcripts_pending += 1
+            self.pipeline.run(wav_path, duration, rms_values,
+                              paste=wants_paste, focus=focus)
+
+    def _on_stage(self, message):
+        # The corner belongs to the recording when one is on: the previous
+        # run's progress must not wipe the waveform mid-sentence.
+        if self.state != RECORDING:
+            self.overlay.show_busy(message)
+
+    def _transcript_settled(self, payload):
+        """One run out of the pipeline; where dictation stands now.
+
+        A request that asked to wait is answered once the queue is empty: with
+        runs finishing in the order they were spoken, the one it stopped is the
+        last of them, and an earlier run's result would be the wrong answer.
+        """
+        self._transcripts_pending -= 1
+        if self.state != RECORDING:
+            self._set_state(BUSY if self._transcripts_pending else IDLE)
+        if not self._transcripts_pending:
+            self._settle(DICTATION, payload)
 
     def _on_finished(self, _raw, text, warning):
         if warning:
             # The text was still pasted, but cleanup did not run. Say so loudly:
             # a rejected key otherwise looks exactly like working dictation.
-            self.overlay.show_warning(
-                t("Pasted raw, cleanup failed: {error}", error=warning.splitlines()[0])
-            )
+            if self.state != RECORDING:
+                self.overlay.show_warning(
+                    t("Pasted raw, cleanup failed: {error}",
+                      error=warning.splitlines()[0])
+                )
             self.tray.showMessage(
                 t("Dikte: cleanup failed"), warning,
                 QSystemTrayIcon.MessageIcon.Warning, 10000,
             )
-        else:
+        elif self.state != RECORDING:
+            # While a new recording is on, the flash is skipped: the text
+            # arriving where the cursor is says everything it would have.
             action = t("Pasted") if self.conf["auto_paste"] else t("Copied")
             self.overlay.show_done(
                 t("{action}: {preview}", action=action, preview=_preview(text))
             )
-        self._set_state(IDLE)
-        self._settle(DICTATION, {"ok": True, "text": text, "raw": _raw,
-                                 "warning": warning})
+        self._transcript_settled({"ok": True, "text": text, "raw": _raw,
+                                  "warning": warning})
 
     def _on_ask_finished(self, _raw, text, warning):
         agent = assistant.display_name(self.conf)
@@ -933,12 +1112,43 @@ class Dikte:
     def _on_recorder_error(self, message):
         """The microphone itself could not run, so it belongs to whoever asked."""
         owner, self.recorder_owner = self.recorder_owner, None
+        self.paste_override.pop(owner, None)
         self.ticker.stop()
         (self._on_ask_error if owner == ASK else self._on_error)(message)
 
+    def _on_pipeline_failed(self, message):
+        """A run the pipeline gave up on; whatever queued behind it still runs."""
+        if self.state == RECORDING:
+            # The corner belongs to the new recording; the failure still has to
+            # be seen somewhere.
+            self.tray.showMessage("Dikte", message,
+                                  QSystemTrayIcon.MessageIcon.Warning, 8000)
+        else:
+            self._report(message, self.overlay)
+        self._transcript_settled({"ok": False, "error": message})
+
+    def _on_recorder_died(self):
+        """The capture quit under a live recording: keep what it caught.
+
+        Ended the way a key press would end it, so the captured half is
+        transcribed rather than thrown away, and said out loud, because the
+        user is still talking at a microphone nobody is reading.
+        """
+        owner = self.recorder_owner
+        self.tray.showMessage(
+            "Dikte",
+            t("The recording stopped on its own; transcribing what was captured."),
+            QSystemTrayIcon.MessageIcon.Warning, 8000,
+        )
+        if owner == ASK and self.ask_state == RECORDING:
+            self.stop_ask()
+        elif self.state == RECORDING:
+            self.stop()
+
     def _on_error(self, message):
+        """The recorder or the key listener failed; no run reached the pipeline."""
         self._report(message, self.overlay)
-        self._set_state(IDLE)
+        self._set_state(BUSY if self._transcripts_pending else IDLE)
         self._settle(DICTATION, {"ok": False, "error": message})
 
     def _on_ask_error(self, message):
@@ -1001,18 +1211,57 @@ class Dikte:
 
     def open_settings(self):
         if self.settings_window is None:
-            self.settings_window = SettingsWindow(self.conf, self.meetings)
-            self.settings_window.applied.connect(self._apply_settings)
-            self.settings_window.update_found.connect(self._found_update)
-            self.settings_window.finished.connect(self._settings_closed)
+            self._make_settings()
         self.settings_window.show()
         self.settings_window.raise_()
         self.settings_window.activateWindow()
 
+    def _make_settings(self):
+        """Build the window without showing it, so a caller that knows where
+        it belongs can place it first."""
+        self.settings_window = SettingsWindow(self.conf, self.meetings)
+        self.settings_window.applied.connect(self._apply_settings)
+        self.settings_window.language_changed.connect(self._reopen_settings)
+        self.settings_window.update_found.connect(self._found_update)
+        self.settings_window.finished.connect(self._settings_closed)
+
     def _settings_closed(self, *_):
         # Don't drop the object while its own signal is still being delivered.
         QTimer.singleShot(0, lambda: setattr(self, "settings_window", None))
 
+    def _reopen_settings(self):
+        """Replace the settings window, so a language change reaches it too.
+
+        A save switches the language everywhere strings are made at the moment
+        they are shown: the tray is rebuilt, the indicator and the message box
+        translate as they speak. The settings window is the one place written
+        once, at construction, so the window that took the new language is the
+        one place still showing the old one. A fresh window comes up where the
+        old one stood, on the same tab.
+        """
+        old = self.settings_window
+        if old is None:
+            return
+        tab = old.tabs.currentIndex()
+        geometry = old.geometry()
+        # Replaced rather than merely closed: left connected, _settings_closed
+        # would drop the reference to the new window a moment after it is made.
+        old.finished.disconnect(self._settings_closed)
+        old.close()
+        # No deleteLater: a daemon thread of the old window's may still be
+        # running, and a closure holding self is what keeps the object alive
+        # until the thread is done. Dropping the reference is how the ordinary
+        # close path lets a window go, and it is enough here too.
+        self.settings_window = None
+        self._make_settings()
+        # Placed and turned to the old tab before it is shown, so the new
+        # window does not come up at the default size and jump.
+        self.settings_window.setGeometry(geometry)
+        self.settings_window.tabs.setCurrentIndex(tab)
+        self.settings_window.show()
+        self.settings_window.raise_()
+        self.settings_window.activateWindow()
+
     def _apply_local(self):
         """Pass the local settings on, and hold the models ready if asked to.
 
@@ -1049,14 +1298,19 @@ class Dikte:
 
     def _apply_settings(self):
         self.overlay.corner = self.conf["overlay_corner"]
+        self.overlay.screen_name = self.conf["overlay_screen"]
         self.ask_overlay.corner = self.conf["overlay_corner"]
+        self.ask_overlay.screen_name = self.conf["overlay_screen"]
         self._apply_local()
         self._build_tray()
         self._refresh_tray()
+        # Taken once here for _external: the answer cannot change under a
+        # running process, and re-deriving it there is a PATH scan per press.
+        self._registry_shortcuts = hotkey.installs_shortcuts()
         # Where the desktop has no shortcut registry of its own, the listener is
         # not the fallback the setting offers to turn on: it is the only way the
         # keys arrive at all, so it runs whatever the setting says.
-        if self.conf["evdev_hotkey"] or not hotkey.installs_shortcuts():
+        if self.conf["evdev_hotkey"] or not self._registry_shortcuts:
             self.evdev.start({name: self.conf[spec.setting]
                               for name, spec in hotkey.SHORTCUTS.items()})
         else:
@@ -1078,22 +1332,25 @@ class Dikte:
         if self.server is not None:
             self.server.close()
         QLocalServer.removeServer(SERVER_NAME)
-        args = ipc.launcher() + ["--gui"]
-        if sys.platform == "win32":
-            # execv on Windows mangles arguments with spaces and leaves the two
-            # processes sharing a console; a detached start does neither.
-            subprocess.Popen(
-                args,
-                creationflags=(subprocess.DETACHED_PROCESS
-                               | subprocess.CREATE_NEW_PROCESS_GROUP),
-                close_fds=True,
-            )
-            QApplication.instance().quit()
-            return
-        os.execv(args[0], args)
+        # The lock too, or the replacement would take this restart for a
+        # double start and hand the attention back to a process on its way out.
+        if self.instance_lock is not None:
+            self.instance_lock.unlock()
+        ipc.respawn(["--gui"])
+        # respawn only returns on Windows, where the replacement was started
+        # detached and this process still has to leave on its own.
+        QApplication.instance().quit()
 
     def shutdown(self):
         self._quitting = True
+        # Waiters first, while the connections still work: a `--wait` left
+        # unanswered reads to the terminal as an instance too old to answer,
+        # which points the user at a version problem that does not exist.
+        # In one language, like every other error that goes over the socket:
+        # a script reads these as often as a person does.
+        for kind in (DICTATION, ASK, MEETING):
+            self._settle(kind, {"ok": False,
+                                "error": "the instance is shutting down"})
         self.evdev.stop()
         if self.recording:
             self.recorder.cancel()
@@ -1216,9 +1473,44 @@ def _stay_out_of_the_dock():
         pass
 
 
+def _hand_over(command):
+    """Give the running instance the attention this start was asking for.
+
+    A start carrying a verb forwards only that verb; a bare double start asks
+    for the Settings window as the sign of life the click was looking for.
+    Retried for a moment, because the copy that won the lock may not be
+    listening yet.
+    """
+    verb = command or "settings"
+    deadline = time.monotonic() + 5
+    while time.monotonic() < deadline:
+        if ipc.send(verb) is not None:
+            return
+        time.sleep(0.2)
+    print("dikte: another copy holds the lock but never answered")
+
+
 def run_app(args):
     command = args[0] if args else ""
 
+    # One Dikte per user. The lock closes the simultaneous-start window two
+    # probes would both fall through; the probe still runs behind it, because
+    # an instance from before the lock existed holds only the socket. Both
+    # sit before the QApplication, so a second copy costs a moment and not a
+    # second tray icon. The lock lives in this frame, which app.exec() below
+    # keeps alive for exactly the process's lifetime.
+    lock = ipc.instance_lock()
+    if lock is not None and not lock.tryLock(0):
+        _hand_over(command)
+        return 0
+    if ipc.already_serving():
+        print("dikte: already running; handing it the attention")
+        if command:
+            ipc.send(command)
+        else:
+            ipc.send("settings")
+        return 0
+
     app = QApplication(sys.argv)
     app.setApplicationName("Dikte")
     app.setDesktopFileName("dikte")
@@ -1246,6 +1538,9 @@ def run_app(args):
         print("dikte: no system tray found, running anyway")
 
     dikte = Dikte(app)
+    # Handed over so that restart() can let go of it before the replacement
+    # tries to take it.
+    dikte.instance_lock = lock
 
     server = QLocalServer()
     # Qt puts the socket in /tmp, so keep it to this user: commands like

dikte/assistant.py

@@ -1,22 +1,29 @@
 """Handing a dictation to an agent as a command, and pasting back its answer.
 
-Three of them, because not everyone has the same one installed:
+Five of them, because not everyone has the same one installed:
 
   Claude Code   `claude -p`, the session you would have opened yourself
   Codex         `codex exec`, the same idea from the other shop
+  Antigravity   `agy -p`, Google's, with a browser of its own attached
   OpenRouter    a plain chat request, over the key that is already configured
+  OpenCode Go   a plain chat request, over a subscription to open coding models
 
-The first two are the whole machine: they run commands, read files, and reach
+The first three are the whole machine: they run commands, read files, and reach
 whatever skills and services you have connected, which is what makes "put that
-in my calendar on Thursday" a thing you can say. OpenRouter cannot touch any of
-that, and is there so that a question still gets an answer on a machine with
-neither CLI installed.
+in my calendar on Thursday" a thing you can say. The two chat requests cannot
+touch any of that, and are there so that a question still gets an answer on a
+machine with no CLI installed at all.
+
+What each of the three is allowed to do without asking is settled where that
+program keeps its own permissions, not here. Dikte hands Claude Code the mode
+chosen in Settings because it has a flag for one; Codex gets a sandbox for the
+same reason; Antigravity has neither, and reads its own allow-rules instead.
 
 Whichever it is, the reply is pasted exactly where the transcript would have
 been, and the conversation carries across dictations so that "and move that to
 Friday" knows what "that" is.
 
-The two CLIs are read as they stream rather than waited out. A command that
+The three CLIs are read as they stream rather than waited out. A command that
 reaches for the calendar or the web takes long enough that a still indicator is
 indistinguishable from a hang, so every tool they pick up is named in the corner
 while they work.
@@ -24,21 +31,30 @@ while they work.
 
 import json
 import os
+import re
 import shutil
+import signal
 import subprocess
+import tempfile
 import threading
 import time
 
 from . import api
 from . import config as cfg
+from . import paths
 from .i18n import t
 
 SESSION_FILE = cfg.DATA_DIR / "assistant.json"
-PROVIDERS = ("claude", "codex", "openrouter")
+PROVIDERS = ("claude", "codex", "agy", "openrouter", "opencode")
 
-# How many messages of an OpenRouter conversation are carried forward. The two
-# CLIs keep their own history and need no such number; here every turn is resent
-# in full, so the window has to end somewhere.
+# What each one is called where a person reads it: the tray, the corner of
+# the screen, and the line an error is written in.
+SERVICES = {"claude": "Claude", "codex": "Codex", "agy": "Antigravity",
+            "openrouter": "OpenRouter", "opencode": "OpenCode Go"}
+
+# How many messages of a chat provider's conversation are carried forward. The
+# two CLIs keep their own history and need no such number; here every turn is
+# resent in full, so the window has to end somewhere.
 MAX_HISTORY = 24
 
 # What to say in the indicator for a tool, keyed by the name the CLI uses.
@@ -66,6 +82,29 @@ CODEX_ITEMS = {
     "patch_apply": "Editing a file…",
     "todo_list": "Planning…",
 }
+# Antigravity carries a browser around with it, so the handful of names below
+# stand in for the couple of dozen browser_* tools it can pick up; being told
+# which mouse button moved is not what the corner of the screen is for.
+AGY_TOOLS = {
+    "run_command": "Running a command…",
+    "command_status": "Running a command…",
+    "send_command_input": "Running a command…",
+    "view_file": "Reading…",
+    "read_url_content": "Reading a web page…",
+    "list_dir": "Looking through files…",
+    "find_by_name": "Looking through files…",
+    "grep_search": "Searching the files…",
+    "search_web": "Searching the web…",
+    "replace_file_content": "Editing a file…",
+    "multi_replace_file_content": "Editing a file…",
+    "sed_file": "Editing a file…",
+    "notebook_edit": "Editing a file…",
+    "write_to_file": "Writing a file…",
+    "generate_image": "Drawing…",
+    "manage_task": "Planning…",
+    "invoke_subagent": "Handing it to a subagent…",
+    "browser_subagent": "Handing it to a subagent…",
+}
 
 
 # How hard to think, in each provider's own vocabulary. The setting is one
@@ -81,6 +120,12 @@ CLAUDE_EFFORT = {"none": "low", "minimal": "low", "low": "low",
 CODEX_EFFORT = {"none": "low", "minimal": "low", "low": "low",
                 "medium": "medium", "high": "high", "xhigh": "high",
                 "max": "high"}
+# agy has three rungs and no word for off, so the bottom of the ladder lands on
+# "low" and the top two on "high". Shared with cleanup, which runs the same
+# program for the smaller job.
+AGY_EFFORT = {"none": "low", "minimal": "low", "low": "low",
+              "medium": "medium", "high": "high", "xhigh": "high",
+              "max": "high"}
 
 
 class AssistantError(Exception):
@@ -98,12 +143,30 @@ def provider(conf):
 
 def executable(name):
     """The CLI a provider runs, or "" when it needs none."""
-    return {"claude": "claude", "codex": "codex"}.get(name, "")
+    return {"claude": "claude", "codex": "codex", "agy": "agy"}.get(name, "")
+
+
+def model(conf):
+    """Which model answered, for the history to record.
+
+    Each provider keeps its own setting, and the one a CLI is left on has no id
+    to report, only a name — the same arrangement cleanup.model() makes.
+    """
+    name = provider(conf)
+    if name == "codex":
+        return conf["assistant_codex_model"].strip() or "codex"
+    if name == "agy":
+        return conf["assistant_agy_model"].strip() or "agy"
+    if name == "openrouter":
+        return conf["assistant_openrouter_model"]
+    if name == "opencode":
+        return conf["assistant_opencode_model"]
+    return conf["assistant_model"]
 
 
 def display_name(conf):
     """What to call the thing being asked, in the tray and in the corner."""
-    return {"claude": "Claude", "codex": "Codex"}.get(provider(conf), "OpenRouter")
+    return SERVICES.get(provider(conf), "OpenRouter")
 
 
 # --- the conversation -----------------------------------------------------
@@ -114,13 +177,19 @@ def display_name(conf):
 # one along costs tokens and invites an answer to the wrong question. Switching
 # provider drops it too, since none of them can pick up another's thread.
 
-def _read_row(name, max_age_seconds):
+def _read_session():
+    """The stored conversation row, or {} however the file fails to read."""
     try:
         with open(SESSION_FILE, encoding="utf-8") as fh:
             row = json.load(fh)
     except (OSError, json.JSONDecodeError, ValueError):
         return {}
-    if not isinstance(row, dict) or row.get("provider") != name:
+    return row if isinstance(row, dict) else {}
+
+
+def _read_row(name, max_age_seconds):
+    row = _read_session()
+    if row.get("provider") != name:
         return {}
     if max_age_seconds and time.time() - row.get("ts", 0) > max_age_seconds:
         return {}
@@ -159,22 +228,13 @@ def clear_session():
 
 def stored_provider():
     """Whose conversation is on disk, whatever the setting says now."""
-    try:
-        with open(SESSION_FILE, encoding="utf-8") as fh:
-            row = json.load(fh)
-    except (OSError, json.JSONDecodeError, ValueError):
-        return ""
-    return str(row.get("provider", "")) if isinstance(row, dict) else ""
+    return str(_read_session().get("provider", ""))
 
 
 def session_age():
     """Seconds since the stored conversation was last used, or None."""
-    try:
-        with open(SESSION_FILE, encoding="utf-8") as fh:
-            row = json.load(fh)
-    except (OSError, json.JSONDecodeError, ValueError):
-        return None
-    if not isinstance(row, dict) or not (row.get("session") or row.get("messages")):
+    row = _read_session()
+    if not (row.get("session") or row.get("messages")):
         return None
     return time.time() - row.get("ts", 0)
 
@@ -196,8 +256,8 @@ def ask(prompt, conf, on_stage=None, should_stop=None):
     one, and only the denial explains why it did not do what it was asked to.
     """
     name = provider(conf)
-    if name == "openrouter":
-        return _ask_openrouter(prompt, conf, on_stage)
+    if name in ("openrouter", "opencode"):
+        return _ask_chat(name, SERVICES[name], prompt, conf, on_stage)
 
     binary = executable(name)
     if not shutil.which(binary):
@@ -206,7 +266,7 @@ def ask(prompt, conf, on_stage=None, should_stop=None):
             "Settings → Agent.", binary=binary,
         ))
 
-    run = _ask_claude if name == "claude" else _ask_codex
+    run = {"claude": _ask_claude, "codex": _ask_codex, "agy": _ask_agy}[name]
     session = read_session(name, conf["assistant_session_minutes"] * 60)
     try:
         return run(prompt, conf, session, on_stage, should_stop)
@@ -258,7 +318,7 @@ def _ask_claude(prompt, conf, session, on_stage, should_stop):
             found["warning"] = _denial_warning(event)
 
     code, stderr = _stream(cmd, conf, on_event, should_stop)
-    return _conclude(found, code, stderr, session, "Claude")
+    return _conclude(found, code, stderr, session, "claude")
 
 
 def _claude_label(block):
@@ -326,7 +386,7 @@ def _ask_codex(prompt, conf, session, on_stage, should_stop):
                                 else str(error)) or t("Codex ended with an error.")
 
     code, stderr = _stream(cmd, conf, on_event, should_stop)
-    return _conclude(found, code, stderr, session, "Codex")
+    return _conclude(found, code, stderr, session, "codex")
 
 
 def _codex_label(item):
@@ -338,29 +398,152 @@ def _codex_label(item):
     return t("Using {name}…", name=item_type or "a tool")
 
 
-# --- OpenRouter -----------------------------------------------------------
+def codex_models():
+    """The models Codex itself would offer right now, best first.
 
-def _ask_openrouter(prompt, conf, on_stage):
-    """No tools, no files, no calendar: a question and an answer.
+    `codex debug models` prints the catalog the CLI's own model picker reads,
+    fetched from OpenAI and cached beside Codex's config, so the list is as
+    current as the installed Codex and there is no second list to keep up to
+    date here. Entries the picker hides are internal and stay hidden. A machine
+    without Codex, or one too old to have the command, answers with nothing and
+    the caller keeps its built-in list.
+    """
+    if not shutil.which("codex"):
+        return []
+    try:
+        proc = subprocess.run(["codex", "debug", "models"],
+                              capture_output=True, text=True, timeout=30)
+        catalog = json.loads(proc.stdout or "null")
+    except (OSError, subprocess.SubprocessError, ValueError):
+        return []
+    if not isinstance(catalog, dict):
+        return []
+    rows = [row for row in catalog.get("models") or []
+            if isinstance(row, dict) and row.get("slug")
+            and row.get("visibility") != "hide"]
+    rows.sort(key=lambda row: row.get("priority") or 0)
+    return [row["slug"] for row in rows]
 
-    It is the fallback for a machine with neither CLI on it, so it says what it
-    knows and nothing else. The conversation is ours to keep here, since there
-    is no session on the other end to resume.
+
+# --- Antigravity ----------------------------------------------------------
+
+def _ask_agy(prompt, conf, session, on_stage, should_stop):
+    # Antigravity takes no system prompt of its own either, so the instruction
+    # rides in front of the command, kept apart from it so the two are not read
+    # as one.
+    body = f"{conf.assistant_prompt()}\n\n---\n\n{prompt}"
+    cmd = [
+        "agy", "-p", body,
+        "--output-format", "stream-json",
+        # agy stops after five minutes unless it is told otherwise, which is
+        # shorter than the timeout this setting offers.
+        "--print-timeout", f"{conf['assistant_timeout']}s",
+    ]
+    # One or the other, always: left with neither, agy picks up whichever
+    # project it was last in and works in that project's directory rather than
+    # the one _stream is about to start it in.
+    cmd += ["--conversation", session] if session else ["--new-project"]
+    if conf["assistant_agy_model"].strip():
+        cmd += ["--model", conf["assistant_agy_model"].strip()]
+    effort = AGY_EFFORT.get(conf["assistant_reasoning"], "")
+    if effort:
+        # Most of agy's own model ids carry the effort in their suffix already;
+        # this is for the ones that do not.
+        cmd += ["--effort", effort]
+
+    found = {"answer": "", "warning": "", "session": "", "failure": ""}
+
+    def on_event(event):
+        kind = event.get("event")
+        if kind == "init":
+            found["session"] = event.get("conversation_id") or found["session"]
+        elif kind == "step_update":
+            step = event.get("step_update") or {}
+            # A tool is reported twice, once when it starts and once when it is
+            # done; the corner wants the first of those.
+            if (on_stage and step.get("step_type") == "tool"
+                    and step.get("state") == "ACTIVE"):
+                on_stage(_agy_label(step))
+        elif kind == "result":
+            result = event.get("result") or {}
+            found["session"] = result.get("conversation_id") or found["session"]
+            answer = (result.get("response") or "").strip()
+            if result.get("status") == "SUCCESS":
+                found["answer"] = answer
+            else:
+                found["failure"] = answer or t("{service} ended with an error.",
+                                               service="Antigravity")
+
+    code, stderr = _stream(cmd, conf, on_event, should_stop)
+    return _conclude(found, code, stderr, session, "agy")
+
+
+def _agy_label(step):
+    name = step.get("tool_name", "")
+    if name in AGY_TOOLS:
+        return t(AGY_TOOLS[name])
+    if name.startswith("browser_") or name.startswith("capture_browser"):
+        return t("Working in the browser…")
+    if name == "call_mcp_tool":
+        server = (step.get("tool_info") or {}).get("parameters") or {}
+        return t("Using {name}…", name=server.get("server") or "a tool")
+    return t("Using {name}…", name=name or "a tool")
+
+
+def agy_models():
+    """The models Antigravity itself would offer right now, in its own order.
+
+    `agy models` prints one `id<TAB>display name` line per model, so the list
+    is as current as the account behind the CLI. Unlike Codex it asks Google
+    rather than a cache on disk, a couple of seconds the caller spends off the
+    interface thread. A machine without agy, or a call that fails, answers
+    with nothing and the caller keeps its built-in list.
+    """
+    if not shutil.which("agy"):
+        return []
+    try:
+        proc = subprocess.run(["agy", "models"],
+                              capture_output=True, text=True, timeout=30)
+    except (OSError, subprocess.SubprocessError):
+        return []
+    if proc.returncode != 0:
+        return []
+    ids = []
+    for line in (proc.stdout or "").splitlines():
+        model_id, tab, _ = line.partition("\t")
+        if tab and model_id.strip():
+            ids.append(model_id.strip())
+    return ids
+
+
+# --- OpenRouter and OpenCode Go -------------------------------------------
+
+def _ask_chat(name, service, prompt, conf, on_stage):
+    """A plain question and answer, over a chat provider's key.
+
+    No tools, no files, no calendar. It is the fallback for a machine with
+    neither CLI on it, so it says what it knows and nothing else. The
+    conversation is ours to keep here, since there is no session on the other
+    end to resume.
     """
     if on_stage:
         on_stage(t("Thinking…"))
-    history = read_messages("openrouter", conf["assistant_session_minutes"] * 60)
+    history = read_messages(name, conf["assistant_session_minutes"] * 60)
     messages = history + [{"role": "user", "content": prompt}]
+    model = (conf["assistant_openrouter_model"] if name == "openrouter"
+             else conf["assistant_opencode_model"])
+    base_url = (conf["openrouter_base_url"] if name == "openrouter"
+                else conf["opencode_base_url"])
+    key = conf.openrouter_key() if name == "openrouter" else conf.opencode_key()
     try:
         answer = api.chat(
-            messages, conf.openrouter_key(), conf["assistant_openrouter_model"],
-            conf.assistant_prompt(), reasoning=conf["assistant_reasoning"],
-            base_url=conf["openrouter_base_url"],
-            timeout=conf["assistant_timeout"],
+            messages, key, model, conf.assistant_prompt(),
+            reasoning=conf["assistant_reasoning"], base_url=base_url,
+            timeout=conf["assistant_timeout"], provider=name, service=service,
         )
     except api.ApiError as exc:
         raise AssistantError(str(exc)) from exc
-    write_session("openrouter",
+    write_session(name,
                   messages=messages + [{"role": "assistant", "content": answer}])
     return answer, ""
 
@@ -373,14 +556,24 @@ def _stream(cmd, conf, on_event, should_stop):
     Returns (exit code, stderr). Raises Cancelled when the stop was asked for,
     and AssistantError when the clock ran out.
     """
+    # stderr lands in a file rather than a pipe: nobody drains it while stdout
+    # is being read, and a CLI chatty enough on stderr would fill the pipe's
+    # buffer and wedge both of us. A file has no such limit, and is read once
+    # at the end, which is the only moment stderr matters.
+    stderr_file = tempfile.TemporaryFile()
+    # On POSIX the run gets its own session, so that ending it can take down
+    # every subprocess it started, not just the CLI itself.
+    grouped = {"start_new_session": True} if os.name == "posix" else {}
     try:
         proc = subprocess.Popen(
             cmd, cwd=working_dir(conf), stdin=subprocess.DEVNULL,
-            stdout=subprocess.PIPE, stderr=subprocess.PIPE,
+            stdout=subprocess.PIPE, stderr=stderr_file,
             text=True, encoding="utf-8", errors="replace", bufsize=1,
-            creationflags=getattr(subprocess, "CREATE_NO_WINDOW", 0),
+            creationflags=paths.NO_WINDOW,
+            **grouped,
         )
     except OSError as exc:
+        stderr_file.close()
         raise AssistantError(t("Could not run {binary}: {error}",
                                binary=cmd[0], error=exc)) from exc
 
@@ -409,7 +602,7 @@ def _stream(cmd, conf, on_event, should_stop):
             if isinstance(event, dict):
                 on_event(event)
     finally:
-        stderr = _finish(proc)
+        stderr = _finish(proc, stderr_file)
         watchdog.join(timeout=1)
 
     if ended["cancelled"]:
@@ -420,30 +613,42 @@ def _stream(cmd, conf, on_event, should_stop):
     return proc.returncode, stderr
 
 
-def _conclude(found, code, stderr, session, service):
+# Failures that a fresh session cannot cure: an exhausted quota, a signed-out
+# CLI, a network that is down. A resumed run that dies with one of these is
+# reported as what it is, not retried without the session, because the retry
+# would fail the same way after making the user wait through a second run.
+_API_TROUBLE = re.compile(
+    r"(?i)rate.?limit|quota|overloaded|too many requests|credit|billing|"
+    r"insufficient|unauthorized|forbidden|authentication|invalid.{0,8}key|"
+    r"log ?in|logged.?out|network|connection|ECONN|ENOTFOUND|ETIMEDOUT|"
+    r"\b(401|403|429|5\d\d)\b")
+
+
+def _conclude(found, code, stderr, session, name):
     """Turn what the stream said into an answer, or into the reason there is none."""
+    service = SERVICES.get(name, name)
     if code != 0 and not found["answer"]:
-        if session and _session_missing(stderr):
+        # A resumed run that died with nothing to show is treated as the
+        # session being gone, whatever the wording: this code used to look for
+        # "session ... not found" in stderr, but a CLI update or another
+        # language rewords that and the recovery stops working. Retrying costs
+        # one clean start, and cannot loop because the retry resumes nothing.
+        # Recognised API trouble is the exception: it is not the session's
+        # fault, and the retry would only repeat it.
+        blame = last_line(stderr) or found["failure"] or ""
+        if session and not _API_TROUBLE.search(blame):
             raise _SessionGone()
-        raise AssistantError(last_line(stderr) or found["failure"] or t(
+        raise AssistantError(blame or t(
             "{service} exited with code {code}.", service=service, code=code))
     if found["failure"] and not found["answer"]:
         raise AssistantError(found["failure"])
     if not found["answer"]:
         raise AssistantError(t("{service} answered with nothing.", service=service))
     if found["session"]:
-        write_session("claude" if service == "Claude" else "codex", found["session"])
+        write_session(name, found["session"])
     return found["answer"], found["warning"]
 
 
-def _session_missing(stderr):
-    lowered = (stderr or "").lower()
-    if "session" in lowered or "thread" in lowered or "conversation" in lowered:
-        return any(word in lowered for word in ("not found", "no such", "unknown",
-                                                "does not exist", "no conversation"))
-    return False
-
-
 def _watch(proc, deadline, should_stop, ended):
     while proc.poll() is None:
         if should_stop is not None and should_stop():
@@ -454,29 +659,60 @@ def _watch(proc, deadline, should_stop, ended):
             break
         time.sleep(0.25)
     if ended["cancelled"] or ended["timed_out"]:
-        _kill(proc)
+        kill_tree(proc)
 
 
-def _kill(proc):
+def kill_tree(proc):
+    """End the process and everything it started.
+
+    A CLI runs tools as subprocesses of its own, and ending only the CLI would
+    leave those behind, still working on a question nobody is waiting for.
+    Shared with cleanup, which runs the same two programs. Every failure here
+    is swallowed: the process being already gone is the outcome being asked for.
+    """
+    if os.name == "nt":
+        # There is no process group to signal on Windows; taskkill walks the
+        # tree instead. The wait after it is best-effort, so a tree that will
+        # not die does not hang the caller on top of everything else.
+        subprocess.run(
+            ["taskkill", "/T", "/F", "/PID", str(proc.pid)],
+            capture_output=True,
+            creationflags=paths.NO_WINDOW,
+        )
+        try:
+            proc.wait(timeout=3)
+        except (subprocess.TimeoutExpired, OSError):
+            pass
+        return
+    # The Popen was started with start_new_session=True, so the pid names a
+    # whole session to signal. SIGTERM first for a clean exit, SIGKILL for a
+    # tree that ignored it.
+    try:
+        os.killpg(proc.pid, signal.SIGTERM)
+    except (ProcessLookupError, PermissionError, OSError):
+        return
     try:
-        proc.terminate()
         proc.wait(timeout=3)
     except subprocess.TimeoutExpired:
-        proc.kill()
-    except OSError:
-        pass
+        try:
+            os.killpg(proc.pid, signal.SIGKILL)
+        except (ProcessLookupError, PermissionError, OSError):
+            pass
 
 
-def _finish(proc):
-    try:
-        stderr = proc.stderr.read() or ""
-    except (OSError, ValueError):
-        stderr = ""
+def _finish(proc, stderr_file):
     try:
         proc.wait(timeout=5)
     except subprocess.TimeoutExpired:
-        proc.kill()
-    for stream in (proc.stdout, proc.stderr):
+        kill_tree(proc)
+    # Read back what the CLI wrote to its stderr file, decoded leniently: a
+    # dying CLI is exactly the one likely to print something half-encoded.
+    try:
+        stderr_file.seek(0)
+        stderr = stderr_file.read().decode("utf-8", "replace")
+    except (OSError, ValueError):
+        stderr = ""
+    for stream in (proc.stdout, stderr_file):
         try:
             stream.close()
         except OSError:

dikte/audio.py

@@ -31,11 +31,21 @@ import wave
 
 from PyQt6.QtCore import QObject, pyqtSignal
 
+from . import paths
 from .i18n import t
 
-# Console programs started from a windowless process would otherwise each open
-# a console window of their own on Windows.
-NO_WINDOW = getattr(subprocess, "CREATE_NO_WINDOW", 0) if sys.platform == "win32" else 0
+# Squaring a chunk sample by sample in Python is the most expensive thing the
+# level meter does, and it does it for every chunk of every recording. sumprod
+# stays in C for the whole sum; it arrived in 3.12 and the floor here is 3.11,
+# so the plain loop remains as the fallback. Both produce the same integer.
+try:
+    from math import sumprod
+except ImportError:
+    sumprod = None
+
+# See paths.NO_WINDOW; re-exported here because this module's callers and
+# tests have always read it under this name.
+NO_WINDOW = paths.NO_WINDOW
 
 RATE = 16000
 CHANNELS = 1
@@ -79,12 +89,15 @@ class Recorder(QObject):
 
     level = pyqtSignal(float)              # 0.0 - 1.0, for the waveform
     stopped = pyqtSignal(str, float, object)  # wav path, duration (s), per-chunk RMS
+    died = pyqtSignal()                    # the capture quit mid-recording
     failed = pyqtSignal(str)
 
     def __init__(self, parent=None):
         super().__init__(parent)
         self._proc = None
         self._thread = None
+        self._log = None
+        self._run = None
         self._buffer = bytearray()
         self._rms = []
         self._cancelled = False
@@ -126,12 +139,17 @@ class Recorder(QObject):
             self.failed.emit(t(sound().missing))
             return
 
+        # The recorder keeps talking to stderr for as long as it runs; a pipe
+        # nobody drains would eventually block it, so it writes to a file.
+        self._drop_log()
+        self._log = tempfile.TemporaryFile()
         try:
             self._proc = subprocess.Popen(
-                cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, bufsize=0,
+                cmd, stdout=subprocess.PIPE, stderr=self._log, bufsize=0,
                 creationflags=NO_WINDOW,
             )
         except OSError as exc:
+            self._drop_log()
             self.failed.emit(t("Could not start recording: {error}", error=exc))
             return
 
@@ -141,15 +159,21 @@ class Recorder(QObject):
         self._stopping = False
         self._paused = False
         self._max_bytes = int(max_seconds * RATE * SAMPLE_WIDTH * CHANNELS)
-        self._thread = threading.Thread(target=self._pump, daemon=True)
+        # The pump is handed this run's objects rather than reading them off
+        # self, and a token to say whose run it still is: a pump that outlives
+        # its 2 s join must not touch the recording that comes after it.
+        self._run = object()
+        self._thread = threading.Thread(
+            target=self._pump, daemon=True,
+            args=(self._run, self._proc, self._proc.stdout,
+                  self._buffer, self._rms, self._max_bytes),
+        )
         self._thread.start()
 
-    def _pump(self):
-        proc = self._proc
-        stdout = proc.stdout
+    def _pump(self, run, proc, stdout, buffer, rms, max_bytes):
         try:
             while True:
-                chunk = stdout.read(CHUNK_BYTES)
+                chunk = _read_exact(stdout, CHUNK_BYTES)
                 if not chunk:
                     break
                 if self._paused:
@@ -157,33 +181,58 @@ class Recorder(QObject):
                     # nobody empties fills up, and the capture program blocks on
                     # a full one instead of waiting quietly for the resume.
                     continue
-                peak, rms = chunk_levels(chunk)
+                peak, chunk_rms = chunk_levels(chunk)
                 with self._lock:
-                    self._buffer.extend(chunk)
-                    self._rms.append(rms)
-                    too_long = len(self._buffer) >= self._max_bytes
+                    buffer.extend(chunk)
+                    rms.append(chunk_rms)
+                    too_long = len(buffer) >= max_bytes
+                if self._run is not run:
+                    # This recording was given up on; whatever happens now
+                    # belongs to the run that replaced it, not to this one.
+                    return
                 self.level.emit(peak)
                 if too_long:
                     self._terminate()
                     break
         except (OSError, ValueError):
             pass
+        if self._run is not run:
+            return
+        if self._stopping or self._cancelled:
+            return
+        with self._lock:
+            captured = bool(buffer)
+        if captured:
+            # Sound had already arrived and nobody asked it to end: the device
+            # went away, or the recorder fell over mid-dictation. That has to
+            # be said while there is still something worth keeping.
+            self.died.emit()
+            return
         # Nobody asked it to end and it captured nothing: the recorder is not
         # installed properly, or the device was refused. Said out loud here,
         # because stop() would otherwise report it as a recording that was too
         # short, which sends the user looking in the wrong place.
-        with self._lock:
-            captured = bool(self._buffer)
-        if self._stopping or self._cancelled or captured:
-            return
-        try:
-            detail = proc.stderr.read().decode("utf-8", "replace").strip()
-        except (AttributeError, OSError):
-            detail = ""
-        self.failed.emit(t(
-            "Audio recorder stopped before receiving sound: {error}",
-            error=detail or f"exit code {proc.returncode}",
-        ))
+        detail = self._error_tail()
+        # poll() first, because returncode stays None until somebody reaps the
+        # process, and "exit code None" answers nothing.
+        code = proc.poll()
+        if not detail and code is not None:
+            detail = f"exit code {code}"
+        if detail:
+            self.failed.emit(t(
+                "Audio recorder stopped before receiving sound: {error}",
+                error=detail,
+            ))
+        else:
+            self.failed.emit(t("Audio recorder stopped before receiving sound"))
+
+    def _error_tail(self):
+        log = self._log
+        return _last_log_line(log) if log is not None else ""
+
+    def _drop_log(self):
+        log, self._log = self._log, None
+        _close_log(log)
 
     def _terminate(self):
         self._stopping = True
@@ -195,7 +244,10 @@ class Recorder(QObject):
             except (subprocess.TimeoutExpired, OSError):
                 try:
                     proc.kill()
-                except OSError:
+                    # Reaped even after a kill, or the child stays a zombie
+                    # holding its slot in the process table.
+                    proc.wait(timeout=1)
+                except (subprocess.TimeoutExpired, OSError):
                     pass
 
     def cancel(self):
@@ -205,6 +257,8 @@ class Recorder(QObject):
             self._thread.join(timeout=2)
         self._thread = None
         self._proc = None
+        self._run = None
+        self._drop_log()
         with self._lock:
             self._buffer = bytearray()
 
@@ -217,7 +271,11 @@ class Recorder(QObject):
             self._thread.join(timeout=2)
         self._thread = None
         self._proc = None
+        self._run = None
+        self._drop_log()
 
+        # The same buffer object the pump was handed, harvested under the same
+        # lock it appends with.
         with self._lock:
             pcm = bytes(self._buffer)
             rms = list(self._rms)
@@ -231,7 +289,13 @@ class Recorder(QObject):
             self.failed.emit(t("Recording too short, speak for at least 0.3 s"))
             return
 
-        path = write_wav(pcm)
+        try:
+            path = write_wav(pcm)
+        except (OSError, wave.Error) as exc:
+            # A full disk or an unwritable temp directory costs this recording
+            # either way; a message beats a traceback in the journal.
+            self.failed.emit(t("Could not write the recording: {error}", error=exc))
+            return
         self.stopped.emit(path, frames / RATE, rms)
 
 
@@ -284,6 +348,7 @@ class MeetingRecorder(QObject):
     def __init__(self, parent=None):
         super().__init__(parent)
         self._procs = []
+        self._interrupted = set()
         self._thread = None
         self._wav = None
         self._logs = []
@@ -336,6 +401,7 @@ class MeetingRecorder(QObject):
             # nobody drains would eventually block it, so it writes to a file.
             self._logs = [tempfile.TemporaryFile() for _ in commands]
             self._procs = []
+            self._interrupted = set()
             for command, log in zip(commands, self._logs):
                 self._procs.append(subprocess.Popen(
                     command, stdout=subprocess.PIPE, stderr=log, bufsize=0,
@@ -440,14 +506,20 @@ class MeetingRecorder(QObject):
             try:
                 _interrupt(proc)
             except OSError:
-                pass
+                continue
+            # ffmpeg reports being interrupted as a failure; stop() needs to
+            # know which exits were our own doing and which were real deaths.
+            self._interrupted.add(proc)
         for proc in running:
             try:
                 proc.wait(timeout=2)
             except (subprocess.TimeoutExpired, OSError):
                 try:
                     proc.kill()
-                except OSError:
+                    # Reaped even after a kill, or the child stays a zombie
+                    # holding its slot in the process table.
+                    proc.wait(timeout=1)
+                except (subprocess.TimeoutExpired, OSError):
                     pass
 
     def _close_file(self):
@@ -460,24 +532,23 @@ class MeetingRecorder(QObject):
                 pass
 
     def _error_tail(self):
-        tails = []
-        for log in self._logs:
-            try:
-                log.seek(0)
-                text = log.read().decode("utf-8", "replace").strip()
-            except OSError:
-                continue
-            lines = [line for line in text.splitlines() if line.strip()]
-            if lines:
-                tails.append(lines[-1])
-        return " | ".join(tails)
+        tails = [_last_log_line(log) for log in self._logs]
+        return " | ".join(tail for tail in tails if tail)
 
     def _finish_process(self):
         self._terminate()
         if self._thread:
             self._thread.join(timeout=3)
         self._thread = None
-        codes = [proc.poll() for proc in self._procs]
+        codes = []
+        for proc in self._procs:
+            code = proc.poll()
+            # A nonzero exit from a process we interrupted ourselves is ffmpeg
+            # complaining about our own stop; one that had already died on its
+            # own keeps its code, because that one is the story.
+            if code and proc in self._interrupted:
+                code = 0
+            codes.append(code)
         code = next((value for value in codes if value), 0)
         self._procs = []
         self._close_file()
@@ -534,13 +605,30 @@ class MeetingRecorder(QObject):
 
     def _drop_log(self):
         for log in self._logs:
-            try:
-                log.close()
-            except OSError:
-                pass
+            _close_log(log)
         self._logs = []
 
 
+def _last_log_line(log):
+    """The last thing a recorder said before it ended, or ''."""
+    try:
+        log.seek(0)
+        text = log.read().decode("utf-8", "replace").strip()
+    except (OSError, ValueError):
+        return ""
+    lines = [line for line in text.splitlines() if line.strip()]
+    return lines[-1] if lines else ""
+
+
+def _close_log(log):
+    if log is None:
+        return
+    try:
+        log.close()
+    except OSError:
+        pass
+
+
 def chunk_levels(chunk):
     """(peak, rms) in 0..1. Peak drives the waveform, RMS drives the silence check."""
     samples = array.array("h")
@@ -549,7 +637,9 @@ def chunk_levels(chunk):
         return 0.0, 0.0
     samples.frombytes(chunk[:usable])
     peak = max(abs(min(samples)), abs(max(samples))) / 32768.0
-    rms = math.sqrt(sum(s * s for s in samples) / len(samples)) / 32768.0
+    power = (sumprod(samples, samples) if sumprod is not None
+             else sum(s * s for s in samples))
+    rms = math.sqrt(power / len(samples)) / 32768.0
     return min(1.0, peak), min(1.0, rms)
 
 
@@ -638,6 +728,13 @@ MERGE_FILTER = (
 )
 
 
+# Whether pw-record takes --raw, asked of the binary once per process: the
+# probe costs a subprocess, and the answer cannot change under a running
+# application. Kept here rather than inside _pw_record_raw_option so the probe
+# itself stays testable against different binaries.
+_PW_RAW = None
+
+
 def _pulse_record(target):
     """parec, or pw-record where PulseAudio's tools were left out.
 
@@ -645,6 +742,7 @@ def _pulse_record(target):
     service, and its source names are the same ones shown by list_sources().
     Keep pw-record as the fallback for minimal native-PipeWire installations.
     """
+    global _PW_RAW
     if shutil.which("parec"):
         cmd = [
             "parec", "--record", "--raw", f"--rate={RATE}",
@@ -660,8 +758,10 @@ def _pulse_record(target):
             cmd.append(f"--device={target}")
         return cmd
     if shutil.which("pw-record"):
+        if _PW_RAW is None:
+            _PW_RAW = _pw_record_raw_option()
         cmd = [
-            "pw-record", *_pw_record_raw_option(), f"--rate={RATE}",
+            "pw-record", *_PW_RAW, f"--rate={RATE}",
             f"--channels={CHANNELS}", "--format=s16",
         ]
         if target:
@@ -682,8 +782,11 @@ def _pw_record_raw_option():
     that line, so ask the installed binary which form it understands.
     """
     try:
+        # utf-8 spelled out: subprocess otherwise decodes with the locale's
+        # codec, and help text through a codec it was not written in raises.
         result = subprocess.run(
-            ["pw-record", "--help"], capture_output=True, text=True, timeout=2
+            ["pw-record", "--help"], capture_output=True, text=True,
+            encoding="utf-8", errors="replace", timeout=2,
         )
         help_text = (result.stdout or "") + (result.stderr or "")
     except (subprocess.SubprocessError, OSError):
@@ -708,9 +811,12 @@ def _pactl_sources():
     if not shutil.which("pactl"):
         return []
     try:
+        # utf-8 spelled out: device descriptions carry whatever alphabet the
+        # machine speaks, and the locale's codec is not always able to say so.
         out = subprocess.run(
             ["pactl", "-f", "json", "list", "sources"],
-            capture_output=True, text=True, timeout=5, check=True,
+            capture_output=True, text=True, encoding="utf-8", errors="replace",
+            timeout=5, check=True,
         ).stdout
         return json.loads(out)
     except (subprocess.SubprocessError, OSError, json.JSONDecodeError):
@@ -739,7 +845,8 @@ def _pulse_default_output():
     try:
         sink = subprocess.run(
             ["pactl", "get-default-sink"],
-            capture_output=True, text=True, timeout=5, check=True,
+            capture_output=True, text=True, encoding="utf-8", errors="replace",
+            timeout=5, check=True,
         ).stdout.strip()
     except (subprocess.SubprocessError, OSError):
         return ""

dikte/cleanup.py

@@ -1,7 +1,8 @@
 """Who rewrites the transcript once it has been heard.
 
-Normally a small model on OpenRouter: one request, a second, a few tenths of a
-cent. A machine with Claude Code or Codex on it is already paying for a model
+Normally a small model over one HTTP request: a second, and a few tenths of a
+cent on OpenRouter or nothing at all on Google AI Studio's free tier. A machine
+with Claude Code, Codex or Antigravity on it is already paying for a model
 though, and the subscription that answers "put that in my calendar on Thursday"
 can just as well take the "eee"s out of a sentence. No second key, no second
 bill. It costs seconds rather than one, because a CLI opens a whole session to
@@ -10,7 +11,11 @@ do it, which is the trade.
 Whoever does it, the job is the same one: no tools, no files, no memory of the
 last dictation. There is nothing here to look up and nothing to carry over, and
 a transcript is text from a microphone rather than an instruction, so the less
-the agent can reach while it reads one, the better.
+the agent can reach while it reads one, the better. Claude Code is handed an
+empty tool list and Codex a read-only sandbox. Antigravity has neither switch,
+and this is worth saying plainly rather than implying parity: there the
+transcript is read by an agent that could go and do something. What can be done
+is done — a project of its own, the home directory, and its slash commands off.
 """
 
 import os
@@ -21,9 +26,10 @@ import tempfile
 from . import api
 from . import assistant
 from . import ggml
+from . import paths
 from .i18n import t
 
-PROVIDERS = ("openrouter", "local", "claude", "codex")
+PROVIDERS = ("openrouter", "gemini", "opencode", "local", "claude", "codex", "agy")
 
 
 class CleanupError(api.ApiError):
@@ -42,7 +48,7 @@ def provider(conf):
 
 def executable(name):
     """The CLI a provider runs, or "" when it needs none."""
-    return {"claude": "claude", "codex": "codex"}.get(name, "")
+    return {"claude": "claude", "codex": "codex", "agy": "agy"}.get(name, "")
 
 
 def model(conf):
@@ -56,13 +62,20 @@ def model(conf):
         # Codex is left on whatever it is set to unless a model is typed in, so
         # here there is only the name of the thing that did it.
         return conf["cleanup_codex_model"].strip() or "codex"
+    if name == "agy":
+        # The same arrangement as Codex, and the same reason for it.
+        return conf["cleanup_agy_model"].strip() or "agy"
+    if name == "gemini":
+        return conf["cleanup_gemini_model"]
+    if name == "opencode":
+        return conf["cleanup_opencode_model"]
     return conf["cleanup_model"]
 
 
 def run(text, conf, system_prompt, timeout=180, aborter=None):
     """Hand the transcript to whoever is set to clean it up.
 
-    `aborter` is only of use to the two that answer over HTTP; a CLI is stopped
+    `aborter` is only of use to the three that answer over HTTP; a CLI is stopped
     between blocks instead, which is close enough when a block is seconds.
     """
     name = provider(conf)
@@ -73,9 +86,23 @@ def run(text, conf, system_prompt, timeout=180, aborter=None):
             base_url=conf["openrouter_base_url"], timeout=timeout,
             aborter=aborter,
         )
+    if name == "gemini":
+        return api.cleanup(
+            text, conf.gemini_key(), conf["cleanup_gemini_model"], system_prompt,
+            reasoning=conf["cleanup_reasoning"],
+            base_url=conf["gemini_base_url"], timeout=timeout,
+            provider="gemini", service="Google AI Studio", aborter=aborter,
+        )
+    if name == "opencode":
+        return api.cleanup(
+            text, conf.opencode_key(), conf["cleanup_opencode_model"], system_prompt,
+            reasoning=conf["cleanup_reasoning"],
+            base_url=conf["opencode_base_url"], timeout=timeout,
+            provider="opencode", service="OpenCode Go", aborter=aborter,
+        )
     if name == "local":
         return _local(text, conf, system_prompt, timeout, aborter)
-    runner = _claude if name == "claude" else _codex
+    runner = {"claude": _claude, "codex": _codex, "agy": _agy}[name]
     return runner(text, conf, system_prompt, timeout)
 
 
@@ -171,6 +198,40 @@ def _codex(text, conf, system_prompt, timeout):
     return answer
 
 
+# --- Antigravity ----------------------------------------------------------
+
+def _agy(text, conf, system_prompt, timeout):
+    # Antigravity takes no system prompt of its own either, so the rules ride in
+    # front of the transcript, kept apart from it so the two are not read as one.
+    body = f"{system_prompt}\n\n---\n\n{_wrap(text)}"
+    cmd = [
+        "agy", "-p", body,
+        "--output-format", "text",       # the answer, and nothing around it
+        # Left to itself agy picks up whichever project it was last in and works
+        # in that project's directory rather than this one. A dictation belongs
+        # to no project, so each one starts on a project of its own.
+        "--new-project",
+        # A transcript that happens to begin with a slash is still a transcript.
+        "--disable-slash-commands",
+        # agy gives up after five minutes of its own accord, which would have it
+        # killed from outside rather than answering.
+        "--print-timeout", f"{timeout}s",
+    ]
+    if conf["cleanup_agy_model"].strip():
+        cmd += ["--model", conf["cleanup_agy_model"].strip()]
+    effort = assistant.AGY_EFFORT.get(conf["cleanup_reasoning"], "")
+    if effort:
+        # agy's own model ids carry the effort in their suffix, so this only
+        # matters for the ones that do not, and for a model typed in by hand.
+        cmd += ["--effort", effort]
+
+    answer = _output(cmd, timeout, "Antigravity")
+    if not answer:
+        raise CleanupError(t("{service} answered with nothing.",
+                             service="Antigravity"))
+    return answer
+
+
 def _read(path):
     try:
         with open(path, encoding="utf-8", errors="replace") as fh:
@@ -194,21 +255,43 @@ def _output(cmd, timeout, service):
             "{binary} not found. Install it, or have OpenRouter clean up "
             "instead, under Settings → API and models.", binary=binary,
         ))
+    # Both streams land in files rather than pipes: nobody drains a pipe while
+    # the process is being waited out, and a CLI chatty enough would fill the
+    # buffer and wedge. And a timeout must end the CLI's tool subprocesses too,
+    # not just the CLI, which subprocess.run's timeout does not do; hence the
+    # own session on POSIX and assistant.kill_tree on the way out.
+    out_file = tempfile.TemporaryFile()
+    err_file = tempfile.TemporaryFile()
+    grouped = {"start_new_session": True} if os.name == "posix" else {}
     try:
-        done = subprocess.run(
-            cmd, cwd=os.path.expanduser("~"), stdin=subprocess.DEVNULL,
-            capture_output=True, text=True, encoding="utf-8", errors="replace",
-            timeout=timeout,
-            creationflags=getattr(subprocess, "CREATE_NO_WINDOW", 0),
-        )
-    except subprocess.TimeoutExpired:
-        raise CleanupError(t("{service} did not finish within {seconds} seconds.",
-                             service=service, seconds=timeout)) from None
-    except OSError as exc:
-        raise CleanupError(t("Could not run {binary}: {error}",
-                             binary=binary, error=exc)) from exc
-    if done.returncode != 0:
-        raise CleanupError(assistant.last_line(done.stderr) or t(
+        try:
+            proc = subprocess.Popen(
+                cmd, cwd=os.path.expanduser("~"), stdin=subprocess.DEVNULL,
+                stdout=out_file, stderr=err_file,
+                creationflags=paths.NO_WINDOW,
+                **grouped,
+            )
+        except OSError as exc:
+            raise CleanupError(t("Could not run {binary}: {error}",
+                                 binary=binary, error=exc)) from exc
+        try:
+            proc.wait(timeout=timeout)
+        except subprocess.TimeoutExpired:
+            assistant.kill_tree(proc)
+            raise CleanupError(t("{service} did not finish within {seconds} seconds.",
+                                 service=service, seconds=timeout)) from None
+        out_file.seek(0)
+        stdout = out_file.read().decode("utf-8", "replace")
+        err_file.seek(0)
+        stderr = err_file.read().decode("utf-8", "replace")
+    finally:
+        for handle in (out_file, err_file):
+            try:
+                handle.close()
+            except OSError:
+                pass
+    if proc.returncode != 0:
+        raise CleanupError(assistant.last_line(stderr) or t(
             "{service} exited with code {code}.",
-            service=service, code=done.returncode))
-    return (done.stdout or "").strip()
+            service=service, code=proc.returncode))
+    return stdout.strip()

dikte/cli.py

@@ -136,21 +136,10 @@ def _ask_instance(opts, cmd, wait=False, **args):
 
 def launch_gui(verb=""):
     """No instance running, so become the application itself."""
-    args = ipc.launcher()
-    if verb:
-        args.append(verb)
-    args.append("--gui")
-    if sys.platform == "win32":
-        # execv on Windows mangles arguments with spaces and would leave the
-        # application tied to this console; start it detached instead.
-        subprocess.Popen(
-            args,
-            creationflags=(subprocess.DETACHED_PROCESS
-                           | subprocess.CREATE_NEW_PROCESS_GROUP),
-            close_fds=True,
-        )
-        sys.exit(0)
-    os.execv(args[0], args)
+    ipc.respawn(([verb] if verb else []) + ["--gui"])
+    # respawn only returns on Windows, where the application was started
+    # detached and this console process's job is over.
+    sys.exit(0)
 
 
 def _not_running(opts):
@@ -227,7 +216,9 @@ def cmd_ask(opts):
         conf["assistant_provider"] = opts.provider
     if opts.model:
         key = {"claude": "assistant_model", "codex": "assistant_codex_model",
-               "openrouter": "assistant_openrouter_model"}[assistant.provider(conf)]
+               "openrouter": "assistant_openrouter_model",
+               "agy": "assistant_agy_model",
+               "opencode": "assistant_opencode_model"}[assistant.provider(conf)]
         conf[key] = opts.model
     if opts.dir:
         conf["assistant_dir"] = opts.dir
@@ -266,7 +257,8 @@ def cmd_ask(opts):
         "cleanup_error": warning,
         "mode": "ask",
         "question": text,
-        "assistant_model": conf["assistant_model"],
+        "assistant": assistant.provider(conf),
+        "assistant_model": assistant.model(conf),
         "raw": text,
         "text": answer,
     })
@@ -519,7 +511,8 @@ def cmd_history_clear(opts):
 
 # --- settings ---------------------------------------------------------------
 
-SECRET_KEYS = ("openai_api_key", "openrouter_api_key")
+SECRET_KEYS = ("openai_api_key", "groq_api_key", "openrouter_api_key",
+               "gemini_api_key", "opencode_api_key")
 
 
 def _mask(key, value):
@@ -709,6 +702,22 @@ def cmd_test_key(opts):
             results[name] = {"ok": True, "message": message}
         except api.ApiError as exc:
             results[name] = {"ok": False, "message": str(exc)}
+    if opts.which in ("gemini", "all"):
+        try:
+            count = len(api.gemini_models(conf.gemini_key(),
+                                          conf["gemini_base_url"]))
+            message = f"connection works, {count} models visible"
+            results["gemini"] = {"ok": True, "message": message}
+        except api.ApiError as exc:
+            results["gemini"] = {"ok": False, "message": str(exc)}
+    if opts.which in ("opencode", "all"):
+        try:
+            count = len(api.openai_models(conf.opencode_key(),
+                                          conf["opencode_base_url"], "OpenCode Go"))
+            results["opencode"] = {"ok": True,
+                                   "message": f"connection works, {count} models visible"}
+        except api.ApiError as exc:
+            results["opencode"] = {"ok": False, "message": str(exc)}
     everything_ok = all(item["ok"] for item in results.values())
     lines = [f"{'✓' if item['ok'] else '✗'} {name}: {item['message']}"
              for name, item in results.items()]
@@ -869,27 +878,60 @@ def cmd_doctor(opts):
     programs = {name: shutil.which(name) or "" for name in wanted if name}
     target = conf.transcribe_target()
     cleaner = cleanup.provider(conf)
+    # What each provider actually needs: the local ones have no key to check,
+    # and marking them by the key they do not use reported every fully local
+    # setup as broken.
+    transcribe_ready = conf.transcribe_ready()
+    # Only the ones that answer over HTTP have a key worth looking at. A CLI
+    # has a program to find instead, and the model on this machine has neither,
+    # so "no key" there has to read as beside the point rather than as one that
+    # has gone missing.
+    cleanup_service, cleanup_key = {
+        "openrouter": ("OpenRouter", conf.openrouter_key()),
+        "gemini": ("Google AI Studio", conf.gemini_key()),
+        "opencode": ("OpenCode Go", conf.opencode_key()),
+    }.get(cleaner, ("", ""))
+    if cleanup_service:
+        cleanup_ready = bool(cleanup_key)
+    elif cleaner == "local":
+        cleanup_ready = conf.local_llm_ready()
+    else:
+        cleanup_ready = bool(programs.get(cleanup.executable(cleaner), ""))
     checks = {
         "programs": programs,
         "transcription": {"provider": target.provider, "model": target.model,
-                          "key": bool(target.api_key)},
+                          "key": bool(target.api_key),
+                          "ready": transcribe_ready},
         "cleanup": {"enabled": conf["cleanup_enabled"], "provider": cleaner,
                     "model": cleanup.model(conf),
-                    "key": bool(conf.openrouter_key())},
+                    "key": bool(cleanup_key) if cleanup_service else None,
+                    "ready": cleanup_ready},
         "agent": {"provider": assistant.provider(conf),
                   "directory": assistant.working_dir(conf)},
         "running": ipc.send("status") is not None,
     }
+    if target.provider == "local":
+        transcribe_line = (f"{'✓' if transcribe_ready else '✗'} {target.service}, "
+                           f"transcribing on {target.model or 'no model yet'}")
+    else:
+        transcribe_line = (f"{'✓' if transcribe_ready else '✗'} {target.service} "
+                           f"key, transcribing on {target.model}")
+    if cleanup_service:
+        cleanup_line = (f"{'✓' if cleanup_ready else '✗'} {cleanup_service} key, "
+                        f"cleaning up on {cleanup.model(conf)}")
+    elif cleaner == "local":
+        cleanup_line = (f"{'✓' if cleanup_ready else '✗'} Local model, "
+                        f"cleaning up on {conf['local_llm_model'] or 'no model yet'}")
+    else:
+        # Cleanup on a CLI needs no key, so what is checked is the program.
+        cleanup_line = (f"{'✓' if cleanup_ready else '✗'} "
+                        f"{cleanup.executable(cleaner)}, cleaning up on "
+                        f"{cleanup.model(conf)}")
     lines = [f"{'✓' if path else '✗'} {name:14} {path or 'not on your PATH'}"
              for name, path in programs.items()]
     lines += [
-        f"{'✓' if target.api_key else '✗'} {target.service} key, transcribing on "
-        f"{target.model}",
-        # Cleanup on a CLI needs no key, so what is checked is the program.
-        (f"{'✓' if conf.openrouter_key() else '✗'} OpenRouter key, cleaning up on "
-         f"{conf['cleanup_model']}") if cleaner == "openrouter" else
-        (f"{'✓' if programs[cleanup.executable(cleaner)] else '✗'} "
-         f"{cleanup.executable(cleaner)}, cleaning up on {cleanup.model(conf)}"),
+        transcribe_line,
+        cleanup_line,
         f"{'✓' if checks['running'] else '·'} application "
         + ("running" if checks["running"] else "not running"),
     ]
@@ -977,7 +1019,7 @@ def build_parser():
     ask = leaf(subs, "ask", "put a command to the agent")
     ask.add_argument("text", nargs="*", help="the command; read from stdin, or "
                                              "recorded when there is none")
-    ask.add_argument("--provider", choices=("claude", "codex", "openrouter"),
+    ask.add_argument("--provider", choices=assistant.PROVIDERS,
                      help="just for this run")
     ask.add_argument("--model", help="just for this run")
     ask.add_argument("--dir", help="working directory, just for this run")
@@ -1012,13 +1054,14 @@ def build_parser():
     transcribe.set_defaults(func=cmd_transcribe)
 
     # --- meetings ---------------------------------------------------------
-    for name, help_text in (("meeting", "start a meeting, or end it and write it up"),
-                            ("meeting-cancel", "")):
-        page = leaf(subs, name, help_text)
-        page.add_argument("--wait", action="store_true",
-                          help="wait for the minutes to be written")
-        page.add_argument("--timeout", type=float, default=0)
-        page.set_defaults(func=cmd_meeting)
+    page = leaf(subs, "meeting", "start a meeting, or end it and write it up")
+    page.add_argument("--wait", action="store_true",
+                      help="wait for the minutes to be written")
+    page.add_argument("--timeout", type=float, default=0)
+    page.set_defaults(func=cmd_meeting)
+    # No --wait here: a cancel is answered on the spot, and a flag the server
+    # would ignore is a promise the help text cannot keep.
+    leaf(subs, "meeting-cancel", "").set_defaults(func=cmd_meeting)
 
     meetings = leaf(subs, "meetings", "recorded meetings and their minutes")
     inner = meetings.add_subparsers(dest="meetings", metavar="")
@@ -1037,12 +1080,14 @@ def build_parser():
     delete.add_argument("which", nargs="+")
     delete.set_defaults(func=cmd_meetings_delete)
     for name, verb, help_text in (("start", "meeting-start", "start recording one"),
-                                  ("stop", "meeting-stop", "end it and write it up"),
-                                  ("cancel", "meeting-cancel", "throw the recording away")):
+                                  ("stop", "meeting-stop", "end it and write it up")):
         page = leaf(inner, name, help_text)
         page.add_argument("--wait", action="store_true")
         page.add_argument("--timeout", type=float, default=0)
         page.set_defaults(func=cmd_meeting, verb=verb)
+    # cancel takes no --wait: see the top-level meeting-cancel.
+    leaf(inner, "cancel", "throw the recording away").set_defaults(
+        func=cmd_meeting, verb="meeting-cancel")
 
     # --- history ----------------------------------------------------------
     history = leaf(subs, "history", "past dictations")
@@ -1100,7 +1145,7 @@ def build_parser():
     models.set_defaults(func=cmd_models)
     test = leaf(subs, "test-key", "check the API keys")
     test.add_argument("which", nargs="?", default="all",
-                      choices=("all", *cfg.TRANSCRIBERS))
+                      choices=("all", *cfg.TRANSCRIBERS, "gemini", "opencode"))
     test.set_defaults(func=cmd_test_key)
     leaf(subs, "doctor", "keys, programs, and what is missing").set_defaults(func=cmd_doctor)
 
@@ -1153,6 +1198,16 @@ def _needs_subcommand(parser):
 
 def run(argv):
     global _app
+    # A redirected stdout on Windows falls back to the console codepage,
+    # strict, and a transcript (or doctor's ✓) with a character outside it
+    # would then fail the run after the work succeeded. Interactively nothing
+    # changes: the console is written through its own Unicode API.
+    if sys.platform == "win32" and not os.environ.get("PYTHONIOENCODING"):
+        for stream in (sys.stdout, sys.stderr):
+            try:
+                stream.reconfigure(errors="replace")
+            except (AttributeError, OSError):
+                pass
     parser = build_parser()
     opts = parser.parse_args(argv)
     # No verb at all is the plain `dikte`, which means the settings window.

dikte/config.py

@@ -5,6 +5,8 @@ import hashlib
 import json
 import os
 import sys
+import threading
+import time
 
 from . import api
 from . import ggml
@@ -385,10 +387,19 @@ DEFAULTS = {
     "groq_base_url": "https://api.groq.com/openai/v1",
     "openrouter_api_key": "",
     "openrouter_base_url": "https://openrouter.ai/api/v1",
+    "gemini_api_key": "",
+    # Google's OpenAI-compatible endpoint. Cleanup only: there is no
+    # /audio/transcriptions behind it, so it is not one of the TRANSCRIBERS.
+    "gemini_base_url": "https://generativelanguage.googleapis.com/v1beta/openai",
+    "opencode_api_key": "",
+    "opencode_base_url": "https://opencode.ai/zen/go/v1",
     "transcribe_provider": "local",  # "local", or a key of TRANSCRIBERS
     "transcribe_model": "gpt-4o-transcribe",           # used when provider is openai
     "groq_transcribe_model": "whisper-large-v3-turbo",
     "openrouter_transcribe_model": "openai/gpt-4o-transcribe",
+    # What a timestamped run (subtitles) asks OpenRouter for: not every model
+    # there returns segment times. Empty -> openai/whisper-1.
+    "openrouter_file_model": "",
     "language": "tr",
     "transcribe_prompt": "",
 
@@ -410,6 +421,9 @@ DEFAULTS = {
     "cleanup_model": "google/gemini-3.5-flash-lite",
     "cleanup_claude_model": "haiku",   # Claude Code: an alias, or a full model id
     "cleanup_codex_model": "",         # empty -> whatever Codex is set to
+    "cleanup_gemini_model": "gemini-3.5-flash-lite",
+    "cleanup_agy_model": "",           # empty -> whatever Antigravity is set to
+    "cleanup_opencode_model": "deepseek-v4-flash",
     "cleanup_reasoning": "",        # empty -> whatever the model does by default
 
     # --- llama.cpp, on this machine -----------------------------------------
@@ -455,6 +469,7 @@ DEFAULTS = {
     "pause_shortcut": "",
     "evdev_hotkey": False,
     "overlay_corner": "bottom-left",
+    "overlay_screen": "",
     "keep_audio": False,
     "history_limit": 200,
     # A look at the releases page once a day, and nothing more than a look:
@@ -482,12 +497,14 @@ DEFAULTS = {
 
     # --- speaking a command to an agent -------------------------------------
     "assistant_shortcut": "",       # empty -> tray only
-    "assistant_provider": "claude",  # claude | codex | openrouter
+    "assistant_provider": "claude",  # claude | codex | agy | openrouter
     "assistant_model": "sonnet",    # Claude Code: an alias, or a full model id
     "assistant_permission_mode": "auto",
     "assistant_codex_model": "",    # empty -> whatever Codex is set to
     "assistant_codex_sandbox": "workspace-write",
     "assistant_openrouter_model": "google/gemini-3.5-flash",
+    "assistant_agy_model": "",      # empty -> whatever Antigravity is set to
+    "assistant_opencode_model": "deepseek-v4-flash",
     "assistant_reasoning": "",      # empty -> the model's own default
     "assistant_dir": "",            # empty -> the home directory
     "assistant_prompt": "",         # empty -> language-specific default
@@ -528,6 +545,30 @@ TRANSCRIBERS = {
                               "openrouter_base_url", "openrouter_transcribe_model"),
 }
 
+# One lock for the history file and the meeting index both, rather than one
+# each: the files are a few kilobytes, the writes happen a handful of times an
+# hour, and a second lock would only add a way to take them in the wrong order.
+_FILES_LOCK = threading.Lock()
+
+
+def _replace_with_retry(tmp, target):
+    """The atomic swap, tried again briefly when the target is held.
+
+    On Windows an antivirus or sync tool opens a freshly written file to look
+    at it, and a rename over the file fails for as long as it is held. The
+    hold lasts milliseconds, so three tries with a short sleep cover it; a
+    file held longer than that is a real error and is raised as one.
+    """
+    for attempt in range(3):
+        try:
+            tmp.replace(target)
+            return
+        except OSError:
+            if attempt == 2:
+                raise
+            time.sleep(0.05)
+
+
 # Corners used to be stored with Turkish names.
 _CORNER_MIGRATION = {
     "sol-alt": "bottom-left", "sağ-alt": "bottom-right",
@@ -548,7 +589,19 @@ class Config:
                 self.data.update({k: v for k, v in stored.items() if k in DEFAULTS})
         except FileNotFoundError:
             pass
-        except (json.JSONDecodeError, OSError) as exc:
+        except json.JSONDecodeError as exc:
+            # Set aside rather than left in place: the next save would write
+            # the defaults over it, and whatever broke the file deserves to
+            # still be there to look at. Best effort; a rename that fails
+            # changes nothing about falling back to the defaults.
+            broken = CONFIG_FILE.with_suffix(".json.broken")
+            try:
+                CONFIG_FILE.replace(broken)
+            except OSError:
+                pass
+            print(f"dikte: could not read settings ({exc}), using defaults; "
+                  f"the unreadable file was kept as {broken}")
+        except OSError as exc:
             print(f"dikte: could not read settings ({exc}), using defaults")
         self.data["overlay_corner"] = _CORNER_MIGRATION.get(
             self.data["overlay_corner"], self.data["overlay_corner"]
@@ -563,8 +616,13 @@ class Config:
         tmp = CONFIG_FILE.with_suffix(".json.tmp")
         with open(tmp, "w", encoding="utf-8") as fh:
             json.dump(self.data, fh, ensure_ascii=False, indent=2)
+            # Pushed to the disk before the rename: swapping in a file that
+            # still lives in the page cache turns a power cut into a settings
+            # wipe, which the atomic replace exists to prevent.
+            fh.flush()
+            os.fsync(fh.fileno())
         os.chmod(tmp, 0o600)
-        tmp.replace(CONFIG_FILE)
+        _replace_with_retry(tmp, CONFIG_FILE)
         i18n.set_language(self.data["ui_language"])
 
     def __getitem__(self, key):
@@ -589,6 +647,12 @@ class Config:
     def openrouter_key(self):
         return self.api_key("openrouter_api_key")
 
+    def gemini_key(self):
+        return self.api_key("gemini_api_key")
+
+    def opencode_key(self):
+        return self.api_key("opencode_api_key")
+
     def transcribe_target(self):
         """Key, endpoint and model for whichever provider does speech to text.
 
@@ -608,8 +672,9 @@ class Config:
             # to land on rather than reading it from there.
             name = "openai"
         who = TRANSCRIBERS[name]
+        file_model = self["openrouter_file_model"] if name == "openrouter" else ""
         return api.Target(name, who.service, self.api_key(who.key),
-                          self[who.url], self[who.model])
+                          self[who.url], self[who.model], file_model.strip())
 
     def transcribe_ready(self):
         """Whether speech to text could run right now, without opening Settings."""
@@ -730,12 +795,22 @@ def default_assistant_prompt():
 
 def append_history(entry):
     DATA_DIR.mkdir(parents=True, exist_ok=True)
-    with open(HISTORY_FILE, "a", encoding="utf-8") as fh:
-        fh.write(json.dumps(entry, ensure_ascii=False) + "\n")
+    with _FILES_LOCK:
+        with open(HISTORY_FILE, "a", encoding="utf-8") as fh:
+            fh.write(json.dumps(entry, ensure_ascii=False) + "\n")
 
 
 def read_history(limit=None):
     """Newest last. A limit of None (or 0) reads the whole file."""
+    # Locked even though the rewrites are atomic: it costs nothing, and a read
+    # that waits out a rewrite in flight hands back the settled file rather
+    # than whichever side of the swap it happened to land on.
+    with _FILES_LOCK:
+        return _read_history(limit)
+
+
+def _read_history(limit=None):
+    """The body of read_history, for callers already holding the lock."""
     try:
         with open(HISTORY_FILE, encoding="utf-8") as fh:
             lines = fh.readlines()
@@ -758,40 +833,66 @@ def _write_history(lines):
     tmp = HISTORY_FILE.with_suffix(".jsonl.tmp")
     with open(tmp, "w", encoding="utf-8") as fh:
         fh.writelines(lines)
-    tmp.replace(HISTORY_FILE)
+        fh.flush()
+        os.fsync(fh.fileno())
+    _replace_with_retry(tmp, HISTORY_FILE)
 
 
 def trim_history(limit):
     """Drop the oldest entries once the file passes `limit` rows. 0 means keep all."""
     if not limit or limit < 0:
         return
-    try:
-        with open(HISTORY_FILE, encoding="utf-8") as fh:
-            lines = fh.readlines()
-    except OSError:
-        return
-    if len(lines) <= limit:
-        return
-    _write_history(lines[-limit:])
+    # Read and rewrite under one lock, so a dictation appended in between the
+    # two is not erased by a rewrite that never saw it.
+    with _FILES_LOCK:
+        try:
+            with open(HISTORY_FILE, encoding="utf-8") as fh:
+                lines = fh.readlines()
+        except OSError:
+            return
+        if len(lines) <= limit:
+            return
+        _write_history(lines[-limit:])
 
 
 def _row_key(row):
     return json.dumps(row, ensure_ascii=False, sort_keys=True)
 
 
+def amend_history(entry, **changes):
+    """Patch one entry in place, matched on its whole content like delete_history.
+
+    For the caller that learns something after its row is already written: the
+    row goes in before the paste is attempted, and a paste that then fails
+    still has to end up in the record. None when the row is gone, which a trim
+    in between can legitimately make true."""
+    wanted = _row_key(entry)
+    with _FILES_LOCK:
+        rows = _read_history()
+        for row in rows:
+            if _row_key(row) == wanted:
+                row.update(changes)
+                _write_history([json.dumps(r, ensure_ascii=False) + "\n"
+                                for r in rows])
+                return row
+    return None
+
+
 def delete_history(rows):
     """Remove the given entries, matched on their whole content rather than on a
     line number: the worker may have appended a new one since the list was read."""
     doomed = {_row_key(row) for row in rows}
     if not doomed:
         return
-    kept = [json.dumps(row, ensure_ascii=False) + "\n"
-            for row in read_history() if _row_key(row) not in doomed]
-    _write_history(kept)
+    with _FILES_LOCK:
+        kept = [json.dumps(row, ensure_ascii=False) + "\n"
+                for row in _read_history() if _row_key(row) not in doomed]
+        _write_history(kept)
 
 
 def clear_history():
-    HISTORY_FILE.unlink(missing_ok=True)
+    with _FILES_LOCK:
+        HISTORY_FILE.unlink(missing_ok=True)
 
 
 # --- meetings -------------------------------------------------------------
@@ -807,6 +908,12 @@ def meeting_paths(base):
 
 def read_meetings():
     """Newest last."""
+    with _FILES_LOCK:
+        return _read_meetings()
+
+
+def _read_meetings():
+    """The body of read_meetings, for callers already holding the lock."""
     try:
         with open(MEETINGS_FILE, encoding="utf-8") as fh:
             lines = fh.readlines()
@@ -829,29 +936,33 @@ def _write_meetings(rows):
     with open(tmp, "w", encoding="utf-8") as fh:
         for row in rows:
             fh.write(json.dumps(row, ensure_ascii=False) + "\n")
-    tmp.replace(MEETINGS_FILE)
+        fh.flush()
+        os.fsync(fh.fileno())
+    _replace_with_retry(tmp, MEETINGS_FILE)
 
 
 def save_meeting(entry):
     """Insert the row, or replace the one with the same base."""
-    rows = read_meetings()
-    for index, row in enumerate(rows):
-        if row["base"] == entry["base"]:
-            rows[index] = entry
-            break
-    else:
-        rows.append(entry)
-    _write_meetings(rows)
+    with _FILES_LOCK:
+        rows = _read_meetings()
+        for index, row in enumerate(rows):
+            if row["base"] == entry["base"]:
+                rows[index] = entry
+                break
+        else:
+            rows.append(entry)
+        _write_meetings(rows)
 
 
 def update_meeting(base, **changes):
     """Patch one row and hand it back, or None when it is gone."""
-    rows = read_meetings()
-    for row in rows:
-        if row["base"] == base:
-            row.update(changes)
-            _write_meetings(rows)
-            return row
+    with _FILES_LOCK:
+        rows = _read_meetings()
+        for row in rows:
+            if row["base"] == base:
+                row.update(changes)
+                _write_meetings(rows)
+                return row
     return None
 
 
@@ -860,7 +971,9 @@ def delete_meetings(bases):
     doomed = set(bases)
     if not doomed:
         return
-    _write_meetings([row for row in read_meetings() if row["base"] not in doomed])
+    with _FILES_LOCK:
+        _write_meetings([row for row in _read_meetings()
+                         if row["base"] not in doomed])
     for base in doomed:
         for path in meeting_paths(base):
             try:

dikte/filetranscribe.py

@@ -31,6 +31,7 @@ from PyQt6.QtCore import QObject, pyqtSignal
 from . import api
 from . import cleanup
 from . import ggml
+from . import paths
 from .i18n import t
 
 UPLOAD_LIMIT = 24 * 1024 * 1024  # the APIs take 25 MB; leave the form its room
@@ -337,8 +338,11 @@ def _ffmpeg(args, out, aborter=None):
     proc = subprocess.Popen(
         ["ffmpeg", "-nostdin", "-y", *args],
         stdin=subprocess.DEVNULL, stdout=subprocess.PIPE, stderr=subprocess.PIPE,
-        text=True,
-        creationflags=getattr(subprocess, "CREATE_NO_WINDOW", 0),
+        # ffmpeg writes UTF-8 whatever the locale says; read as the Windows
+        # codepage its messages mojibake, and a byte the codepage cannot place
+        # raises from inside communicate itself.
+        text=True, encoding="utf-8", errors="replace",
+        creationflags=paths.NO_WINDOW,
     )
     # A two hour film is a minute of ffmpeg, which is a minute of a Stop button
     # doing nothing unless the abort reaches the process itself.

dikte/ggml.py

@@ -26,6 +26,7 @@ interface already knows how to show.
 
 import atexit
 import collections
+import ctypes
 import ctypes.util
 import hashlib
 import http.client
@@ -63,6 +64,10 @@ MODELS_DIR = DATA_DIR / "models"
 # Loading a large model onto a GPU is the slow part of a start, and on a cold
 # page cache a large LLM read from a spinning disk is slower still.
 STARTUP_TIMEOUT = 180.0
+# A child that loses the bind race fails and exits at once; a model that fails
+# to load takes longer than this to be read in first. The line between "worth
+# another port" and "would fail the same way again" is drawn on time.
+EARLY_EXIT_WINDOW = 5.0
 DOWNLOAD_CHUNK = 1 << 20
 
 # `health` is the path that answers only once the model is in memory. whisper
@@ -72,6 +77,13 @@ Program = collections.namedtuple("Program", "name repo binary health")
 
 WHISPER = Program("whisper", "ggml-org/whisper.cpp", "whisper-server", "")
 LLAMA = Program("llama", "ggml-org/llama.cpp", "llama-server", "/health")
+DIKTE_REPO = "yusufipk/dikte"
+MANAGED_WHISPER_RELEASE = "whisper.cpp-v1.9.3"
+MANAGED_WHISPER_VERSION = "v1.9.3"
+MANAGED_WHISPER_VULKAN = "whisper-bin-ubuntu-vulkan-x64.tar.gz"
+MANAGED_WHISPER_SHA256 = (
+    "c25ca76504144da488eb74441390a7b9aa7ce547e5f2f391cbd831253c9b54d8"
+)
 
 # Where the models are listed. Neither list is written into Dikte: a catalogue
 # in the source means a release of Dikte for every model somebody else
@@ -79,32 +91,139 @@ LLAMA = Program("llama", "ggml-org/llama.cpp", "llama-server", "/health")
 WHISPER_MODELS_REPO = "ggerganov/whisper.cpp"
 LLM_AUTHOR = "ggml-org"
 
+# The file llama.cpp attaches to its version releases in place of the binaries:
+# a line naming the nightly tag those are published under.
+NIGHTLY_TAG = "nightly-tag.txt"
+
 # What the whisper repository holds besides models: Core ML encoders for Apple
 # hardware and the odd loose file.
 WHISPER_PREFIX = "ggml-"
 WHISPER_SUFFIX = ".bin"
+# The mark on the whisper models trained on English alone. They are half of the
+# list, and they belong under the model they are a variant of rather than
+# scattered through it by size.
+ENGLISH_ONLY = ".en"
+
+# Full-precision weights, however they are spelled. Several times the memory of
+# a quantisation of the same model, for a difference dictation and cleanup
+# cannot see, so nothing here ever points at one.
+SIXTEEN_BIT = ("bf16", "f16", "fp16")
+
+# How many bits a weight is stored in, read off the file name. Every one of
+# these lists spells it differently, `q5_1` and `Q4_K_M` and `MXFP4` and
+# `BF16`, and the only part of that anybody choosing between two rows needs is
+# the number. Longest mark first, so `bf16` is not read as `f16`.
+BIT_DEPTHS = (("mxfp4", 4), ("bf16", 16), ("fp16", 16), ("f16", 16),
+              ("q2", 2), ("q3", 3), ("q4", 4), ("q5", 5), ("q6", 6), ("q8", 8))
 
 # What a GGUF repository holds besides the model: mmproj is the vision half of a
-# multimodal model, mtp a draft head for speculative decoding. Neither is a model
-# a server can be started on, and offering them is offering a failure.
-GGUF_SKIP = ("mmproj", "mtp-")
+# multimodal model, and mtp, dflash, dspark and eagle3 are draft heads for
+# speculative decoding. None of them is a model a server can be started on, and
+# they are the small files in the repository, so a list sorted by size puts them
+# at the top where they are likeliest to be clicked.
+GGUF_SKIP = ("mmproj", "mtp-", "dflash-", "dspark-", "eagle3-", "draft-")
 # Big enough for a 12B at Q4 and far past anything cleanup wants; the point is
 # to keep a 400 GB frontier model out of a list somebody might click.
 GGUF_MAX_BYTES = 16 << 30
 
+# Repositories that carry GGUF files but nothing a cleanup server can be started
+# on: a vision or audio tower with no text half worth running, a speech model,
+# and the base models, which continue text rather than following an instruction
+# and answer a cleanup prompt by carrying on writing the transcript.
+# Matched as plain substrings, so every one of these carries its own
+# delimiters: an unanchored "test-" is also inside "Latest-" and would drop a
+# publisher that is perfectly usable.
+LLM_REPO_SKIP = ("-Base-GGUF", "-VL-", "-Vision-", "-Omni-", "-Video-",
+                 "-TTS-", "parakeet", "/test-")
+
+GB = 1 << 30
+
 # Suggestions, not a catalogue: the list itself is fetched, and these are only
-# the rows that float to the top of it. Small instruction-following models,
-# because cleanup is punctuation and filler words rather than anything that
-# wants thinking about.
+# the rows that float to the top of it. Cleanup is punctuation, capitals and
+# filler words rather than anything that wants thinking about, so what it is
+# picked on is instruction following at a size a desktop can spare. Gemma 4
+# scores 94.6 on IFEval at E2B and 96.7 at E4B, and E2B leads here rather than
+# E4B because two points of instruction following is not worth twice the
+# weights on a job that runs while somebody waits for their sentence to appear.
+# SmolLM3 and Gemma 3 are the older pair below them. Qwen3.5 0.8B is for the
+# machines nothing else fits on; it thinks before it answers, which is what the
+# Thinking box in the settings window turns off.
 SUGGESTED_LLM = (
-    "ggml-org/gemma-3-4b-it-GGUF",
     "ggml-org/gemma-4-E2B-it-GGUF",
     "ggml-org/gemma-4-E4B-it-GGUF",
+    "ggml-org/gemma-3-4b-it-GGUF",
     "ggml-org/SmolLM3-3B-GGUF",
+    "ggml-org/Qwen3.5-0.8B-GGUF",
 )
+# Roughly what each of those weighs at the quantisation cleanup would run, to
+# the nearest half gigabyte. Not a catalogue of files: the sizes on the rows
+# come from the publisher, and this only decides which suggestion is offered
+# first on a machine that has room for some of them and not others.
+SUGGESTED_LLM_SIZE = {
+    "ggml-org/gemma-4-E2B-it-GGUF": 3 * GB,
+    "ggml-org/gemma-4-E4B-it-GGUF": 5 * GB,
+    "ggml-org/gemma-3-4b-it-GGUF": 5 * GB // 2,
+    "ggml-org/SmolLM3-3B-GGUF": 2 * GB,
+    "ggml-org/Qwen3.5-0.8B-GGUF": GB // 2,
+}
+
+# What each of them is, in the words somebody choosing between them would
+# use. A repository id says the publisher, the parameter count, the shape of
+# the weights and nothing at all about whether it is the one to click, and
+# `ggml-org/gemma-4-E2B-it-GGUF` reads as four pieces of jargon to everybody
+# who has not been reading model cards all year.
+SUGGESTED_LLM_NOTE = {
+    "ggml-org/gemma-4-E2B-it-GGUF":
+        "Google Gemma 4, the small one. The default: nothing else this size "
+        "follows an instruction as closely, and cleanup is all instruction.",
+    "ggml-org/gemma-4-E4B-it-GGUF":
+        "The same model one size up. A little more accurate, about twice the "
+        "weights and twice the wait.",
+    "ggml-org/gemma-3-4b-it-GGUF":
+        "The previous Gemma. Still good, and the smallest of the Gemmas here.",
+    "ggml-org/SmolLM3-3B-GGUF":
+        "Hugging Face's own small model, for a machine the Gemmas crowd.",
+    "ggml-org/Qwen3.5-0.8B-GGUF":
+        "The smallest of them, for a machine nothing else fits on. It thinks "
+        "before it answers unless Thinking below is off.",
+}
+
 # Turbo at q5_0 is smaller than `small` and better than it, which makes the
-# usual "start small" advice point at the same file as "start good".
+# usual "start small" advice point at the same file as "start good". It is
+# large-v3 with the decoder cut from 32 layers to 4: several times faster, at
+# one to two points of word error in English and about two and a half in the
+# other languages.
 SUGGESTED_WHISPER = "ggml-large-v3-turbo-q5_0.bin"
+# Those two and a half points back, for twice the file and several times the
+# work per second. Only suggested where there is a card to do the work and
+# memory to hold it, because that is where the trade stops costing anything a
+# person waiting for a dictation would notice.
+ACCURATE_WHISPER = "ggml-large-v3-q5_0.bin"
+# What to point at instead on a machine the turbo model would crowd. Same
+# quantisation ladder, one rung down in size and in accuracy.
+SMALL_MACHINE_WHISPER = "ggml-small-q5_1.bin"
+# Under this much system memory, a 600 MB model plus the rest of a desktop is
+# already tight, so the suggestion drops to the smaller one. Over the other,
+# the accurate model is the one to point at.
+SMALL_MACHINE = 4 * GB
+# Fifteen and not sixteen: what the machine reports is what is left after the
+# firmware and the graphics have taken their reservations out of it, and a
+# 16 GB machine answers about 15.4. A threshold written at the number on the
+# box is one no machine sold as that size ever reaches.
+ROOMY_MACHINE = 15 * GB
+# What a model may take of this machine's memory before it is called too big:
+# half of it, less a gigabyte for the context and the runtime around the
+# weights. A rule of thumb rather than a measurement, and deliberately a
+# cautious one, because the failure it is guarding against is a machine that
+# swaps itself to a standstill rather than a model that refuses to load.
+MEMORY_SHARE = 0.5
+MEMORY_OVERHEAD = GB
+# What is left to offer on a machine too small for the sum above to leave
+# anything. Enough for the smallest whisper models and for a sub-billion
+# cleanup model, which is what such a machine can run.
+MEMORY_FLOOR = GB // 2
+# What total_memory() read the one time it asked. None until it has.
+_MEMORY = None
 
 
 class LocalError(Exception):
@@ -193,7 +312,16 @@ def download(item, target, on_progress=None, should_stop=None, require_hash=True
             part.unlink(missing_ok=True)
             raise LocalError(t("{name} does not match its published checksum. "
                                "Nothing was installed.", name=item.name))
-        part.replace(target)
+        try:
+            part.replace(target)
+        except PermissionError as exc:
+            # Windows refuses to replace a file something has open, and a
+            # running server holds its model and its binary open. The bytes
+            # are complete and verified: keeping the .part costs a retry,
+            # deleting it costs the whole download again.
+            raise LocalError(t("{name} downloaded, but the old file is held "
+                               "open by the running server. Stop it and try "
+                               "again.", name=item.name)) from exc
         return True
     except urllib.error.HTTPError as exc:
         part.unlink(missing_ok=True)
@@ -264,26 +392,138 @@ def _wanted_assets(program):
     return (f"bin-ubuntu-{arch}.tar.gz",)
 
 
+def _managed_whisper(program, tag=""):
+    """Whether this machine is one Dikte publishes its own whisper-server for.
+
+    Linux x86_64 with a Vulkan loader on it, and no version asked for by hand:
+    a pinned version is upstream's to answer.
+    """
+    return (not tag and program is WHISPER and sys.platform == "linux"
+            and platform.machine().lower() in ("x86_64", "amd64")
+            and _has_vulkan())
+
+
+def _managed_asset(refresh=False):
+    """The Vulkan whisper-server Dikte builds itself, or None.
+
+    Taken only when the archive's digest is the reviewed one. Anything else,
+    a release that is not there yet, a GitHub that cannot be reached, a file
+    that is not the reviewed bytes, leaves upstream's processor build as the
+    answer, and the install record says which of the two landed.
+    """
+    try:
+        _, assets = hub.release(DIKTE_REPO, MANAGED_WHISPER_RELEASE,
+                                refresh=refresh)
+    except hub.HubError:
+        return None
+    return next((a for a in assets
+                 if a.name.endswith(MANAGED_WHISPER_VULKAN)
+                 and a.sha256 == MANAGED_WHISPER_SHA256), None)
+
+
+def _matching_asset(program, assets):
+    """The archive this machine wants out of one release's files, or None."""
+    for ending in _wanted_assets(program):
+        item = next((a for a in assets if a.name.endswith(ending)), None)
+        if item:
+            return item
+    return None
+
+
+def _pick_asset(program, tag="", refresh=False):
+    """(tag, Item) for the release archive to install. Item is None when there
+    is none for this machine.
+
+    Dikte's own Vulkan whisper-server comes before upstream's where this
+    machine is one it is built for, because whisper.cpp publishes no Vulkan
+    archive for Linux at all.
+
+    A named tag is taken as given. For the newest, what GitHub answers is not
+    always where the builds are: llama.cpp's latest release is a version marker
+    carrying a single nightly-tag.txt, which names the tag the archives are
+    actually attached to, and those are prereleases that "latest" never points
+    at. The pointer is followed when it is there, and when it is not, the newest
+    release that does carry a build for this machine is taken instead.
+    """
+    if _managed_whisper(program, tag):
+        item = _managed_asset(refresh=refresh)
+        if item:
+            return MANAGED_WHISPER_VERSION, item
+    named = bool(tag) and tag != "latest"
+    missing = None
+    try:
+        tag, assets = hub.release(program.repo, tag or "latest", refresh=refresh)
+    except hub.HubError as exc:
+        # A release carrying no files at all is the case the search below exists
+        # for, not a reason to stop before it: the build for this machine may be
+        # attached to a prerelease that "latest" never points at. The failure is
+        # kept rather than dropped, because an unreachable GitHub arrives here
+        # the same way and that one is the message the caller wants.
+        if named:
+            raise
+        missing, assets = exc, []
+    item = _matching_asset(program, assets)
+    if item or named:
+        return tag, item
+    # Best effort from here on: a machine this project publishes nothing for is
+    # not a failed lookup, and the caller's message about that is the useful
+    # one. Whatever goes wrong while looking further leaves it standing.
+    try:
+        pointer = next((a for a in assets if a.name == NIGHTLY_TAG), None)
+        if pointer:
+            nightly = hub.text(pointer.url).strip()
+            if nightly:
+                found, assets = hub.release(program.repo, nightly, refresh=refresh)
+                item = _matching_asset(program, assets)
+                if item:
+                    return found, item
+        for found, assets in hub.releases(program.repo, refresh=refresh):
+            item = _matching_asset(program, assets)
+            if item:
+                return found, item
+    except hub.HubError:
+        pass
+    if missing is not None:
+        raise missing
+    return tag, None
+
+
 def _install_record(program):
     return BIN_DIR / program.name / "installed.json"
 
 
-def installed_program(program):
-    """The binary Dikte downloaded, or "" when there is none that still runs."""
+def _read_record(program):
+    """The install record, or {} however it fails to read."""
     try:
         record = json.loads(_install_record(program).read_text(encoding="utf-8"))
-        path = record.get("binary") or ""
     except (OSError, ValueError):
-        return ""
+        return {}
+    return record if isinstance(record, dict) else {}
+
+
+def installed_program(program):
+    """The binary Dikte downloaded, or "" when there is none that still runs."""
+    path = _read_record(program).get("binary") or ""
     return path if os.path.isfile(path) and os.access(path, os.X_OK) else ""
 
 
 def installed_version(program):
-    try:
-        record = json.loads(_install_record(program).read_text(encoding="utf-8"))
-        return record.get("tag") or ""
-    except (OSError, ValueError):
-        return ""
+    return _read_record(program).get("tag") or ""
+
+
+def vulkan_missing(program):
+    """Whether what Dikte installed is the processor build on a machine the
+    Vulkan one was fetched for.
+
+    The Vulkan whisper-server is a release of Dikte's own, published by hand
+    once the reviewed archive is built, and the install falls back to the
+    upstream processor build whenever that release, the file in it, or its
+    reviewed digest is not there. Nothing is wrong with the fallback except
+    that it is invisible: a graphics card sitting idle looks exactly like a
+    graphics card being used.
+    """
+    return (bool(installed_program(program))
+            and _read_record(program).get("backend") == "processor")
 
 
 def program_path(program, custom=""):
@@ -339,24 +579,34 @@ def _extract(archive, into):
                            name=os.path.basename(str(archive)), error=exc)) from exc
 
 
+def _under(path, root):
+    """Whether `path` lies inside `root`, symlinks and case resolved.
+
+    Resolved on both sides, because the same directory can be reached under
+    two spellings and this answer decides whether a server gets stopped.
+    """
+    try:
+        pathlib.Path(path).resolve().relative_to(pathlib.Path(root).resolve())
+        return True
+    except (OSError, ValueError):
+        return False
+
+
 def install_program(program, tag="", on_progress=None, should_stop=None,
                     refresh=False):
     """Fetch and unpack a release. The path to the binary, or "" when stopped.
 
     `tag` is empty for whatever the project released last, which is the point:
     a version pinned in Dikte's source would mean a release of Dikte every time
-    whisper.cpp has one.
+    whisper.cpp has one. The Linux Vulkan whisper-server is the exception, and
+    _pick_asset says why.
     """
+    managed = _managed_whisper(program, tag)
     try:
-        tag, assets = hub.release(program.repo, tag or "latest", refresh=refresh)
+        tag, item = _pick_asset(program, tag, refresh=refresh)
     except hub.HubError as exc:
         raise LocalError(str(exc)) from exc
 
-    item = None
-    for ending in _wanted_assets(program):
-        item = next((a for a in assets if a.name.endswith(ending)), None)
-        if item:
-            break
     if item is None:
         # Nothing to download and nothing to install for you: whisper.cpp
         # publishes no macOS binary, and Homebrew's whisper-cpp is configured
@@ -374,19 +624,61 @@ def install_program(program, tag="", on_progress=None, should_stop=None,
                            repo=program.repo, tag=tag))
 
     into = BIN_DIR / program.name / tag
-    shutil.rmtree(into, ignore_errors=True)
+    fresh = into.with_name(tag + ".new")
     archive = BIN_DIR / program.name / item.name
+
     try:
         if not download(item, archive, on_progress, should_stop):
             return ""
-        _extract(archive, into)
-        binary = _find_binary(into, _binary_file(program))
-        if binary is None:
-            raise LocalError(t("{name} was not in the download.",
-                               name=program.binary))
-        binary.chmod(binary.stat().st_mode | 0o111)
-        _install_record(program).write_text(
-            json.dumps({"tag": tag, "binary": str(binary)}), encoding="utf-8")
+        try:
+            # Unpacked into a sibling and swapped in only once the binary is
+            # known to be inside: a failure anywhere in here leaves the
+            # previous install, and its record, exactly as they were.
+            shutil.rmtree(fresh, ignore_errors=True)
+            _extract(archive, fresh)
+            binary = _find_binary(fresh, _binary_file(program))
+            if binary is None:
+                raise LocalError(t("{name} was not in the download.",
+                                   name=program.binary))
+            binary.chmod(binary.stat().st_mode | 0o111)
+            # A running server holds its binary open, and Windows will not
+            # delete an open file: whichever of our servers runs out of this
+            # program's directory is stopped here, after the download and the
+            # unpack are known good, so the outage is the swap and not the
+            # whole transfer.
+            for server in SERVERS:
+                current = program_path(server.program,
+                                       server.settings().get("binary", ""))
+                if current and _under(current, BIN_DIR / program.name):
+                    server.stop()
+            if into.exists():
+                try:
+                    shutil.rmtree(into)
+                except OSError as exc:
+                    # Not ignore_errors: silently losing this would rename the
+                    # new version somewhere it can never land, and the user can
+                    # actually fix it by closing whatever holds the directory.
+                    raise LocalError(t(
+                        "Could not replace {path}: a file in it is still "
+                        "open: {error}", path=into, error=exc)) from exc
+            fresh.rename(into)
+        except BaseException:
+            # Half an unpacked sibling is not worth keeping, and the swap
+            # never ran, so the previous install is still whole.
+            shutil.rmtree(fresh, ignore_errors=True)
+            raise
+        # Found under the sibling, run from the final directory.
+        binary = into / binary.relative_to(fresh)
+        # Written last, so the record never points at anything half-made.
+        record = {"tag": tag, "binary": str(binary)}
+        if managed:
+            # Which of the two builds this machine ended up with. Only written
+            # where both were on offer, so an install that never had the
+            # choice is not made to look like a fallback.
+            record["backend"] = (
+                "vulkan" if item.name.endswith(MANAGED_WHISPER_VULKAN)
+                else "processor")
+        _install_record(program).write_text(json.dumps(record), encoding="utf-8")
     except OSError as exc:
         raise LocalError(t("Could not install {name}: {error}",
                            name=program.name, error=exc)) from exc
@@ -404,15 +696,225 @@ def _drop_old_versions(program, keep):
     root = BIN_DIR / program.name
     try:
         for path in root.iterdir():
-            if path.is_dir() and path.name != keep:
-                shutil.rmtree(path, ignore_errors=True)
+            if not path.is_dir() or path.name == keep:
+                continue
+            # A ".new" sibling belongs to an install mid-swap; housekeeping
+            # must not pull it out from under it.
+            if path.name.endswith(".new"):
+                continue
+            # ignore_errors on purpose: this is housekeeping, and a locked old
+            # version is a little wasted disk rather than a failed install.
+            shutil.rmtree(path, ignore_errors=True)
     except OSError:
         pass
 
 
+# --- what this machine can run --------------------------------------------
+
+
+def total_memory():
+    """Bytes of memory on this machine, or 0 when it cannot be read.
+
+    Zero is a real answer and not a failure: every caller treats an unknown
+    machine as one big enough for whatever it is looking at, because a wrong
+    "too big" is worse advice than none.
+
+    Read once and kept. The memory in a machine does not change while Dikte
+    runs, and a list of thirty rows asks this question seventy times: on the
+    Mac path below, where the answer comes from a program rather than a
+    library call, that was seventy processes started on the interface thread
+    every time a list was drawn.
+    """
+    global _MEMORY
+    if _MEMORY is None:
+        _MEMORY = max(_read_memory(), 0)
+    return _MEMORY
+
+
+def _read_memory():
+    """What the system says, which on a bad day is a negative number.
+
+    sysconf answers -1 for a limit it holds to be indeterminate, and CPython
+    hands that straight back rather than raising, so the product below can
+    come out negative. The caller floors it at zero, which is the answer for
+    a machine nothing could be read from: a 64 GB workstation whose sysconf
+    shrugged was otherwise being told every model past 512 MB was too big
+    for it.
+    """
+    try:
+        return os.sysconf("SC_PAGE_SIZE") * os.sysconf("SC_PHYS_PAGES")
+    except (AttributeError, ValueError, OSError):
+        pass
+    if sys.platform == "darwin":
+        # Not every build of Python on a Mac has SC_PHYS_PAGES in its sysconf
+        # table, and this is the number the system itself is asked for.
+        try:
+            out = subprocess.run(["sysctl", "-n", "hw.memsize"], check=True,
+                                 capture_output=True, text=True, timeout=5)
+            return int(out.stdout.strip())
+        except (OSError, ValueError, subprocess.SubprocessError):
+            return 0
+    if sys.platform != "win32":
+        return 0
+
+    class Status(ctypes.Structure):
+        _fields_ = [("dwLength", ctypes.c_ulong),
+                    ("dwMemoryLoad", ctypes.c_ulong),
+                    ("ullTotalPhys", ctypes.c_ulonglong),
+                    ("ullAvailPhys", ctypes.c_ulonglong),
+                    ("ullTotalPageFile", ctypes.c_ulonglong),
+                    ("ullAvailPageFile", ctypes.c_ulonglong),
+                    ("ullTotalVirtual", ctypes.c_ulonglong),
+                    ("ullAvailVirtual", ctypes.c_ulonglong),
+                    ("ullAvailExtendedVirtual", ctypes.c_ulonglong)]
+
+    try:
+        status = Status()
+        status.dwLength = ctypes.sizeof(Status)
+        if ctypes.windll.kernel32.GlobalMemoryStatusEx(ctypes.byref(status)):
+            return int(status.ullTotalPhys)
+    except (AttributeError, OSError, ValueError):
+        pass
+    return 0
+
+
+def accelerator():
+    """The graphics interface this machine offers, or "".
+
+    The machine's half of the answer only. Whether a card is actually reached
+    also depends on which build landed, and the program line above says that:
+    a processor build ignores the card whatever is installed here. What this
+    is for is the other half, which nothing else on the window says at all.
+    """
+    if sys.platform == "darwin":
+        return "Metal"
+    return "Vulkan" if _has_vulkan() else ""
+
+
+def memory_budget(memory=None):
+    """What a model may weigh on this machine, or 0 when that is unknown.
+
+    Floored rather than allowed to reach zero: on a 2 GB machine the share
+    less the overhead is nothing at all, and a budget of nothing is the same
+    number this returns for a machine it could not read, which would turn the
+    tightest machine there is into the one where everything is offered.
+    """
+    memory = total_memory() if memory is None else memory
+    if not memory:
+        return 0
+    return max(int(memory * MEMORY_SHARE) - MEMORY_OVERHEAD, MEMORY_FLOOR)
+
+
+def fits(size, memory=None):
+    """Whether a model of this size is worth offering on this machine."""
+    budget = memory_budget(memory)
+    return not budget or size <= budget
+
+
+def suggested_whisper(memory=None, graphics=None):
+    """The whisper model to point at here, by name.
+
+    Three machines. One with no room, which gets the model that leaves some.
+    One with a card and memory to spare, which gets the accurate model, because
+    the several times the work it is per second is several times a fraction of
+    a second there. Everything in between gets turbo, which is the answer
+    almost every time somebody asks.
+
+    A Vulkan or Metal loader is not proof of a fast card, so the accurate model
+    waits on the memory as well: a machine with 16 GB in it and a driver
+    installed is one that will not notice either way.
+    """
+    memory = total_memory() if memory is None else memory
+    graphics = accelerator() if graphics is None else graphics
+    if memory and memory < SMALL_MACHINE:
+        return SMALL_MACHINE_WHISPER
+    if graphics and memory >= ROOMY_MACHINE:
+        return ACCURATE_WHISPER
+    return SUGGESTED_WHISPER
+
+
+def suggested_llm(memory=None):
+    """The suggested cleanup repositories, the ones that fit here first.
+
+    The order they are written in is the order they are worth having. What
+    this changes is only which of them a machine that cannot hold the best one
+    is shown first, and nothing is dropped: a model that does not fit today
+    fits once something else is closed.
+    """
+    return sorted(SUGGESTED_LLM,
+                  key=lambda repo: not fits(SUGGESTED_LLM_SIZE.get(repo, 0),
+                                            memory))
+
+
+def recommended(items, want="", memory=None):
+    """The one row out of `items` worth pointing at here, or "".
+
+    `want` is taken when it is on offer and fits. Without it, which is the
+    cleanup list, the smallest file that does is taken: q4 is where these
+    lists start, and every rung above it is roughly twice the memory and twice
+    the wait for a difference neither dictation nor cleanup can see. The
+    16-bit weights are left out for the same reason, twice over.
+    """
+    fitting = [i for i in items if fits(i.size, memory)]
+    if want and any(i.name == want for i in fitting):
+        return want
+    usable = [i for i in fitting
+              if not any(mark in i.name.lower() for mark in SIXTEEN_BIT)]
+    return min(usable, key=lambda i: i.size).name if usable else ""
+
+
 # --- the models -----------------------------------------------------------
 
 
+def bit_depth(name):
+    """The bits per weight the file name says, or 0 when it says nothing."""
+    lowered = name.lower()
+    for mark, bits in BIT_DEPTHS:
+        if mark in lowered:
+            return bits
+    return 0
+
+
+def whisper_family(name):
+    """The model a whisper file belongs to: ggml-small.en-q5_1.bin is `small`.
+
+    The list arrives sorted by size and nothing else, which interleaves the
+    families: `large-v3-turbo-q5_0` lands between the two `medium`
+    quantisations, half a screen from the turbo model it is a copy of. Grouping
+    is what puts the choice between models above the choice of quantisation,
+    which is the order somebody actually makes them in.
+    """
+    stem = name
+    if stem.startswith(WHISPER_PREFIX):
+        stem = stem[len(WHISPER_PREFIX):]
+    if stem.endswith(WHISPER_SUFFIX):
+        stem = stem[:-len(WHISPER_SUFFIX)]
+    head, _, last = stem.rpartition("-")
+    # q5_0, q5_1, q8_0. `turbo` is the other thing a last chunk can be, and it
+    # is part of the model's name rather than a quantisation of it.
+    if head and last.startswith("q") and last[1:].replace("_", "").isdigit():
+        stem = head
+    return stem[:-len(ENGLISH_ONLY)] if stem.endswith(ENGLISH_ONLY) else stem
+
+
+def whisper_groups(items):
+    """[(family, [Item])] for a whisper list: one group per model.
+
+    Groups by how big the model gets rather than by a ladder written down
+    here, so a family published next year sorts itself. Inside one, the
+    multilingual files come before the English-only ones and the small
+    quantisations before the large.
+    """
+    groups = {}
+    for item in items:
+        groups.setdefault(whisper_family(item.name), []).append(item)
+    ordered = sorted(groups.items(),
+                     key=lambda pair: (max(i.size for i in pair[1]), pair[0]))
+    return [(family, sorted(files,
+                            key=lambda i: (ENGLISH_ONLY in i.name, i.size)))
+            for family, files in ordered]
+
+
 def whisper_models(refresh=False):
     """[hub.Item] for every whisper model on offer, smallest first."""
     try:
@@ -425,10 +927,23 @@ def whisper_models(refresh=False):
     return sorted(models, key=lambda f: f.size)
 
 
+def can_clean(repo):
+    """Whether a repository could hold a model cleanup can be started on.
+
+    By name, because the alternative is a file listing per repository and the
+    list is forty of them. It catches the kinds that are never a cleanup model
+    rather than the ones that are too big, which the file sizes answer exactly
+    once a publisher is chosen.
+    """
+    lowered = repo.lower()
+    return not any(mark.lower() in lowered for mark in LLM_REPO_SKIP)
+
+
 def llm_repos(refresh=False):
     """Repository ids for the GGUF models on offer, suggestions first."""
     try:
-        found = [r.id for r in hub.repos(author=LLM_AUTHOR, refresh=refresh)]
+        found = [r.id for r in hub.repos(author=LLM_AUTHOR, refresh=refresh)
+                 if can_clean(r.id)]
     except hub.HubError:
         # A menu rather than a catalogue: with nothing to show, the suggestions
         # are still worth showing, and whatever is wrong with the network will
@@ -436,6 +951,14 @@ def llm_repos(refresh=False):
         found = []
     if not found:
         return list(SUGGESTED_LLM)
+    # Gemma publishes its base models under the instruction-tuned one's name
+    # with the `-it` taken out, so the two sit next to each other in the list
+    # and the wrong one answers a cleanup prompt by carrying on writing the
+    # transcript. Dropped only where the tuned sibling is here to drop it for.
+    tuned = set(found)
+    found = [r for r in found
+             if not r.endswith("-GGUF")
+             or r[:-len("-GGUF")] + "-it-GGUF" not in tuned]
     first = [r for r in SUGGESTED_LLM if r in found]
     return first + [r for r in found if r not in first]
 
@@ -538,7 +1061,12 @@ def _tail(path, lines=3):
 
 
 def _win_image_name(pid):
-    """The lower-cased file name of the process's executable, or ''."""
+    """The full, lower-cased path of the process's executable, or ''.
+
+    The full path rather than the base name, because the name alone is anyone's
+    whisper-server.exe and this answer decides what gets killed. MAX_PATH is a
+    convention rather than a limit, so the buffer grows until the query fits.
+    """
     import ctypes
     kernel32 = ctypes.WinDLL("kernel32", use_last_error=True)
     kernel32.OpenProcess.restype = ctypes.c_void_p
@@ -548,11 +1076,18 @@ def _win_image_name(pid):
     if not handle:
         return ""
     try:
-        buffer = ctypes.create_unicode_buffer(260)
-        size = ctypes.c_uint32(len(buffer))
-        ok = kernel32.QueryFullProcessImageNameW(
-            ctypes.c_void_p(handle), 0, buffer, ctypes.byref(size))
-        return os.path.basename(buffer.value).lower() if ok else ""
+        length = 260
+        while length <= 32768:
+            buffer = ctypes.create_unicode_buffer(length)
+            size = ctypes.c_uint32(len(buffer))
+            ok = kernel32.QueryFullProcessImageNameW(
+                ctypes.c_void_p(handle), 0, buffer, ctypes.byref(size))
+            if ok:
+                return buffer.value.lower()
+            if ctypes.get_last_error() != 122:   # ERROR_INSUFFICIENT_BUFFER
+                return ""
+            length *= 2
+        return ""
     finally:
         kernel32.CloseHandle(handle)
 
@@ -578,6 +1113,9 @@ class Server:
         self._port = 0
         self._log = ""
         self._key = None
+        # The pid this instance last wrote to its pid file, so _forget never
+        # removes a file some other Dikte wrote after us.
+        self._pid = 0
 
     # ---- settings --------------------------------------------------------
 
@@ -626,7 +1164,9 @@ class Server:
             ready = self._current_url()
             if ready:
                 return ready
-            self.stop()
+            # _stop_now rather than stop(): this thread already holds
+            # _starting, and the public stop() waits for it.
+            self._stop_now()
             with self._lock:
                 settings, key = dict(self._settings), self._settings_key()
             proc, port, log = self._launch(settings)
@@ -659,7 +1199,7 @@ class Server:
                         stdout=sink, stderr=subprocess.STDOUT,
                         stdin=subprocess.DEVNULL,
                         # No console window of its own on Windows.
-                        creationflags=getattr(subprocess, "CREATE_NO_WINDOW", 0),
+                        creationflags=paths.NO_WINDOW,
                     )
             except OSError as exc:
                 raise LocalError(t("Could not start {name}: {error}",
@@ -668,8 +1208,9 @@ class Server:
             # Written before it is ready rather than after, so that a kill
             # during the model load leaves something for the sweep to find.
             self._remember(proc.pid)
+            began = time.monotonic()
             try:
-                ready = self._wait_ready(proc, port)
+                reason, listened = self._wait_ready(proc, port)
             except BaseException:
                 # Whatever went wrong while waiting, the process is ours and
                 # nothing else is left holding a reference to it. Leaving it
@@ -679,31 +1220,52 @@ class Server:
                 self._kill(proc)
                 self._forget()
                 raise
-            if ready:
+            if reason == "ready":
                 return proc, port, str(log)
             last = _tail(log)
             self._forget()
-            # A port taken between the probe and the bind is the one failure
-            # worth another go; anything else will fail the same way again.
-            if "address" not in last.lower() and "bind" not in last.lower():
+            # Losing the port between the probe and the bind is the one
+            # failure another port fixes, and it has a shape rather than a
+            # message: the child died at once without the port ever having
+            # answered as its own. Grepping the log for "bind" would tie this
+            # to one program's wording in one language.
+            early = time.monotonic() - began < EARLY_EXIT_WINDOW
+            if reason != "exited" or listened or not early:
                 break
         raise LocalError(t("{name} did not start: {error}",
                            name=self.program.binary, error=last or t("no output")))
 
     def _wait_ready(self, proc, port):
+        """("ready" | "exited" | "timeout", whether the port answered as ours).
+
+        whisper binds after the model is loaded, so the open port is the
+        answer; llama binds first and answers /health with 503 until it is
+        ready. For the health-less case the open port alone is not proof: a
+        child that lost the bind race exits at once while the winner keeps the
+        port open, so "ready" also wants our child alive a beat after the port
+        was first seen open.
+        """
         deadline = time.monotonic() + STARTUP_TIMEOUT
+        seen_open = False
+        listened = False
         while time.monotonic() < deadline:
             if proc.poll() is not None:
-                return False
+                return "exited", listened
+            if seen_open:
+                # Port open on the last pass and our child still alive now:
+                # an imposter's port would have left our child dead by here.
+                return "ready", True
             if _listening(port):
-                # whisper binds after the model is loaded, so the open port is
-                # the answer. llama binds first and answers /health with 503
-                # until it is ready.
-                if not self.program.health or _healthy(port, self.program.health):
-                    return True
+                if self.program.health:
+                    # llama did the binding itself, so the port is its.
+                    listened = True
+                    if _healthy(port, self.program.health):
+                        return "ready", True
+                else:
+                    seen_open = True
             time.sleep(0.1)
         self._kill(proc)
-        return False
+        return "timeout", listened
 
     @staticmethod
     def _kill(proc, gently=False):
@@ -724,6 +1286,14 @@ class Server:
             pass
 
     def stop(self):
+        # Taking _starting means a stop cannot slide past a launch in flight:
+        # serve() finishes registering its child first, and the child is then
+        # killed here rather than surviving the shutdown unowned.
+        with self._starting:
+            self._stop_now()
+
+    def _stop_now(self):
+        """stop() for a thread that already holds _starting."""
         with self._lock:
             proc, self._proc = self._proc, None
             self._port, self._log, self._key = 0, "", None
@@ -737,6 +1307,8 @@ class Server:
         return DATA_DIR / f"{self.program.name}-server.pid"
 
     def _remember(self, pid):
+        with self._lock:
+            self._pid = pid
         try:
             path = self._pid_file()
             path.parent.mkdir(parents=True, exist_ok=True)
@@ -744,28 +1316,62 @@ class Server:
         except OSError:
             pass      # the sweep is a safety net, not something to fail a run over
 
-    def _forget(self):
+    def _forget(self, pid=None):
+        """Remove the pid file, but only while it still holds our own pid.
+
+        Another Dikte started after us writes its pid over ours, and removing
+        that file would hide its server from every future sweep.
+        """
+        if pid is None:
+            with self._lock:
+                pid = self._pid
         try:
-            self._pid_file().unlink()
-        except OSError:
+            if int(self._pid_file().read_text().strip()) == pid:
+                self._pid_file().unlink()
+        except (OSError, ValueError):
             pass
 
     def _is_ours(self, pid):
-        """Whether that pid is still the server this Dikte started.
+        """True: still our server. False: definitely not. None: cannot tell now.
 
         Asked because pids are handed out again: by the time anyone looks, the
         number could belong to something else entirely, and killing it would be
-        a good deal worse than the leak being cleaned up. The program name alone
-        could be somebody else's copy; the name together with Dikte's own data
-        directory on the command line could not. Windows offers no command line
-        to read, so the executable's name is the whole of the answer there.
+        a good deal worse than the leak being cleaned up. The tri-state matters
+        for the pid file: a definitive "not ours" means the file is stale and
+        safe to drop, while "cannot tell" means it has to stay so a later start
+        can ask again.
+
+        On Linux the program name alone could be somebody else's copy; the name
+        together with Dikte's own data directory on the command line could not.
+        Windows offers no command line to read, so the executable's full path
+        is the answer there: under our bin directory, or exactly the binary the
+        settings point this server at. Never the base name alone, which is
+        anyone's whisper-server.exe.
         """
         if sys.platform == "win32":
-            return _win_image_name(pid) == _binary_file(self.program).lower()
+            path = _win_image_name(pid)
+            if not path:
+                # OpenProcess said nothing: the process may be gone or merely
+                # unreadable from here, and the difference decides whether the
+                # pid file may be dropped, so no verdict rather than a wrong one.
+                return None
+            path = os.path.normcase(path)
+            if path.startswith(os.path.normcase(str(BIN_DIR)) + os.sep):
+                return True
+            with self._lock:
+                custom = self._settings.get("binary", "")
+            configured = program_path(self.program, custom)
+            if configured:
+                resolved = os.path.normcase(str(pathlib.Path(configured).resolve()))
+                if path == resolved:
+                    return True
+            return False
         try:
             blob = pathlib.Path(f"/proc/{pid}/cmdline").read_bytes()
+        except (FileNotFoundError, ProcessLookupError):
+            return False          # the process is definitively gone
         except OSError:
-            return False
+            return None           # /proc would not answer just now
         return (self.program.binary.encode() in blob
                 and str(DATA_DIR).encode() in blob)
 
@@ -781,13 +1387,21 @@ class Server:
             pid = int(self._pid_file().read_text().strip())
         except (OSError, ValueError):
             return False
-        self._forget()
-        if not self._is_ours(pid):
+        owned = self._is_ours(pid)
+        if owned is None:
+            # Could not be verified rather than known stale: the file stays,
+            # so the next start asks again instead of losing track of a server
+            # that may still be holding a model.
+            return False
+        if not owned:
+            self._forget(pid)
             return False
         try:
             os.kill(pid, signal.SIGTERM)
         except OSError:
+            self._forget(pid)
             return False
+        self._forget(pid)
         return True
 
 

dikte/hotkey.py

@@ -970,9 +970,7 @@ def conflicting_shortcuts(shortcut, desktop_id=DESKTOP_ID):
         if "=" not in line or desktop_id in section:
             continue
         key, _, value = line.partition("=")
-        if shortcut.lower() in value.lower().split(","):
-            hits.append(f"{section} → {key}")
-        elif any(shortcut.lower() == part.strip().lower()
-                 for part in re.split(r"[,\t]", value)):
+        if any(shortcut.lower() == part.strip().lower()
+               for part in re.split(r"[,\t]", value)):
             hits.append(f"{section} → {key}")
     return hits

dikte/hub.py

@@ -12,19 +12,18 @@ means one for every whisper.cpp release; both of those are somebody else's news,
 not Dikte's. Answers are cached for a few hours, and a cache that has gone stale
 is still a better answer than none when the network is down.
 
-Nothing here imports the rest of Dikte apart from the string table: this module
-knows two websites and nothing about dictation.
+Nothing here imports the rest of Dikte apart from two leaves, the string table
+and the path map: this module knows two websites and nothing about dictation.
 """
 
 import collections
 import json
-import os
-import pathlib
 import time
 import urllib.error
 import urllib.parse
 import urllib.request
 
+from . import paths
 from .i18n import t
 
 GITHUB_API = "https://api.github.com"
@@ -32,8 +31,7 @@ HF_API = "https://huggingface.co/api"
 HF_FILES = "https://huggingface.co"
 USER_AGENT = "dikte/1.0 (+https://github.com/yusufipk/dikte)"
 
-CACHE_DIR = (pathlib.Path(os.environ.get("XDG_CACHE_HOME")
-                          or os.path.expanduser("~/.cache")) / "dikte")
+CACHE_DIR = paths.cache_dir()
 # Long enough that opening the settings window twice in an evening asks nobody
 # anything, short enough that a model published this morning is offered today.
 CACHE_TTL = 6 * 3600
@@ -123,6 +121,13 @@ def _digest(value):
     return value.split(":", 1)[1] if value.startswith("sha256:") else value
 
 
+def _assets(data):
+    return [Item(a.get("name") or "", a.get("browser_download_url") or "",
+                 int(a.get("size") or 0), _digest(a.get("digest")))
+            for a in (data.get("assets") or [])
+            if a.get("browser_download_url")]
+
+
 def release(repo, tag="latest", refresh=False):
     """(tag, [Item]) for one GitHub release, newest when no tag is given."""
     where = "latest" if tag in ("", "latest") else f"tags/{tag}"
@@ -130,10 +135,47 @@ def release(repo, tag="latest", refresh=False):
                   f"{GITHUB_API}/repos/{repo}/releases/{where}", refresh=refresh)
     if not isinstance(data, dict) or not data.get("assets"):
         raise HubError(t("{repo} has no downloadable release.", repo=repo))
-    assets = [Item(a.get("name") or "", a.get("browser_download_url") or "",
-                   int(a.get("size") or 0), _digest(a.get("digest")))
-              for a in data["assets"] if a.get("browser_download_url")]
-    return data.get("tag_name") or tag, assets
+    return data.get("tag_name") or tag, _assets(data)
+
+
+def releases(repo, limit=20, refresh=False):
+    """[(tag, [Item])] for the recent releases, newest first, with their files.
+
+    "latest" is one release and this is the list behind it, prereleases
+    included: a project that attaches its builds to a prerelease is invisible
+    to release() above, and its newest usable build is in here.
+    """
+    data = _fetch(f"gh-list-{repo}-{limit}",
+                  f"{GITHUB_API}/repos/{repo}/releases?per_page={limit}",
+                  refresh=refresh)
+    if not isinstance(data, list):
+        raise HubError(t("{repo} has no downloadable release.", repo=repo))
+    out = []
+    for entry in data:
+        tag, items = entry.get("tag_name") or "", _assets(entry)
+        if tag and items:
+            out.append((tag, items))
+    return out
+
+
+def text(url, limit=4096, timeout=20):
+    """A small text file from a release, as a string.
+
+    Not cached and not checksummed, because what it carries is a pointer: a few
+    bytes naming the release the actual archives are attached to, read once on
+    the way to a download that is checked in full.
+    """
+    request = urllib.request.Request(url, headers={"User-Agent": USER_AGENT})
+    try:
+        with urllib.request.urlopen(request, timeout=timeout) as response:
+            return response.read(limit).decode("utf-8", "replace")
+    except urllib.error.HTTPError as exc:
+        exc.close()
+        raise HubError(t("{url} answered HTTP {code}.",
+                         url=urllib.parse.urlsplit(url).netloc, code=exc.code)) from exc
+    except (urllib.error.URLError, OSError, ValueError) as exc:
+        raise HubError(t("Could not reach {url}: {error}",
+                         url=urllib.parse.urlsplit(url).netloc, error=exc)) from exc
 
 
 def newest_release(repo, refresh=False):

dikte/i18n.py

@@ -40,8 +40,16 @@ def t(text, /, **kwargs):
 # by the sentence, so it arrives already inflected. English takes the name as it
 # is and puts the preposition in the sentence, where it belongs.
 _TR_CASES = {
-    "dative": {"Claude": "Claude'a", "Codex": "Codex'e", "OpenRouter": "OpenRouter'a"},
-    "accusative": {"Claude": "Claude'u", "Codex": "Codex'i", "OpenRouter": "OpenRouter'ı"},
+    "dative": {
+        "Claude": "Claude'a", "Codex": "Codex'e", "OpenRouter": "OpenRouter'a",
+        "Google AI Studio": "Google AI Studio'ya", "Antigravity": "Antigravity'ye",
+        "OpenCode Go": "OpenCode Go'ya",
+    },
+    "accusative": {
+        "Claude": "Claude'u", "Codex": "Codex'i", "OpenRouter": "OpenRouter'ı",
+        "Google AI Studio": "Google AI Studio'yu", "Antigravity": "Antigravity'yi",
+        "OpenCode Go": "OpenCode Go'yu",
+    },
 }
 
 
@@ -69,6 +77,7 @@ TR = {
 
     # --- overlay / pipeline -------------------------------------------
     "Transcribing…": "Yazıya çevriliyor…",
+    "Waiting for the one before it…": "Öncekinin bitmesi bekleniyor…",
     "Cleaning up…": "Temizleniyor…",
     "Pasting…": "Yapıştırılıyor…",
     "Pasted": "Yapıştırıldı",
@@ -150,6 +159,7 @@ TR = {
     # --- settings: tabs and general ------------------------------------
     "Dikte Settings": "Dikte Ayarları",
     "General": "Genel",
+    "Display": "Ekran",
     "API and models": "API ve modeller",
     "Cleanup rules": "Temizleme kuralları",
     "Audio file": "Ses dosyası",
@@ -162,8 +172,6 @@ TR = {
     "Automatic (system)": "Otomatik (sistem)",
     "Turkish": "Türkçe",
     "English": "İngilizce",
-    "Restart Dikte for the language change to reach every window.":
-        "Dil değişikliğinin her pencereye işlemesi için Dikte'yi yeniden başlat.",
     "Microphone": "Mikrofon",
     "Default microphone": "Varsayılan mikrofon",
     "Speech language": "Konuşma dili",
@@ -180,6 +188,9 @@ TR = {
         "macOS bu ilk gönderildiğinde Erişilebilirlik izni ister.",
     "Restore the previous clipboard after pasting":
         "Yapıştırdıktan sonra eski pano içeriğini geri koy",
+    "Indicator screen": "Gösterge ekranı",
+    "Follow the mouse pointer": "Fare imlecini takip et",
+    "{name} (not connected)": "{name} (bağlı değil)",
     "Indicator corner": "Gösterge köşesi",
     "bottom-left": "sol-alt",
     "bottom-right": "sağ-alt",
@@ -217,30 +228,34 @@ TR = {
     "Transcript cleanup": "Transkripti temizleme",
     "API key": "API anahtarı",
     "Model": "Model",
+    "Audio file model": "Ses dosyası modeli",
+    "The model a timestamped audio file (subtitles) is sent to. Not every model on "
+    "OpenRouter returns segment times; empty means openai/whisper-1.":
+        "Zaman damgalı bir ses dosyasının (altyazı) gönderildiği model. OpenRouter'daki her "
+        "model segment zamanı döndürmez; boşsa openai/whisper-1 kullanılır.",
     "Provider": "Sağlayıcı",
     "sk-… (falls back to OPENAI_API_KEY)": "sk-… (boşsa OPENAI_API_KEY kullanılır)",
     "gsk_… (falls back to GROQ_API_KEY)": "gsk_… (boşsa GROQ_API_KEY kullanılır)",
     "sk-or-… (falls back to OPENROUTER_API_KEY)": "sk-or-… (boşsa OPENROUTER_API_KEY kullanılır)",
+    "(falls back to GEMINI_API_KEY)": "(boşsa GEMINI_API_KEY kullanılır)",
+    "(falls back to OPENCODE_API_KEY)": "(boşsa OPENCODE_API_KEY kullanılır)",
     "Test": "Test et",
     "Trying…": "Deneniyor…",
     "Runs on OpenRouter.": "OpenRouter üzerinde çalışır.",
+    "Runs on Google AI Studio.": "Google AI Studio üzerinde çalışır.",
+    "Runs on OpenCode Go.": "OpenCode Go üzerinde çalışır.",
     "Connection works. {count} audio models visible.":
         "Bağlantı tamam. {count} ses modeli görünüyor.",
+    "Connection works. {count} models visible.":
+        "Bağlantı tamam. {count} model görünüyor.",
     "Clean the transcript with a model": "Transkripti bir modelle temizle",
-    "OpenRouter is the quickest and the only one that needs nothing installed. "
-    "Claude Code and Codex clean up on the subscription you already have, "
-    "without a second key, and take a few seconds longer because each one opens "
-    "a session to do it.":
-        "En hızlısı OpenRouter'dır ve kurulu bir program istemeyen tek seçenektir. "
-        "Claude Code ile Codex, temizliği hâlihazırda ödediğin abonelik üzerinden "
-        "yapar, ikinci bir anahtar istemez; her biri bunun için bir oturum açtığından "
-        "birkaç saniye daha uzun sürer.",
     "{binary} is not on your PATH, so cleanup would fail and the raw transcript "
     "would be pasted. Install it, or pick another one above.":
         "{binary} PATH'te değil; temizleme başarısız olur ve ham transkript "
         "yapıştırılır. Kur ya da yukarıdan başka birini seç.",
     "Thinking": "Düşünme",
     "Model's own default": "Modelin kendi varsayılanı",
+    "Antigravity's own default": "Antigravity'nin kendi varsayılanı",
     "Off": "Kapalı",
     "Minimal": "En az",
     "Low": "Düşük",
@@ -503,18 +518,6 @@ TR = {
 
     # --- settings: the agent ------------------------------------------------
     "Agent": "Ajan",
-    "This shortcut records the same way dictation does, but the transcript is "
-    "not what gets pasted. It goes to an agent as a command, and what comes "
-    "back is pasted instead: the answer to a question, or a sentence saying "
-    "what was done. Claude Code and Codex run as the session you would have "
-    "opened yourself, with your skills, your connected services and your "
-    "account.":
-        "Bu kısayol dikte ile aynı şekilde kaydeder, ama yapıştırılan şey "
-        "transkript değildir. Transkript bir ajana komut olarak gider ve yerine "
-        "oradan döneni yapıştırılır: bir sorunun cevabı ya da ne yapıldığını "
-        "söyleyen bir cümle. Claude Code ve Codex, kendi açacağın oturumun "
-        "aynısı olarak çalışır: skill'lerinle, bağlı servislerinle ve kendi "
-        "hesabınla.",
     "How it runs": "Nasıl çalışıyor",
     "Runs on": "Şunun üstünde çalışır",
     "More thinking is slower, and you are standing in front of the screen while "
@@ -541,6 +544,24 @@ TR = {
         "Yukarıdaki çalışma dizini ve izinler burada bir şey ifade etmez.",
     "Needs no program installed, only the OpenRouter key.":
         "Kurulu bir programa değil, yalnızca OpenRouter anahtarına ihtiyaç duyar.",
+    "A plain question and a plain answer, over the OpenCode Go key you already "
+    "have. It runs no commands, opens no files and reaches none of your "
+    "services, so it can tell you what the capital of Peru is but not what is "
+    "in your calendar. Working directory and permissions above mean nothing "
+    "here.":
+        "Elindeki OpenCode Go anahtarı üzerinden düz bir soru ve düz bir cevap. "
+        "Komut çalıştırmaz, dosya açmaz, servislerinin hiçbirine erişmez; yani "
+        "Peru'nun başkentini söyler ama takviminde ne olduğunu söyleyemez. "
+        "Yukarıdaki çalışma dizini ve izinler burada bir şey ifade etmez.",
+    "Needs no program installed, only an OpenCode Go key.":
+        "Kurulu bir programa değil, yalnızca bir OpenCode Go anahtarına ihtiyaç duyar.",
+    "Antigravity has neither a permission mode nor a sandbox to hand it, so "
+    "what it may do without asking is whatever its own allow-rules say. The "
+    "Permissions and Sandbox boxes above belong to the other two; the working "
+    "directory still applies.":
+        "Antigravity'ye verilebilecek bir izin kipi ya da sandbox yok; sormadan "
+        "ne yapabileceğini kendi allow-rule'ları belirler. Yukarıdaki İzinler ve "
+        "Sandbox kutuları diğer ikisine ait; çalışma dizini burada da geçerli.",
     "{binary} is not on your PATH, so this cannot run yet. Install it, or pick "
     "another one above.":
         "{binary} PATH'te değil, dolayısıyla bu henüz çalışamaz. Kur ya da "
@@ -556,6 +577,10 @@ TR = {
         "“sonnet” gibi bir ad her zaman o serinin en yenisini seçer. Opus daha "
         "çok düşünür ve daha geç cevaplar; bu da en çok burada hissedilir, "
         "çünkü ekranın başında bekliyorsun.",
+    "The list is a starting point, not a fence: any model name {name} accepts "
+    "can be typed straight in.":
+        "Liste başlangıç için, sınır değil: {name} hangi model adını kabul "
+        "ediyorsa buraya elle yazılabilir.",
     "Permissions": "İzinler",
     "Decide on its own, with the safety checks on":
         "Kendi karar versin, güvenlik denetimleri açık",
@@ -595,7 +620,7 @@ TR = {
     "configuration already says.":
         "Her komutla birlikte ajana söylenir, kendi yapılandırmanın zaten "
         "söylediklerinin üstüne eklenir.",
-    "  ·  asked Claude: {question}": "  ·  Claude'a soruldu: {question}",
+    "  ·  asked {who}: {question}": "  ·  {who} soruldu: {question}",
 
     # --- meetings: tray and pipeline ---------------------------------------
     "Record a meeting": "Toplantı kaydet",
@@ -665,12 +690,6 @@ TR = {
 
     # --- settings: meeting --------------------------------------------------
     "Minutes": "Tutanaklar",
-    "A meeting is recorded from two devices at once: your microphone and "
-    "whatever comes out of your speakers. Nothing has to guess who was "
-    "speaking, because the two never share a channel.":
-        "Toplantı iki aygıttan aynı anda kaydedilir: mikrofonun ve hoparlöründen "
-        "çıkan ses. Kimin konuştuğunun tahmin edilmesi gerekmez, çünkü ikisi hiç "
-        "aynı kanala girmez.",
     "Sound": "Ses",
     "Same as dictation": "Diktedekiyle aynı",
     "Current output": "Geçerli çıkış",
@@ -748,4 +767,204 @@ TR = {
     "This one is being written up right now.": "Bunun tutanağı şu anda çıkarılıyor.",
     "Delete this meeting, its minutes and its recording?":
         "Bu toplantı, tutanağı ve ses kaydı silinsin mi?",
+
+    # --- local models and downloads ------------------------------------
+    # The whole box was born after the last translation pass, which left the
+    # first-run screen half English on a Turkish machine.
+    "Download": "İndir",
+    "Delete": "Sil",
+    "Program": "Program",
+    "Publisher": "Yayıncı",
+    "Automatic": "Otomatik",
+    "Threads": "İş parçacığı",
+    "On this machine": "Bu makinede",
+    "Use the graphics card": "Ekran kartını kullan",
+    "Load the model when Dikte starts": "Modeli Dikte açılırken yükle",
+    "Local whisper": "Yerel whisper",
+    "Local model": "Yerel model",
+    "Not installed.": "Kurulu değil.",
+    "Installed on the system: {path}": "Sistemde kurulu: {path}",
+    "Download again": "Yeniden indir",
+    "Downloaded, version {version}.": "İndirildi, sürüm {version}.",
+    "Downloaded, version {version}. There was no Vulkan build, "
+    "so this one runs on the processor.":
+        "İndirildi, sürüm {version}. Vulkan sürümü yoktu, bu sürüm işlemcide çalışıyor.",
+    "Fetching the model list…": "Model listesi çekiliyor…",
+    "Downloading…": "İndiriliyor…",
+    "Downloading: {done} of {total}{share}": "İndiriliyor: {done} / {total}{share}",
+    "Download stopped.": "İndirme durduruldu.",
+    "Ready: {name}.": "Hazır: {name}.",
+    "Nothing downloaded yet.": "Henüz bir şey indirilmedi.",
+    "{name} has not been downloaded yet.": "{name} henüz indirilmedi.",
+    "{name} is here, but the program above is not. Download it first.":
+        "{name} burada, ama yukarıdaki program değil. Önce onu indirin.",
+    "{name} is not on this machine and this publisher does not offer it. "
+    "Choose another model, or another publisher.":
+        "{name} bu makinede yok ve bu yayıncı da sunmuyor. Başka bir model, "
+        "ya da başka bir yayıncı seçin.",
+    "downloaded": "indirildi",
+    "not downloaded": "indirilmedi",
+    "recommended": "önerilen",
+    "{bits}-bit": "{bits} bit",
+    "English only": "yalnızca İngilizce",
+    "All": "Tümü",
+    "Everything ggml-org publishes, including the models that are too big to "
+    "run here and the ones that are not for cleaning up text.":
+        "ggml-org'un yayımladığı her şey; burada çalıştırılamayacak kadar "
+        "büyük olanlar ve metin temizlemek için olmayanlar dahil.",
+    "Google Gemma 4, the small one. The default: nothing else this size "
+    "follows an instruction as closely, and cleanup is all instruction.":
+        "Google Gemma 4'ün küçüğü. Varsayılan: bu boyutta verilen yönergeyi "
+        "bu kadar iyi izleyen başka bir model yok, temizleme de baştan sona "
+        "yönerge demek.",
+    "The same model one size up. A little more accurate, about twice the "
+    "weights and twice the wait.":
+        "Aynı modelin bir boy büyüğü. Biraz daha isabetli, yaklaşık iki katı "
+        "ağırlık ve iki katı bekleyiş.",
+    "The previous Gemma. Still good, and the smallest of the Gemmas here.":
+        "Bir önceki Gemma. Hâlâ iyi ve buradaki Gemma'ların en küçüğü.",
+    "Hugging Face's own small model, for a machine the Gemmas crowd.":
+        "Hugging Face'in kendi küçük modeli; Gemma'ların sıkıştırdığı bir "
+        "makine için.",
+    "The smallest of them, for a machine nothing else fits on. It thinks "
+    "before it answers unless Thinking below is off.":
+        "En küçükleri; başka hiçbir şeyin sığmadığı bir makine için. "
+        "Aşağıdaki Düşünme kapalı değilse cevaplamadan önce düşünür.",
+    "too big for this machine": "bu makine için fazla büyük",
+    "This machine": "Bu makine",
+    "Graphics: {name}.": "Ekran kartı: {name}.",
+    "No graphics interface found, so this runs on the processor.":
+        "Ekran kartı arayüzü bulunamadı, bu yüzden işlemcide çalışıyor.",
+    "Memory: {size}.": "Bellek: {size}.",
+    "A model may take half of this memory, less a gigabyte for the context "
+    "around the weights. Anything past that is marked too big; it may still "
+    "load, on a machine with nothing else open.":
+        "Bir model bu belleğin yarısını, ağırlıkların çevresindeki bağlam için "
+        "bir gigabayt düşülerek kullanabilir. Bunu aşan modeller fazla büyük "
+        "diye işaretlenir; başka hiçbir şeyin açık olmadığı bir makinede yine "
+        "de yüklenebilirler.",
+    "Recommended for this machine": "Bu makine için önerilen",
+    "Everything this publisher offers": "Bu yayıncının sunduğu her şey",
+    "Already on this machine": "Bu makinede zaten var",
+    "Chosen, but not downloaded": "Seçili, ama indirilmedi",
+    "{repo} publishes nothing that can be run here. Its models are split "
+    "across files, larger than {cap}, or pieces of a model rather than one. "
+    "Choose another publisher.":
+        "{repo} burada çalıştırılabilecek bir şey yayımlamıyor. Modelleri "
+        "birden çok dosyaya bölünmüş, {cap} boyutundan büyük ya da modelin "
+        "kendisi değil parçaları. Başka bir yayıncı seçin.",
+    "large-v3 makes the fewest mistakes and is the slowest of them. "
+    "large-v3-turbo is that model with a four layer decoder in place of a "
+    "thirty-two layer one: several times faster, at one to two points of word "
+    "error in English and about two and a half in the other languages. Below "
+    "those, every step down the list trades accuracy for size, and the .en "
+    "models are trained on English alone.":
+        "En az hatayı large-v3 yapar, en yavaşı da odur. large-v3-turbo, aynı "
+        "modelin otuz iki katmanlı çözücüsü yerine dört katmanlı bir çözücü "
+        "konmuş hâli: birkaç kat hızlı, karşılığında İngilizcede bir iki "
+        "puan, diğer dillerde yaklaşık iki buçuk puan kelime hatası. Bunların "
+        "altında listede her basamak, doğruluğu boyuta değişir; .en modelleri "
+        "ise yalnızca İngilizce ile eğitilmiştir.",
+    "Cleanup is punctuation, capitals and filler words, so what these are "
+    "picked on is following an instruction rather than knowing anything. "
+    "Start at a q4 file; the 16-bit ones are several times the memory for a "
+    "difference this job cannot see.":
+        "Temizleme; noktalama, büyük harf ve dolgu sözcükleri demek, yani bu "
+        "modeller bir şey bilmelerine değil verilen yönergeyi izlemelerine "
+        "göre seçilir. Bir q4 dosyasından başlayın; 16 bitlik olanlar, bu işin "
+        "göremeyeceği bir fark için kat kat bellek ister.",
+    "Delete model": "Modeli sil",
+    "Delete {name} from this machine?": "{name} bu makineden silinsin mi?",
+    "Runs on this machine, on llama.cpp.": "Bu makinede, llama.cpp üzerinde çalışır.",
+    "A Hugging Face repository of GGUF files. The list is fetched; any other "
+    "one can be typed in.":
+        "GGUF dosyaları içeren bir Hugging Face deposu. Liste internetten "
+        "çekilir; başka bir depo da yazılabilir.",
+    "A large model takes a second or two to load. Loading it up front spends "
+    "that once instead of on the first dictation, at the cost of the memory "
+    "it sits in.":
+        "Büyük bir modelin yüklenmesi bir iki saniye sürer. Baştan yüklemek bu "
+        "bedeli ilk diktede değil bir kez öder; karşılığı, modelin oturduğu "
+        "bellektir.",
+    "An LLM is slower to load than a whisper model and sits in more memory. "
+    "Off means it is loaded on the first cleanup instead.":
+        "Bir LLM, whisper modelinden daha geç yüklenir ve daha çok bellekte "
+        "oturur. Kapalı, ilk temizlemede yüklenmesi demektir.",
+    "A model trained to think will think unless it is told not to, and "
+    "spending 300 tokens of reasoning on a comma is 300 tokens of waiting. "
+    "Off is what cleanup wants.":
+        "Düşünmeye eğitilmiş bir model, aksi söylenmedikçe düşünür; bir virgül "
+        "için 300 token akıl yürütmek 300 token'lık bekleyiştir. Temizleme için "
+        "doğrusu Kapalı.",
+    "whisper.cpp reaches the card through CUDA, ROCm or Vulkan when the build "
+    "it is running was made with one. A build without any of them runs on the "
+    "processor whatever this says.":
+        "whisper.cpp karta CUDA, ROCm ya da Vulkan üzerinden ulaşır; koştuğu "
+        "derleme bunlardan biriyle yapılmışsa. Hiçbiri olmadan derlenmiş bir "
+        "kopya, bu ne derse desin işlemcide çalışır.",
+    "whisper.cpp is not installed. Settings → API and models → Download.":
+        "whisper.cpp kurulu değil. Ayarlar → API ve modeller → İndir.",
+    "llama.cpp is not installed. Settings → API and models → Download.":
+        "llama.cpp kurulu değil. Ayarlar → API ve modeller → İndir.",
+    "No whisper model has been downloaded yet. Settings → API and models → "
+    "Download.":
+        "Henüz whisper modeli indirilmedi. Ayarlar → API ve modeller → İndir.",
+    "No local cleanup model has been downloaded yet. Settings → API and "
+    "models → Download.":
+        "Henüz yerel temizleme modeli indirilmedi. Ayarlar → API ve modeller "
+        "→ İndir.",
+    "Hugging Face did not return a model list.":
+        "Hugging Face model listesi döndürmedi.",
+    "{repo} did not return a file list.": "{repo} dosya listesi döndürmedi.",
+    "{repo} has no downloadable release.":
+        "{repo} deposunun indirilebilir bir sürümü yok.",
+    "{repo} {tag} has no build for this machine.":
+        "{repo} {tag} bu makine için derleme içermiyor.",
+    "{url} answered HTTP {code}.": "{url} HTTP {code} yanıtı verdi.",
+    "Could not reach {url}: {error}": "{url} adresine ulaşılamadı: {error}",
+    "Could not read the answer from {url}: {error}":
+        "{url} yanıtı okunamadı: {error}",
+    "Could not create {path}: {error}": "{path} oluşturulamadı: {error}",
+    "Could not download {name}: HTTP {code}":
+        "{name} indirilemedi: HTTP {code}",
+    "Could not download {name}: {error}": "{name} indirilemedi: {error}",
+    "Could not write {name}: {error}": "{name} yazılamadı: {error}",
+    "Could not unpack {name}: {error}": "{name} açılamadı: {error}",
+    "Could not install {name}: {error}": "{name} kurulamadı: {error}",
+    "Could not start {name}: {error}": "{name} başlatılamadı: {error}",
+    "Could not delete the model: {error}": "Model silinemedi: {error}",
+    "Could not replace {path}: a file in it is still open: {error}":
+        "{path} değiştirilemedi: içindeki bir dosya hâlâ açık: {error}",
+    "{name} did not start: {error}": "{name} başlamadı: {error}",
+    "no output": "çıktı yok",
+    "The download stopped early ({done} of {total}).":
+        "İndirme erken kesildi ({done} / {total}).",
+    "{name} is longer than it said it would be.":
+        "{name} bildirdiğinden daha uzun çıktı.",
+    "{name} does not match its published checksum. Nothing was installed.":
+        "{name} yayımlanan sağlama toplamıyla uyuşmuyor. Hiçbir şey kurulmadı.",
+    "{name} is published without a checksum, so there is no way to tell what "
+    "arrived. Nothing was installed.":
+        "{name} sağlama toplamı olmadan yayımlanmış; gelenin ne olduğu "
+        "doğrulanamaz. Hiçbir şey kurulmadı.",
+    "{name} was not in the download.": "{name} indirilenin içinde yoktu.",
+    "{name} downloaded, but the old file is held open by the running server. "
+    "Stop it and try again.":
+        "{name} indirildi ama eski dosyayı çalışan sunucu açık tutuyor. "
+        "Sunucuyu durdurup yeniden dene.",
+    "The cleanup model spent its whole reply on thinking. Set Thinking to "
+    "“Off”.":
+        "Temizleme modeli bütün yanıtını düşünmeye harcadı. Düşünme'yi "
+        "“Kapalı” yap.",
+
+    # --- this pass's new messages ---------------------------------------
+    "Audio recorder stopped before receiving sound":
+        "Ses kayıt aracı veri alamadan kapandı",
+    "Could not write the recording: {error}": "Kayıt dosyası yazılamadı: {error}",
+    "Copied, but pasting failed: {error}":
+        "Kopyalandı ama yapıştırma başarısız: {error}",
+    "The recording was kept: {path}": "Kayıt saklandı: {path}",
+    "The recording stopped on its own; transcribing what was captured.":
+        "Kayıt kendi kendine durdu; yakalanan kısım yazıya dökülüyor.",
+    "Could not save the settings: {error}": "Ayarlar kaydedilemedi: {error}",
 }

dikte/integrate.py

@@ -56,6 +56,17 @@ def packaged():
     return bool(getattr(sys, "frozen", False))
 
 
+def windowed_executable(executable=None):
+    """The windowed executable installed beside this one, or None.
+
+    Beside rather than at a known place, because the setup program lays the two
+    executables into the same directory wherever that directory was put: asking
+    from either of them finds the other without knowing where the install is.
+    """
+    windowed = pathlib.Path(executable or sys.executable).with_name(WINDOWS_APP)
+    return windowed if windowed.is_file() else None
+
+
 def target():
     """The file a launcher has to name to start this build again.
 
@@ -74,8 +85,8 @@ def target():
         # The windowed executable, whichever of the two is running: the console
         # one is what the `dikte` command names, and a sign-in that started
         # that one would open a console window nobody asked for.
-        windowed = executable.with_name(WINDOWS_APP)
-        if windowed.is_file():
+        windowed = windowed_executable(executable)
+        if windowed is not None:
             return windowed
     return executable
 
@@ -572,24 +583,60 @@ def _run_entry_name():
     return f"HKCU\\{RUN_KEY}\\{RUN_VALUE}"
 
 
+def _run_target(value):
+    """The executable a Run value names, out of the quoting the setup wrote.
+
+    Only the first word matters here: it is the file whose existence says
+    whether the entry still starts anything.
+    """
+    if value.startswith('"'):
+        closing = value.find('"', 1)
+        return value[1:closing] if closing > 0 else ""
+    return value.split(" ", 1)[0]
+
+
+def _startup_shortcut():
+    """Where install.ps1 -Autostart puts a checkout's sign-in entry."""
+    appdata = os.environ.get("APPDATA")
+    if not appdata:
+        return None
+    return (pathlib.Path(appdata) / "Microsoft" / "Windows" / "Start Menu"
+            / "Programs" / "Startup" / "Dikte.lnk")
+
+
 def _windows_install(app, force=False):
     """Point the autostart entry at this build. What changed.
 
     Only `force`, which is what typing `dikte integrate` means, creates one.
     The call on every start repairs an entry that is already there and names an
-    executable somewhere else, which is what an installation moved to another
-    drive or reinstalled into another directory leaves behind; somebody who
-    unticked the box in the wizard, or turned it off since, is not asked again
-    by every start.
+    executable that is gone, which is what an installation moved to another
+    drive or reinstalled into another directory leaves behind. An entry naming
+    an executable that still exists is another installation that still works,
+    and is stood aside for the way the Linux half stands aside for another
+    menu entry; somebody who unticked the box in the wizard, or turned it off
+    since, is not asked again by every start either.
     """
     command = f'"{app}"'
     current = _run_entry()
+    changed = []
+    if force:
+        # install.ps1 -Autostart wrote this for a checkout. The Run value
+        # written below replaces it, and both left in place would be two
+        # Diktes at every sign-in. Only on force: the silent call on every
+        # start has not been asked to move the machine off its checkout.
+        shortcut = _startup_shortcut()
+        if shortcut is not None and shortcut.is_file():
+            shortcut.unlink()
+            changed.append(shortcut)
     if not current and not force:
-        return []
-    if current == command:
-        return []
-    _write_run_entry(command)
-    return [_run_entry_name()]
+        return changed
+    if current != command:
+        theirs = _run_target(current) if current else ""
+        if not force and theirs and theirs != str(app) and os.path.exists(theirs):
+            return changed
+        _write_run_entry(command)
+        changed.append(_run_entry_name())
+    return changed
 
 
 def _windows_remove():

dikte/ipc.py

@@ -11,11 +11,14 @@ shortcut may still send.
 import json
 import os
 import shlex
+import subprocess
 import sys
 
+from PyQt6.QtCore import QLockFile
 from PyQt6.QtNetwork import QLocalSocket
 
 from . import integrate
+from . import paths
 
 SERVER_NAME = "dikte-" + (
     str(os.getuid()) if hasattr(os, "getuid")
@@ -54,10 +57,9 @@ def launcher():
     if not getattr(sys, "frozen", False):
         return [sys.executable, script_path()]
     if sys.platform == "win32":
-        windowed = os.path.join(os.path.dirname(sys.executable),
-                                integrate.WINDOWS_APP)
-        if os.path.isfile(windowed):
-            return [windowed]
+        windowed = integrate.windowed_executable()
+        if windowed is not None:
+            return [str(windowed)]
     return [os.environ.get("APPIMAGE") or sys.executable]
 
 
@@ -72,6 +74,60 @@ def command_for(verb):
     return shlex.join(launcher() + ([verb] if verb else []))
 
 
+def already_serving():
+    """Whether a running instance answers on this user's name.
+
+    Asked before an instance opens a server of its own, because listen() is
+    not the check: a Windows named pipe takes a second server on the same name
+    rather than refusing it, and everywhere else removeServer() would first
+    take the live socket away from the instance holding it. Either way two
+    whole Diktes then run, and the newer one's sweep() kills the whisper the
+    older one is answering dictations with. The probe is "status" and nothing
+    else: a verb with a side effect here would fire it during the relaunch a
+    slow-to-answer instance provokes, on top of the verb being forwarded.
+    """
+    return send("status") is not None
+
+
+def instance_lock():
+    """This user's one-Dikte lock, taken before anything else is built.
+
+    The probe above has a hole: two copies started in the same moment both ask
+    before either listens, and both come up. A lock file closes it, and
+    QLockFile writes the holder's pid into it, so a lock a killed instance
+    left behind identifies itself as stale and clears. None when the data
+    directory cannot be made, which a start should survive: the probe still
+    stands guard, just without the simultaneous-start case.
+    """
+    try:
+        paths.DATA_DIR.mkdir(parents=True, exist_ok=True)
+    except OSError:
+        return None
+    lock = QLockFile(str(paths.DATA_DIR / "dikte.lock"))
+    # Never presume a lock is stale by age alone; the pid check is the truth.
+    lock.setStaleLockTime(0)
+    return lock
+
+
+def respawn(arguments):
+    """Start this installation again with `arguments`, leaving this process.
+
+    execv everywhere it works the way it says: the new process takes this
+    pid and nothing is left behind. On Windows execv mangles arguments with
+    spaces and leaves the two processes sharing a console, so the replacement
+    is started detached instead and the caller exits on its own.
+    """
+    args = launcher() + list(arguments)
+    if sys.platform == "win32":
+        # By value where the names are missing, so the Windows half of this is
+        # testable from the suite's other platforms too.
+        detached = (getattr(subprocess, "DETACHED_PROCESS", 0x00000008)
+                    | getattr(subprocess, "CREATE_NEW_PROCESS_GROUP", 0x00000200))
+        subprocess.Popen(args, creationflags=detached, close_fds=True)
+        return
+    os.execv(args[0], args)
+
+
 def send(cmd, wait=False, timeout=0, **args):
     """Send one request; the reply, or None when no instance is running.
 

dikte/mac_window.py

@@ -0,0 +1,203 @@
+"""The parts of AppKit a dictation needs on macOS and Qt does not reach.
+
+Two jobs, both about staying out of the user's way.
+
+The first is keeping the indicator on screen. Qt draws it as a tool window,
+which on macOS is an NSPanel, and an NSPanel is hidden by the system the moment
+its application stops being the active one. For a dictation indicator that is
+exactly backwards: you press the shortcut inside some other program, so Dikte is
+never the active application, and the one window that has something to say
+disappears as soon as you look away from it. Three settings AppKit has and Qt
+does not expose:
+
+  hidesOnDeactivate = NO      stay put when another application comes forward
+  collectionBehavior          show on whichever desktop is in front, including
+                              over a full screen window, and stay out of Cmd+Tab
+  nonactivating panel         come to the front without bringing Dikte with it
+
+The second is putting the front back. Opening the microphone activates Dikte
+whatever the indicator does, so app.py watches for that and calls activate()
+here. See _give_the_front_back() there for the measurement.
+
+Done through the Objective-C runtime rather than a binding, because Dikte has no
+third party Python packages and this is a handful of messages to three objects.
+The runtime is loaded in _appkit() rather than at import, the way paste.py loads
+its frameworks in _macos_api(): that is the one function a test fakes, and it is
+what lets the tests below run on a machine that has no AppKit at all.
+"""
+
+import ctypes
+import ctypes.util
+import os
+
+from PyQt6.QtGui import QGuiApplication
+
+# NSWindowCollectionBehavior, as of the macOS these names come from:
+CAN_JOIN_ALL_SPACES = 1 << 0
+IGNORES_CYCLE = 1 << 6        # not a window Cmd+Tab should ever land on
+FULL_SCREEN_AUXILIARY = 1 << 8
+BEHAVIOUR = CAN_JOIN_ALL_SPACES | IGNORES_CYCLE | FULL_SCREEN_AUXILIARY
+
+# NSWindowStyleMaskNonactivatingPanel. Without it, ordering the indicator to
+# the front brings Dikte to the front with it, and the application the user was
+# typing in loses focus the moment they start dictating: the Cmd+V at the end
+# then lands on the indicator instead of their document. Qt has no flag for
+# this; WA_ShowWithoutActivating governs the show, not what the panel does to
+# the application afterwards.
+NONACTIVATING_PANEL = 1 << 7
+
+_appkit_runtime = None
+
+
+class _AppKit:
+    """objc_msgSend under the signatures this file sends it through.
+
+    It has no fixed signature of its own, and calling it through the wrong
+    argument or return types is how a Mac crashes rather than raises, so each
+    one is spelled out once here and used by name below.
+    """
+
+    def __init__(self, objc):
+        self.objc = objc
+        objc.sel_registerName.restype = ctypes.c_void_p
+        objc.sel_registerName.argtypes = [ctypes.c_char_p]
+        objc.objc_getClass.restype = ctypes.c_void_p
+        objc.objc_getClass.argtypes = [ctypes.c_char_p]
+        self.ask = self._as(ctypes.c_void_p)
+        self.ask_bool = self._as(ctypes.c_bool)
+        self.ask_pid = self._as(ctypes.c_int)          # pid_t is an int32
+        self.ask_unsigned = self._as(ctypes.c_ulong)   # NSUInteger
+        self.tell_bool = self._as(None, ctypes.c_bool)
+        self.tell_unsigned = self._as(None, ctypes.c_ulong)
+        self.ask_of_class = self._as(ctypes.c_bool, ctypes.c_void_p)
+        self.ask_of_pid = self._as(ctypes.c_void_p, ctypes.c_int)
+        self.ask_with_options = self._as(ctypes.c_bool, ctypes.c_ulong)
+
+    def _as(self, returns, *arguments):
+        return ctypes.cast(self.objc.objc_msgSend, ctypes.CFUNCTYPE(
+            returns, ctypes.c_void_p, ctypes.c_void_p, *arguments))
+
+    def selector(self, name):
+        return self.objc.sel_registerName(name)
+
+    def shared(self, class_name, selector):
+        """A class's singleton, e.g. +[NSWorkspace sharedWorkspace]."""
+        return ctypes.c_void_p(self.ask(
+            ctypes.c_void_p(self.objc.objc_getClass(class_name)),
+            self.selector(selector)))
+
+
+def _appkit():
+    """The Objective-C runtime, loaded the first time something needs it.
+
+    Loaded here rather than at import so that this module can be imported on a
+    machine that has no AppKit: the tests stand on macOS from a Linux machine
+    and back, and this is the one function they replace to do it.
+    """
+    global _appkit_runtime
+    if _appkit_runtime is None:
+        _appkit_runtime = _AppKit(
+            ctypes.cdll.LoadLibrary(ctypes.util.find_library("objc")))
+    return _appkit_runtime
+
+
+def frontmost_pid():
+    """Which application is in front, by process id, or None when unasked.
+
+    A process id rather than the object itself: the object would have to be
+    retained to survive the trip, and a number needs nothing looking after it.
+    """
+    try:
+        api = _appkit()
+        workspace = api.shared(b"NSWorkspace", b"sharedWorkspace")
+        running = ctypes.c_void_p(api.ask(
+            workspace, api.selector(b"frontmostApplication")))
+        if not running:
+            return None
+        return int(api.ask_pid(running, api.selector(b"processIdentifier")))
+    except Exception:
+        return None
+
+
+def activate(pid):
+    """Put the application with that process id back in front.
+
+    False when it has gone away in the meantime, or when the message could not
+    be sent at all: a dictation is not worth failing over the window behind it.
+    """
+    if not pid:
+        return False
+    try:
+        api = _appkit()
+        running = ctypes.c_void_p(api.ask_of_pid(
+            ctypes.c_void_p(api.objc.objc_getClass(b"NSRunningApplication")),
+            api.selector(b"runningApplicationWithProcessIdentifier:"),
+            int(pid)))
+        if not running:
+            return False
+        # activateWithOptions: rather than the deprecated activate, and with no
+        # options: bringing every one of its windows forward is not asked for,
+        # only the application it was before Dikte took the front from it.
+        return bool(api.ask_with_options(
+            running, api.selector(b"activateWithOptions:"), 0))
+    except Exception:
+        return False
+
+
+def is_frontmost():
+    """Whether Dikte itself is the application in front.
+
+    By process id rather than -[NSRunningApplication isActive] on our own
+    process, which stays true once the application has ever been activated:
+    measured True with another application plainly in front.
+    """
+    pid = frontmost_pid()
+    return pid is not None and pid == os.getpid()
+
+
+def _is_panel(api, window):
+    """Whether this window is an NSPanel, which is the only kind the
+    nonactivating bit is legal on: setting it on a plain NSWindow raises an
+    Objective-C exception, and an exception through ctypes takes the process
+    down with it."""
+    panel = api.objc.objc_getClass(b"NSPanel")
+    if not panel:
+        return False
+    return bool(api.ask_of_class(window, api.selector(b"isKindOfClass:"),
+                                 ctypes.c_void_p(panel)))
+
+
+def keep_on_screen(widget):
+    """Ask the window behind `widget` to stay while other programs are used.
+
+    Silent when anything is not as expected: an indicator that cannot be made
+    to linger is still an indicator, and a dictation should not fail over the
+    window it is drawn in.
+    """
+    # Only the Cocoa backend hands out a real NSView. Under the offscreen
+    # platform the tests run on, winId() is a number that means something else
+    # entirely, and sending an Objective-C message to it is how a test run
+    # turns into a crash.
+    if QGuiApplication.platformName() != "cocoa":
+        return False
+    try:
+        api = _appkit()
+        view = ctypes.c_void_p(int(widget.winId()))
+        window = api.ask(view, api.selector(b"window"))
+        if not window:
+            return False
+        window = ctypes.c_void_p(window)
+        api.tell_bool(window, api.selector(b"setHidesOnDeactivate:"), False)
+        api.tell_unsigned(window, api.selector(b"setCollectionBehavior:"),
+                          BEHAVIOUR)
+        # Only a panel may carry the nonactivating bit, and only a panel is
+        # asked to: on anything else the message raises, and an Objective-C
+        # exception through ctypes takes the process with it.
+        if _is_panel(api, window):
+            mask = api.ask_unsigned(window, api.selector(b"styleMask"))
+            if not mask & NONACTIVATING_PANEL:
+                api.tell_unsigned(window, api.selector(b"setStyleMask:"),
+                                  mask | NONACTIVATING_PANEL)
+        return True
+    except (AttributeError, OSError, RuntimeError, ValueError):
+        return False

dikte/overlay.py

@@ -7,6 +7,8 @@ from PyQt6.QtCore import Qt, QTimer, QRectF, QPointF
 from PyQt6.QtGui import QColor, QCursor, QFont, QPainter, QPainterPath, QPen, QFontMetrics
 from PyQt6.QtWidgets import QWidget, QApplication
 
+from . import mac_window
+
 BARS = 22
 HEIGHT = 56
 MIN_WIDTH = 210
@@ -40,9 +42,11 @@ class Overlay(QWidget):
     of covering it, which is what lets a dictation and a command to the agent be
     under way at the same time and still both be visible."""
 
-    def __init__(self, corner="bottom-left", below=None, dismissable=False):
+    def __init__(self, corner="bottom-left", below=None, dismissable=False,
+                 screen_name=""):
         super().__init__(None)
         self.corner = corner
+        self.screen_name = screen_name
         self.below = below
         # A job that can run for ten minutes should not have to be watched for
         # ten minutes. Clicking such an indicator puts the progress away; the
@@ -202,6 +206,11 @@ class Overlay(QWidget):
         self._reposition()
         if not self.isVisible():
             self.show()
+        if sys.platform == "darwin":
+            # After show(), because the window it works on does not exist until
+            # then, and every time, because a window Qt rebuilt has the setting
+            # again at its default.
+            mac_window.keep_on_screen(self)
         if self._concealed:
             self.raise_()
             self._concealed = False
@@ -244,8 +253,15 @@ class Overlay(QWidget):
         self.resize(width, HEIGHT)
 
     def _reposition(self):
-        # On a multi-monitor setup, show up where the user actually is.
-        screen = QApplication.screenAt(QCursor.pos()) or QApplication.primaryScreen()
+        # The screen the settings name, or, when none is named or it is not
+        # plugged in right now, where the user actually is. Names are connector
+        # names on X11 and model names on macOS, where two identical monitors
+        # can share one; the first then wins.
+        screen = next(
+            (item for item in QApplication.screens() if item.name() == self.screen_name),
+            None,
+        )
+        screen = screen or QApplication.screenAt(QCursor.pos()) or QApplication.primaryScreen()
         area = screen.availableGeometry()
         left = "left" in self.corner
         top = "top" in self.corner

dikte/paste.py

@@ -147,7 +147,9 @@ def _program_keyboard(program, command, hint=""):
     def ready():
         return shutil.which(program) is not None
 
-    def press(shortcut, delay):
+    def press(shortcut, delay, _focus=None):
+        # Nothing here takes the front from the window being dictated into, so
+        # there is nothing to hand back: the process id is a macOS concern.
         if not ready():
             raise PasteError(t("{tool} not found, cannot paste automatically.",
                                tool=program))
@@ -296,11 +298,27 @@ def _ask_for_permission():
         pass
 
 
-def _macos_press(shortcut, delay):
+def _macos_press(shortcut, delay, focus=None):
     """Post the key down and up straight into the window system.
 
     Nothing is typed anywhere until macOS has been told to trust Dikte, and it
     only asks once, when the paste it was granted for is first tried.
+
+    `focus` is the application that was in front when the recording began. The
+    keys land wherever the window system is pointing, so a Dikte that has ended
+    up in front would swallow its own transcript; when that has happened the
+    front is handed back before pressing. Nothing is taken from anyone else: an
+    application the user went to while the transcription ran is where they want
+    the text now.
+
+    This runs on the transcription's own thread rather than the main one, and
+    the two calls it makes are the kind AppKit documents as answering
+    atomically wherever they are asked from: NSRunningApplication is thread
+    safe by its own header, and the workspace lookup behind it returns a
+    reference rather than anything that has to be held. Stressed with four
+    threads and 32000 lookups against a running main loop without a fault; if
+    one ever does happen, mac_window answers None and the press goes ahead
+    where it would have gone anyway.
     """
     keycode, flags = _macos_keys(shortcut)
     services, core = _macos_api()
@@ -310,6 +328,12 @@ def _macos_press(shortcut, delay):
             "macOS has not been told to let Dikte press keys. Turn Dikte on "
             "under System Settings → Privacy & Security → Accessibility."
         ))
+    if focus:
+        # Imported here rather than at the top: it reaches for QtGui, and a
+        # terminal that only wants the clipboard should not pay for that.
+        from . import mac_window
+        if mac_window.is_frontmost():
+            mac_window.activate(focus)
 
     time.sleep(delay)  # let the selection settle and focus come back
     down = services.CGEventCreateKeyboardEvent(None, keycode, True)
@@ -464,11 +488,14 @@ class _WinInput(ctypes.Structure):
     _fields_ = [("type", ctypes.c_ulong), ("union", _WinInputUnion)]
 
 
-def _win_press(shortcut, delay):
+def _win_press(shortcut, delay, _focus=None):
     """Post the presses and releases straight into the input queue.
 
     No permission stands in front of SendInput the way Accessibility does on
     macOS: whatever window has focus receives the combination.
+
+    Nothing here takes the front from the window being dictated into, so the
+    remembered process id has nothing to hand back to: it is a macOS concern.
     """
     codes = _win_keys(shortcut)
     user32, _ = _win_api()
@@ -675,7 +702,11 @@ def paste_ready():
     return desktop().ready()
 
 
-def press(shortcut="", delay=0.12):
-    """Press a paste combination, e.g. 'ctrl+v', or this desktop's own."""
+def press(shortcut="", delay=0.12, focus=None):
+    """Press a paste combination, e.g. 'ctrl+v', or this desktop's own.
+
+    `focus` is the process the keys are meant for, remembered when the
+    recording started: see the macOS press for what is done with it.
+    """
     here = desktop()
-    here.press(shortcut or here.shortcuts[0], delay)
+    here.press(shortcut or here.shortcuts[0], delay, focus)

dikte/paths.py

@@ -13,10 +13,18 @@ platform as an argument so that a test can stand on the other one.
 
 import os
 import pathlib
+import subprocess
 import sys
 
+# The one other platform constant every subprocess site needs, kept in this
+# leaf so no caller has to pull the audio stack in for it: console programs
+# started from a windowless process would otherwise each open a console window
+# of their own on Windows.
+NO_WINDOW = (getattr(subprocess, "CREATE_NO_WINDOW", 0)
+             if sys.platform == "win32" else 0)
 
-def _env(var, default):
+
+def env_path(var, default):
     """The directory a variable names, or the one it stands in for."""
     return pathlib.Path(os.environ.get(var) or os.path.expanduser(default))
 
@@ -34,11 +42,28 @@ def directories(platform=None):
         support = pathlib.Path.home() / "Library/Application Support/Dikte"
         return support, support
     if here == "win32":
-        roaming = _env("APPDATA", "~/AppData/Roaming")
-        local = _env("LOCALAPPDATA", "~/AppData/Local")
+        roaming = env_path("APPDATA", "~/AppData/Roaming")
+        local = env_path("LOCALAPPDATA", "~/AppData/Local")
         return roaming / "Dikte", local / "Dikte"
-    return (_env("XDG_CONFIG_HOME", "~/.config") / "dikte",
-            _env("XDG_DATA_HOME", "~/.local/share") / "dikte")
+    return (env_path("XDG_CONFIG_HOME", "~/.config") / "dikte",
+            env_path("XDG_DATA_HOME", "~/.local/share") / "dikte")
+
+
+def cache_dir(platform=None):
+    """The directory for answers worth keeping but never worth backing up.
+
+    A third place because a cache is neither settings nor data: losing it costs
+    a network request, not a model or a preference, and every system sets aside
+    a directory for exactly that kind of file, one that backups skip and
+    cleanup tools may empty. Storing it with the data would ask a backup to
+    carry files whose whole point is that they can be thrown away.
+    """
+    here = platform or sys.platform
+    if here == "darwin":
+        return pathlib.Path.home() / "Library/Caches/Dikte"
+    if here == "win32":
+        return env_path("LOCALAPPDATA", "~/AppData/Local") / "Dikte" / "cache"
+    return env_path("XDG_CACHE_HOME", "~/.cache") / "dikte"
 
 
 CONFIG_DIR, DATA_DIR = directories()

dikte/vad.py

@@ -17,13 +17,15 @@ import unicodedata
 
 # Stock phrases the models produce when handed silence. Kept deliberately
 # narrow: only sentences nobody dictates on purpose in a two-second clip.
+# Whisper does invent "you" and "bye" too, but people dictate both as whole
+# answers, so a single word never belongs here.
 HALLUCINATIONS = {
     "altyazi mk", "altyazi m k", "altyazi", "altyazilar",
     "abone olmayi unutmayin", "izlediginiz icin tesekkurler",
     "izlediginiz icin tesekkur ederim", "izlediginiz icin tesekkur ederiz",
     "kanalima abone olmayi unutmayin", "altyazi mk altyazi mk",
     "thanks for watching", "thank you for watching", "thanks for watching!",
-    "please subscribe", "subscribe to my channel", "you", "bye",
+    "please subscribe", "subscribe to my channel",
     "mbc masr", "sous titres realises par la communaute damara org",
     "amara org community", "sous titrage st 501",
 }

dikte/worker.py

@@ -6,6 +6,7 @@ whatever came of it: an answer to a question, or a sentence saying what was
 done.
 """
 
+import collections
 import os
 import shutil
 import sys
@@ -45,25 +46,50 @@ class Pipeline(QObject):
         self.conf = conf
         self._thread = None
         self._stop = threading.Event()
+        # Recordings waiting their turn, and whether a thread is working them
+        # off. The flag rather than the thread's own liveness, because a thread
+        # stays alive for a moment after deciding it is done, and a job arriving
+        # in that moment would be left in the queue with nobody coming back.
+        self._jobs = collections.deque()
+        self._draining = False
+        self._jobs_lock = threading.Lock()
 
     @property
     def busy(self):
         return self._thread is not None and self._thread.is_alive()
 
-    def run(self, wav_path, duration, rms_values=(), ask=False, paste=None):
+    def run(self, wav_path, duration, rms_values=(), ask=False, paste=None,
+            focus=None):
         """`paste` overrides the setting for this one run, which is what a
         dictation asked for from a terminal wants: the text comes back down the
-        socket, and pasting it into whatever had focus is nobody's intention."""
-        if self.busy:
-            return
+        socket, and pasting it into whatever had focus is nobody's intention.
+
+        `focus` is the application that was in front when the recording began,
+        as a process id, and is where the paste is meant to land.
+
+        A run started while one is going waits its turn rather than being
+        dropped: the next dictation can be spoken while the last one is still
+        being cleaned up, and each one is finished, pasted and reported in the
+        order it was spoken."""
+        with self._jobs_lock:
+            self._jobs.append((wav_path, duration, list(rms_values), ask, paste,
+                               focus))
+            if self._draining:
+                return
+            self._draining = True
         self._stop.clear()
-        self._thread = threading.Thread(
-            target=self._work,
-            args=(wav_path, duration, list(rms_values), ask, paste),
-            daemon=True,
-        )
+        self._thread = threading.Thread(target=self._drain, daemon=True)
         self._thread.start()
 
+    def _drain(self):
+        while True:
+            with self._jobs_lock:
+                if not self._jobs:
+                    self._draining = False
+                    return
+                job = self._jobs.popleft()
+            self._work(*job)
+
     def cancel(self):
         """Give up on a job already under way.
 
@@ -73,7 +99,8 @@ class Pipeline(QObject):
         """
         self._stop.set()
 
-    def _work(self, wav_path, duration, rms_values, ask, paste_override=None):
+    def _work(self, wav_path, duration, rms_values, ask, paste_override=None,
+              focus=None):
         conf = self.conf
         started = time.monotonic()
         raw = ""
@@ -107,11 +134,16 @@ class Pipeline(QObject):
 
             text = raw
             warning = ""
+            # Remembered rather than re-derived at the history write below: the
+            # ask path runs cleanup under a different setting, and the record
+            # should say what happened, not what one of the two gates implies.
+            cleaned = False
             # Claude reads through “eee” and “hani” without help, so a dictation
             # on its way there is normally sent as it was heard, one API call and
             # a second or two lighter.
             if (conf["assistant_cleanup"] if ask else conf["cleanup_enabled"]):
                 self.stage.emit(t("Cleaning up…"))
+                cleaned = True
                 try:
                     text = cleanup.run(raw, conf, conf.cleanup_prompt())
                 except api.ApiError as exc:
@@ -137,62 +169,112 @@ class Pipeline(QObject):
             if paste_override is not None:
                 wants_paste = paste_override
 
-            with _paste_lock:
-                previous = (paste.read_clipboard()
-                            if conf["restore_clipboard"] and wants_paste else None)
-                try:
-                    paste.copy(text)
-                    if wants_paste:
-                        self.stage.emit(t("Pasting…"))
-                        paste.press(conf["paste_shortcut"])
-                finally:
-                    if previous is not None:
-                        # Let the focused application consume the temporary
-                        # transcription before putting every old clipboard type
-                        # back.  This also runs when key injection fails.
-                        time.sleep(0.35)
-                        paste.copy_bytes(previous)
-
-            cfg.append_history({
+            # Into the history before the paste is attempted: the record says
+            # what was dictated, not whether a key press landed, and a paste
+            # that fails must not take the transcript down with it.
+            record = {
                 "ts": time.strftime("%Y-%m-%d %H:%M:%S"),
                 "duration": round(duration, 1),
                 "elapsed": round(time.monotonic() - started, 1),
                 "model": target.model,
-                "cleanup_model": cleanup.model(conf) if conf["cleanup_enabled"] else "",
+                "cleanup_model": cleanup.model(conf) if cleaned else "",
                 "cleanup_error": warning,
                 "mode": "ask" if ask else "",
                 "question": question,
-                "assistant_model": conf["assistant_model"] if ask else "",
+                "assistant": assistant.provider(conf) if ask else "",
+                "assistant_model": assistant.model(conf) if ask else "",
                 "raw": raw,
                 "text": text,
-            })
+            }
+            cfg.append_history(record)
             try:
                 cfg.trim_history(conf["history_limit"])
             except OSError as exc:
                 print(f"dikte: could not trim the history: {exc}", file=sys.stderr)
+
+            with _paste_lock:
+                previous = (paste.read_clipboard()
+                            if conf["restore_clipboard"] and wants_paste else None)
+                paste.copy(text)
+                if wants_paste:
+                    self.stage.emit(t("Pasting…"))
+                    try:
+                        paste.press(conf["paste_shortcut"], focus=focus)
+                    except paste.PasteError as exc:
+                        # The transcript is on the clipboard and in the history;
+                        # a key press that would not land is a warning, not a
+                        # failure, and the old clipboard is NOT put back over
+                        # the text the user now has to paste by hand.
+                        previous = None
+                        warning = "\n".join(x for x in (
+                            warning,
+                            t("Copied, but pasting failed: {error}", error=exc),
+                        ) if x)
+                        # The row above was written before the paste, so it
+                        # has to be told what the paste then did.
+                        record = cfg.amend_history(
+                            record, cleanup_error=warning) or record
+                if previous is not None:
+                    # Let the focused application consume the temporary
+                    # transcription before putting every old clipboard type
+                    # back.
+                    time.sleep(0.35)
+                    paste.copy_bytes(previous)
+
             self.finished.emit(raw, text, warning)
 
         except assistant.Cancelled:
             self.cancelled.emit()
         except (api.ApiError, paste.PasteError, assistant.AssistantError) as exc:
             print(f"dikte: {exc}", file=sys.stderr)
-            self.failed.emit(str(exc))
+            self.failed.emit(self._keeping(wav_path, str(exc)))
         except Exception as exc:  # never fail silently
             traceback.print_exc()
-            self.failed.emit(t("Unexpected error: {error}", error=exc))
+            self.failed.emit(self._keeping(wav_path, t("Unexpected error: {error}",
+                                                       error=exc)))
         finally:
             self._discard(wav_path)
 
+    def _keeping(self, wav_path, message):
+        """Put the failed run's audio somewhere a retry can find it.
+
+        A dictation that died on the way to the model is speech the user cannot
+        say again from memory; deleting it because a server was down turns one
+        failure into two. Kept regardless of the keep_audio setting, which is
+        about the runs that succeeded.
+        """
+        kept = self._keep(wav_path)
+        if not kept:
+            return message
+        return message + "\n" + t("The recording was kept: {path}", path=kept)
+
+    def _keep(self, wav_path):
+        """Move the WAV into the recordings directory; its new path, or ''."""
+        try:
+            cfg.RECORDINGS_DIR.mkdir(parents=True, exist_ok=True)
+            base = time.strftime("%Y%m%d-%H%M%S")
+            # Two runs can finish inside the same second; the first one kept
+            # must not be overwritten by the second.
+            for suffix in ("",) + tuple(f"-{n}" for n in range(1, 100)):
+                target = cfg.RECORDINGS_DIR / f"{base}{suffix}.wav"
+                if not target.exists():
+                    shutil.move(wav_path, target)
+                    return str(target)
+            return ""
+        except OSError as exc:
+            print(f"dikte: could not keep the audio: {exc}", file=sys.stderr)
+            return ""
+
     def _discard(self, wav_path):
         if not os.path.exists(wav_path):
             return
         if self.conf["keep_audio"]:
-            try:
-                cfg.RECORDINGS_DIR.mkdir(parents=True, exist_ok=True)
-                shutil.move(wav_path, cfg.RECORDINGS_DIR / (time.strftime("%Y%m%d-%H%M%S") + ".wav"))
+            if self._keep(wav_path):
                 return
-            except OSError:
-                pass
+            # The move failing is no reason to delete what the user asked to
+            # keep: the temporary file stays where it is, named in the log.
+            print(f"dikte: the audio stays at {wav_path}", file=sys.stderr)
+            return
         try:
             os.unlink(wav_path)
         except OSError:

install.ps1

@@ -23,9 +23,20 @@ $autostartLink = Join-Path $startup "Dikte.lnk"
 $cmdShim = Join-Path $env:LOCALAPPDATA "Microsoft\WindowsApps\dikte.cmd"
 
 if ($Uninstall) {
-    foreach ($path in @($shortcut, $autostartLink, $cmdShim)) {
+    foreach ($path in @($shortcut, $autostartLink)) {
         if (Test-Path $path) { Remove-Item $path -Force; Write-Host "removed: $path" }
     }
+    # The packaged install writes the same shim, naming its own dikte-cli.exe.
+    # Only the one naming this checkout is ours to delete; taking the other
+    # would break the `dikte` command of an install this script never made.
+    if (Test-Path $cmdShim) {
+        if ((Get-Content $cmdShim -Raw).Contains($entry)) {
+            Remove-Item $cmdShim -Force
+            Write-Host "removed: $cmdShim"
+        } else {
+            Write-Host "left alone: $cmdShim (it names another install, not this checkout)"
+        }
+    }
     Write-Host "Dikte's shortcuts are gone. The repository and your settings are not."
     exit 0
 }
@@ -65,6 +76,15 @@ foreach ($path in @($shortcut) + $(if ($Autostart) { @($autostartLink) } else {
     Write-Host "shortcut: $path"
 }
 
+if ($Autostart) {
+    # The packaged build keeps its sign-in entry in the registry. Left there
+    # beside the shortcut written above, both would start a Dikte at sign-in,
+    # and this install is the one being asked for.
+    Remove-ItemProperty -Path "HKCU:\Software\Microsoft\Windows\CurrentVersion\Run" `
+        -Name "Dikte" -ErrorAction SilentlyContinue
+    Write-Host "autostart: the Startup shortcut replaces any registry Run entry a packaged install left"
+}
+
 # --- the dikte command ------------------------------------------------------
 # The interpreter by its full path rather than by name: the one checked above is
 # the one the command line should run, whatever a later PATH change puts first.

packaging/dikte.iss

@@ -59,6 +59,13 @@ Name: "autostart"; Description: "Start Dikte when I sign in"
 [Files]
 Source: "{#Source}\*"; DestDir: "{app}"; Flags: recursesubdirs ignoreversion
 
+[InstallDelete]
+; A checkout's install.ps1 -Autostart is a shortcut in the Startup folder. The
+; registry entry this setup writes replaces it, and both left in place would be
+; two Diktes at every sign-in. Only under the autostart task: somebody who
+; unticked the box has not asked for their checkout's entry to go.
+Type: files; Name: "{userstartup}\Dikte.lnk"; Tasks: autostart
+
 [Icons]
 Name: "{autoprograms}\Dikte"; Filename: "{app}\Dikte.exe"
 
@@ -111,7 +118,14 @@ begin
 end;
 
 procedure CurUninstallStepChanged(CurUninstallStep: TUninstallStep);
+var
+  Shim: AnsiString;
 begin
+  { Only the shim this setup wrote, which is the one naming its dikte-cli.exe.
+    install.ps1 writes the same file for a checkout, naming that checkout's
+    Python, and a shim somebody else wrote is not this uninstaller's to take. }
   if CurUninstallStep = usUninstall then
-    DeleteFile(ShimPath());
+    if LoadStringFromFile(ShimPath(), Shim)
+       and (Pos(ExpandConstant('{app}\dikte-cli.exe'), Shim) > 0) then
+      DeleteFile(ShimPath());
 end;

packaging/dikte.spec

@@ -56,6 +56,18 @@ analysis = Analysis(                            # noqa: F821
     noarchive=False,
 )
 
+# Qt's xcb platform plugin uses libxkbcommon in two halves: the core library
+# and libxkbcommon-x11, which allocates keymap objects and hands them to the
+# core half to use and free, so the two must come from the same build. The
+# build machine has only the core half installed, which had PyInstaller
+# bundling that one while the other kept coming from the user's system, and a
+# 22.04-era core freeing what a current x11 half allocated is the startup
+# crash of issue #57. Ship neither: any desktop that can show a window
+# carries both, from one build.
+if not (MACOS or WINDOWS):
+    analysis.binaries = [entry for entry in analysis.binaries
+                         if "libxkbcommon" not in entry[0]]
+
 archive = PYZ(analysis.pure)                    # noqa: F821
 
 executable = EXE(                               # noqa: F821

packaging/whisper-vulkan/Dockerfile.build

@@ -0,0 +1,44 @@
+FROM ubuntu@sha256:2edbbc5dc405e9612ba3584ce95480277e3eb374407b5505fe26f17df77c7dbc
+
+ARG DEBIAN_FRONTEND=noninteractive
+ARG CMAKE_VERSION=3.31.6
+ARG CMAKE_SHA256=5a1133ff103c71eb5120e2cc3de922733e7d8a26a98ae716397e8676adb367bf
+
+COPY lunarg-signing-key-pub.asc /tmp/lunarg.asc
+
+RUN set -eux; \
+    test "$(sha256sum /tmp/lunarg.asc | cut -d' ' -f1)" = aa1c3c29673140e77f0d6a9aaeed5d9b5621e305ead51c59fae4458bbb4df92b; \
+    apt-get update; \
+    apt-get install --no-install-recommends -y \
+      build-essential=12.9ubuntu3 \
+      ca-certificates \
+      curl \
+      file \
+      git \
+      gnupg \
+      ninja-build=1.10.1-1 \
+      patchelf=0.14.3-1 \
+      python3 \
+      xz-utils; \
+    install -d -m 0755 /usr/share/keyrings; \
+    gpg --dearmor -o /usr/share/keyrings/lunarg.gpg /tmp/lunarg.asc; \
+    printf '%s\n' 'deb [signed-by=/usr/share/keyrings/lunarg.gpg] https://packages.lunarg.com/vulkan jammy main' \
+      > /etc/apt/sources.list.d/lunarg-vulkan.list; \
+    apt-get update; \
+    apt-get install --no-install-recommends -y \
+      libvulkan-dev=1.4.313.0~rc1-1lunarg22.04-1 \
+      vulkan-headers=1.4.313.0~rc1-1lunarg22.04-1 \
+      shaderc=2025.2~rc1-1lunarg22.04-1 \
+      spirv-headers=1.6.1+1.4.313.0~rc1-1lunarg22.04-1; \
+    curl --fail --location --retry 3 \
+      "https://github.com/Kitware/CMake/releases/download/v${CMAKE_VERSION}/cmake-${CMAKE_VERSION}-linux-x86_64.tar.gz" \
+      -o /tmp/cmake.tar.gz; \
+    test "$(sha256sum /tmp/cmake.tar.gz | cut -d' ' -f1)" = "$CMAKE_SHA256"; \
+    tar -xzf /tmp/cmake.tar.gz --strip-components=1 -C /usr/local; \
+    rm -rf /var/lib/apt/lists/* /tmp/cmake.tar.gz /tmp/lunarg.asc; \
+    cmake --version; \
+    glslc --version; \
+    test -f /usr/include/vulkan/vulkan.h; \
+    test -f /usr/share/cmake/SPIRV-Headers/SPIRV-HeadersConfig.cmake
+
+WORKDIR /work

packaging/whisper-vulkan/Dockerfile.runtime-cpu

@@ -0,0 +1,6 @@
+FROM ubuntu@sha256:2edbbc5dc405e9612ba3584ce95480277e3eb374407b5505fe26f17df77c7dbc
+ARG DEBIAN_FRONTEND=noninteractive
+RUN apt-get update \
+ && apt-get install --no-install-recommends -y ca-certificates curl libstdc++6 \
+ && rm -rf /var/lib/apt/lists/*
+WORKDIR /bundle

packaging/whisper-vulkan/Dockerfile.runtime-noicd

@@ -0,0 +1,9 @@
+FROM ubuntu@sha256:2edbbc5dc405e9612ba3584ce95480277e3eb374407b5505fe26f17df77c7dbc
+ARG DEBIAN_FRONTEND=noninteractive
+# The loader and nothing behind it: the machine that has libvulkan because
+# something else pulled it in, and no driver to go with it.
+RUN apt-get update \
+ && apt-get install --no-install-recommends -y \
+      ca-certificates curl libstdc++6 libvulkan1 \
+ && rm -rf /var/lib/apt/lists/*
+WORKDIR /bundle

packaging/whisper-vulkan/Dockerfile.runtime-vulkan

@@ -0,0 +1,7 @@
+FROM ubuntu@sha256:2edbbc5dc405e9612ba3584ce95480277e3eb374407b5505fe26f17df77c7dbc
+ARG DEBIAN_FRONTEND=noninteractive
+RUN apt-get update \
+ && apt-get install --no-install-recommends -y \
+      ca-certificates curl libstdc++6 libvulkan1 mesa-vulkan-drivers vulkan-tools \
+ && rm -rf /var/lib/apt/lists/*
+WORKDIR /bundle

packaging/whisper-vulkan/README.md

@@ -0,0 +1,49 @@
+# The Vulkan whisper-server bundle
+
+whisper.cpp publishes a CPU-only archive for Linux, so the graphics card on a
+Linux machine is out of reach through the Download button. This directory
+builds the archive upstream does not: `whisper-server` with a dynamic Vulkan
+backend next to the CPU ones, for x86_64, against the Ubuntu 22.04 runtime
+contract.
+
+It is published as a release of Dikte's own, `whisper.cpp-v<version>`, marked
+as a prerelease and kept off Latest so that neither the update check nor the
+download page picks it up. `dikte/ggml.py` fetches it by tag and installs it
+only when the archive's digest is the reviewed one; anything else falls back
+to upstream's CPU archive, and the settings window says when it did.
+
+## Publishing a new bundle
+
+1. Enable GitHub's immutable releases setting for the repository, and give the
+   `dependency-release` environment a required reviewer. Both are repository
+   settings, not something this workflow can do for itself.
+2. Run **whisper.cpp Vulkan bundle** on `master` with the new version and its
+   peeled commit, `expected_sha256` empty and `publish: false`. The run builds
+   the archive and reports its digest; without a reviewed digest it refuses to
+   publish, which is what the first run is for.
+3. Review that digest against a build of your own, then run the workflow again
+   with the same version and commit, `expected_sha256` set to it, and
+   `publish: true`. Approve the environment when it asks.
+4. Write the same version, tag and digest into `MANAGED_WHISPER_RELEASE`,
+   `MANAGED_WHISPER_VERSION` and `MANAGED_WHISPER_SHA256` in `dikte/ggml.py`,
+   and into `REVIEWED_WHISPER_VERSION` and `REVIEWED_WHISPER_SHA256` in the
+   workflow. `tests/test_packaging.py` holds the two sides together.
+5. Ship a Dikte release. Until one goes out, nobody's Dikte knows the new
+   bundle exists.
+
+## What this costs, and what it does not promise
+
+The digest lives in Dikte's source, so a backend update is a Dikte release.
+Linux x86_64 machines with a Vulkan loader stay on the pinned whisper.cpp
+version until step 5 happens, while every other platform follows upstream's
+newest release on its own. That is the deliberate trade: an executable Dikte
+downloads is not allowed to change without a reviewed digest behind it.
+
+The build is deterministic between two runs of the same builder, not across
+time. The base image, the CMake tarball, the LunarG packages and the direct
+apt packages are pinned by digest or version, but the Ubuntu and LunarG
+repository metadata behind them is not, and LunarG drops superseded packages.
+A rebuild months later can fail to resolve, or resolve to something that
+produces a different digest. Treat the published archive as the artifact, not
+as something reproducible on demand: a version bump means building,
+validating, reviewing the new digest and updating the pinned tuple together.

packaging/whisper-vulkan/build-package.sh

@@ -0,0 +1,128 @@
+#!/usr/bin/env bash
+set -euo pipefail
+shopt -s nullglob
+
+: "${SOURCE_DIR:=/src}"
+: "${OUT_DIR:=/work/out}"
+: "${WHISPER_VERSION:=1.9.3}"
+: "${WHISPER_COMMIT:=371b5a7561823ab2bb32142d2751e35e7534727b}"
+: "${SOURCE_DATE_EPOCH:=1787219223}"
+
+export SOURCE_DATE_EPOCH TZ=UTC LC_ALL=C LANG=C
+asset=whisper-bin-ubuntu-vulkan-x64
+build=/work/build
+source_copy=/work/source
+root="$OUT_DIR/root/$asset"
+
+rm -rf "$build" "$source_copy" "$OUT_DIR"
+mkdir -p "$build" "$root/LICENSES"
+# Upstream configures bindings/javascript/package.json in the source directory.
+# Build a private copy so the checked-out, verified source remains untouched.
+cp -a "$SOURCE_DIR" "$source_copy"
+chmod -R u+w "$source_copy"
+git config --global --add safe.directory "$source_copy"
+
+cmake -S "$source_copy" -B "$build" -G Ninja \
+  -DCMAKE_BUILD_TYPE=Release \
+  -DCMAKE_BUILD_RPATH='$ORIGIN' \
+  -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+  -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+  -DCMAKE_C_FLAGS="-ffile-prefix-map=$source_copy=. -fdebug-prefix-map=$source_copy=. -fmacro-prefix-map=$source_copy=." \
+  -DCMAKE_CXX_FLAGS="-ffile-prefix-map=$source_copy=. -fdebug-prefix-map=$source_copy=. -fmacro-prefix-map=$source_copy=." \
+  -DBUILD_SHARED_LIBS=ON \
+  -DGGML_BACKEND_DL=ON \
+  -DGGML_CPU_ALL_VARIANTS=ON \
+  -DGGML_NATIVE=OFF \
+  -DGGML_CCACHE=OFF \
+  -DGGML_OPENMP=OFF \
+  -DGGML_VULKAN=ON \
+  -DWHISPER_BUILD_EXAMPLES=ON \
+  -DWHISPER_BUILD_SERVER=ON \
+  -DWHISPER_BUILD_TESTS=OFF \
+  -DWHISPER_BUILD_IS_DEV=OFF \
+  -DWHISPER_CURL=OFF \
+  -DWHISPER_SDL2=OFF \
+  -DWHISPER_COMMON_FFMPEG=OFF \
+  -DWHISPER_BUILD_COMMIT="$WHISPER_COMMIT" \
+  -DWHISPER_BUILD_NUMBER=0
+cmake --build "$build" --target whisper-server --parallel "$(nproc)"
+
+# Package an allowlist, not everything examples/ happens to build in the future.
+cp -a "$build/bin/whisper-server" "$root/"
+cp -a "$build/bin"/libwhisper.so* "$root/"
+cp -a "$build/bin"/libggml.so* "$root/"
+cp -a "$build/bin"/libggml-base.so* "$root/"
+cp -a "$build/bin"/libggml-cpu*.so* "$root/"
+cp -a "$build/bin"/libggml-vulkan.so* "$root/"
+
+# Strip real ELF files only; preserve the SONAME symlink chains.
+while IFS= read -r -d '' file; do
+  if file "$file" | grep -q ELF; then
+    strip --strip-unneeded "$file"
+    patchelf --set-rpath '$ORIGIN' "$file"
+  fi
+done < <(find "$root" -type f -print0)
+
+cp "$SOURCE_DIR/LICENSE" "$root/LICENSES/whisper.cpp-MIT.txt"
+cp /packaging/licenses/cpp-httplib-MIT.txt "$root/LICENSES/"
+cp /packaging/licenses/nlohmann-json-MIT.txt "$root/LICENSES/"
+
+cat > "$root/BUILD-INFO.json" <<EOF
+{
+  "asset": "$asset.tar.gz",
+  "source": "https://github.com/ggml-org/whisper.cpp",
+  "source_version": "v$WHISPER_VERSION",
+  "source_commit": "$WHISPER_COMMIT",
+  "source_date_epoch": $SOURCE_DATE_EPOCH,
+  "build_platform": "ubuntu-22.04-x86_64",
+  "base_image": "ubuntu@sha256:2edbbc5dc405e9612ba3584ce95480277e3eb374407b5505fe26f17df77c7dbc",
+  "cmake": "3.31.6",
+  "cmake_flags": [
+    "BUILD_SHARED_LIBS=ON",
+    "C/CXX_FILE_PREFIX_MAP=/work/source=.",
+    "GGML_BACKEND_DL=ON",
+    "GGML_CPU_ALL_VARIANTS=ON",
+    "GGML_NATIVE=OFF",
+    "GGML_CCACHE=OFF",
+    "GGML_OPENMP=OFF",
+    "GGML_VULKAN=ON",
+    "WHISPER_BUILD_EXAMPLES=ON",
+    "WHISPER_BUILD_SERVER=ON",
+    "WHISPER_BUILD_TESTS=OFF",
+    "WHISPER_BUILD_IS_DEV=OFF",
+    "WHISPER_CURL=OFF",
+    "WHISPER_SDL2=OFF",
+    "WHISPER_COMMON_FFMPEG=OFF"
+  ],
+  "runtime_contract": {
+    "minimum_glibc": "2.34",
+    "minimum_glibcxx": "3.4.30",
+    "required": ["x86_64 Linux", "glibc", "libstdc++.so.6", "libgcc_s.so.1"],
+    "optional_gpu": ["libvulkan.so.1", "a working Vulkan ICD"],
+    "cpu_fallback": "dynamic CPU backends are included; -ng forces CPU"
+  }
+}
+EOF
+
+# A deterministic CycloneDX sidecar generated from the files actually shipped.
+ROOT="$root" VERSION="$WHISPER_VERSION" COMMIT="$WHISPER_COMMIT" EPOCH="$SOURCE_DATE_EPOCH" \
+python3 /packaging/make-sbom.py > "$root/$asset.cdx.json"
+
+(
+  cd "$root"
+  find . -type f ! -name SHA256SUMS -print0 \
+    | sort -z \
+    | xargs -0 sha256sum
+) > "$root/SHA256SUMS"
+
+mkdir -p "$OUT_DIR"
+tar --sort=name --owner=0 --group=0 --numeric-owner \
+  --mtime="@$SOURCE_DATE_EPOCH" \
+  --pax-option=delete=atime,delete=ctime \
+  -C "$OUT_DIR/root" -cf - "$asset" \
+  | gzip -n -9 > "$OUT_DIR/$asset.tar.gz"
+(
+  cd "$OUT_DIR"
+  sha256sum "$asset.tar.gz" > "$asset.tar.gz.sha256"
+)
+cp "$root/$asset.cdx.json" "$OUT_DIR/$asset.cdx.json"

packaging/whisper-vulkan/licenses/cpp-httplib-MIT.txt

@@ -0,0 +1,21 @@
+The MIT License (MIT)
+
+Copyright (c) 2017 yhirose
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

packaging/whisper-vulkan/licenses/nlohmann-json-MIT.txt

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2013-2022 Niels Lohmann
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

packaging/whisper-vulkan/lunarg-signing-key-pub.asc

@@ -0,0 +1,31 @@
+-----BEGIN PGP PUBLIC KEY BLOCK-----
+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+=nXAe
+-----END PGP PUBLIC KEY BLOCK-----

packaging/whisper-vulkan/make-sbom.py

@@ -0,0 +1,116 @@
+#!/usr/bin/env python3
+import datetime
+import hashlib
+import json
+import os
+import uuid
+from pathlib import Path
+
+root = Path(os.environ["ROOT"])
+version = os.environ["VERSION"]
+commit = os.environ["COMMIT"]
+epoch = int(os.environ["EPOCH"])
+asset = "whisper-bin-ubuntu-vulkan-x64"
+sbom_path = root / f"{asset}.cdx.json"
+
+def digest(path):
+    h = hashlib.sha256()
+    with path.open("rb") as stream:
+        for block in iter(lambda: stream.read(1024 * 1024), b""):
+            h.update(block)
+    return h.hexdigest()
+
+files = []
+for path in sorted(root.rglob("*")):
+    if path != sbom_path and path.is_file() and not path.is_symlink():
+        rel = path.relative_to(root).as_posix()
+        files.append({
+            "type": "file",
+            "bom-ref": f"file:{rel}",
+            "name": rel,
+            "hashes": [{"alg": "SHA-256", "content": digest(path)}],
+        })
+
+ts = datetime.datetime.fromtimestamp(
+    epoch, datetime.timezone.utc,
+).isoformat().replace("+00:00", "Z")
+root_ref = f"pkg:github/ggml-org/whisper.cpp@{version}?commit={commit}"
+ggml_ref = "pkg:github/ggml-org/[email protected]"
+httplib_ref = "pkg:github/yhirose/[email protected]"
+json_ref = "pkg:github/nlohmann/[email protected]"
+
+sbom = {
+    "bomFormat": "CycloneDX",
+    "specVersion": "1.6",
+    "serialNumber": f"urn:uuid:{uuid.uuid5(uuid.NAMESPACE_URL, root_ref)}",
+    "version": 1,
+    "metadata": {
+        "timestamp": ts,
+        "tools": {"components": [
+            {"type": "application", "name": "make-sbom.py", "version": "1"},
+            {"type": "application", "name": "CMake", "version": "3.31.6"},
+            {"type": "application", "name": "glslc", "version": "2025.2"},
+        ]},
+        "component": {
+            "type": "application",
+            "bom-ref": root_ref,
+            "group": "ggml-org",
+            "name": "whisper-server",
+            "version": version,
+            "purl": root_ref,
+            "licenses": [{"expression": "MIT"}],
+            "externalReferences": [{
+                "type": "vcs",
+                "url": f"https://github.com/ggml-org/whisper.cpp/tree/{commit}",
+            }],
+            "properties": [
+                {"name": "dikte:asset-name", "value": f"{asset}.tar.gz"},
+                {"name": "dikte:source-commit", "value": commit},
+                {"name": "dikte:runtime:glibc-minimum", "value": "2.34"},
+                {"name": "dikte:runtime:glibcxx-minimum", "value": "3.4.30"},
+                {"name": "dikte:runtime:vulkan-loader", "value": "optional; libvulkan.so.1"},
+            ],
+        },
+    },
+    "components": [
+        {
+            "type": "library",
+            "bom-ref": ggml_ref,
+            "group": "ggml-org",
+            "name": "ggml",
+            "version": "0.20.2",
+            "purl": ggml_ref,
+            "licenses": [{"expression": "MIT"}],
+            "properties": [{
+                "name": "dikte:source",
+                "value": "vendored by the pinned whisper.cpp commit",
+            }],
+        },
+        {
+            "type": "library",
+            "bom-ref": httplib_ref,
+            "group": "yhirose",
+            "name": "cpp-httplib",
+            "version": "0.20.0",
+            "purl": httplib_ref,
+            "licenses": [{"expression": "MIT"}],
+        },
+        {
+            "type": "library",
+            "bom-ref": json_ref,
+            "group": "nlohmann",
+            "name": "json",
+            "version": "3.11.2",
+            "purl": json_ref,
+            "licenses": [{"expression": "MIT"}],
+        },
+        *files,
+    ],
+    "dependencies": [{
+        "ref": root_ref,
+        "dependsOn": [ggml_ref, httplib_ref, json_ref]
+                     + [item["bom-ref"] for item in files],
+    }],
+}
+json.dump(sbom, fp=os.sys.stdout, indent=2, sort_keys=True)
+print()

packaging/whisper-vulkan/smoke-runtime.sh

@@ -0,0 +1,76 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+mode=${1:?usage: smoke-runtime.sh cpu|noicd|vulkan}
+: "${OUT_DIR:=work/out}"
+: "${FIXTURE_SOURCE:=vendor/whisper.cpp}"
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+OUT_DIR="$(realpath "$OUT_DIR")"
+FIXTURE_SOURCE="$(realpath "$FIXTURE_SOURCE")"
+asset=whisper-bin-ubuntu-vulkan-x64
+case "$mode" in
+  cpu) dockerfile=Dockerfile.runtime-cpu; image=dikte-whisper-runtime-cpu:spike ;;
+  noicd) dockerfile=Dockerfile.runtime-noicd; image=dikte-whisper-runtime-noicd:spike ;;
+  vulkan) dockerfile=Dockerfile.runtime-vulkan; image=dikte-whisper-runtime-vulkan:spike ;;
+  *) echo "unknown mode: $mode" >&2; exit 2 ;;
+esac
+
+tmp=$(mktemp -d)
+trap 'rm -rf "$tmp"' EXIT
+tar -xzf "$OUT_DIR/$asset.tar.gz" -C "$tmp"
+docker build --pull=false -f "$SCRIPT_DIR/$dockerfile" -t "$image" "$SCRIPT_DIR"
+
+args=("/bundle/$asset/whisper-server" -m /fixtures/model.bin
+      --host 127.0.0.1 --port 8080
+      --inference-path /v1/audio/transcriptions -l auto -sns -nlp)
+env_args=()
+# No -ng anywhere: Dikte passes it only when its GPU setting is off, so the
+# run that has to survive a missing loader or a missing device is this one,
+# where the backend registry actually goes looking for them.
+if [[ "$mode" == vulkan ]]; then
+  env_args=(-e LIBGL_ALWAYS_SOFTWARE=1
+            -e VK_ICD_FILENAMES=/usr/share/vulkan/icd.d/lvp_icd.x86_64.json)
+fi
+
+docker run --rm --name "dikte-whisper-$mode-smoke" \
+  -e SMOKE_MODE="$mode" \
+  "${env_args[@]}" \
+  -v "$tmp/$asset:/bundle/$asset:ro" \
+  -v "$FIXTURE_SOURCE/models/for-tests-ggml-base.en.bin:/fixtures/model.bin:ro" \
+  -v "$FIXTURE_SOURCE/samples/jfk.wav:/fixtures/jfk.wav:ro" \
+  "$image" bash -ec '
+    if [ "$SMOKE_MODE" = cpu ] && ldconfig -p | grep -q libvulkan.so.1; then
+      echo "CPU smoke image unexpectedly has a Vulkan loader" >&2
+      exit 1
+    fi
+    if [ "$SMOKE_MODE" = noicd ]; then
+      if ! ldconfig -p | grep -q libvulkan.so.1; then
+        echo "no-ICD smoke image has no Vulkan loader to load" >&2
+        exit 1
+      fi
+      if compgen -G "/usr/share/vulkan/icd.d/*.json" >/dev/null; then
+        echo "no-ICD smoke image has a driver after all" >&2
+        exit 1
+      fi
+    fi
+    "$@" >/tmp/server.log 2>&1 &
+    pid=$!
+    trap "kill $pid 2>/dev/null || true" EXIT
+    for _ in $(seq 1 120); do
+      kill -0 "$pid" 2>/dev/null || { cat /tmp/server.log; exit 1; }
+      if curl --silent --show-error --fail --max-time 180 \
+          -F file=@/fixtures/jfk.wav -F response_format=json \
+          http://127.0.0.1:8080/v1/audio/transcriptions >/tmp/response.json; then
+        grep -q "\"text\"" /tmp/response.json
+        if [ "$SMOKE_MODE" = vulkan ]; then
+          grep -q "loaded Vulkan backend" /tmp/server.log
+        fi
+        cat /tmp/response.json
+        cat /tmp/server.log
+        exit 0
+      fi
+      sleep 1
+    done
+    cat /tmp/server.log
+    exit 1
+  ' bash "${args[@]}"

packaging/whisper-vulkan/validate-package.sh

@@ -0,0 +1,145 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+: "${OUT_DIR:=work/out}"
+: "${SOURCE_DIR:=whisper.cpp}"
+asset=whisper-bin-ubuntu-vulkan-x64
+archive="$OUT_DIR/$asset.tar.gz"
+tmp=$(mktemp -d)
+trap 'rm -rf "$tmp"' EXIT
+
+test -s "$archive"
+(cd "$OUT_DIR" && sha256sum --check "$asset.tar.gz.sha256")
+ARCHIVE="$archive" ASSET="$asset" python3 - <<'PY'
+import os
+import posixpath
+import tarfile
+
+archive = os.environ["ARCHIVE"]
+asset = os.environ["ASSET"]
+
+
+def under_root(name):
+    normalized = posixpath.normpath(name)
+    return (not posixpath.isabs(normalized)
+            and normalized != ".."
+            and not normalized.startswith("../")
+            and normalized.split("/", 1)[0] == asset)
+
+
+with tarfile.open(archive, "r:gz") as bundle:
+    for member in bundle:
+        if not under_root(member.name):
+            raise SystemExit(f"unsafe archive member: {member.name}")
+        if member.isdev() or member.isfifo():
+            raise SystemExit(f"special archive member: {member.name}")
+        if not (member.isdir() or member.isfile()
+                or member.issym() or member.islnk()):
+            raise SystemExit(f"unsupported archive member: {member.name}")
+        if member.issym():
+            target = posixpath.join(posixpath.dirname(member.name),
+                                    member.linkname)
+            if not under_root(target):
+                raise SystemExit(f"unsafe symlink: {member.name}")
+        if member.islnk() and not under_root(member.linkname):
+            raise SystemExit(f"unsafe hardlink: {member.name}")
+PY
+tar -xzf "$archive" -C "$tmp"
+root="$tmp/$asset"
+
+test -x "$root/whisper-server"
+test -f "$root/libwhisper.so"
+test -f "$root/libggml.so"
+test -f "$root/libggml-base.so"
+test -f "$root/libggml-vulkan.so"
+compgen -G "$root/libggml-cpu-*.so" >/dev/null
+test -f "$root/LICENSES/whisper.cpp-MIT.txt"
+test -f "$root/LICENSES/cpp-httplib-MIT.txt"
+test -f "$root/LICENSES/nlohmann-json-MIT.txt"
+(cd "$root" && sha256sum --check SHA256SUMS)
+
+# All shipped ELF objects must be relocatable and must not remember /work.
+while IFS= read -r -d '' file; do
+  file "$file" | grep -q ELF || continue
+  dynamic=$(readelf -d "$file")
+  if ! grep -Fq 'Library runpath: [$ORIGIN]' <<<"$dynamic"; then
+    echo "runpath is not \$ORIGIN in $file" >&2
+    exit 1
+  fi
+  if grep -Eq '/(home|tmp|work)/' <<<"$dynamic"; then
+    echo "build path remains in $file" >&2
+    exit 1
+  fi
+done < <(find "$root" -type f -print0)
+
+# Vulkan remains a plugin dependency. The executable must start without a loader.
+if readelf -d "$root/whisper-server" | grep -q 'libvulkan.so'; then
+  echo "whisper-server links Vulkan instead of loading it as a plugin" >&2
+  exit 1
+fi
+readelf -d "$root/libggml-vulkan.so" | grep -q 'libvulkan.so.1'
+
+# Ubuntu 22.04 establishes the glibc ceiling promised by this artifact.
+ROOT="$root" python3 - <<'PY'
+import os, pathlib, re, subprocess
+root = pathlib.Path(os.environ['ROOT'])
+seen = {'GLIBC': set(), 'GLIBCXX': set(), 'CXXABI': set()}
+external = {
+    'libc.so.6', 'libgcc_s.so.1', 'libm.so.6', 'libstdc++.so.6',
+    'libvulkan.so.1', 'ld-linux-x86-64.so.2',
+}
+for path in root.iterdir():
+    if not path.is_file() or path.is_symlink():
+        continue
+    header = subprocess.run(['readelf', '-h', path], text=True,
+                            stdout=subprocess.PIPE,
+                            stderr=subprocess.DEVNULL).stdout
+    if not header:
+        continue
+    if 'Machine:                           Advanced Micro Devices X86-64' not in header:
+        raise SystemExit(f'wrong ELF architecture: {path.name}')
+    dynamic = subprocess.run(['readelf', '-d', path], text=True,
+                             stdout=subprocess.PIPE,
+                             stderr=subprocess.DEVNULL).stdout
+    needed = re.findall(r'\(NEEDED\).*\[(.*?)\]', dynamic)
+    unexpected = [name for name in needed
+                  if name not in external
+                  and not re.fullmatch(
+                      r'lib(?:whisper|ggml(?:-base)?)\.so\.\d+', name)]
+    if unexpected:
+        raise SystemExit(
+            f'unexpected DT_NEEDED in {path.name}: {unexpected}')
+    if path.name != 'libggml-vulkan.so' and 'libvulkan.so.1' in needed:
+        raise SystemExit(f'Vulkan is not plugin-only in {path.name}')
+    contents = path.read_bytes()
+    for marker in (b'/home/', b'/tmp/', b'/work/'):
+        if marker in contents:
+            raise SystemExit(
+                f'build path {marker!r} remains in {path.name}')
+    text = subprocess.run(['objdump', '-T', path], text=True,
+                          stdout=subprocess.PIPE, stderr=subprocess.DEVNULL).stdout
+    for family in seen:
+        pattern = rf'{family}_([0-9]+(?:\.[0-9]+)+)'
+        seen[family].update(tuple(map(int, version.split('.')))
+                            for version in re.findall(pattern, text))
+assert seen['GLIBC'] and max(seen['GLIBC']) <= (2, 34), max(seen['GLIBC'])
+assert seen['GLIBCXX'] and max(seen['GLIBCXX']) <= (3, 4, 30), max(seen['GLIBCXX'])
+assert seen['CXXABI'] and max(seen['CXXABI']) <= (1, 3, 13), max(seen['CXXABI'])
+for family, versions in seen.items():
+    print(f'maximum {family} symbol:', '.'.join(map(str, max(versions))))
+PY
+
+python3 - "$root/$asset.cdx.json" <<'PY'
+import json, sys
+with open(sys.argv[1], encoding='utf-8') as stream:
+    doc = json.load(stream)
+assert doc['bomFormat'] == 'CycloneDX'
+assert doc['specVersion'] == '1.6'
+assert doc['metadata']['component']['name'] == 'whisper-server'
+assert len(doc['components']) >= 3
+print('SBOM components:', len(doc['components']))
+PY
+
+LD_LIBRARY_PATH='' "$root/whisper-server" --help >/dev/null 2>&1
+
+echo "structure: PASS"

tests/__init__.py

@@ -17,6 +17,7 @@ os.environ.setdefault("QT_QPA_PLATFORM", "offscreen")
 _SANDBOX = tempfile.mkdtemp(prefix="dikte-tests-")
 os.environ["XDG_CONFIG_HOME"] = os.path.join(_SANDBOX, "config")
 os.environ["XDG_DATA_HOME"] = os.path.join(_SANDBOX, "data")
+os.environ["XDG_CACHE_HOME"] = os.path.join(_SANDBOX, "cache")
 # Home goes with them: the shortcut file, the applications directory and every
 # macOS path start from it rather than from an XDG variable, and a test run is
 # not allowed to touch the real one.

tests/support.py

@@ -24,6 +24,7 @@ from unittest import mock
 
 from dikte import assistant
 from dikte import config as cfg
+from dikte import ggml
 from dikte import i18n
 from dikte import update
 
@@ -95,6 +96,10 @@ class DikteTest(unittest.TestCase):
         i18n.set_language("en")
         self.addCleanup(i18n.set_language, "en")
 
+        # Read once and kept for the life of the process, which across a test
+        # run means one test's machine answering for the next one's.
+        self.patch_attr(ggml, "_MEMORY", None)
+
         # cli.launch_gui replaces this process with the application when no
         # instance is running. A test that reaches it would take the whole run
         # with it and hang, so it fails loudly here instead.

tests/test_api.py

@@ -53,6 +53,16 @@ class TimestampModel(unittest.TestCase):
         self.assertEqual(api.timestamp_model("openai", "gpt-4o-transcribe"),
                          "whisper-1")
 
+    def test_openrouter_takes_the_file_model_that_was_set(self):
+        self.assertEqual(
+            api.timestamp_model("openrouter", "openai/gpt-4o-transcribe",
+                                "openai/whisper-large-v3"),
+            "openai/whisper-large-v3")
+
+    def test_openrouter_with_no_file_model_falls_back_to_whisper(self):
+        self.assertEqual(api.timestamp_model("openrouter", "openai/gpt-4o-transcribe", ""),
+                         "openai/whisper-1")
+
 
 class Explain(DikteTest):
     def error(self, status):
@@ -119,9 +129,22 @@ class ExtractError(unittest.TestCase):
         body = json.dumps({"error": {"code": 42}})
         self.assertIn("42", api._extract_error(body))
 
+    def test_an_error_wrapped_in_an_array(self):
+        """Google's 503 arrives this way, and .get() on a list raises."""
+        body = json.dumps([{"error": {"code": 503,
+                                      "message": "The model is overloaded."}}])
+        self.assertEqual(api._extract_error(body), "The model is overloaded.")
+
     def test_a_body_that_is_not_json(self):
         self.assertEqual(api._extract_error("<html>502</html>"), "<html>502</html>")
 
+    def test_no_shape_at_all_still_comes_back_as_a_string(self):
+        """It runs while an ApiError is being raised: throwing here would
+        escape the `except ApiError` holding the raw transcript."""
+        for body in ("[]", "[1, 2]", '"a string"', "null", "17"):
+            with self.subTest(body=body):
+                self.assertIsInstance(api._extract_error(body), str)
+
     def test_a_wall_of_html_is_cut_short(self):
         self.assertEqual(len(api._extract_error("x" * 5000)), 300)
 
@@ -305,6 +328,13 @@ class TranscribeSegments(DikteTest):
             api.transcribe_segments(OPENROUTER, self.wav)
         self.assertEqual(multipart_fields(calls[0])["model"], "openai/whisper-1")
 
+    def test_openrouter_asks_for_the_file_model_when_one_is_set(self):
+        target = OPENROUTER._replace(file_model="mistralai/voxtral-mini-transcribe")
+        with fake_urlopen(self.reply([{"start": 0, "end": 1, "text": "hi"}])) as calls:
+            api.transcribe_segments(target, self.wav)
+        self.assertEqual(multipart_fields(calls[0])["model"],
+                         "mistralai/voxtral-mini-transcribe")
+
     def test_groq_stays_on_the_model_it_was_given(self):
         target = GROQ._replace(model="whisper-large-v3")
         with fake_urlopen(self.reply([{"start": 0, "end": 1, "text": "hi"}])) as calls:
@@ -384,6 +414,37 @@ class Cleanup(DikteTest):
         self.assertEqual(sent_json(calls[0])["reasoning"],
                          {"effort": "high", "exclude": True})
 
+    def test_gemini_takes_openai_s_flat_field_rather_than_the_object(self):
+        _, calls = self.call(chat_reply("Hello."), reasoning="low",
+                             provider="gemini", service="Google AI Studio")
+        payload = sent_json(calls[0])
+        self.assertEqual(payload["reasoning_effort"], "low")
+        self.assertNotIn("reasoning", payload)
+
+    def test_off_is_asked_for_as_the_lowest_rung_google_actually_has(self):
+        """Sending "none" is a 400, and Flash left alone thinks."""
+        _, calls = self.call(chat_reply("Hello."), reasoning="none",
+                             provider="gemini", service="Google AI Studio")
+        self.assertEqual(sent_json(calls[0])["reasoning_effort"], "minimal")
+
+    def test_a_rung_google_does_not_have_lands_on_the_nearest_one(self):
+        for asked in ("xhigh", "max"):
+            with self.subTest(asked=asked):
+                _, calls = self.call(chat_reply("Hello."), reasoning=asked,
+                                     provider="gemini", service="Google AI Studio")
+                self.assertEqual(sent_json(calls[0])["reasoning_effort"], "high")
+
+    def test_gemini_left_on_the_model_s_own_default_is_told_nothing(self):
+        _, calls = self.call(chat_reply("Hello."), provider="gemini",
+                             service="Google AI Studio")
+        self.assertNotIn("reasoning_effort", sent_json(calls[0]))
+
+    def test_a_missing_gemini_key_says_google_ai_studio(self):
+        with self.assertRaises(api.ApiError) as caught:
+            api.cleanup("hello", "", "gemini-3.5-flash-lite", "prompt",
+                        provider="gemini", service="Google AI Studio")
+        self.assertIn("Google AI Studio", str(caught.exception))
+
     def test_a_local_base_url(self):
         _, calls = self.call(chat_reply("Hello."), base_url="http://localhost:1234/v1")
         self.assertEqual(calls[0].full_url, "http://localhost:1234/v1/chat/completions")
@@ -513,6 +574,40 @@ class ModelLists(DikteTest):
             api.openai_models("", api.GROQ_URL, "Groq")
         self.assertIn("Groq", str(caught.exception))
 
+    def test_gemini_keeps_only_the_models_that_answer_a_chat_request(self):
+        with fake_urlopen({"data": [{"id": "gemini-3.5-flash"},
+                                    {"id": "text-embedding-004"},
+                                    {"id": "imagen-4.0"},
+                                    {"id": "gemini-2.5-flash-lite"}]}) as calls:
+            models = api.gemini_models("AIza-test")
+        self.assertEqual(calls[0].full_url,
+                         "https://generativelanguage.googleapis.com/v1beta/openai/models")
+        self.assertEqual(models, ["gemini-2.5-flash-lite", "gemini-3.5-flash"])
+
+    def test_the_long_form_of_an_id_is_shortened_to_what_a_request_wants(self):
+        with fake_urlopen({"data": [{"id": "models/gemini-3.5-flash-lite"}]}):
+            self.assertEqual(api.gemini_models("AIza-test"),
+                             ["gemini-3.5-flash-lite"])
+
+    def test_a_gemini_id_that_is_not_a_chat_model_is_left_out(self):
+        """Google names its pictures and its voices `gemini` too."""
+        with fake_urlopen({"data": [{"id": "gemini-3.5-flash"},
+                                    {"id": "gemini-embedding-001"},
+                                    {"id": "gemini-2.5-flash-image"},
+                                    {"id": "gemini-2.5-flash-preview-tts"},
+                                    {"id": "gemini-2.5-native-audio"}]}):
+            self.assertEqual(api.gemini_models("AIza-test"), ["gemini-3.5-flash"])
+
+    def test_gemini_sends_the_key_as_a_bearer_token(self):
+        with fake_urlopen({"data": []}) as calls:
+            api.gemini_models("AIza-test")
+        self.assertEqual(calls[0].get_header("Authorization"), "Bearer AIza-test")
+
+    def test_a_missing_gemini_key_says_google_ai_studio(self):
+        with self.assertRaises(api.ApiError) as caught:
+            api.gemini_models("")
+        self.assertIn("Google AI Studio", str(caught.exception))
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_assistant.py

@@ -10,21 +10,28 @@ import io
 import json
 import os
 import subprocess
+import sys
+import threading
 import time
 import unittest
 from unittest import mock
 
 from dikte import assistant
-from tests.support import DikteTest, fake_urlopen, only_these_tools
+from tests.support import (DikteTest, FakeCompleted, fake_urlopen,
+                           only_these_tools)
 
 
 class FakeCli:
-    """A CLI that prints the given events and exits."""
+    """A CLI that prints the given events and exits.
 
-    def __init__(self, events=(), code=0, stderr="", noise=()):
+    Its stderr is not modelled: _stream hands the process a temporary file for
+    that, and a mocked Popen leaves the file empty, which is what a quiet CLI
+    writes anyway.
+    """
+
+    def __init__(self, events=(), code=0, noise=()):
         lines = list(noise) + [json.dumps(event) for event in events]
         self.stdout = io.StringIO("\n".join(lines) + "\n")
-        self.stderr = io.StringIO(stderr)
         self.returncode = code
         self.killed = False
 
@@ -41,6 +48,38 @@ class FakeCli:
         self.killed = True
 
 
+class WedgedCli:
+    """A CLI whose stdout produces nothing, the way a hung process's does.
+
+    Iterating its stdout blocks until the kill arrives, because that is what
+    reading a silent pipe does: only the process ending closes the stream.
+    """
+
+    def __init__(self):
+        self.pid = 4242
+        self.returncode = None
+        self.released = threading.Event()
+        self.stdout = self
+
+    def __iter__(self):
+        return self
+
+    def __next__(self):
+        # The 5 second cap is a safety net for the test itself; the kill is
+        # what is supposed to end the wait.
+        self.released.wait(timeout=5)
+        raise StopIteration
+
+    def poll(self):
+        return self.returncode
+
+    def wait(self, timeout=None):
+        return self.returncode
+
+    def close(self):
+        pass
+
+
 class Provider(DikteTest):
     def test_the_default(self):
         self.assertEqual(assistant.provider(self.config()), "claude")
@@ -58,15 +97,38 @@ class Provider(DikteTest):
     def test_what_each_one_runs(self):
         self.assertEqual(assistant.executable("claude"), "claude")
         self.assertEqual(assistant.executable("codex"), "codex")
+        self.assertEqual(assistant.executable("agy"), "agy")
         self.assertEqual(assistant.executable("openrouter"), "")
+        self.assertEqual(assistant.executable("opencode"), "")
+
+    def test_the_model_recorded_is_the_one_that_answered(self):
+        """The history used to write Claude's setting whoever had answered."""
+        self.assertEqual(assistant.model(self.config()), "sonnet")
+        self.assertEqual(
+            assistant.model(self.config(assistant_provider="codex")), "codex")
+        self.assertEqual(
+            assistant.model(self.config(assistant_provider="agy")), "agy")
+        self.assertEqual(
+            assistant.model(self.config(assistant_provider="agy",
+                                        assistant_agy_model="gemini-3.1-pro-low")),
+            "gemini-3.1-pro-low")
+        self.assertEqual(
+            assistant.model(self.config(assistant_provider="openrouter")),
+            "google/gemini-3.5-flash")
 
     def test_what_each_one_is_called(self):
         self.assertEqual(assistant.display_name(self.config()), "Claude")
-        self.assertEqual(
-            assistant.display_name(self.config(assistant_provider="codex")), "Codex")
-        self.assertEqual(
-            assistant.display_name(self.config(assistant_provider="openrouter")),
-            "OpenRouter")
+        for name, called in (("codex", "Codex"), ("agy", "Antigravity"),
+                             ("openrouter", "OpenRouter"),
+                             ("opencode", "OpenCode Go")):
+            with self.subTest(name=name):
+                self.assertEqual(
+                    assistant.display_name(self.config(assistant_provider=name)),
+                    called)
+
+    def test_every_provider_has_a_name_to_be_called_by(self):
+        """_conclude writes its errors in it, so a gap here is a bare id."""
+        self.assertEqual(set(assistant.SERVICES), set(assistant.PROVIDERS))
 
 
 class Effort(unittest.TestCase):
@@ -74,21 +136,26 @@ class Effort(unittest.TestCase):
 
     def test_the_scales_cover_the_same_settings(self):
         self.assertEqual(set(assistant.CLAUDE_EFFORT), set(assistant.CODEX_EFFORT))
+        self.assertEqual(set(assistant.CLAUDE_EFFORT), set(assistant.AGY_EFFORT))
 
-    def test_codex_has_no_rung_above_high(self):
-        self.assertEqual(assistant.CODEX_EFFORT["xhigh"], "high")
-        self.assertEqual(assistant.CODEX_EFFORT["max"], "high")
+    def test_neither_codex_nor_agy_has_a_rung_above_high(self):
+        for scale in (assistant.CODEX_EFFORT, assistant.AGY_EFFORT):
+            self.assertEqual(scale["xhigh"], "high")
+            self.assertEqual(scale["max"], "high")
 
-    def test_neither_one_asks_for_a_rung_below_low(self):
+    def test_none_of_them_asks_for_a_rung_below_low(self):
         # Claude has none; Codex has one, but calls it "minimal" on the older
-        # models and "none" on the newer ones, and refuses the wrong word.
-        for scale in (assistant.CLAUDE_EFFORT, assistant.CODEX_EFFORT):
+        # models and "none" on the newer ones, and refuses the wrong word; agy
+        # has three rungs and no word for off at all.
+        for scale in (assistant.CLAUDE_EFFORT, assistant.CODEX_EFFORT,
+                      assistant.AGY_EFFORT):
             self.assertEqual(scale["none"], "low")
             self.assertEqual(scale["minimal"], "low")
 
     def test_an_empty_setting_asks_for_nothing(self):
-        self.assertEqual(assistant.CLAUDE_EFFORT.get("", ""), "")
-        self.assertEqual(assistant.CODEX_EFFORT.get("", ""), "")
+        for scale in (assistant.CLAUDE_EFFORT, assistant.CODEX_EFFORT,
+                      assistant.AGY_EFFORT):
+            self.assertEqual(scale.get("", ""), "")
 
 
 class Session(DikteTest):
@@ -221,19 +288,7 @@ class Denials(DikteTest):
         self.assertIn("Write", warning)
 
 
-class SessionMissing(unittest.TestCase):
-    def test_a_session_that_is_gone(self):
-        for text in ("Error: session abc not found",
-                     "No conversation with that id",
-                     "unknown thread: abc"):
-            with self.subTest(text=text):
-                self.assertTrue(assistant._session_missing(text))
-
-    def test_an_unrelated_failure(self):
-        for text in ("", "network unreachable", "session limit exceeded"):
-            with self.subTest(text=text):
-                self.assertFalse(assistant._session_missing(text))
-
+class LastLine(unittest.TestCase):
     def test_the_last_line_is_the_one_worth_showing(self):
         self.assertEqual(assistant.last_line("warning\n\nreal error\n"),
                          "real error")
@@ -249,56 +304,129 @@ class Conclude(DikteTest):
 
     def test_an_answer_and_its_session(self):
         answer, warning = assistant._conclude(
-            self.found(answer="done", session="abc"), 0, "", "", "Claude")
+            self.found(answer="done", session="abc"), 0, "", "", "claude")
         self.assertEqual(answer, "done")
         self.assertEqual(warning, "")
         self.assertEqual(assistant.read_session("claude", 1800), "abc")
 
-    def test_codex_stores_under_its_own_name(self):
-        assistant._conclude(self.found(answer="done", session="t-1"), 0, "",
-                            "", "Codex")
-        self.assertEqual(assistant.read_session("codex", 1800), "t-1")
+    def test_each_one_stores_under_its_own_name(self):
+        for name, session in (("codex", "t-1"), ("agy", "c-9")):
+            with self.subTest(name=name):
+                assistant._conclude(self.found(answer="done", session=session),
+                                    0, "", "", name)
+                self.assertEqual(assistant.read_session(name, 1800), session)
+
+    def test_the_error_is_written_in_the_provider_s_own_name(self):
+        with self.assertRaises(assistant.AssistantError) as caught:
+            assistant._conclude(self.found(), 1, "", "", "agy")
+        self.assertIn("Antigravity", str(caught.exception))
 
     def test_a_non_zero_exit_with_nothing_to_show_for_it(self):
         with self.assertRaises(assistant.AssistantError) as caught:
-            assistant._conclude(self.found(), 1, "it all went wrong\n", "", "Claude")
+            assistant._conclude(self.found(), 1, "it all went wrong\n", "", "claude")
         self.assertIn("it all went wrong", str(caught.exception))
 
     def test_a_session_that_is_gone_is_raised_apart(self):
         with self.assertRaises(assistant._SessionGone):
             assistant._conclude(self.found(), 1, "session abc not found",
-                                "abc", "Claude")
+                                "abc", "claude")
+
+    def test_the_recovery_no_longer_hangs_on_the_words_the_cli_chose(self):
+        # The complaint used to be matched by substring, which a CLI update or
+        # another language broke. A resumed run that died with nothing to show
+        # is now enough on its own.
+        for stderr in ("Oturum bulunamadı", "something else entirely", ""):
+            with self.subTest(stderr=stderr):
+                with self.assertRaises(assistant._SessionGone):
+                    assistant._conclude(self.found(), 1, stderr, "abc", "Claude")
+
+    def test_api_trouble_on_a_resumed_run_is_not_blamed_on_the_session(self):
+        # A fresh session cannot cure a spent quota, a signed-out CLI or a dead
+        # network: the retry would fail the same way after a second wait, and
+        # the user would lose the conversation thread on top.
+        for stderr in ("Rate limit exceeded",
+                       "You are not logged in. Please run /login.",
+                       "API Error: 401 Unauthorized",
+                       "fetch failed: ECONNREFUSED 127.0.0.1"):
+            with self.subTest(stderr=stderr):
+                with self.assertRaises(assistant.AssistantError) as caught:
+                    assistant._conclude(self.found(), 1, stderr, "abc", "Claude")
+                self.assertIn(stderr, str(caught.exception))
 
     def test_a_session_that_is_gone_only_matters_when_one_was_resumed(self):
         with self.assertRaises(assistant.AssistantError):
             assistant._conclude(self.found(), 1, "session abc not found",
-                                "", "Claude")
+                                "", "claude")
 
     def test_an_answer_survives_a_non_zero_exit(self):
         answer, _ = assistant._conclude(self.found(answer="done"), 1, "noise",
-                                        "", "Claude")
+                                        "", "claude")
+        self.assertEqual(answer, "done")
+
+    def test_an_answer_on_a_resumed_session_is_kept_rather_than_retried(self):
+        answer, _ = assistant._conclude(self.found(answer="done"), 1, "noise",
+                                        "abc", "Claude")
         self.assertEqual(answer, "done")
 
     def test_a_reported_failure_with_no_answer(self):
         with self.assertRaises(assistant.AssistantError) as caught:
             assistant._conclude(self.found(failure="the model refused"), 0, "",
-                                "", "Claude")
+                                "", "claude")
         self.assertIn("refused", str(caught.exception))
 
     def test_a_run_that_said_nothing_at_all(self):
         with self.assertRaises(assistant.AssistantError) as caught:
-            assistant._conclude(self.found(), 0, "", "", "Codex")
+            assistant._conclude(self.found(), 0, "", "", "codex")
         self.assertIn("Codex", str(caught.exception))
 
 
+class Stream(DikteTest):
+    def test_a_cli_that_floods_stderr_still_finishes(self):
+        # A real subprocess, because the wedge being tested is real plumbing:
+        # with stderr on a pipe nobody drains, 200 KB fills the pipe's buffer,
+        # the child blocks writing it, and the run hangs until the watchdog
+        # timeout. With stderr on a file the run completes at once.
+        script = (
+            "import sys\n"
+            "sys.stderr.write('x' * 200000)\n"
+            "sys.stderr.flush()\n"
+            "print('{\"type\": \"result\", \"result\": \"done\"}')\n"
+        )
+        conf = self.config(assistant_timeout=15)
+        events = []
+        code, stderr = assistant._stream(
+            [sys.executable, "-c", script], conf, events.append, None)
+        self.assertEqual(code, 0)
+        self.assertEqual(len(stderr), 200000)
+        self.assertEqual(events[-1]["result"], "done")
+
+    def test_the_watchdog_takes_the_whole_tree_down_on_timeout(self):
+        proc = WedgedCli()
+
+        def killed(target):
+            # What the real kill does, as far as _stream can see: the process
+            # ends, and its closing stream releases the blocked read.
+            target.returncode = 1
+            target.released.set()
+
+        conf = self.config(assistant_timeout=0)
+        with mock.patch.object(subprocess, "Popen", return_value=proc), \
+                mock.patch.object(assistant, "kill_tree",
+                                  side_effect=killed) as kill:
+            with self.assertRaises(assistant.AssistantError) as caught:
+                assistant._stream(["claude"], conf, lambda event: None, None)
+        kill.assert_called_once_with(proc)
+        self.assertIn("did not finish", str(caught.exception))
+
+
 class AskClaude(DikteTest):
-    def run_ask(self, conf=None, events=None, code=0, stderr="", noise=(),
+    def run_ask(self, conf=None, events=None, code=0, noise=(),
                 session=""):
         conf = conf or self.config()
         proc = FakeCli(events or [
             {"type": "system", "subtype": "init", "session_id": "abc"},
             {"type": "result", "session_id": "abc", "result": "  done  "},
-        ], code=code, stderr=stderr, noise=noise)
+        ], code=code, noise=noise)
         stages = []
         with only_these_tools("claude", "codex"), \
                 mock.patch.object(subprocess, "Popen", return_value=proc) as popen:
@@ -472,6 +600,109 @@ class AskCodex(DikteTest):
         self.assertIn("quota", str(caught.exception))
 
 
+class AskAgy(DikteTest):
+    """agy's stream is shaped nothing like the other two: the key is `event`,
+    the answer arrives whole in `result.response`, and the conversation to
+    resume is named in the first line rather than the last."""
+
+    def run_ask(self, conf=None, events=None, session=""):
+        conf = conf or self.config(assistant_provider="agy")
+        proc = FakeCli(events or [
+            {"event": "init", "conversation_id": "c-9", "init": {"cwd": "/home"}},
+            {"event": "result",
+             "result": {"conversation_id": "c-9", "status": "SUCCESS",
+                        "response": "  done  "}},
+        ])
+        stages = []
+        with only_these_tools("agy"), \
+                mock.patch.object(subprocess, "Popen", return_value=proc) as popen:
+            result = assistant._ask_agy("book it", conf, session,
+                                        stages.append, None)
+        return result, popen.call_args.args[0], stages
+
+    def test_the_answer_comes_back_stripped(self):
+        (answer, warning), _, _ = self.run_ask()
+        self.assertEqual(answer, "done")
+        self.assertEqual(warning, "")
+
+    def test_the_instruction_is_kept_apart_from_the_command(self):
+        """agy takes no system prompt, so the two must not read as one."""
+        conf = self.config(assistant_provider="agy")
+        _, cmd, _ = self.run_ask(conf)
+        body = cmd[cmd.index("-p") + 1]
+        self.assertTrue(body.startswith(conf.assistant_prompt()))
+        self.assertIn("\n\n---\n\n", body)
+        self.assertTrue(body.endswith("book it"))
+
+    def test_a_first_command_starts_a_project_of_its_own(self):
+        """Without it agy works in whichever project it was last in."""
+        _, cmd, _ = self.run_ask()
+        self.assertIn("--new-project", cmd)
+        self.assertNotIn("--conversation", cmd)
+
+    def test_a_second_command_carries_the_conversation_rather_than_starting_one(self):
+        _, cmd, _ = self.run_ask(session="c-9")
+        self.assertEqual(cmd[cmd.index("--conversation") + 1], "c-9")
+        self.assertNotIn("--new-project", cmd)
+
+    def test_the_conversation_is_kept_under_agy_s_own_name(self):
+        self.run_ask()
+        self.assertEqual(assistant.read_session("agy", 1800), "c-9")
+
+    def test_it_is_not_left_to_give_up_before_the_caller_does(self):
+        conf = self.config(assistant_provider="agy", assistant_timeout=90)
+        _, cmd, _ = self.run_ask(conf)
+        self.assertEqual(cmd[cmd.index("--print-timeout") + 1], "90s")
+
+    def test_no_model_named_means_whatever_agy_is_set_to(self):
+        _, cmd, _ = self.run_ask()
+        self.assertNotIn("--model", cmd)
+
+    def test_a_model_of_your_own(self):
+        _, cmd, _ = self.run_ask(
+            self.config(assistant_provider="agy",
+                        assistant_agy_model="gemini-3.1-pro-low"))
+        self.assertEqual(cmd[cmd.index("--model") + 1], "gemini-3.1-pro-low")
+
+    def test_a_tool_is_named_in_the_corner_as_it_starts(self):
+        _, _, stages = self.run_ask(events=[
+            {"event": "step_update",
+             "step_update": {"step_type": "tool", "state": "ACTIVE",
+                             "tool_name": "run_command"}},
+            {"event": "step_update",
+             "step_update": {"step_type": "tool", "state": "DONE",
+                             "tool_name": "run_command"}},
+            {"event": "result",
+             "result": {"status": "SUCCESS", "response": "done"}},
+        ])
+        self.assertEqual(stages, ["Running a command…"])
+
+    def test_the_two_dozen_browser_tools_are_one_line_between_them(self):
+        _, _, stages = self.run_ask(events=[
+            {"event": "step_update",
+             "step_update": {"step_type": "tool", "state": "ACTIVE",
+                             "tool_name": "browser_click_element"}},
+            {"event": "result",
+             "result": {"status": "SUCCESS", "response": "done"}},
+        ])
+        self.assertEqual(stages, ["Working in the browser…"])
+
+    def test_a_turn_that_did_not_succeed_is_a_failure_rather_than_an_answer(self):
+        with self.assertRaises(assistant.AssistantError) as caught:
+            self.run_ask(events=[
+                {"event": "result",
+                 "result": {"status": "ERROR", "response": "the model refused"}},
+            ])
+        self.assertIn("refused", str(caught.exception))
+
+    def test_a_failure_with_nothing_to_say_is_still_named(self):
+        with self.assertRaises(assistant.AssistantError) as caught:
+            self.run_ask(events=[
+                {"event": "result", "result": {"status": "ERROR"}},
+            ])
+        self.assertIn("Antigravity", str(caught.exception))
+
+
 class AskOpenRouter(DikteTest):
     def test_a_question_and_an_answer(self):
         conf = self.config(assistant_provider="openrouter",
@@ -507,6 +738,41 @@ class AskOpenRouter(DikteTest):
             assistant.ask("when is it", conf)
 
 
+class AskOpenCode(DikteTest):
+    def test_a_question_and_an_answer(self):
+        conf = self.config(assistant_provider="opencode",
+                           opencode_api_key="opencode-test-key")
+        with fake_urlopen({"choices": [{"message": {"content": "on Thursday"}}]}):
+            answer, warning = assistant.ask("when is it", conf)
+        self.assertEqual(answer, "on Thursday")
+        self.assertEqual(warning, "")
+
+    def test_the_conversation_is_ours_to_keep(self):
+        conf = self.config(assistant_provider="opencode",
+                           opencode_api_key="opencode-test-key")
+        with fake_urlopen({"choices": [{"message": {"content": "on Thursday"}}]}):
+            assistant.ask("when is it", conf)
+        stored = assistant.read_messages("opencode", 1800)
+        self.assertEqual([row["content"] for row in stored],
+                         ["when is it", "on Thursday"])
+
+    def test_the_model_and_endpoint_are_opencode_s_own(self):
+        conf = self.config(assistant_provider="opencode",
+                           opencode_api_key="opencode-test-key",
+                           assistant_opencode_model="glm-5.3")
+        with fake_urlopen({"choices": [{"message": {"content": "on Thursday"}}]}) as calls:
+            assistant.ask("when is it", conf)
+        sent = json.loads(calls[0].data.decode("utf-8"))
+        self.assertEqual(sent["model"], "glm-5.3")
+        self.assertIn("https://opencode.ai/zen/go/v1/chat/completions",
+                      calls[0].full_url)
+
+    def test_an_api_failure_reads_as_an_assistant_failure(self):
+        conf = self.config(assistant_provider="opencode")
+        with self.assertRaises(assistant.AssistantError):
+            assistant.ask("when is it", conf)
+
+
 class Ask(DikteTest):
     def test_a_cli_that_is_not_installed_says_where_to_change_it(self):
         with only_these_tools(), \
@@ -533,6 +799,129 @@ class Ask(DikteTest):
         self.assertEqual(attempts, ["stale-id", ""])
         self.assertEqual(assistant.stored_provider(), "")
 
+    def test_a_resumed_run_that_dies_is_retried_without_the_session_flag(self):
+        # All the way through the stream this time: the first run exits 1 with
+        # no answer and whatever stderr it liked, and the recovery must not
+        # depend on those words.
+        conf = self.config()
+        assistant.write_session("claude", "stale-id")
+        procs = iter([
+            FakeCli(code=1),
+            FakeCli(events=[{"type": "result", "result": "done"}]),
+        ])
+        cmds = []
+
+        def popen(cmd, **kwargs):
+            cmds.append(cmd)
+            return next(procs)
+
+        with only_these_tools("claude"), \
+                mock.patch.object(subprocess, "Popen", side_effect=popen):
+            answer, _ = assistant.ask("hi", conf)
+        self.assertEqual(answer, "done")
+        self.assertEqual(cmds[0][cmds[0].index("--resume") + 1], "stale-id")
+        self.assertNotIn("--resume", cmds[1])
+
+    def test_a_run_that_dies_with_an_answer_in_hand_is_not_retried(self):
+        conf = self.config()
+        assistant.write_session("claude", "stale-id")
+        calls = []
+
+        def popen(cmd, **kwargs):
+            calls.append(cmd)
+            return FakeCli(events=[{"type": "result", "result": "done"}], code=1)
+
+        with only_these_tools("claude"), \
+                mock.patch.object(subprocess, "Popen", side_effect=popen):
+            answer, _ = assistant.ask("hi", conf)
+        self.assertEqual(answer, "done")
+        self.assertEqual(len(calls), 1)
+
+
+class CodexModels(DikteTest):
+    """The model list read off `codex debug models`."""
+
+    CATALOG = {"models": [
+        {"slug": "gpt-6-mini", "visibility": "list", "priority": 9},
+        {"slug": "gpt-6", "visibility": "list", "priority": 1},
+        {"slug": "codex-auto-review", "visibility": "hide", "priority": 3},
+    ]}
+
+    def models(self, reply, code=0):
+        with only_these_tools("codex"), \
+                mock.patch.object(subprocess, "run",
+                                  return_value=FakeCompleted(
+                                      returncode=code, stdout=reply)) as run:
+            found = assistant.codex_models()
+        self.run_call = run
+        return found
+
+    def test_the_catalog_arrives_best_first_without_the_hidden_ones(self):
+        found = self.models(json.dumps(self.CATALOG))
+        self.assertEqual(found, ["gpt-6", "gpt-6-mini"])
+        self.assertEqual(self.run_call.call_args.args[0],
+                         ["codex", "debug", "models"])
+
+    def test_a_codex_that_is_not_installed_is_not_run(self):
+        with only_these_tools(), \
+                mock.patch.object(subprocess, "run") as run:
+            self.assertEqual(assistant.codex_models(), [])
+        run.assert_not_called()
+
+    def test_a_codex_too_old_to_have_the_command(self):
+        self.assertEqual(self.models("error: unknown subcommand", code=2), [])
+
+    def test_a_catalog_that_is_not_what_was_expected(self):
+        self.assertEqual(self.models(json.dumps(["gpt-6"])), [])
+        self.assertEqual(self.models(""), [])
+
+    def test_a_codex_that_hangs_is_given_up_on(self):
+        with only_these_tools("codex"), \
+                mock.patch.object(subprocess, "run",
+                                  side_effect=subprocess.TimeoutExpired(
+                                      ["codex"], 30)):
+            self.assertEqual(assistant.codex_models(), [])
+
+
+class AgyModels(DikteTest):
+    """The model list read off `agy models`: one id, a tab, a display name."""
+
+    LISTING = ("gemini-4-flash-high\tGemini 4 Flash (High)\n"
+               "gemini-4-flash-low\tGemini 4 Flash (Low)\n"
+               "a line with no tab is not a model\n"
+               "\ta tab with no id in front of it is not one either\n")
+
+    def models(self, reply, code=0):
+        with only_these_tools("agy"), \
+                mock.patch.object(subprocess, "run",
+                                  return_value=FakeCompleted(
+                                      returncode=code, stdout=reply)) as run:
+            found = assistant.agy_models()
+        self.run_call = run
+        return found
+
+    def test_the_listing_arrives_in_agy_s_own_order(self):
+        found = self.models(self.LISTING)
+        self.assertEqual(found, ["gemini-4-flash-high", "gemini-4-flash-low"])
+        self.assertEqual(self.run_call.call_args.args[0], ["agy", "models"])
+
+    def test_an_agy_that_is_not_installed_is_not_run(self):
+        with only_these_tools(), \
+                mock.patch.object(subprocess, "run") as run:
+            self.assertEqual(assistant.agy_models(), [])
+        run.assert_not_called()
+
+    def test_a_call_that_failed_answers_with_nothing(self):
+        self.assertEqual(self.models("error: not logged in", code=1), [])
+        self.assertEqual(self.models(""), [])
+
+    def test_an_agy_that_hangs_is_given_up_on(self):
+        with only_these_tools("agy"), \
+                mock.patch.object(subprocess, "run",
+                                  side_effect=subprocess.TimeoutExpired(
+                                      ["agy"], 30)):
+            self.assertEqual(assistant.agy_models(), [])
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_audio.py

@@ -81,6 +81,14 @@ class ChunkLevels(unittest.TestCase):
         self.assertEqual(peak, 1.0)
         self.assertEqual(rms, 1.0)
 
+    def test_the_fast_and_plain_rms_paths_agree(self):
+        """sumprod is a speedup, not a different sum: on a 3.11 machine the
+        loop must land on the same integers."""
+        chunk = tone(0.1)
+        with mock.patch.object(audio, "sumprod", None):
+            plain = audio.chunk_levels(chunk)
+        self.assertEqual(audio.chunk_levels(chunk), plain)
+
 
 class StereoLevels(unittest.TestCase):
     def test_the_channels_are_read_apart(self):
@@ -280,6 +288,48 @@ class _StalledStream:
         self._released.set()
 
 
+class _DribblingStream:
+    """A pipe that never fills a whole chunk in one read, the way an unbuffered
+    pipe hands data over under load."""
+
+    def __init__(self, data, piece):
+        self._data = io.BytesIO(data)
+        self._piece = piece
+
+    def read(self, size):
+        return self._data.read(min(size, self._piece))
+
+
+class _RunSwappingStream:
+    """A pipe whose recorder moves on mid-read, the way a new recording starts
+    while a stale pump is still draining the old one."""
+
+    def __init__(self, data, recorder, swap_at, new_proc):
+        self._data = io.BytesIO(data)
+        self._recorder = recorder
+        self._swap_at = swap_at
+        self._new_proc = new_proc
+        self.reads = 0
+
+    def read(self, size):
+        if self.reads == self._swap_at:
+            self._recorder._run = object()
+            self._recorder._proc = self._new_proc
+        self.reads += 1
+        return self._data.read(size)
+
+
+class _SignalCrashingProcess(FakeProcess):
+    """An ffmpeg that calls being interrupted a failure, the way ffmpeg does."""
+
+    def __init__(self, data, code=255):
+        super().__init__(data)
+        self._code = code
+
+    def poll(self):
+        return None if self._alive else self._code
+
+
 class _HeldStream:
     """A capture that is paused and taken up again partway through, the way a
     key press lands in the middle of a recording rather than between two."""
@@ -307,7 +357,10 @@ class RecordingCommand(OnLinux, DikteTest):
         super().setUp()
         # Whether pw-record takes --raw is read off the installed binary, and
         # what is being tested here is the command rather than the machine the
-        # test is running on. PwRecordRawOption covers the reading itself.
+        # test is running on. PwRecordRawOption covers the reading itself. The
+        # answer is remembered between calls, so it cannot be remembered
+        # between tests.
+        self.enterContext(mock.patch.object(audio, "_PW_RAW", None))
         self.enterContext(mock.patch.object(
             audio, "_pw_record_raw_option", return_value=["--raw"]))
 
@@ -392,11 +445,30 @@ class PwRecordRawOption(DikteTest):
         self.assertEqual(["--raw"], self.option(return_value=FakeCompleted()))
 
 
+class PwRawMemo(OnLinux, DikteTest):
+    """The --raw probe runs once per process, not once per key press."""
+
+    def setUp(self):
+        super().setUp()
+        self.enterContext(mock.patch.object(audio, "_PW_RAW", None))
+
+    def test_the_probe_is_asked_once_and_remembered(self):
+        with only_these_tools("pw-record"), \
+                mock.patch.object(audio, "_pw_record_raw_option",
+                                  return_value=["--raw"]) as probe:
+            first = audio.recording_command()
+            second = audio.recording_command()
+        probe.assert_called_once_with()
+        self.assertIn("--raw", first)
+        self.assertEqual(first, second)
+
+
 class RecorderChain(OnLinux, DikteTest):
     """Start to WAV, with pw-record faked out."""
 
     def setUp(self):
         super().setUp()
+        self.enterContext(mock.patch.object(audio, "_PW_RAW", None))
         self.enterContext(mock.patch.object(
             audio, "_pw_record_raw_option", return_value=["--raw"]))
 
@@ -476,42 +548,121 @@ class RecorderChain(OnLinux, DikteTest):
         self.assertEqual(len(failures), 1)
         self.assertIn("pulseaudio-utils", failures[0])
 
-    def pump(self, data=b"", stderr=b"", stopping=False, cancelled=False):
+    def pump(self, data=b"", stderr=b"", stopping=False, cancelled=False,
+             alive=False):
         """Run the pump in this thread, where a queued signal would need an
         event loop nobody is running here."""
         recorder = audio.Recorder()
         failures = []
+        deaths = []
         recorder.failed.connect(failures.append)
+        recorder.died.connect(lambda: deaths.append(True))
         proc = FakeProcess(data)
-        proc.stderr = io.BytesIO(stderr)
-        proc._alive = False
+        proc._alive = alive
         recorder._proc = proc
-        recorder._max_bytes = 10 ** 9
+        recorder._log = io.BytesIO(stderr)
         recorder._stopping = stopping
         recorder._cancelled = cancelled
-        recorder._pump()
-        return failures
+        recorder._run = run = object()
+        recorder._pump(run, proc, proc.stdout, recorder._buffer,
+                       recorder._rms, 10 ** 9)
+        return failures, deaths
 
     def test_a_recorder_that_died_on_its_own_says_so(self):
         """parec refused the device, or the sound server went away."""
-        failures = self.pump(stderr=b"connection refused\n")
+        failures, _ = self.pump(stderr=b"connection refused\n")
         self.assertEqual(len(failures), 1)
         self.assertIn("connection refused", failures[0])
 
     def test_a_death_with_nothing_on_stderr_still_names_the_exit_code(self):
-        failures = self.pump()
+        failures, _ = self.pump()
         self.assertIn("exit code", failures[0])
 
+    def test_a_death_that_left_no_exit_code_is_not_named_none(self):
+        """A process nobody managed to reap has no code to show, and "exit
+        code None" would only puzzle the person reading it."""
+        failures, _ = self.pump(alive=True)
+        self.assertEqual(len(failures), 1)
+        self.assertNotIn("None", failures[0])
+
     def test_a_recording_we_ended_ourselves_is_not_a_death(self):
         """Otherwise a stray keypress produces two errors, and the first one
         sends the user looking for a broken sound server."""
-        self.assertEqual(self.pump(stopping=True), [])
+        self.assertEqual(self.pump(stopping=True), ([], []))
 
     def test_a_cancelled_recording_is_not_a_death(self):
-        self.assertEqual(self.pump(cancelled=True), [])
+        self.assertEqual(self.pump(cancelled=True), ([], []))
 
-    def test_a_recorder_that_captured_something_first_is_not_a_death(self):
-        self.assertEqual(self.pump(data=silence(0.5)), [])
+    def test_a_capture_that_ends_mid_recording_dies_rather_than_fails(self):
+        """Sound had already arrived, so this is not a broken installation:
+        the app is told the recording died and can rescue what there is."""
+        failures, deaths = self.pump(data=silence(0.5))
+        self.assertEqual(failures, [])
+        self.assertEqual(deaths, [True])
+
+    def test_short_pipe_reads_are_gathered_into_whole_chunks(self):
+        """Every RMS entry must stand for one full chunk, or the silence check
+        weighs a half-filled read as its own stretch of room tone."""
+        half = audio.CHUNK_BYTES // 2
+        data = pcm([1000] * (3 * half // 2))   # three half-chunk reads
+        recorder = audio.Recorder()
+        proc = FakeProcess(b"")
+        proc.stdout = _DribblingStream(data, half)
+        proc._alive = False
+        recorder._proc = proc
+        recorder._log = io.BytesIO(b"")
+        recorder._run = run = object()
+        buffer, rms = bytearray(), []
+        recorder._pump(run, proc, proc.stdout, buffer, rms, 10 ** 9)
+        self.assertEqual(len(buffer), len(data))
+        self.assertEqual(len(rms), 2)   # one whole chunk, then the tail
+
+    def test_a_stale_pump_cannot_touch_the_recording_that_replaced_it(self):
+        """A pump that outlives its join must not meter the next run, stop its
+        process, push audio into its buffer, or speak on its behalf."""
+        recorder = audio.Recorder()
+        levels, failures, deaths = [], [], []
+        recorder.level.connect(levels.append)
+        recorder.failed.connect(failures.append)
+        recorder.died.connect(lambda: deaths.append(True))
+        new_proc = FakeProcess(b"")
+        old_proc = FakeProcess(b"")
+        old_proc.stdout = _RunSwappingStream(tone(0.192), recorder,
+                                             swap_at=1, new_proc=new_proc)
+        recorder._proc = old_proc
+        recorder._log = io.BytesIO(b"")
+        old_run = object()
+        recorder._run = old_run
+        recorder._buffer = bytearray()   # the next recording's buffer
+        old_buffer, old_rms = bytearray(), []
+        recorder._pump(old_run, old_proc, old_proc.stdout, old_buffer, old_rms,
+                       2 * audio.CHUNK_BYTES)
+        # Metered once, then the new run took over: the over-length cutoff hit
+        # on the next chunk and had to stand down instead of stopping a
+        # process that was never its own.
+        self.assertEqual(len(levels), 1)
+        self.assertEqual(len(old_buffer), 2 * audio.CHUNK_BYTES)
+        self.assertEqual(new_proc.signals, [])
+        self.assertEqual(old_proc.signals, [])
+        self.assertEqual(recorder._buffer, bytearray())
+        self.assertEqual((failures, deaths), ([], []))
+
+    def test_a_wav_that_cannot_be_written_is_reported_not_raised(self):
+        recorder = audio.Recorder()
+        results, failures = [], []
+        recorder.stopped.connect(lambda *args: results.append(args))
+        recorder.failed.connect(failures.append)
+        proc = FakeProcess(tone(1.0))
+        with only_these_tools("pw-record"), \
+                mock.patch.object(subprocess, "Popen", return_value=proc), \
+                mock.patch.object(audio, "write_wav",
+                                  side_effect=OSError("disk full")):
+            recorder.start()
+            recorder._thread.join(timeout=5)
+            recorder.stop()
+        self.assertEqual(results, [])
+        self.assertEqual(len(failures), 1)
+        self.assertIn("disk full", failures[0])
 
     def test_a_short_recording_reports_only_that(self):
         _, results, failures, _ = self.record(silence(0.1))
@@ -722,6 +873,42 @@ class MacMeetingRecorder(OnMacOS, DikteTest):
         _, _, _, _, processes, _ = self.record(tone(0.5), tone(0.5))
         self.assertTrue(all(process.signals for process in processes))
 
+    def test_a_stop_we_asked_for_is_not_reported_as_an_ffmpeg_failure(self):
+        """ffmpeg exits 255 when interrupted, and the interruption was our own
+        stop: a meeting ended at once must say "too short", not "ffmpeg → 255"."""
+        path = str(self.path("meeting.wav"))
+        recorder = audio.MeetingRecorder()
+        failed = []
+        recorder.failed.connect(failed.append)
+        processes = [_SignalCrashingProcess(tone(0.1)),
+                     _SignalCrashingProcess(tone(0.1))]
+        with only_these_tools("ffmpeg"), self.devices(), \
+                mock.patch.object(subprocess, "Popen", side_effect=processes):
+            recorder.start(path, "MacBook Pro Microphone", "BlackHole 2ch")
+            recorder._thread.join(timeout=5)
+            recorder.stop()
+        self.assertEqual(len(failed), 1)
+        self.assertIn("0.3", failed[0])
+        self.assertNotIn("255", failed[0])
+
+    def test_an_ffmpeg_that_died_on_its_own_keeps_its_exit_code(self):
+        """A process nobody interrupted has a story to tell, and its code is
+        the only lead the user gets."""
+        path = str(self.path("meeting.wav"))
+        recorder = audio.MeetingRecorder()
+        failed = []
+        recorder.failed.connect(failed.append)
+        dead = _SignalCrashingProcess(tone(0.1))
+        dead._alive = False   # it fell over before stop() reached it
+        processes = [dead, FakeProcess(tone(0.1))]
+        with only_these_tools("ffmpeg"), self.devices(), \
+                mock.patch.object(subprocess, "Popen", side_effect=processes):
+            recorder.start(path, "MacBook Pro Microphone", "BlackHole 2ch")
+            recorder._thread.join(timeout=5)
+            recorder.stop()
+        self.assertEqual(len(failed), 1)
+        self.assertIn("255", failed[0])
+
     def test_a_legacy_numeric_target_fails_before_recording(self):
         recorder = audio.MeetingRecorder()
         failed = []
@@ -953,9 +1140,11 @@ class WindowsDevices(OnWindows, DikteTest):
     def setUp(self):
         super().setUp()
         # The listing is remembered between calls, so that a dictation does not
-        # run ffmpeg of its own. It cannot be remembered between tests.
+        # run ffmpeg of its own. It cannot be remembered between tests, and
+        # neither can the pw-record probe's answer.
         audio._DSHOW_SEEN.clear()
         self.addCleanup(audio._DSHOW_SEEN.clear)
+        self.enterContext(mock.patch.object(audio, "_PW_RAW", None))
 
     @contextlib.contextmanager
     def listing(self, stderr=None, tools=("ffmpeg",)):

tests/test_cleanup.py

@@ -1,9 +1,9 @@
-"""Who cleans the transcript up, and what they are asked.
+"""Who cleans the transcript up, and what they are asked.
 
-The CLIs are faked at subprocess.run: what the tests read is the argument list
-each one is given, where the answer is picked up from, and what happens to the
-chain when the program is missing, slow or unhappy. The OpenRouter path is the
-one that was always there and is checked here only for still being taken.
+The CLIs are faked at subprocess.Popen: what the tests read is the argument
+list each one is given, where the answer is picked up from, and what happens to
+the chain when the program is missing, slow or unhappy. The OpenRouter path is
+the one that was always there and is checked here only for still being taken.
 """
 
 import os
@@ -18,18 +18,26 @@ from tests.support import DikteTest, fake_urlopen, sent_json, url_error
 from tests.test_api import FakeServer, chat_reply
 
 
-def fake_run(stdout="", code=0, stderr="", last_message=""):
-    """Stand in for subprocess.run, writing the file Codex would have written."""
+def fake_cli(stdout="", code=0, stderr="", last_message=""):
+    """Stand in for subprocess.Popen.
+
+    _output hands the process a temporary file for each stream, so the fake
+    writes into those, plus the file Codex would have written on its way out.
+    """
     calls = []
 
-    def run(cmd, **kwargs):
+    def popen(cmd, **kwargs):
         calls.append(cmd)
+        kwargs["stdout"].write(stdout.encode("utf-8"))
+        kwargs["stderr"].write(stderr.encode("utf-8"))
         if last_message and "-o" in cmd:
             with open(cmd[cmd.index("-o") + 1], "w", encoding="utf-8") as fh:
                 fh.write(last_message)
-        return subprocess.CompletedProcess(cmd, code, stdout, stderr)
+        proc = mock.Mock()
+        proc.returncode = code
+        return proc
 
-    return mock.patch.object(subprocess, "run", side_effect=run), calls
+    return mock.patch.object(subprocess, "Popen", side_effect=popen), calls
 
 
 class Provider(DikteTest):
@@ -49,7 +57,10 @@ class Provider(DikteTest):
     def test_what_each_one_runs(self):
         self.assertEqual(cleanup.executable("claude"), "claude")
         self.assertEqual(cleanup.executable("codex"), "codex")
+        self.assertEqual(cleanup.executable("agy"), "agy")
         self.assertEqual(cleanup.executable("openrouter"), "")
+        self.assertEqual(cleanup.executable("gemini"), "")
+        self.assertEqual(cleanup.executable("opencode"), "")
 
     def test_the_model_named_in_the_history_is_the_one_that_did_it(self):
         self.assertEqual(cleanup.model(self.config(cleanup_model="some/model")),
@@ -65,6 +76,19 @@ class Provider(DikteTest):
         self.assertEqual(
             cleanup.model(self.config(cleanup_provider="codex",
                                       cleanup_codex_model="gpt-5.4")), "gpt-5.4")
+        self.assertEqual(
+            cleanup.model(self.config(cleanup_provider="gemini")),
+            "gemini-3.5-flash-lite")
+        # Antigravity is left on its own default the way Codex is.
+        self.assertEqual(
+            cleanup.model(self.config(cleanup_provider="agy")), "agy")
+        self.assertEqual(
+            cleanup.model(self.config(cleanup_provider="agy",
+                                      cleanup_agy_model="gemini-3.7-flash-low")),
+            "gemini-3.7-flash-low")
+        self.assertEqual(
+            cleanup.model(self.config(cleanup_provider="opencode",
+                                      cleanup_opencode_model="glm-5.3")), "glm-5.3")
 
 
 class OpenRouter(DikteTest):
@@ -80,12 +104,78 @@ class OpenRouter(DikteTest):
 
     def test_no_cli_is_started_for_it(self):
         conf = self.config(openrouter_api_key="sk-or-test")
-        patcher, calls = fake_run(stdout="never")
+        patcher, calls = fake_cli(stdout="never")
         with patcher, mock.patch.object(api, "cleanup", return_value="Done."):
             cleanup.run("uh, done", conf, "the rules")
         self.assertEqual(calls, [])
 
 
+class OpenCode(DikteTest):
+    def test_it_is_one_request_with_the_settings_as_they_were(self):
+        conf = self.config(cleanup_provider="opencode",
+                           opencode_api_key="opencode-test-key",
+                           cleanup_opencode_model="some/model",
+                           cleanup_reasoning="low")
+        with mock.patch.object(api, "cleanup", return_value="Done.") as call:
+            self.assertEqual(cleanup.run("uh, done", conf, "the rules"), "Done.")
+        text, key, model, prompt = call.call_args.args
+        self.assertEqual((text, key, model, prompt),
+                         ("uh, done", "opencode-test-key", "some/model", "the rules"))
+        self.assertEqual(call.call_args.kwargs["reasoning"], "low")
+        self.assertEqual(call.call_args.kwargs["provider"], "opencode")
+        self.assertEqual(call.call_args.kwargs["service"], "OpenCode Go")
+        self.assertEqual(call.call_args.kwargs["base_url"],
+                         "https://opencode.ai/zen/go/v1")
+
+    def test_no_cli_is_started_for_it(self):
+        conf = self.config(cleanup_provider="opencode",
+                           opencode_api_key="opencode-test-key")
+        patcher, calls = fake_cli(stdout="never")
+        with patcher, mock.patch.object(api, "cleanup", return_value="Done."):
+            cleanup.run("uh, done", conf, "the rules")
+        self.assertEqual(calls, [])
+
+
+class GoogleAiStudio(DikteTest):
+    """Cleanup over Google's OpenAI-compatible endpoint: one request, no CLI."""
+
+    def setUp(self):
+        super().setUp()
+        self.conf = self.config(cleanup_provider="gemini",
+                                gemini_api_key="AIza-test")
+
+    def test_it_goes_to_google_with_the_settings_as_they_were(self):
+        self.conf["cleanup_reasoning"] = "none"
+        with fake_urlopen(chat_reply("Done.")) as calls:
+            self.assertEqual(cleanup.run("uh, done", self.conf, "the rules"),
+                             "Done.")
+        self.assertEqual(
+            calls[0].full_url,
+            "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions")
+        payload = sent_json(calls[0])
+        self.assertEqual(payload["model"], "gemini-3.5-flash-lite")
+        self.assertEqual(payload["reasoning_effort"], "minimal")
+        self.assertIn("uh, done", payload["messages"][1]["content"])
+
+    def test_the_key_travels_as_a_bearer_token(self):
+        with fake_urlopen(chat_reply("Done.")) as calls:
+            cleanup.run("uh, done", self.conf, "the rules")
+        self.assertEqual(calls[0].get_header("Authorization"), "Bearer AIza-test")
+
+    def test_a_missing_key_names_google_rather_than_openrouter(self):
+        self.conf["gemini_api_key"] = ""
+        with mock.patch.dict(os.environ, {}, clear=True), \
+                self.assertRaises(api.ApiError) as caught:
+            cleanup.run("uh, done", self.conf, "the rules")
+        self.assertIn("Google AI Studio", str(caught.exception))
+
+    def test_no_cli_is_started_for_it(self):
+        patcher, calls = fake_cli(stdout="never")
+        with patcher, fake_urlopen(chat_reply("Done.")):
+            cleanup.run("uh, done", self.conf, "the rules")
+        self.assertEqual(calls, [])
+
+
 class ClaudeCode(DikteTest):
     def setUp(self):
         super().setUp()
@@ -93,7 +183,7 @@ class ClaudeCode(DikteTest):
         self.patch_attr(cleanup.shutil, "which", lambda name: f"/usr/bin/{name}")
 
     def run_cleanup(self, text="uh, book it", **kwargs):
-        patcher, calls = fake_run(**kwargs)
+        patcher, calls = fake_cli(**kwargs)
         with patcher:
             answer = cleanup.run(text, self.conf, "the rules")
         return answer, calls[0]
@@ -146,14 +236,18 @@ class ClaudeCode(DikteTest):
             self.run_cleanup(stdout="Book it.")
         self.assertIn("claude", str(caught.exception))
 
-    def test_a_run_that_never_ends(self):
-        def run(cmd, **kwargs):
-            raise subprocess.TimeoutExpired(cmd, 180)
+    def test_a_run_that_never_ends_is_killed_with_its_whole_tree(self):
+        def popen(cmd, **kwargs):
+            proc = mock.Mock()
+            proc.wait.side_effect = subprocess.TimeoutExpired(cmd, 180)
+            return proc
 
-        with mock.patch.object(subprocess, "run", side_effect=run):
+        with mock.patch.object(subprocess, "Popen", side_effect=popen), \
+                mock.patch.object(cleanup.assistant, "kill_tree") as kill:
             with self.assertRaises(cleanup.CleanupError) as caught:
                 cleanup.run("uh, book it", self.conf, "the rules")
         self.assertIn("180", str(caught.exception))
+        kill.assert_called_once()
 
 
 class Codex(DikteTest):
@@ -163,7 +257,7 @@ class Codex(DikteTest):
         self.patch_attr(cleanup.shutil, "which", lambda name: f"/usr/bin/{name}")
 
     def run_cleanup(self, text="uh, book it", **kwargs):
-        patcher, calls = fake_run(**kwargs)
+        patcher, calls = fake_cli(**kwargs)
         with patcher:
             answer = cleanup.run(text, self.conf, "the rules")
         return answer, calls[0]
@@ -209,6 +303,63 @@ class Codex(DikteTest):
             self.run_cleanup(stdout="tokens used 400", last_message="")
 
 
+class Antigravity(DikteTest):
+    def setUp(self):
+        super().setUp()
+        self.conf = self.config(cleanup_provider="agy")
+        self.patch_attr(cleanup.shutil, "which", lambda name: f"/usr/bin/{name}")
+
+    def run_cleanup(self, text="uh, book it", **kwargs):
+        patcher, calls = fake_cli(**kwargs)
+        with patcher:
+            answer = cleanup.run(text, self.conf, "the rules")
+        return answer, calls[0]
+
+    def test_the_rules_ride_in_front_of_the_transcript(self):
+        answer, cmd = self.run_cleanup(stdout="Book it.\n")
+        self.assertEqual(answer, "Book it.")
+        self.assertEqual(cmd[0], "agy")
+        self.assertEqual(cmd[cmd.index("-p") + 1],
+                         "the rules\n\n---\n\n<transcript>\nuh, book it\n</transcript>")
+
+    def test_it_starts_somewhere_of_its_own_and_takes_no_slash_commands(self):
+        """Without --new-project agy works in whichever project it was last in."""
+        _, cmd = self.run_cleanup(stdout="Book it.")
+        self.assertIn("--new-project", cmd)
+        self.assertIn("--disable-slash-commands", cmd)
+        self.assertEqual(cmd[cmd.index("--output-format") + 1], "text")
+
+    def test_it_is_not_left_to_give_up_before_the_caller_does(self):
+        _, cmd = self.run_cleanup(stdout="Book it.")
+        self.assertEqual(cmd[cmd.index("--print-timeout") + 1], "180s")
+
+    def test_the_model_is_left_alone_until_one_is_typed_in(self):
+        _, cmd = self.run_cleanup(stdout="Book it.")
+        self.assertNotIn("--model", cmd)
+        self.conf["cleanup_agy_model"] = "gemini-3.7-flash-low"
+        _, cmd = self.run_cleanup(stdout="Book it.")
+        self.assertEqual(cmd[cmd.index("--model") + 1], "gemini-3.7-flash-low")
+
+    def test_the_thinking_setting_lands_on_the_nearest_rung_agy_has(self):
+        self.conf["cleanup_reasoning"] = "max"
+        _, cmd = self.run_cleanup(stdout="Book it.")
+        self.assertEqual(cmd[cmd.index("--effort") + 1], "high")
+
+    def test_no_thinking_setting_means_no_flag(self):
+        _, cmd = self.run_cleanup(stdout="Book it.")
+        self.assertNotIn("--effort", cmd)
+
+    def test_an_answer_of_nothing_is_a_failure_rather_than_an_empty_paste(self):
+        with self.assertRaises(cleanup.CleanupError):
+            self.run_cleanup(stdout="   ")
+
+    def test_a_program_that_is_not_installed_says_so_before_running_anything(self):
+        self.patch_attr(cleanup.shutil, "which", lambda name: "")
+        with self.assertRaises(cleanup.CleanupError) as caught:
+            self.run_cleanup(stdout="Book it.")
+        self.assertIn("agy", str(caught.exception))
+
+
 if __name__ == "__main__":
     unittest.main()
 
@@ -280,7 +431,8 @@ class Here(DikteTest):
         self.assertIn("out of memory", str(caught.exception))
 
     def test_no_cli_is_started_for_it(self):
-        patcher, calls = fake_run(stdout="never")
+        patcher, calls = fake_cli(stdout="never")
         with patcher, fake_urlopen(chat_reply("Done.")):
             cleanup.run("uh, done", self.conf, "the rules")
         self.assertEqual(calls, [])
+

tests/test_cli.py

@@ -9,12 +9,14 @@ socket is faked, and everything that runs locally runs for real.
 import contextlib
 import io
 import json
+import sys
 import unittest
 import webbrowser
 from typing import ClassVar
 from unittest import mock
 
 from dikte import audio
+from dikte import cleanup
 from dikte import cli
 from dikte import config as cfg
 from dikte import ggml
@@ -423,6 +425,16 @@ class Providers(DikteTest):
         self.assertIn("groq", out)
         self.assertIn("Groq", out)
 
+    def test_opencode_is_a_choice_and_reports_under_its_own_name(self):
+        parser = cli.build_parser()
+        self.assertEqual(
+            parser.parse_args(["test-key", "opencode"]).which, "opencode")
+        self.write_config({"opencode_api_key": "opencode-test"})
+        with fake_urlopen({"data": [{"id": "deepseek-v4-flash"}]}):
+            code, out, _ = self.run_cmd(cli.cmd_test_key, which="opencode")
+        self.assertEqual(code, 0)
+        self.assertIn("opencode: connection works, 1 models visible", out)
+
 
 class Updates(DikteTest):
     """`dikte update` looks, says what it found, and installs nothing."""
@@ -503,6 +515,35 @@ class Doctor(DikteTest):
         self.assertIn("OpenRouter key, cleaning up on some/model",
                       self.run_doctor(as_json=False, cleanup_model="some/model"))
 
+    def test_cleanup_on_opencode_is_a_question_about_its_own_key(self):
+        reply = self.run_doctor(cleanup_provider="opencode",
+                                cleanup_opencode_model="glm-5.3")
+        self.assertEqual(reply["cleanup"]["provider"], "opencode")
+        self.assertEqual(reply["cleanup"]["model"], "glm-5.3")
+        self.assertIn("OpenCode Go key, cleaning up on glm-5.3",
+                      self.run_doctor(as_json=False, cleanup_provider="opencode",
+                                      cleanup_opencode_model="glm-5.3"))
+
+    def test_it_survives_every_provider_cleanup_can_be_set_to(self):
+        """It used to raise KeyError on the local model, whose executable is ""."""
+        for name in cleanup.PROVIDERS:
+            with self.subTest(provider=name):
+                reply = self.run_doctor(cleanup_provider=name)
+                self.assertEqual(reply["cleanup"]["provider"], name)
+                self.run_doctor(as_json=False, cleanup_provider=name)
+
+    def test_a_provider_with_no_key_to_check_says_so_rather_than_no(self):
+        """A CLI needs none, so `false` there would read as one gone missing."""
+        self.assertIsNone(self.run_doctor(cleanup_provider="claude")["cleanup"]["key"])
+        self.assertIsNone(self.run_doctor(cleanup_provider="local")["cleanup"]["key"])
+        self.assertIs(self.run_doctor(cleanup_provider="gemini")["cleanup"]["key"],
+                      False)
+
+    def test_cleanup_on_google_is_a_question_about_its_own_key(self):
+        line = self.run_doctor(as_json=False, cleanup_provider="gemini",
+                               cleanup_gemini_model="gemini-2.5-flash")
+        self.assertIn("Google AI Studio key, cleaning up on gemini-2.5-flash", line)
+
     def test_it_asks_after_the_programs_this_desktop_actually_uses(self):
         """A missing ydotool on a Mac is a red mark with nothing behind it."""
         with mock.patch.object(cli.paste, "desktop", return_value=paste.MACOS):
@@ -628,7 +669,11 @@ class WithoutAnInstance(DikteTest):
     def run_verb(self, argv):
         # launch_gui replaces this process with the application, so it never
         # comes back in real use and must not be allowed to here.
+        # `ask` with no text reads what was piped in, and the runner's own
+        # stdin is not that: under pytest it is an object that refuses to be
+        # read at all.
         with mock.patch.object(ipc, "send", return_value=None), \
+                mock.patch.object(sys, "stdin", io.StringIO()), \
                 mock.patch.object(cli, "launch_gui") as launch, \
                 captured() as (out, err):
             code = cli.run(argv)

tests/test_config.py

@@ -8,7 +8,10 @@ config and now shadows the default.
 
 import json
 import os
+import pathlib
 import sys
+import threading
+import time
 import unittest
 from unittest import mock
 
@@ -44,6 +47,21 @@ class Loading(DikteTest):
             conf = cfg.Config()
         self.assertEqual(conf["cleanup_model"], cfg.DEFAULTS["cleanup_model"])
 
+    def test_a_config_that_is_not_json_is_set_aside_as_evidence(self):
+        """Left in place it would be overwritten by the very next save."""
+        cfg.CONFIG_DIR.mkdir(parents=True, exist_ok=True)
+        cfg.CONFIG_FILE.write_text("{not json", encoding="utf-8")
+        broken = cfg.CONFIG_FILE.with_suffix(".json.broken")
+        with mock.patch("builtins.print") as told:
+            conf = cfg.Config()
+        self.assertEqual(broken.read_text(encoding="utf-8"), "{not json")
+        self.assertFalse(cfg.CONFIG_FILE.exists())
+        self.assertIn(str(broken), told.call_args[0][0])
+        conf.save()
+        self.assertEqual(broken.read_text(encoding="utf-8"), "{not json")
+        self.assertEqual(self.read_config_file()["cleanup_model"],
+                         cfg.DEFAULTS["cleanup_model"])
+
     def test_a_config_that_is_json_but_not_an_object(self):
         cfg.CONFIG_DIR.mkdir(parents=True, exist_ok=True)
         cfg.CONFIG_FILE.write_text("[1, 2]", encoding="utf-8")
@@ -113,6 +131,41 @@ class Saving(DikteTest):
         conf.save()
         self.assertEqual(i18n.language(), "tr")
 
+    def test_the_settings_hit_the_disk_before_the_swap(self):
+        """Renaming a file still in the page cache into place makes a power
+        cut a settings wipe, which is what the atomic replace exists to stop."""
+        with mock.patch("os.fsync") as fsync:
+            cfg.Config().save()
+        fsync.assert_called_once()
+
+    def test_a_file_held_briefly_by_a_scanner_does_not_fail_the_save(self):
+        """Antivirus and sync tools on Windows hold a fresh file for a moment,
+        and the rename over it fails until they let go."""
+        attempts = []
+        real_replace = pathlib.Path.replace
+
+        def flaky(path, target):
+            attempts.append(str(target))
+            if len(attempts) < 3:
+                raise PermissionError("held by a scanner")
+            return real_replace(path, target)
+
+        with mock.patch.object(pathlib.Path, "replace", flaky), \
+                mock.patch("time.sleep"):
+            cfg.Config().save()
+        self.assertEqual(len(attempts), 3)
+        self.assertEqual(self.read_config_file()["language"],
+                         cfg.DEFAULTS["language"])
+
+    def test_a_file_held_for_good_still_raises(self):
+        def held(path, target):
+            raise PermissionError("never let go")
+
+        with mock.patch.object(pathlib.Path, "replace", held), \
+                mock.patch("time.sleep"):
+            with self.assertRaises(PermissionError):
+                cfg.Config().save()
+
 
 class Keys(DikteTest):
     def test_a_stored_key_is_used(self):
@@ -134,6 +187,10 @@ class Keys(DikteTest):
     def test_every_provider_falls_back_to_the_variable_of_its_own_name(self):
         with mock.patch.dict(os.environ, {"GROQ_API_KEY": "gsk-env"}):
             self.assertEqual(cfg.Config().groq_key(), "gsk-env")
+        with mock.patch.dict(os.environ, {"GEMINI_API_KEY": "AIza-env"}):
+            self.assertEqual(cfg.Config().gemini_key(), "AIza-env")
+        with mock.patch.dict(os.environ, {"OPENCODE_API_KEY": "opencode-env"}):
+            self.assertEqual(cfg.Config().opencode_key(), "opencode-env")
 
 
 class TranscribeTarget(DikteTest):
@@ -163,6 +220,19 @@ class TranscribeTarget(DikteTest):
         self.assertEqual(target.service, "OpenRouter")
         self.assertEqual(target.api_key, "sk-or-test")
         self.assertEqual(target.model, "openai/whisper-1")
+        self.assertEqual(target.file_model, "")
+
+    def test_openrouter_carries_its_file_model(self):
+        conf = self.config(transcribe_provider="openrouter",
+                           openrouter_api_key="sk-or-test",
+                           openrouter_file_model=" openai/whisper-large-v3 ")
+        self.assertEqual(conf.transcribe_target().file_model,
+                         "openai/whisper-large-v3")
+
+    def test_only_openrouter_has_a_file_model(self):
+        conf = self.config(transcribe_provider="openai", openai_api_key="sk-test",
+                           openrouter_file_model="openai/whisper-large-v3")
+        self.assertEqual(conf.transcribe_target().file_model, "")
 
     def test_groq_when_it_is_picked(self):
         conf = self.config(transcribe_provider="groq", groq_api_key="gsk-test",
@@ -350,6 +420,23 @@ class History(DikteTest):
         cfg.delete_history([])
         self.assertEqual(len(cfg.read_history()), 1)
 
+    def test_amending_matches_on_content_and_patches_in_place(self):
+        rows = [self.entry("a"), self.entry("b")]
+        for row in rows:
+            cfg.append_history(row)
+        patched = cfg.amend_history(rows[0], cleanup_error="could not paste")
+        self.assertEqual(patched["cleanup_error"], "could not paste")
+        kept = cfg.read_history()
+        self.assertEqual([row["text"] for row in kept], ["a", "b"])
+        self.assertEqual(kept[0]["cleanup_error"], "could not paste")
+
+    def test_amending_a_row_a_trim_took_away_is_a_no_op(self):
+        row = self.entry("gone")
+        cfg.append_history(row)
+        cfg.clear_history()
+        self.assertIsNone(cfg.amend_history(row, cleanup_error="x"))
+        self.assertEqual(cfg.read_history(), [])
+
     def test_clearing(self):
         cfg.append_history(self.entry("a"))
         cfg.clear_history()
@@ -358,6 +445,40 @@ class History(DikteTest):
     def test_clearing_a_history_that_is_not_there(self):
         cfg.clear_history()   # must not raise
 
+    def test_an_append_during_a_trim_is_not_lost(self):
+        """Trim is read, cut, rewrite; a dictation appended between the read
+        and the rewrite must wait rather than be erased by a rewrite that
+        never saw it. The rewrite is slowed down to hold the race open."""
+        for index in range(10):
+            cfg.append_history(self.entry(str(index)))
+        real_write = cfg._write_history
+        rewriting = threading.Event()
+
+        def slow_write(lines):
+            rewriting.set()
+            time.sleep(0.1)
+            real_write(lines)
+
+        with mock.patch.object(cfg, "_write_history", slow_write):
+            trimmer = threading.Thread(target=cfg.trim_history, args=(3,))
+            trimmer.start()
+            # The trim now holds the lock inside its read-cut-rewrite window.
+            self.assertTrue(rewriting.wait(5))
+            appender = threading.Thread(target=cfg.append_history,
+                                        args=(self.entry("late"),))
+            appender.start()
+            trimmer.join()
+            appender.join()
+        self.assertEqual([row["text"] for row in cfg.read_history()],
+                         ["7", "8", "9", "late"])
+
+    def test_the_rewrite_hits_the_disk_before_the_swap(self):
+        for index in range(5):
+            cfg.append_history(self.entry(str(index)))
+        with mock.patch("os.fsync") as fsync:
+            cfg.trim_history(2)
+        fsync.assert_called_once()
+
 
 class Meetings(DikteTest):
     def entry(self, base, **changes):
@@ -430,6 +551,27 @@ class Meetings(DikteTest):
         cfg.delete_meetings([])
         self.assertEqual(len(cfg.read_meetings()), 1)
 
+    def test_the_index_hits_the_disk_before_the_swap(self):
+        with mock.patch("os.fsync") as fsync:
+            cfg.save_meeting(self.entry("a"))
+        fsync.assert_called_once()
+
+    def test_an_index_held_briefly_by_a_scanner_is_still_written(self):
+        real_replace = pathlib.Path.replace
+        attempts = []
+
+        def flaky(path, target):
+            attempts.append(str(target))
+            if len(attempts) < 3:
+                raise PermissionError("held by a scanner")
+            return real_replace(path, target)
+
+        with mock.patch.object(pathlib.Path, "replace", flaky), \
+                mock.patch("time.sleep"):
+            cfg.save_meeting(self.entry("a"))
+        self.assertEqual(len(attempts), 3)
+        self.assertEqual([row["base"] for row in cfg.read_meetings()], ["a"])
+
 
 class Defaults(unittest.TestCase):
     """The table itself, which every command line and settings tab reads."""
@@ -449,6 +591,19 @@ class Defaults(unittest.TestCase):
     def test_the_keys_ship_empty(self):
         self.assertEqual(cfg.DEFAULTS["openai_api_key"], "")
         self.assertEqual(cfg.DEFAULTS["openrouter_api_key"], "")
+        self.assertEqual(cfg.DEFAULTS["gemini_api_key"], "")
+        self.assertEqual(cfg.DEFAULTS["opencode_api_key"], "")
+
+    def test_google_ai_studio_is_a_cleanup_provider_and_not_a_transcriber(self):
+        """Its compatible endpoint has no /audio/transcriptions behind it."""
+        self.assertNotIn("gemini", cfg.TRANSCRIBERS)
+        self.assertIn("gemini", cleanup.PROVIDERS)
+
+    def test_opencode_ships_on_its_own_endpoint(self):
+        self.assertEqual(cfg.DEFAULTS["opencode_base_url"],
+                         "https://opencode.ai/zen/go/v1")
+        self.assertEqual(cfg.DEFAULTS["cleanup_opencode_model"], "deepseek-v4-flash")
+        self.assertEqual(cfg.DEFAULTS["assistant_opencode_model"], "deepseek-v4-flash")
 
     def test_every_language_specific_prompt_has_both_languages(self):
         for name in ("CLEANUP_PROMPT", "FILE_CLEANUP_PROMPT", "MEETING_PROMPT",

tests/test_filetranscribe.py

@@ -235,6 +235,28 @@ class ChunkSeconds(DikteTest):
         self.assertEqual(ft.chunk_seconds(self.file(ft.UPLOAD_LIMIT * 2), 0), 0.0)
 
 
+class Ffmpeg(DikteTest):
+    """How the converter process is started."""
+
+    def test_its_output_is_read_as_utf8_whatever_the_locale_says(self):
+        """ffmpeg writes UTF-8; read as the locale codepage its messages
+        mojibake, and a byte the codepage cannot place raises from inside
+        communicate itself."""
+        out = str(self.path("out.wav"))
+        with open(out, "wb") as fh:
+            fh.write(b"\x00")
+        proc = mock.Mock()
+        proc.communicate.return_value = ("", "")
+        proc.returncode = 0
+        proc.poll.return_value = 0
+        with mock.patch.object(ft.subprocess, "Popen", return_value=proc) as popen:
+            ft._ffmpeg(["-i", "in.mp4", out], out)
+        kwargs = popen.call_args.kwargs
+        self.assertTrue(kwargs["text"])
+        self.assertEqual(kwargs["encoding"], "utf-8")
+        self.assertEqual(kwargs["errors"], "replace")
+
+
 class Chunks(DikteTest):
     """What each provider is handed, and in how many pieces."""
 

tests/test_ggml.py

@@ -9,6 +9,7 @@ import contextlib
 import hashlib
 import io
 import os
+import pathlib
 import signal
 import sys
 import tarfile
@@ -48,6 +49,11 @@ def item(name, data, url="https://example.invalid/f", sha=True):
                     hashlib.sha256(data).hexdigest() if sha else "")
 
 
+def listed(name, size):
+    """A row as a listing hands it over: a name and a size, no bytes."""
+    return hub.Item(name, f"https://example.invalid/{name}", size, "a" * 64)
+
+
 @contextlib.contextmanager
 def serving(release, archive):
     """Answer by what is being asked for rather than by what came before.
@@ -179,6 +185,21 @@ class Download(Local):
                 ggml.download(item("m.bin", b"x"), target)
         self.assertFalse(target.exists())
 
+    def test_a_target_held_open_keeps_the_finished_download(self):
+        # Windows refuses to replace a file a running server holds open. The
+        # bytes are complete and verified by then, so the .part must survive
+        # the failure rather than being deleted with everything else.
+        data = b"a finished, verified download"
+        target = self.path("data", "models", "m.bin")
+        with fake_urlopen(body(data)):
+            with mock.patch.object(pathlib.Path, "replace",
+                                   side_effect=PermissionError(13, "in use")):
+                with self.assertRaises(ggml.LocalError) as caught:
+                    ggml.download(item("m.bin", data), target)
+        self.assertIn("held open", str(caught.exception))
+        self.assertEqual(target.with_name("m.bin.part").read_bytes(), data)
+        self.assertFalse(target.exists())
+
 
 # --- installing a program -------------------------------------------------
 
@@ -189,6 +210,7 @@ class InstallProgram(Local):
         # These fixtures are Ubuntu release archives.  Keep checking that path
         # on every host, including the Mac that checks the macOS backend.
         self.patch_attr(sys, "platform", "linux")
+        self.patch_attr(ggml.platform, "machine", lambda: "x86_64")
         # Built once, because the release listing has to publish its checksum
         # and a tarball is not the same bytes twice.
         self.archive = tarball({
@@ -205,6 +227,7 @@ class InstallProgram(Local):
 
     def install(self, *names, archive=None):
         self.patch_attr(ggml, "_arch", lambda: "x64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: False)
         blob = self.archive if archive is None else archive
         with serving(self.release(*names, archive=blob), blob) as calls:
             path = ggml.install_program(ggml.WHISPER)
@@ -222,6 +245,162 @@ class InstallProgram(Local):
                                "whisper-bin-ubuntu-x64.tar.gz")
         self.assertTrue(urls[1].endswith("whisper-bin-ubuntu-x64.tar.gz"))
 
+    def test_the_nightly_pointer_is_followed_to_where_the_builds_are(self):
+        """llama.cpp's latest release carries a tag name, not the binaries."""
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: False)
+        marker = self.release(ggml.NIGHTLY_TAG)
+        nightly = dict(self.release("llama-b10809-bin-ubuntu-x64.tar.gz"),
+                       tag_name="b10809")
+
+        def opener(request, timeout=None):
+            url = request.full_url
+            if url.endswith("/releases/latest"):
+                return json_body(marker)
+            if url.endswith("/releases/tags/b10809"):
+                return json_body(nightly)
+            if url.endswith(ggml.NIGHTLY_TAG):
+                return body(b"b10809\n")
+            return body(self.archive)
+
+        with mock.patch("urllib.request.urlopen", side_effect=opener):
+            tag, found = ggml._pick_asset(ggml.LLAMA)
+        self.assertEqual(tag, "b10809")
+        self.assertEqual(found.name, "llama-b10809-bin-ubuntu-x64.tar.gz")
+
+    def test_without_a_pointer_the_newest_release_that_has_a_build_is_taken(self):
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: False)
+        marker = self.release("source.zip")
+        listing = [dict(self.release("llama-b2-bin-win-cpu-x64.zip"), tag_name="b2"),
+                   dict(self.release("llama-b1-bin-ubuntu-x64.tar.gz"), tag_name="b1")]
+
+        def opener(request, timeout=None):
+            url = request.full_url
+            return json_body(listing if "per_page" in url else marker)
+
+        with mock.patch("urllib.request.urlopen", side_effect=opener):
+            tag, found = ggml._pick_asset(ggml.LLAMA)
+        self.assertEqual(tag, "b1")
+        self.assertEqual(found.name, "llama-b1-bin-ubuntu-x64.tar.gz")
+    def test_linux_x64_with_vulkan_takes_diktes_accelerated_build(self):
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: True)
+        listing = self.release("whisper-bin-ubuntu-vulkan-x64.tar.gz")
+        listing["tag_name"] = "whisper.cpp-v1.9.3"
+        managed_sha = hashlib.sha256(self.archive).hexdigest()
+        with mock.patch.object(ggml, "MANAGED_WHISPER_SHA256", managed_sha,
+                               create=True):
+            with fake_urlopen(listing, body(self.archive)) as calls:
+                path = ggml.install_program(ggml.WHISPER)
+        urls = [call.full_url for call in calls]
+        self.assertIn(
+            "/repos/yusufipk/dikte/releases/tags/whisper.cpp-v1.9.3",
+            urls[0],
+        )
+        self.assertTrue(urls[1].endswith(
+            "whisper-bin-ubuntu-vulkan-x64.tar.gz"))
+        self.assertTrue(os.path.isfile(path))
+        self.assertEqual("v1.9.3", ggml.installed_version(ggml.WHISPER))
+        self.assertFalse(ggml.vulkan_missing(ggml.WHISPER))
+
+    def test_an_explicit_whisper_version_still_comes_from_upstream(self):
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: True)
+        listing = self.release("whisper-bin-ubuntu-x64.tar.gz")
+        with fake_urlopen(listing, body(self.archive)) as calls:
+            ggml.install_program(ggml.WHISPER, tag="v1.9.1")
+        self.assertIn(
+            "/repos/ggml-org/whisper.cpp/releases/tags/v1.9.1",
+            calls[0].full_url,
+        )
+
+    def test_linux_arm64_keeps_using_the_upstream_cpu_build(self):
+        self.patch_attr(ggml, "_arch", lambda: "arm64")
+        self.patch_attr(ggml.platform, "machine", lambda: "aarch64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: True)
+        listing = self.release("whisper-bin-ubuntu-arm64.tar.gz")
+        with fake_urlopen(listing, body(self.archive)) as calls:
+            ggml.install_program(ggml.WHISPER)
+        self.assertIn(
+            "/repos/ggml-org/whisper.cpp/releases/latest",
+            calls[0].full_url,
+        )
+
+    def test_linux_non_x86_does_not_try_the_managed_x64_build(self):
+        self.patch_attr(ggml, "_has_vulkan", lambda: True)
+        listing = self.release("whisper-bin-ubuntu-arm64.tar.gz")
+        with mock.patch("platform.machine", return_value="ppc64le"):
+            with fake_urlopen(listing, listing) as calls:
+                with self.assertRaises(ggml.LocalError):
+                    ggml.install_program(ggml.WHISPER)
+        self.assertIn(
+            "/repos/ggml-org/whisper.cpp/releases/latest",
+            calls[0].full_url,
+        )
+
+    def test_a_missing_managed_build_falls_back_to_upstream_cpu(self):
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: True)
+        managed = self.release("Dikte-1.1.0-x86_64.AppImage")
+        managed["tag_name"] = "whisper.cpp-v1.9.3"
+        upstream = self.release("whisper-bin-ubuntu-x64.tar.gz")
+        with fake_urlopen(managed, upstream, body(self.archive)) as calls:
+            path = ggml.install_program(ggml.WHISPER)
+        urls = [call.full_url for call in calls]
+        self.assertIn(
+            "/repos/yusufipk/dikte/releases/tags/whisper.cpp-v1.9.3",
+            urls[0],
+        )
+        self.assertIn("/repos/ggml-org/whisper.cpp/releases/latest", urls[1])
+        self.assertTrue(urls[2].endswith("whisper-bin-ubuntu-x64.tar.gz"))
+        self.assertTrue(os.path.isfile(path))
+
+    def test_a_managed_build_with_an_unreviewed_digest_falls_back(self):
+        self.patch_attr(ggml, "_has_vulkan", lambda: True)
+        managed = self.release("whisper-bin-ubuntu-vulkan-x64.tar.gz")
+        managed["assets"][0]["digest"] = "sha256:" + "0" * 64
+        upstream = self.release("whisper-bin-ubuntu-x64.tar.gz")
+        with fake_urlopen(managed, upstream, body(self.archive)) as calls:
+            try:
+                path = ggml.install_program(ggml.WHISPER)
+            except ggml.LocalError as exc:
+                self.fail(f"unreviewed digest did not fall back: {exc}")
+        urls = [call.full_url for call in calls]
+        self.assertEqual(3, len(urls))
+        self.assertTrue(urls[2].endswith("whisper-bin-ubuntu-x64.tar.gz"))
+        self.assertTrue(os.path.isfile(path))
+
+    def test_an_unavailable_managed_release_falls_back_to_upstream_cpu(self):
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: True)
+        upstream = self.release("whisper-bin-ubuntu-x64.tar.gz")
+        with fake_urlopen(http_error(404), upstream,
+                          body(self.archive)) as calls:
+            path = ggml.install_program(ggml.WHISPER)
+        self.assertEqual(3, len(calls))
+        self.assertTrue(calls[2].full_url.endswith(
+            "whisper-bin-ubuntu-x64.tar.gz"))
+        self.assertTrue(os.path.isfile(path))
+
+    def test_a_fallback_to_the_processor_build_is_there_to_be_shown(self):
+        """Until the Vulkan package is published every download lands the
+        processor build, and a graphics card sitting idle looks exactly like
+        one being used. The window asks this and says so."""
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: True)
+        managed = self.release("Dikte-1.1.0-x86_64.AppImage")
+        managed["tag_name"] = "whisper.cpp-v1.9.3"
+        upstream = self.release("whisper-bin-ubuntu-x64.tar.gz")
+        with fake_urlopen(managed, upstream, body(self.archive)):
+            ggml.install_program(ggml.WHISPER)
+        self.assertTrue(ggml.vulkan_missing(ggml.WHISPER))
+
+    def test_a_machine_with_no_vulkan_is_not_told_it_is_missing_one(self):
+        # Nothing was on offer to fall back from, so there is nothing to say.
+        self.install("whisper-bin-ubuntu-x64.tar.gz")
+        self.assertFalse(ggml.vulkan_missing(ggml.WHISPER))
+
     def test_a_release_with_nothing_for_this_machine_says_so(self):
         self.patch_attr(ggml, "_arch", lambda: "x64")
         with fake_urlopen(self.release("whisper-bin-Win32.zip")):
@@ -275,6 +454,70 @@ class InstallProgram(Local):
             self.install("whisper-bin-ubuntu-x64.tar.gz", archive=empty)
         self.assertIn("whisper-server", str(caught.exception))
 
+    def test_a_failed_update_leaves_the_working_install_alone(self):
+        # The old install used to be deleted before the new bytes had even
+        # arrived, so a bad download left no local server at all.
+        path, _ = self.install("whisper-bin-ubuntu-x64.tar.gz")
+        with self.assertRaises(ggml.LocalError):
+            self.install("whisper-bin-ubuntu-x64.tar.gz",
+                         archive=b"not an archive at all")
+        self.assertEqual(ggml.installed_program(ggml.WHISPER), path)
+        self.assertTrue(os.path.isfile(path))
+        self.assertEqual(ggml.installed_version(ggml.WHISPER), "v1.9.1")
+        # And the half-made sibling did not linger either.
+        left = list(self.path("data", "bin", "whisper").glob("*.new"))
+        self.assertEqual(left, [])
+
+    def test_an_update_that_never_downloaded_leaves_the_install_alone(self):
+        path, _ = self.install("whisper-bin-ubuntu-x64.tar.gz")
+        listing = self.release("whisper-bin-ubuntu-x64.tar.gz")
+        with serving(listing, self.archive):
+            with mock.patch.object(ggml, "download",
+                                   side_effect=ggml.LocalError("no route")):
+                with self.assertRaises(ggml.LocalError):
+                    ggml.install_program(ggml.WHISPER)
+        self.assertEqual(ggml.installed_program(ggml.WHISPER), path)
+        self.assertTrue(os.path.isfile(path))
+
+    class StubServer:
+        """Owns the installed binary, remembers when it was told to stop."""
+
+        def __init__(self):
+            self.program = ggml.WHISPER
+            self.stops = 0
+            self.new_version_was_ready = False
+
+        def settings(self):
+            return {"binary": ""}
+
+        def stop(self):
+            self.stops += 1
+            self.new_version_was_ready = any(
+                (ggml.BIN_DIR / "whisper").glob("*.new"))
+
+    def test_the_running_server_is_stopped_only_for_the_swap(self):
+        # The outage is the swap, not the transfer: the server keeps answering
+        # through a download that can take minutes, and is stopped only once
+        # the replacement is unpacked next door and known to be whole.
+        self.install("whisper-bin-ubuntu-x64.tar.gz")
+        server = self.StubServer()
+        with mock.patch.object(ggml, "SERVERS", (server,)):
+            self.install("whisper-bin-ubuntu-x64.tar.gz")
+        self.assertEqual(server.stops, 1)
+        self.assertTrue(server.new_version_was_ready)
+
+    def test_a_download_that_fails_never_stops_the_server(self):
+        self.install("whisper-bin-ubuntu-x64.tar.gz")
+        server = self.StubServer()
+        listing = self.release("whisper-bin-ubuntu-x64.tar.gz")
+        with mock.patch.object(ggml, "SERVERS", (server,)):
+            with serving(listing, self.archive):
+                with mock.patch.object(ggml, "download",
+                                       side_effect=ggml.LocalError("no route")):
+                    with self.assertRaises(ggml.LocalError):
+                        ggml.install_program(ggml.WHISPER)
+        self.assertEqual(server.stops, 0)
+
 
     def test_a_release_without_a_published_checksum_is_refused(self):
         # GitHub did not always publish one, and whisper.cpp v1.8.0 and older
@@ -426,6 +669,53 @@ class Catalogue(Local):
             with self.assertRaises(ggml.LocalError):
                 ggml.whisper_models()
 
+    def test_the_speculative_decoding_heads_are_not_models(self):
+        # They are the small files in a repository, so a list sorted by size
+        # puts them first, where the eye lands and the click goes.
+        tree = GGUF_TREE + [
+            {"type": "file", "path": "dflash-Qwen3-8B-Q8_0.gguf",
+             "size": 1_120_000_000, "lfs": {"oid": "f" * 64}},
+            {"type": "file", "path": "eagle3-gpt-oss-20b-Q8_0.gguf",
+             "size": 920_000_000, "lfs": {"oid": "0" * 64}},
+        ]
+        with fake_urlopen(tree):
+            names = [q.name for q in ggml.llm_quants("ggml-org/x-GGUF")]
+        self.assertEqual(names,
+                         ["gemma-3-4b-it-Q4_K_M.gguf", "gemma-3-4b-it-Q8_0.gguf"])
+
+    def test_a_speech_or_vision_repository_is_not_a_cleanup_publisher(self):
+        listing = [{"id": "ggml-org/parakeet-GGUF"},
+                   {"id": "ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF"},
+                   {"id": "ggml-org/SmolVLM2-256M-Video-Instruct-GGUF"},
+                   {"id": "ggml-org/Qwen3-8B-Base-GGUF"},
+                   {"id": "ggml-org/SmolLM3-3B-GGUF"}]
+        with fake_urlopen(listing):
+            found = ggml.llm_repos()
+        self.assertEqual([r for r in found if r.startswith("ggml-org/Smol")],
+                         ["ggml-org/SmolLM3-3B-GGUF"])
+        self.assertNotIn("ggml-org/parakeet-GGUF", found)
+        self.assertNotIn("ggml-org/Qwen3-8B-Base-GGUF", found)
+
+    def test_a_publisher_is_not_dropped_for_a_word_it_happens_to_contain(self):
+        # The skip marks are matched as plain substrings, and an unanchored
+        # "test-" is also inside "Latest-".
+        self.assertTrue(ggml.can_clean("ggml-org/Qwen3-Latest-GGUF"))
+        self.assertFalse(ggml.can_clean("ggml-org/test-model-router-download"))
+
+    def test_a_base_model_beside_its_tuned_twin_is_dropped(self):
+        # Gemma names the base model after the tuned one with the `-it` taken
+        # out, so the two sit next to each other and the wrong one answers a
+        # cleanup prompt by carrying on writing the transcript.
+        listing = [{"id": "ggml-org/gemma-4-E2B-GGUF"},
+                   {"id": "ggml-org/gemma-4-E2B-it-GGUF"},
+                   {"id": "ggml-org/Qwen3-0.6B-GGUF"}]
+        with fake_urlopen(listing):
+            found = ggml.llm_repos()
+        self.assertNotIn("ggml-org/gemma-4-E2B-GGUF", found)
+        self.assertIn("ggml-org/gemma-4-E2B-it-GGUF", found)
+        # Nothing named it, so nothing says it is the wrong half of a pair.
+        self.assertIn("ggml-org/Qwen3-0.6B-GGUF", found)
+
     def test_what_is_on_disk_is_read_from_disk(self):
         self.assertEqual(ggml.installed_whisper_models(), [])
         path = ggml.whisper_model_path("ggml-base.bin")
@@ -456,12 +746,14 @@ STAND_IN = textwrap.dedent("""
     def opt(name, default=""):
         return args[args.index(name) + 1] if name in args else default
 
+    time.sleep(float(opt("--wait", "0")))
+
+    # After the sleep, so that --wait plus --die is a program that runs for a
+    # while and then crashes, the way a bad model dies mid-load.
     if "--die" in args:
         print("could not load model: no such file")
         sys.exit(2)
 
-    time.sleep(float(opt("--wait", "0")))
-
     started = time.monotonic()
     healthy_after = float(opt("--healthy-after", "0"))
 
@@ -539,6 +831,14 @@ class Servers(Local):
         self.assertTrue(server.running)
         self.assertTrue(second)
 
+    def count_ports(self):
+        """Record every port handed to a launch, one per attempt."""
+        ports = []
+        real = ggml._free_port
+        self.patch_attr(ggml, "_free_port",
+                        lambda: ports.append(real()) or ports[-1])
+        return ports
+
     def test_a_program_that_dies_reports_what_it_printed(self):
         server = self.server(extra=["--die"])
         with self.assertRaises(ggml.LocalError) as caught:
@@ -546,6 +846,46 @@ class Servers(Local):
         self.assertIn("no such file", str(caught.exception))
         self.assertFalse(server.running)
 
+    def test_an_early_death_that_never_listened_is_retried(self):
+        # A child that loses the bind race fails and exits at once, and which
+        # port was lost cannot be told from the log: the shape of the failure,
+        # not its wording, is what earns another port.
+        ports = self.count_ports()
+        server = self.server(extra=["--die"])
+        with self.assertRaises(ggml.LocalError):
+            server.serve()
+        self.assertEqual(len(ports), 3)
+
+    def test_a_late_crash_is_not_retried(self):
+        # A program that ran for a while before dying was not a bind race: it
+        # would die the same way on any port.
+        self.patch_attr(ggml, "EARLY_EXIT_WINDOW", 0.2)
+        ports = self.count_ports()
+        server = self.server(extra=["--wait", "0.5", "--die"])
+        with self.assertRaises(ggml.LocalError) as caught:
+            server.serve()
+        self.assertEqual(len(ports), 1)
+        self.assertIn("no such file", str(caught.exception))
+
+    def test_an_open_port_with_a_dead_child_is_not_ready(self):
+        # Another process winning the bind race leaves the port open while our
+        # child exits: the open port alone must not be read as ready.
+        polls = iter([None, 2])
+        proc = mock.Mock()
+        proc.poll = lambda: next(polls)
+        self.patch_attr(ggml, "_listening", lambda port: True)
+        reason, listened = self.server()._wait_ready(proc, 1)
+        self.assertEqual(reason, "exited")
+        self.assertFalse(listened)
+
+    def test_an_open_port_with_a_child_that_outlives_it_a_beat_is_ready(self):
+        proc = mock.Mock()
+        proc.poll = lambda: None
+        self.patch_attr(ggml, "_listening", lambda port: True)
+        reason, listened = self.server()._wait_ready(proc, 1)
+        self.assertEqual(reason, "ready")
+        self.assertTrue(listened)
+
     def test_a_model_that_is_still_loading_is_not_ready_yet(self):
         # llama binds its port first and answers /health with 503 until the
         # model is in memory, so the open port on its own is not the signal.
@@ -614,6 +954,72 @@ class Servers(Local):
     def test_no_pid_file_is_nothing_to_sweep(self):
         self.assertFalse(self.server().sweep())
 
+    def test_an_unverifiable_pid_is_kept_for_a_later_sweep(self):
+        # Could not be checked is not the same as known stale: dropping the
+        # file here would lose track of a server that may still hold a model.
+        server = self.server()
+        server._remember(4242)
+        with mock.patch.object(server, "_is_ours", return_value=None):
+            self.assertFalse(server.sweep())
+        self.assertTrue(server._pid_file().exists())
+
+    def test_a_pid_known_stale_is_forgotten(self):
+        server = self.server()
+        server._remember(4242)
+        with mock.patch.object(server, "_is_ours", return_value=False):
+            self.assertFalse(server.sweep())
+        self.assertFalse(server._pid_file().exists())
+
+    def test_the_pid_file_goes_once_the_kill_was_attempted(self):
+        server = self.server()
+        server._remember(4242)
+        with mock.patch.object(server, "_is_ours", return_value=True):
+            with mock.patch.object(ggml.os, "kill") as kill:
+                self.assertTrue(server.sweep())
+        kill.assert_called_once_with(4242, signal.SIGTERM)
+        self.assertFalse(server._pid_file().exists())
+
+    def test_forget_leaves_a_pid_file_that_is_no_longer_ours(self):
+        server = self.server()
+        server._remember(111)
+        # A Dikte started after us wrote its own server's pid over ours;
+        # removing the file would hide that server from every future sweep.
+        server._pid_file().write_text("222")
+        server._forget()
+        self.assertEqual(server._pid_file().read_text(), "222")
+
+    def test_forget_removes_the_pid_it_remembered(self):
+        server = self.server()
+        server._remember(111)
+        server._forget()
+        self.assertFalse(server._pid_file().exists())
+
+    def test_stop_waits_out_a_start_in_flight_and_kills_it(self):
+        # stop_all on quit must not slide past a launch that is mid-load: the
+        # child would then survive Dikte with nothing left that knows its pid.
+        children = []
+        real = ggml.subprocess.Popen
+
+        def popen(*args, **kwargs):
+            proc = real(*args, **kwargs)
+            children.append(proc)
+            return proc
+
+        self.patch_attr(ggml.subprocess, "Popen", popen)
+        server = self.server(extra=["--wait", "0.5"])
+        thread = threading.Thread(target=server.serve)
+        thread.start()
+        try:
+            deadline = time.monotonic() + 10
+            while not children and time.monotonic() < deadline:
+                time.sleep(0.01)     # until the launch is truly in flight
+            server.stop()
+        finally:
+            thread.join(timeout=10)
+        self.assertEqual(len(children), 1)
+        self.assertIsNotNone(children[0].poll())
+        self.assertFalse(server.running)
+
     def test_a_start_that_goes_wrong_takes_its_process_with_it(self):
         started = []
 
@@ -752,6 +1158,10 @@ class InstallOnWindows(Local):
         super().setUp()
         self.patch_attr(sys, "platform", "win32")
         self.patch_attr(ggml, "_arch", lambda: "x64")
+        # shutil.which cannot be allowed through to the real one: standing on
+        # win32 from another system, Python 3.12's Windows branch of which()
+        # reaches for the nt module that is not there.
+        self.enterContext(mock.patch("shutil.which", return_value=None))
         self.archive = zipball({
             "Release/whisper-server.exe": b"MZ not really a program",
             "Release/whisper.dll": b"not really a library",
@@ -785,3 +1195,245 @@ class InstallOnWindows(Local):
             with self.assertRaises(ggml.LocalError) as caught:
                 ggml.install_program(ggml.WHISPER)
         self.assertIn("this machine", str(caught.exception))
+
+
+class WindowsOwnership(Local):
+    """Sweeping on Windows goes by the executable's full path, never its name:
+    the base name alone is anyone's whisper-server.exe, and this answer decides
+    what gets killed."""
+
+    def setUp(self):
+        super().setUp()
+        self.patch_attr(sys, "platform", "win32")
+        # See InstallOnWindows: the real which() on win32 wants the nt module.
+        self.enterContext(mock.patch("shutil.which", return_value=None))
+        self.made = ggml.Server(ggml.WHISPER, lambda values: [], {"binary": ""})
+
+    def image(self, path):
+        self.patch_attr(ggml, "_win_image_name", lambda pid: path)
+
+    def test_a_binary_under_our_bin_directory_is_ours(self):
+        self.image(str(ggml.BIN_DIR / "whisper" / "v1.9.1" / "whisper-server.exe"))
+        self.assertIs(self.made._is_ours(1234), True)
+
+    def test_the_configured_binary_is_ours_wherever_it_lives(self):
+        mine = self.path("elsewhere", "whisper-server.exe")
+        mine.parent.mkdir(parents=True)
+        mine.write_bytes(b"MZ")
+        mine.chmod(0o755)
+        self.made._settings["binary"] = str(mine)
+        self.image(str(mine.resolve()))
+        self.assertIs(self.made._is_ours(1234), True)
+
+    def test_the_same_name_somewhere_else_is_not_ours(self):
+        self.image(str(self.path("theirs", "whisper-server.exe")))
+        with mock.patch("shutil.which", return_value=None):
+            self.assertIs(self.made._is_ours(1234), False)
+
+    def test_a_process_that_cannot_be_read_is_no_verdict(self):
+        # OpenProcess answering nothing covers both "gone" and "not readable
+        # from here", and only one of those makes the pid file safe to drop.
+        self.image("")
+        self.assertIsNone(self.made._is_ours(1234))
+
+
+class Machine(Local):
+    """What this machine can hold, and what that makes worth pointing at."""
+
+    def _sysconf(self, phys_pages, page_size=4096):
+        """Stand where sysconf answers whatever this test wants it to.
+
+        `create` because Windows has no os.sysconf at all, and a patch that
+        insists on the real attribute fails there before the test runs. What
+        the code under test does about that absence is two lines down from
+        what these are checking, and it is checked on its own below.
+        """
+        return mock.patch.object(
+            ggml.os, "sysconf", create=True,
+            side_effect=lambda name: (page_size if name == "SC_PAGE_SIZE"
+                                      else phys_pages))
+
+    def test_the_memory_is_read_the_way_each_system_reports_it(self):
+        # Linux and most Macs answer through sysconf.
+        with self._sysconf(4_194_304):
+            self.assertEqual(ggml.total_memory(), 16 * ggml.GB)
+
+    def test_a_mac_without_the_page_count_is_asked_for_the_number(self):
+        # Not every build of Python on a Mac carries SC_PHYS_PAGES, and a Mac
+        # that answered nothing would be a Mac with none of this on it.
+        def answer(args, **kwargs):
+            self.assertEqual(args, ["sysctl", "-n", "hw.memsize"])
+            return mock.Mock(stdout=f"{32 * ggml.GB}\n")
+
+        with mock.patch.object(ggml.os, "sysconf", create=True,
+                               side_effect=ValueError), \
+                mock.patch.object(sys, "platform", "darwin"), \
+                mock.patch.object(ggml.subprocess, "run", answer):
+            self.assertEqual(ggml.total_memory(), 32 * ggml.GB)
+
+    def test_a_sysconf_that_shrugs_is_an_unknown_machine_and_not_a_tiny_one(self):
+        # sysconf answers -1 for a limit it holds to be indeterminate and
+        # CPython hands that back rather than raising, so the product came out
+        # negative: a 64 GB workstation was told every model past 512 MB was
+        # too big for it, and the machine line read "Memory: -4096 B".
+        with self._sysconf(-1):
+            self.assertEqual(ggml.total_memory(), 0)
+        self.assertTrue(ggml.fits(574 << 20, memory=0))
+
+    def test_the_memory_is_read_once_and_kept(self):
+        # A list of thirty rows asks seventy times, and on the Mac path the
+        # answer comes from a program rather than a library call.
+        calls = []
+        with mock.patch.object(ggml, "_read_memory",
+                               lambda: calls.append(1) or 16 * ggml.GB):
+            self.assertEqual(ggml.total_memory(), 16 * ggml.GB)
+            self.assertEqual(ggml.total_memory(), 16 * ggml.GB)
+        self.assertEqual(len(calls), 1)
+
+    def test_a_system_that_answers_nothing_is_an_unknown_machine(self):
+        with mock.patch.object(ggml.os, "sysconf", create=True,
+                               side_effect=ValueError), \
+                mock.patch.object(sys, "platform", "linux"):
+            self.assertEqual(ggml.total_memory(), 0)
+
+    def test_a_mac_is_taken_to_have_a_graphics_interface(self):
+        with mock.patch.object(sys, "platform", "darwin"):
+            self.assertEqual(ggml.accelerator(), "Metal")
+
+    def test_elsewhere_the_vulkan_loader_is_what_says_so(self):
+        with mock.patch.object(sys, "platform", "linux"), \
+                mock.patch.object(ggml.ctypes.util, "find_library",
+                                  lambda name: "/usr/lib/libvulkan.so.1"):
+            self.assertEqual(ggml.accelerator(), "Vulkan")
+        with mock.patch.object(sys, "platform", "linux"), \
+                mock.patch.object(ggml.ctypes.util, "find_library",
+                                  lambda name: None):
+            self.assertEqual(ggml.accelerator(), "")
+
+    def test_a_model_is_measured_against_half_the_memory(self):
+        self.assertTrue(ggml.fits(2 * ggml.GB, memory=8 * ggml.GB))
+        self.assertFalse(ggml.fits(4 * ggml.GB, memory=8 * ggml.GB))
+
+    def test_a_machine_whose_memory_could_not_be_read_holds_anything(self):
+        # A wrong "too big" is worse advice than none.
+        self.assertTrue(ggml.fits(40 * ggml.GB, memory=0))
+
+    def test_the_smallest_machine_is_not_the_one_where_everything_fits(self):
+        # Half of 2 GB less the gigabyte of overhead is nothing, and a budget
+        # of nothing used to read as the unknown machine above.
+        self.assertFalse(ggml.fits(3 * ggml.GB, memory=2 * ggml.GB))
+
+    def test_a_crowded_machine_is_pointed_at_the_smaller_model(self):
+        self.assertEqual(ggml.suggested_whisper(memory=3 * ggml.GB, graphics=""),
+                         ggml.SMALL_MACHINE_WHISPER)
+
+    def test_a_card_and_the_memory_for_it_are_pointed_at_the_accurate_one(self):
+        self.assertEqual(
+            ggml.suggested_whisper(memory=32 * ggml.GB, graphics="Vulkan"),
+            ggml.ACCURATE_WHISPER)
+
+    def test_memory_without_a_card_is_pointed_at_the_fast_one(self):
+        # Several times the work per second is several times a long wait on a
+        # processor, whatever there is room for.
+        self.assertEqual(
+            ggml.suggested_whisper(memory=32 * ggml.GB, graphics=""),
+            ggml.SUGGESTED_WHISPER)
+
+    def test_a_sixteen_gigabyte_machine_counts_as_a_roomy_one(self):
+        # What a machine reports is what the firmware and the graphics left
+        # of it: 16 GB answers about 15.4, and a threshold written at the
+        # number on the box is one no machine ever reaches.
+        self.assertEqual(
+            ggml.suggested_whisper(memory=int(15.4 * ggml.GB), graphics="Metal"),
+            ggml.ACCURATE_WHISPER)
+
+    def test_the_suggestion_that_fits_is_offered_first(self):
+        first = ggml.suggested_llm(memory=6 * ggml.GB)[0]
+        self.assertTrue(ggml.fits(ggml.SUGGESTED_LLM_SIZE[first],
+                                  memory=6 * ggml.GB))
+        # Nothing is dropped: what does not fit today fits once something else
+        # is closed.
+        self.assertEqual(sorted(ggml.suggested_llm(memory=6 * ggml.GB)),
+                         sorted(ggml.SUGGESTED_LLM))
+
+    def test_the_wanted_model_wins_when_there_is_room_for_it(self):
+        items = [listed("ggml-tiny.bin", 70 << 20),
+                 listed("ggml-large-v3-turbo-q5_0.bin", 574 << 20)]
+        self.assertEqual(
+            ggml.recommended(items, "ggml-large-v3-turbo-q5_0.bin",
+                             memory=16 * ggml.GB),
+            "ggml-large-v3-turbo-q5_0.bin")
+
+    def test_a_model_too_big_for_the_machine_is_not_recommended(self):
+        items = [listed("small.gguf", 1 << 30), listed("huge.gguf", 12 * ggml.GB)]
+        self.assertEqual(ggml.recommended(items, "huge.gguf",
+                                          memory=8 * ggml.GB), "small.gguf")
+
+    def test_the_full_precision_weights_are_never_the_recommendation(self):
+        # Twice the memory and twice the wait for a difference this job
+        # cannot see.
+        items = [listed("model-Q4_0.gguf", 2 * ggml.GB),
+                 listed("model-BF16.gguf", 3 * ggml.GB)]
+        self.assertEqual(ggml.recommended(items, memory=32 * ggml.GB),
+                         "model-Q4_0.gguf")
+
+    def test_nothing_is_recommended_when_nothing_fits(self):
+        self.assertEqual(
+            ggml.recommended([listed("huge.gguf", 40 * ggml.GB)],
+                             memory=8 * ggml.GB), "")
+
+
+class Grouping(Local):
+    """One group per model, rather than one long list sorted by size."""
+
+    def test_every_spelling_of_a_quantisation_reads_as_its_number(self):
+        # One list holds q5_1, Q4_K_M, MXFP4 and BF16, and the number is the
+        # whole of what any of them says to somebody choosing a row.
+        self.assertEqual(ggml.bit_depth("ggml-small-q5_1.bin"), 5)
+        self.assertEqual(ggml.bit_depth("SmolLM3-Q4_K_M.gguf"), 4)
+        self.assertEqual(ggml.bit_depth("gpt-oss-20b-MXFP4.gguf"), 4)
+        self.assertEqual(ggml.bit_depth("gemma-4-E2B-it-Q8_0.gguf"), 8)
+        # bf16 is not f16 read badly.
+        self.assertEqual(ggml.bit_depth("gemma-4-E2B-it-BF16.gguf"), 16)
+        self.assertEqual(ggml.bit_depth("mmproj-model-f16.gguf"), 16)
+        # A whisper file with no mark is the full model, and its name is the
+        # one convention here that does not carry the answer.
+        self.assertEqual(ggml.bit_depth("ggml-large-v3-turbo.bin"), 0)
+
+    def test_a_quantisation_belongs_to_the_model_it_is_a_copy_of(self):
+        self.assertEqual(ggml.whisper_family("ggml-small.en-q5_1.bin"), "small")
+        self.assertEqual(ggml.whisper_family("ggml-large-v3-q5_0.bin"),
+                         "large-v3")
+        self.assertEqual(ggml.whisper_family("ggml-large-v3-turbo.bin"),
+                         "large-v3-turbo")
+        self.assertEqual(ggml.whisper_family("ggml-medium.en.bin"), "medium")
+
+    def test_turbo_is_a_model_and_not_a_quantisation(self):
+        # The last chunk of the name is a quantisation for most of the list
+        # and part of the model's name here.
+        self.assertEqual(ggml.whisper_family("ggml-large-v3-turbo-q8_0.bin"),
+                         "large-v3-turbo")
+
+    def test_the_turbo_files_are_not_scattered_through_the_medium_ones(self):
+        # Sorted by size alone, large-v3-turbo-q5_0 lands between the two
+        # medium quantisations, half a screen from the model it is a copy of.
+        models = [listed("ggml-medium-q5_0.bin", 539 << 20),
+                  listed("ggml-large-v3-turbo-q5_0.bin", 574 << 20),
+                  listed("ggml-medium-q8_0.bin", 823 << 20),
+                  listed("ggml-large-v3-turbo.bin", 1624 << 20)]
+        groups = dict(ggml.whisper_groups(models))
+        self.assertEqual([i.name for i in groups["large-v3-turbo"]],
+                         ["ggml-large-v3-turbo-q5_0.bin",
+                          "ggml-large-v3-turbo.bin"])
+        self.assertEqual([i.name for i in groups["medium"]],
+                         ["ggml-medium-q5_0.bin", "ggml-medium-q8_0.bin"])
+
+    def test_the_smallest_model_comes_first_and_the_english_ones_last(self):
+        models = [listed("ggml-small.en-q5_1.bin", 190 << 20),
+                  listed("ggml-small-q5_1.bin", 190 << 20),
+                  listed("ggml-tiny.bin", 77 << 20)]
+        groups = ggml.whisper_groups(models)
+        self.assertEqual([family for family, _ in groups], ["tiny", "small"])
+        self.assertEqual([i.name for _, group in groups for i in group],
+                         ["ggml-tiny.bin", "ggml-small-q5_1.bin",
+                          "ggml-small.en-q5_1.bin"])

tests/test_hub.py

@@ -3,6 +3,7 @@
 import json
 
 from dikte import hub
+from dikte import paths
 from tests.support import DikteTest, fake_urlopen, http_error, url_error
 
 RELEASE = {
@@ -165,6 +166,15 @@ def os_utime(path):
     os.utime(path, (old, old))
 
 
+class CacheLocation(DikteTest):
+    """Resolved at import, like every other path constant."""
+
+    def test_the_cache_lives_in_the_system_cache_directory(self):
+        # One answer for both, the same way ggml and config share DATA_DIR:
+        # hub asked paths once, at import, and kept what it was told.
+        self.assertEqual(hub.CACHE_DIR, paths.cache_dir())
+
+
 class CacheOnDisk(DikteTest):
     def setUp(self):
         super().setUp()

tests/test_integrate.py

@@ -475,6 +475,17 @@ class Windows(unittest.TestCase):
         self.installed = pathlib.Path(self.tmp.name).resolve()
         self.app = self.installed / "Dikte.exe"
         self.app.write_text("")
+        # APPDATA pointed into the sandbox, so that the Startup folder these
+        # tests delete from is never the machine's own.
+        appdata = mock.patch.dict(os.environ,
+                                  {"APPDATA": str(self.installed / "Roaming")})
+        appdata.start()
+        self.addCleanup(appdata.stop)
+
+    def startup_shortcut(self):
+        """Where install.ps1 -Autostart puts a checkout's sign-in entry."""
+        return (self.installed / "Roaming" / "Microsoft" / "Windows"
+                / "Start Menu" / "Programs" / "Startup" / "Dikte.lnk")
 
     def _write(self, command):
         self.value = command
@@ -510,6 +521,45 @@ class Windows(unittest.TestCase):
         self.assertEqual(len(self.install()), 1)
         self.assertEqual(self.value, f'"{self.app}"')
 
+    def test_an_entry_for_another_working_install_is_left_alone(self):
+        """The same courtesy the Linux half pays another menu entry: an entry
+        naming an executable that still exists is an installation that still
+        works, and a start of this one has no business redirecting it."""
+        other = self.installed / "Elsewhere" / "Dikte.exe"
+        other.parent.mkdir()
+        other.write_text("")
+        self.value = f'"{other}"'
+        self.assertEqual(self.install(), [])
+        self.assertEqual(self.value, f'"{other}"')
+
+    def test_asking_outright_overrules_a_working_other_install(self):
+        other = self.installed / "Elsewhere" / "Dikte.exe"
+        other.parent.mkdir()
+        other.write_text("")
+        self.value = f'"{other}"'
+        self.assertEqual(self.install(force=True), [integrate._run_entry_name()])
+        self.assertEqual(self.value, f'"{self.app}"')
+
+    def test_typing_it_sweeps_away_a_checkout_startup_shortcut(self):
+        """install.ps1 -Autostart writes it, the Run value replaces it, and
+        both left in place would be two Diktes at every sign-in."""
+        shortcut = self.startup_shortcut()
+        shortcut.parent.mkdir(parents=True)
+        shortcut.write_text("")
+        changed = self.install(force=True)
+        self.assertIn(shortcut, changed)
+        self.assertFalse(shortcut.exists())
+
+    def test_a_start_leaves_a_checkout_startup_shortcut_alone(self):
+        """The silent call on every start has not been asked to move the
+        machine off its checkout."""
+        shortcut = self.startup_shortcut()
+        shortcut.parent.mkdir(parents=True)
+        shortcut.write_text("")
+        self.value = f'"{self.app}"'
+        self.assertEqual(self.install(), [])
+        self.assertTrue(shortcut.exists())
+
     def test_running_it_again_changes_nothing(self):
         self.install(force=True)
         self.assertEqual(self.install(), [])
@@ -521,6 +571,31 @@ class Windows(unittest.TestCase):
         self.assertEqual(self.remove(), [])
 
 
+class WindowedExecutable(unittest.TestCase):
+    """The windowed executable, looked up beside whichever one is running.
+
+    Beside rather than at a known place: the setup lays both executables into
+    one directory wherever that directory was put, so either can find the
+    other without knowing where the install is.
+    """
+
+    def setUp(self):
+        self.tmp = tempfile.TemporaryDirectory()
+        self.addCleanup(self.tmp.cleanup)
+        self.installed = pathlib.Path(self.tmp.name).resolve()
+
+    def test_found_beside_the_named_executable(self):
+        windowed = self.installed / "Dikte.exe"
+        windowed.write_text("")
+        self.assertEqual(
+            integrate.windowed_executable(str(self.installed / "dikte-cli.exe")),
+            windowed)
+
+    def test_none_when_no_setup_installed_one(self):
+        self.assertIsNone(
+            integrate.windowed_executable(str(self.installed / "dikte-cli.exe")))
+
+
 class WindowsExecutableNames(unittest.TestCase):
     """The two Windows executables, read out of the files that name them.
 

tests/test_ipc.py

@@ -14,6 +14,7 @@ import unittest
 from unittest import mock
 
 from dikte import ipc
+from tests.support import DikteTest
 
 
 class FakeSocket:
@@ -185,5 +186,74 @@ class Send(unittest.TestCase):
         self.assertTrue(sock.disconnected)
 
 
+class AlreadyServing(unittest.TestCase):
+    """The single-instance check, which listen() cannot be: a Windows pipe
+    takes a second server on the same name rather than refusing it."""
+
+    def probe(self, socket):
+        with mock.patch.object(ipc, "QLocalSocket", return_value=socket):
+            return ipc.already_serving()
+
+    def test_nothing_running_means_go_ahead(self):
+        self.assertFalse(self.probe(FakeSocket(connected=False)))
+
+    def test_an_answer_means_yield(self):
+        self.assertTrue(self.probe(FakeSocket(reply=b'{"ok": true}\n')))
+
+    def test_the_probe_has_no_side_effect(self):
+        """A probe that opened a window would open it during the relaunch a
+        slow instance provokes, on top of the verb being forwarded."""
+        sock = FakeSocket(reply=b'{"ok": true}\n')
+        self.probe(sock)
+        self.assertEqual(sock.written.decode("utf-8").strip(), "status")
+
+    def test_an_instance_too_old_to_answer_still_counts_as_running(self):
+        self.assertTrue(self.probe(FakeSocket(reply=b"")))
+
+
+class InstanceLock(DikteTest):
+    def setUp(self):
+        super().setUp()
+        # The lock derives its home from paths, which DikteTest's cfg patches
+        # do not cover; without this the test would write into the real one.
+        from dikte import paths
+        self.patch_attr(paths, "DATA_DIR", self.path("data"))
+
+    def test_one_holder_at_a_time(self):
+        first = ipc.instance_lock()
+        self.assertIsNotNone(first)
+        self.assertTrue(first.tryLock(0))
+        second = ipc.instance_lock()
+        self.assertFalse(second.tryLock(0))
+        first.unlock()
+        self.assertTrue(second.tryLock(0))
+        second.unlock()
+
+    def test_the_lock_lives_in_the_data_directory(self):
+        from dikte import paths
+        lock = ipc.instance_lock()
+        self.assertTrue(lock.tryLock(0))
+        self.assertTrue((paths.DATA_DIR / "dikte.lock").exists())
+        lock.unlock()
+
+
+class Respawn(unittest.TestCase):
+    def test_windows_starts_a_detached_process_and_returns(self):
+        with mock.patch.object(sys, "platform", "win32"), \
+                mock.patch.object(ipc, "launcher", return_value=["py", "x"]), \
+                mock.patch.object(ipc.subprocess, "Popen") as popen:
+            ipc.respawn(["--gui"])
+        self.assertEqual(popen.call_args.args[0], ["py", "x", "--gui"])
+        self.assertEqual(popen.call_args.kwargs["creationflags"],
+                         0x00000008 | 0x00000200)
+
+    def test_everywhere_else_the_process_is_replaced(self):
+        with mock.patch.object(sys, "platform", "linux"), \
+                mock.patch.object(ipc, "launcher", return_value=["py", "x"]), \
+                mock.patch.object(ipc.os, "execv") as execv:
+            ipc.respawn(["toggle", "--gui"])
+        execv.assert_called_once_with("py", ["py", "x", "toggle", "--gui"])
+
+
 if __name__ == "__main__":
     unittest.main()

tests/test_packaging.py

@@ -0,0 +1,217 @@
+"""The release build that makes Linux Vulkan a one-click install."""
+
+import hashlib
+import io
+import json
+import os
+import pathlib
+import shutil
+import subprocess
+import sys
+import tarfile
+import tempfile
+import unittest
+
+from dikte import ggml
+
+
+ROOT = pathlib.Path(__file__).parents[1]
+PACKAGING = ROOT / "packaging" / "whisper-vulkan"
+WORKFLOW = ROOT / ".github" / "workflows" / "whisper-vulkan.yml"
+
+
+class WhisperVulkanPackaging(unittest.TestCase):
+    @unittest.skipUnless(sys.platform != "win32" and shutil.which("bash"),
+                         "bash syntax check is unavailable")
+    def test_the_release_scripts_parse_as_shell(self):
+        for name in ("build-package.sh", "validate-package.sh",
+                     "smoke-runtime.sh"):
+            script = PACKAGING / name
+            checked = subprocess.run(
+                ["bash", "-n", script], capture_output=True, text=True,
+            )
+            self.assertEqual("", checked.stderr)
+            self.assertEqual(0, checked.returncode)
+
+    def test_the_workflow_builds_validates_smokes_and_publishes(self):
+        workflow = WORKFLOW.read_text(encoding="utf-8")
+        for step in ("Build deterministic archive",
+                     "Verify reviewed archive digest",
+                     "Validate archive and ELF contract",
+                     "CPU fallback smoke test (no Vulkan loader)",
+                     "Vulkan loader present, no device smoke test",
+                     "Vulkan plugin-load smoke test (Mesa llvmpipe)",
+                     "Publish dependency release"):
+            self.assertIn(step, workflow)
+        self.assertNotRegex(workflow, r"uses: [^\n]+@v\d+(?:\s|$)")
+
+    def test_publish_is_safe_for_dikte_and_limited_to_reviewed_master(self):
+        workflow = WORKFLOW.read_text(encoding="utf-8")
+        self.assertGreaterEqual(workflow.count("persist-credentials: false"), 2)
+        self.assertIn("github.ref == 'refs/heads/master'", workflow)
+        self.assertIn("--prerelease", workflow)
+        self.assertIn("--latest=false", workflow)
+        self.assertIn("--verify-tag", workflow)
+        self.assertIn("refusing to replace existing tag", workflow)
+        self.assertIn("^[0-9]+\\.[0-9]+\\.[0-9]+$", workflow)
+        self.assertIn("^[0-9a-f]{40}$", workflow)
+        publish_script = workflow.split("      - name: Publish dependency release", 1)[1]
+        publish_script = publish_script.split("        run: |", 1)[1]
+        self.assertNotIn("${{ inputs.", publish_script)
+
+    def test_bundle_ci_runs_only_for_what_the_bundle_is_built_from(self):
+        """A 45 minute build on a README typo is a tax on every other change.
+
+        What ties ggml.py to the release is checked in this file instead, and
+        this file runs on every pull request in milliseconds."""
+        workflow = WORKFLOW.read_text(encoding="utf-8")
+        trigger = workflow.split("workflow_dispatch:", 1)[0]
+        self.assertIn("- packaging/whisper-vulkan/**", trigger)
+        self.assertIn("- .github/workflows/whisper-vulkan.yml", trigger)
+        for path in ("dikte/ggml.py", "tests/test_ggml.py",
+                     "tests/test_packaging.py", "README.md", "README.tr.md"):
+            self.assertNotIn(f"- {path}", trigger)
+
+    def test_the_smoke_tests_run_what_dikte_runs(self):
+        """-ng is what Dikte passes when its GPU setting is off, and a run
+        with it never asks for a backend at all. The three runs that have to
+        hold are the ones without it: no loader, a loader with nothing behind
+        it, and a working device."""
+        script = (PACKAGING / "smoke-runtime.sh").read_text(encoding="utf-8")
+        code = "\n".join(line for line in script.splitlines()
+                         if not line.lstrip().startswith("#"))
+        self.assertNotIn("-ng", code)
+        for mode in ("cpu)", "noicd)", "vulkan)"):
+            self.assertIn(mode, script)
+        self.assertTrue((PACKAGING / "Dockerfile.runtime-noicd").is_file())
+
+    def test_an_unreviewed_version_is_reported_and_never_published(self):
+        """The digest of a version nobody has reviewed cannot be known before
+        it is built, so the gate cannot be the only way through."""
+        workflow = WORKFLOW.read_text(encoding="utf-8")
+        self.assertIn("expected_sha256", workflow)
+        self.assertIn(
+            "refusing to publish an archive whose digest has not been reviewed",
+            workflow)
+
+    def test_the_shape_of_the_inputs_is_checked_before_they_are_used(self):
+        workflow = WORKFLOW.read_text(encoding="utf-8")
+        self.assertLess(workflow.index("- name: Validate source coordinates"),
+                        workflow.index("- name: Check out pinned whisper.cpp"))
+
+    def test_the_validator_checks_tar_links_before_extraction(self):
+        validator = (PACKAGING / "validate-package.sh").read_text(
+            encoding="utf-8")
+        for check in ("member.issym()", "member.islnk()", "member.isdev()"):
+            self.assertIn(check, validator)
+
+    @unittest.skipUnless(sys.platform == "linux" and shutil.which("bash"),
+                         "Linux packaging test is unavailable")
+    def test_the_validator_rejects_an_escaping_symlink(self):
+        asset = "whisper-bin-ubuntu-vulkan-x64"
+        with tempfile.TemporaryDirectory() as temporary:
+            output = pathlib.Path(temporary)
+            archive = output / f"{asset}.tar.gz"
+            with tarfile.open(archive, "w:gz") as bundle:
+                link = tarfile.TarInfo(f"{asset}/whisper-server")
+                link.type = tarfile.SYMTYPE
+                link.linkname = "/etc/passwd"
+                bundle.addfile(link, io.BytesIO())
+            digest = hashlib.sha256(archive.read_bytes()).hexdigest()
+            (output / f"{asset}.tar.gz.sha256").write_text(
+                f"{digest}  {asset}.tar.gz\n", encoding="utf-8",
+            )
+            checked = subprocess.run(
+                ["bash", PACKAGING / "validate-package.sh"],
+                env=os.environ | {"OUT_DIR": str(output)},
+                capture_output=True, text=True,
+            )
+        self.assertNotEqual(0, checked.returncode)
+        self.assertIn("unsafe symlink", checked.stderr)
+
+    def test_the_validator_checks_elf_architecture_dependencies_and_paths(self):
+        validator = (PACKAGING / "validate-package.sh").read_text(
+            encoding="utf-8")
+        for check in ("Advanced Micro Devices X86-64", "unexpected DT_NEEDED",
+                      "path.read_bytes()"):
+            self.assertIn(check, validator)
+
+    def test_the_builder_and_its_downloads_are_pinned(self):
+        dockerfile = (PACKAGING / "Dockerfile.build").read_text(
+            encoding="utf-8")
+        self.assertRegex(dockerfile, r"FROM ubuntu@sha256:[0-9a-f]{64}")
+        self.assertIn("CMAKE_SHA256=", dockerfile)
+        self.assertIn("libvulkan-dev=", dockerfile)
+        self.assertIn("shaderc=", dockerfile)
+        key = (PACKAGING / "lunarg-signing-key-pub.asc").read_bytes()
+        key = key.replace(b"\r\n", b"\n")
+        self.assertEqual(
+            "aa1c3c29673140e77f0d6a9aaeed5d9b5621e305ead51c59fae4458bbb4df92b",
+            hashlib.sha256(key).hexdigest(),
+        )
+
+    def test_the_bundle_has_portable_dynamic_backends(self):
+        script = (PACKAGING / "build-package.sh").read_text(
+            encoding="utf-8")
+        for flag in ("GGML_BACKEND_DL=ON", "GGML_CPU_ALL_VARIANTS=ON",
+                     "GGML_NATIVE=OFF", "GGML_OPENMP=OFF",
+                     "GGML_VULKAN=ON"):
+            self.assertIn(flag, script)
+        self.assertIn("libggml-cpu*.so", script)
+        self.assertIn("libggml-vulkan.so", script)
+
+    def test_the_dependency_release_matches_the_installer(self):
+        workflow = WORKFLOW.read_text(encoding="utf-8")
+        script = (PACKAGING / "build-package.sh").read_text(
+            encoding="utf-8")
+        self.assertEqual("whisper.cpp-v1.9.3",
+                         ggml.MANAGED_WHISPER_RELEASE)
+        self.assertEqual("v1.9.3", ggml.MANAGED_WHISPER_VERSION)
+        self.assertIn("RELEASE_TAG: whisper.cpp-v${{ inputs.whisper_version }}",
+                      workflow)
+        self.assertIn("WHISPER_VERSION:=1.9.3", script)
+        commit = "371b5a7561823ab2bb32142d2751e35e7534727b"
+        self.assertIn(f"WHISPER_COMMIT:={commit}", script)
+        self.assertIn(commit, workflow)
+        self.assertIn(ggml.MANAGED_WHISPER_VULKAN, workflow)
+        self.assertIn(ggml.MANAGED_WHISPER_SHA256, workflow)
+
+    def test_the_bundle_carries_metadata_and_all_required_licenses(self):
+        script = (PACKAGING / "build-package.sh").read_text(
+            encoding="utf-8")
+        for name in ("BUILD-INFO.json", "SHA256SUMS", ".cdx.json"):
+            self.assertIn(name, script)
+        for name in ("cpp-httplib-MIT.txt", "nlohmann-json-MIT.txt"):
+            self.assertTrue((PACKAGING / "licenses" / name).is_file())
+
+    def _make_test_sbom(self):
+        with tempfile.TemporaryDirectory() as temporary:
+            root = pathlib.Path(temporary)
+            (root / "whisper-server").write_bytes(b"elf")
+            sbom = root / "whisper-bin-ubuntu-vulkan-x64.cdx.json"
+            environment = os.environ | {
+                "ROOT": str(root),
+                "VERSION": "1.9.3",
+                "COMMIT": "371b5a7561823ab2bb32142d2751e35e7534727b",
+                "EPOCH": "1787219223",
+            }
+            with sbom.open("w", encoding="utf-8") as output:
+                subprocess.run(
+                    [sys.executable, PACKAGING / "make-sbom.py"],
+                    env=environment, stdout=output, check=True,
+                )
+            return json.loads(sbom.read_text(encoding="utf-8")), sbom.name
+
+    def test_the_sbom_does_not_record_the_file_being_written(self):
+        document, sbom_name = self._make_test_sbom()
+        names = {component["name"] for component in document["components"]}
+        self.assertNotIn(sbom_name, names)
+
+    def test_the_sbom_lists_ggml(self):
+        document, _ = self._make_test_sbom()
+        names = {component["name"] for component in document["components"]}
+        self.assertIn("ggml", names)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_paste.py

@@ -444,6 +444,45 @@ class MacOS(ClipboardContract, DikteTest):
         self.assertFalse(paste.paste_ready())
 
 
+class MacPasteGoesWhereTheDictationStarted(MacOS):
+    """The keys land in the frontmost window, so the front is what decides
+    where a transcript ends up."""
+
+    def setUp(self):
+        super().setUp()
+        from dikte import mac_window
+        self.mac_window = mac_window
+        self.activated = []
+        self.patch_attr(mac_window, "activate", self.activated.append)
+
+    def frontmost(self, dikte_is):
+        self.patch_attr(self.mac_window, "is_frontmost", lambda: dikte_is)
+
+    def test_a_dikte_that_took_the_front_hands_it_back_before_pressing(self):
+        self.frontmost(True)
+        paste.press("cmd+v", focus=4242)
+        self.assertEqual(self.activated, [4242])
+        self.assertEqual([event for _, event in self.api.posted], [1001, 1002])
+
+    def test_another_application_in_front_is_where_the_user_went_and_is_left(self):
+        self.frontmost(False)
+        paste.press("cmd+v", focus=4242)
+        self.assertEqual(self.activated, [])
+
+    def test_a_run_that_remembered_nobody_asks_nothing(self):
+        self.frontmost(True)
+        paste.press("cmd+v")
+        self.assertEqual(self.activated, [])
+
+    def test_the_front_is_handed_back_only_once_macos_trusts_dikte(self):
+        """Pulling the user out of their window and then failing to type would
+        be the worst of both."""
+        self.frontmost(True)
+        self.api.trusted = False
+        with self.assertRaises(paste.PasteError):
+            paste.press("cmd+v", focus=4242)
+        self.assertEqual(self.activated, [])
+
 class MacClipboardSnapshot(DikteTest):
     def test_every_native_type_is_restored_and_the_files_are_removed(self):
         directory = tempfile.mkdtemp(prefix="dikte-test-clipboard-")

tests/test_paths.py

@@ -42,9 +42,12 @@ class Directories(unittest.TestCase):
 
     def test_a_mac_does_not_read_the_xdg_variables(self):
         """A Mac with them set from some other tool still stores in one place."""
-        with mock.patch.dict(os.environ, {"XDG_CONFIG_HOME": "/c"}):
+        # Something no temporary directory can be called: the home this runs
+        # under is a mkdtemp path, and a two-letter needle matched the "/c" in
+        # somebody's TMPDIR rather than the variable being read.
+        with mock.patch.dict(os.environ, {"XDG_CONFIG_HOME": "/xdg-elsewhere"}):
             config_dir, _ = paths.directories("darwin")
-        self.assertNotIn("/c", config_dir.as_posix())
+        self.assertNotIn("xdg-elsewhere", config_dir.as_posix())
 
     def test_windows_keeps_the_models_out_of_the_roaming_profile(self):
         """Settings roam with the account; several gigabytes must not."""
@@ -61,6 +64,34 @@ class Directories(unittest.TestCase):
         self.assertTrue(data_dir.as_posix().endswith("/AppData/Local/Dikte"))
 
 
+class CacheDir(unittest.TestCase):
+    """The third place: files whose whole point is that they can be lost."""
+
+    def test_linux_follows_xdg(self):
+        with mock.patch.dict(os.environ, {"XDG_CACHE_HOME": "/k"}):
+            self.assertEqual(paths.cache_dir("linux").as_posix(), "/k/dikte")
+
+    def test_linux_without_the_variable_set(self):
+        with mock.patch.dict(os.environ, {}, clear=True):
+            self.assertTrue(paths.cache_dir("linux").as_posix()
+                            .endswith("/.cache/dikte"))
+
+    def test_a_mac_caches_under_library_caches(self):
+        """Where Time Machine already knows not to look."""
+        self.assertTrue(paths.cache_dir("darwin").as_posix()
+                        .endswith("/Library/Caches/Dikte"))
+
+    def test_windows_caches_outside_the_roaming_profile(self):
+        with mock.patch.dict(os.environ, {"LOCALAPPDATA": "C:/local"}):
+            self.assertEqual(paths.cache_dir("win32").as_posix(),
+                             "C:/local/Dikte/cache")
+
+    def test_windows_without_the_variable_set(self):
+        with mock.patch.dict(os.environ, {}, clear=True):
+            self.assertTrue(paths.cache_dir("win32").as_posix()
+                            .endswith("/AppData/Local/Dikte/cache"))
+
+
 class OnePlace(unittest.TestCase):
     """The programs and the models go where everything else goes.
 

tests/test_ui.py

@@ -6,8 +6,10 @@ save, so a setting added to one half and not the other is silently reset the
 next time anybody presses Save. That is the failure this catches.
 """
 
+import json
 import os
 import sys
+import time
 import unittest
 from typing import ClassVar
 from unittest import mock
@@ -21,13 +23,20 @@ from dikte import cleanup
 from dikte import config as cfg
 from dikte import ggml
 from dikte import hotkey
+from dikte import hub
 from dikte import ipc
 from dikte import overlay as overlay_module
 from dikte import paste
 from dikte import settings_ui
 from dikte import update
+from dikte.i18n import t
 from tests.support import DikteTest, only_these_tools
 
+# The harness below replaces this method on the class so that opening a window
+# in a test never calls anybody; taken here, before any test runs, so the two
+# tests about what it does when called still have the real one.
+REAL_LOAD_HOSTED_MODELS = settings_ui.SettingsWindow._load_hosted_models
+
 # One application for the whole run; Qt allows no second one.
 _app = QApplication.instance() or QApplication([])
 
@@ -42,6 +51,7 @@ CHANGED = {
     "paste_shortcut": "ctrl+shift+v",
     "restore_clipboard": True,
     "overlay_corner": "top-right",
+    "overlay_screen": "DP-1",
     "max_seconds": 120,
     "skip_silent": False,
     "silence_db": -42.0,
@@ -50,6 +60,7 @@ CHANGED = {
     "openai_api_key": "sk-test-key",
     "groq_api_key": "gsk-test-key",
     "openrouter_api_key": "sk-or-test-key",
+    "gemini_api_key": "AIza-test-key",
     "transcribe_provider": "openrouter",
     "transcribe_model": "whisper-1",
     "groq_transcribe_model": "whisper-large-v3",
@@ -59,6 +70,8 @@ CHANGED = {
     "cleanup_model": "some/other-model",
     "cleanup_claude_model": "opus",
     "cleanup_codex_model": "gpt-5",
+    "cleanup_gemini_model": "gemini-2.5-flash",
+    "cleanup_agy_model": "gemini-3.1-pro-low",
     "cleanup_reasoning": "high",
     "local_model": "ggml-small.bin",
     "local_gpu": False,
@@ -78,6 +91,7 @@ CHANGED = {
     "assistant_codex_model": "gpt-5",
     "assistant_codex_sandbox": "read-only",
     "assistant_openrouter_model": "some/agent-model",
+    "assistant_agy_model": "gemini-3.1-pro-low",
     "assistant_reasoning": "high",
     "assistant_dir": "/tmp",
     "assistant_timeout": 600,
@@ -133,6 +147,12 @@ class Settings(DikteTest):
                                             "_load_models"))
         self.enterContext(mock.patch.object(settings_ui.SettingsWindow,
                                             "_load_transcribe_models"))
+        self.enterContext(mock.patch.object(settings_ui.SettingsWindow,
+                                            "_load_codex_models"))
+        self.enterContext(mock.patch.object(settings_ui.SettingsWindow,
+                                            "_load_agy_models"))
+        self.enterContext(mock.patch.object(settings_ui.SettingsWindow,
+                                            "_load_hosted_models"))
         # The local model boxes fetch their own list the moment they are shown,
         # from a thread, which is nobody's test failing but a real request.
         self.enterContext(mock.patch.object(settings_ui.LocalModelBox,
@@ -159,7 +179,7 @@ class Settings(DikteTest):
     def test_the_window_opens_with_every_tab_on_it(self):
         window = self.window(cfg.Config())
         tabs = window.findChildren(settings_ui.QTabWidget)[0]
-        self.assertEqual(tabs.count(), 9)
+        self.assertEqual(tabs.count(), 10)
         self.assertEqual(window.windowTitle(), "Dikte Settings")
 
     def test_no_tab_can_stretch_the_window_past_a_small_screen(self):
@@ -240,10 +260,26 @@ class Settings(DikteTest):
             with self.subTest(key=key):
                 self.assertEqual(stored[key], value)
 
+    def test_only_a_save_that_changed_the_language_says_so(self):
+        # The owner answers language_changed by replacing the window, so a
+        # save that left the language alone must keep quiet.
+        i18n = settings_ui.i18n
+        self.addCleanup(i18n.set_language, i18n.language())
+        window = self.window(cfg.Config())
+        heard = []
+        window.language_changed.connect(lambda: heard.append(True))
+        window._save()
+        self.assertEqual(heard, [])
+        other = "en" if i18n.language() == "tr" else "tr"
+        window._select_data(window.ui_language, other)
+        window._save()
+        self.assertEqual(heard, [True])
+
     def test_the_model_box_on_screen_belongs_to_whoever_cleans_up(self):
         """An OpenRouter id and a Claude alias are not the same field."""
         window = self.window(cfg.Config())
         boxes = {"openrouter": window.cleanup_model_row,
+                 "opencode": window.cleanup_opencode_model_row,
                  "claude": window.cleanup_claude_model,
                  "codex": window.cleanup_codex_model}
         for provider, box in boxes.items():
@@ -254,6 +290,103 @@ class Settings(DikteTest):
                 self.assertEqual(shown, [provider])
                 self.assertFalse(box.isHidden())
 
+    def test_codex_answering_refills_both_of_its_boxes(self):
+        """The list Codex gave replaces the built-in one, in both places, and
+        neither loses what was already picked."""
+        conf = self.config(cleanup_codex_model="my-own-model")
+        window = self.window(conf)
+        window._on_codex_models_loaded(["gpt-6", "gpt-6-mini"])
+        for combo in (window.cleanup_codex_model, window.assistant_codex_model):
+            with self.subTest(combo=combo.objectName() or "combo"):
+                offered = [combo.itemText(i) for i in range(combo.count())]
+                self.assertEqual(offered[1:], ["gpt-6", "gpt-6-mini"])
+        self.assertEqual(window.cleanup_codex_model.currentText(),
+                         "my-own-model")
+
+    def test_opencode_answering_refills_both_of_its_boxes(self):
+        """The fetched catalog replaces the built-in list in the cleanup and
+        agent boxes alike, and neither loses what was picked."""
+        conf = self.config(cleanup_opencode_model="my-own-model")
+        window = self.window(conf)
+        window._on_opencode_models_loaded(["glm-9", "kimi-k9"], "")
+        for combo in (window.cleanup_opencode_model,
+                      window.assistant_opencode_model):
+            with self.subTest(combo=combo.objectName() or "combo"):
+                offered = [combo.itemText(i) for i in range(combo.count())]
+                self.assertEqual(offered, ["glm-9", "kimi-k9"])
+        self.assertEqual(window.cleanup_opencode_model.currentText(),
+                         "my-own-model")
+
+    def test_opencode_s_list_arriving_at_open_leaves_the_other_boxes_alone(self):
+        conf = self.config(cleanup_opencode_model="my-own-model",
+                           meeting_model="some/meeting-model")
+        window = self.window(conf)
+        before = [window.meeting_model.itemText(i)
+                  for i in range(window.meeting_model.count())]
+        window._on_hosted_models_loaded("opencode", ["glm-5.3", "kimi-k3"])
+        combo = window.cleanup_opencode_model
+        offered = [combo.itemText(i) for i in range(combo.count())]
+        self.assertEqual(offered, ["glm-5.3", "kimi-k3"])
+        self.assertEqual(combo.currentText(), "my-own-model")
+        self.assertEqual([window.meeting_model.itemText(i)
+                          for i in range(window.meeting_model.count())], before)
+
+    def test_opencode_cleanup_offers_a_fetch_button_of_its_own(self):
+        """The OpenRouter button leaves the screen with its box, so OpenCode Go
+        carries its own."""
+        window = self.window(cfg.Config())
+        window._select_data(window.cleanup_provider, "opencode")
+        self.assertFalse(window.cleanup_opencode_model_row.isHidden())
+        self.assertTrue(window.cleanup_model_row.isHidden())
+
+    def test_agy_answering_refills_both_of_its_boxes(self):
+        """The same arrangement as Codex: both boxes, nothing chosen is lost."""
+        conf = self.config(cleanup_agy_model="my-own-model")
+        window = self.window(conf)
+        window._on_agy_models_loaded(["gemini-4-flash-low", "gemini-4-pro-low"])
+        for combo in (window.cleanup_agy_model, window.assistant_agy_model):
+            with self.subTest(combo=combo.objectName() or "combo"):
+                offered = [combo.itemText(i) for i in range(combo.count())]
+                self.assertEqual(offered[1:],
+                                 ["gemini-4-flash-low", "gemini-4-pro-low"])
+        self.assertEqual(window.cleanup_agy_model.currentText(), "my-own-model")
+
+    def test_openrouter_s_list_arriving_at_open_refills_cleanup_and_meetings(self):
+        conf = self.config(cleanup_model="my/own-model")
+        window = self.window(conf)
+        window._on_hosted_models_loaded("openrouter", ["a/one", "b/two"])
+        for combo in (window.cleanup_model, window.meeting_model):
+            with self.subTest(combo=combo.objectName() or "combo"):
+                offered = [combo.itemText(i) for i in range(combo.count())]
+                self.assertEqual(offered, ["a/one", "b/two"])
+        self.assertEqual(window.cleanup_model.currentText(), "my/own-model")
+
+    def test_google_s_list_arriving_at_open_refills_its_own_box_only(self):
+        window = self.window(self.config(cleanup_gemini_model="gemini-x"))
+        before = window.cleanup_model.count()
+        window._on_hosted_models_loaded("gemini", ["gemini-4-flash"])
+        offered = [window.cleanup_gemini_model.itemText(i)
+                   for i in range(window.cleanup_gemini_model.count())]
+        self.assertEqual(offered, ["gemini-4-flash"])
+        self.assertEqual(window.cleanup_gemini_model.currentText(), "gemini-x")
+        self.assertEqual(window.cleanup_model.count(), before)
+
+    def test_no_key_no_call_home_at_open(self):
+        """Opening Settings is not consent to be talked about to two vendors."""
+        window = self.window(self.config())
+        with mock.patch.dict(os.environ, {}, clear=True), \
+                mock.patch.object(settings_ui.threading, "Thread") as thread:
+            REAL_LOAD_HOSTED_MODELS(window)
+        thread.assert_not_called()
+
+    def test_a_key_on_file_is_fetched_with_at_open(self):
+        window = self.window(self.config(openrouter_api_key="sk-or-x",
+                                         gemini_api_key="AIza-x",
+                                         opencode_api_key="opencode-x"))
+        with mock.patch.object(settings_ui.threading, "Thread") as thread:
+            REAL_LOAD_HOSTED_MODELS(window)
+        self.assertEqual(thread.call_count, 3)
+
     def test_the_update_line_names_the_version_that_is_running(self):
         window = self.window(cfg.Config())
         self.assertIn(settings_ui.__version__, window.update_status.text())
@@ -375,6 +508,20 @@ class Settings(DikteTest):
         self.assertEqual(conf["transcribe_model"], "gpt-4o-transcribe")
         self.assertEqual(conf["groq_transcribe_model"], "whisper-large-v3")
 
+    def test_the_file_model_is_saved_and_only_shown_for_openrouter(self):
+        self.write_config({"transcribe_provider": "openrouter",
+                           "openrouter_file_model": "openai/whisper-large-v3"})
+        conf = cfg.Config()
+        window = self.window(conf)
+        self.assertEqual(window.file_model.currentText(), "openai/whisper-large-v3")
+        self.assertTrue(window.stt_form.isRowVisible(window.file_model_row))
+        window.file_model.setCurrentText(" deepgram/nova-3 ")
+        window._save()
+        self.assertEqual(conf["openrouter_file_model"], "deepgram/nova-3")
+        window.transcribe_provider.setCurrentIndex(
+            window.transcribe_provider.findData("openai"))
+        self.assertFalse(window.stt_form.isRowVisible(window.file_model_row))
+
     def test_the_provider_box_offers_every_provider_config_knows(self):
         window = self.window(cfg.Config())
         offered = [window.transcribe_provider.itemData(i)
@@ -518,6 +665,321 @@ class KdeSettings(Settings):
         self.assertFalse(window.evdev_enabled.isHidden())
 
 
+
+
+class FakeAppKit:
+    """The Objective-C runtime, answering rather than being asked.
+
+    Stands in for what mac_window._appkit() loads, so what the indicator's
+    window would have been sent can be read off `told` on any machine. The
+    selectors come back as the names they were registered under, which is what
+    lets the tests below name the message they mean.
+    """
+
+    def __init__(self, window=4242, panel=True, mask=0xe):
+        self.objc = self
+        self.window, self.panel, self.mask = window, panel, mask
+        self.told = {}
+
+    def objc_getClass(self, name):
+        return 1 if name == b"NSPanel" else 0
+
+    def selector(self, name):
+        return name.decode()
+
+    def shared(self, _class_name, _selector):
+        return 1
+
+    def ask(self, _to, selector):
+        return self.window if selector == "window" else 0
+
+    def ask_unsigned(self, _to, _selector):
+        return self.mask
+
+    def ask_of_class(self, _to, _selector, _klass):
+        return self.panel
+
+    def tell_bool(self, _to, selector, value):
+        self.told[selector] = value
+
+    tell_unsigned = tell_bool
+
+
+class MacIndicatorWindow(DikteTest):
+    """Which of the three AppKit settings the indicator's window is sent.
+
+    The nonactivating bit is the one worth a test of its own: it is legal on an
+    NSPanel and nowhere else, and sending it to a plain NSWindow raises an
+    Objective-C exception, which through ctypes is not something Python can
+    catch. It takes the whole process down. `_is_panel` is the only thing
+    standing between the two, so what it answers has to decide what is sent.
+    """
+
+    def setUp(self):
+        super().setUp()
+        from dikte import mac_window
+        self.mac_window = mac_window
+        self.patch_attr(mac_window.QGuiApplication, "platformName",
+                        staticmethod(lambda: "cocoa"))
+
+    def told(self, **kwargs):
+        """What keep_on_screen sends an indicator, against this AppKit."""
+        appkit = FakeAppKit(**kwargs)
+        self.patch_attr(self.mac_window, "_appkit", lambda: appkit)
+        widget = overlay_module.Overlay()
+        self.addCleanup(widget.deleteLater)
+        self.addCleanup(widget.close)
+        self.answered = self.mac_window.keep_on_screen(widget)
+        return appkit.told
+
+    def test_a_window_that_is_not_a_panel_is_never_sent_the_style_mask(self):
+        """The message that would kill the process. The other two still go."""
+        told = self.told(panel=False)
+        self.assertTrue(self.answered)
+        self.assertNotIn("setStyleMask:", told)
+        self.assertIs(told["setHidesOnDeactivate:"], False)
+        self.assertEqual(told["setCollectionBehavior:"],
+                         self.mac_window.BEHAVIOUR)
+
+    def test_a_panel_without_the_bit_is_sent_the_mask_with_it_added(self):
+        told = self.told(panel=True, mask=0xe)
+        self.assertEqual(told["setStyleMask:"],
+                         0xe | self.mac_window.NONACTIVATING_PANEL)
+
+    def test_a_panel_that_already_has_the_bit_is_left_alone(self):
+        """Every dictation runs this again, and a mask Cocoa did not need is a
+        window it rebuilds underneath the indicator."""
+        told = self.told(panel=True,
+                         mask=0xe | self.mac_window.NONACTIVATING_PANEL)
+        self.assertNotIn("setStyleMask:", told)
+
+    def test_a_window_the_view_does_not_have_yet_is_not_messaged(self):
+        self.assertEqual(self.told(window=0), {})
+        self.assertFalse(self.answered)
+
+    def test_nothing_is_sent_anywhere_but_cocoa(self):
+        """Every other platform hands out a winId that means something else
+        entirely, and messaging it crashes the run."""
+        self.patch_attr(self.mac_window.QGuiApplication, "platformName",
+                        staticmethod(lambda: "offscreen"))
+        self.assertEqual(self.told(), {})
+        self.assertFalse(self.answered)
+
+
+class GivingTheFrontBack(DikteTest):
+    """Opening the microphone brings Dikte to the front, and the window the
+    user was dictating into goes inactive with the caret in it. Nothing can be
+    asked of the capture session, so the front is put back afterwards.
+
+    The watch is driven by hand here: what matters is what it decides, not how
+    long Qt takes to tick.
+    """
+
+    def setUp(self):
+        super().setUp()
+        from dikte import app as dikte_module
+        from dikte import mac_window
+        self.dikte = dikte_module
+        self.activated = []
+        self.patch_attr(mac_window, "activate", self.activate)
+        self.ticks = []
+        outer = self
+
+        class FakeTimer:
+            """Records what it was asked to do and hands over the tick."""
+
+            def __init__(self, _parent):
+                self.interval = None
+                self.running = False
+                outer.ticks.append(self)
+
+            def setInterval(self, milliseconds):
+                self.interval = milliseconds
+
+            def start(self):
+                self.running = True
+
+            def stop(self):
+                self.running = False
+
+            @property
+            def timeout(self):
+                return self
+
+            def connect(self, slot):
+                self.tick = slot
+
+        self.patch_attr(dikte_module, "QTimer", FakeTimer)
+
+        class BareDikte:
+            """As much of the application as this one method touches."""
+
+            app = None
+            _front_watch = None
+            _the_front = dikte_module.Dikte._the_front
+            _give_the_front_back = dikte_module.Dikte._give_the_front_back
+            _stop_watching_the_front = dikte_module.Dikte._stop_watching_the_front
+
+        self.bare = BareDikte
+
+    def activate(self, pid):
+        self.activated.append(pid)
+        return True
+
+    def watching(self, was_in_front, dikte_in_front, on=None):
+        from dikte import mac_window
+        self.patch_attr(mac_window, "is_frontmost", lambda: dikte_in_front)
+        dikte = on if on is not None else self.bare()
+        dikte._give_the_front_back(was_in_front)
+        return self.ticks[-1] if self.ticks else None
+
+    def test_the_front_goes_back_to_whoever_had_it(self):
+        watch = self.watching(4242, dikte_in_front=True)
+        watch.tick()
+        self.assertEqual(self.activated, [4242])
+        self.assertTrue(watch.running)    # accepted is not the same as landed
+        self.patch_attr(self.mac_window_module(), "is_frontmost", lambda: False)
+        watch.tick()
+        self.assertFalse(watch.running)
+
+    def test_an_accepted_restore_is_not_sent_again_while_it_is_landing(self):
+        watch = self.watching(4242, dikte_in_front=True)
+        watch.tick()
+        watch.tick()
+        self.assertEqual(self.activated, [4242])
+        self.assertTrue(watch.running)
+
+    def test_an_accepted_restore_that_never_lands_still_times_out(self):
+        watch = self.watching(4242, dikte_in_front=True)
+        watch.tick()
+        with mock.patch.object(self.dikte.time, "monotonic",
+                               return_value=self.dikte.time.monotonic() + 60):
+            watch.tick()
+        self.assertFalse(watch.running)
+        self.assertEqual(self.activated, [4242])
+
+    def test_a_restore_the_system_refused_is_retried(self):
+        from dikte import mac_window
+        self.patch_attr(mac_window, "activate",
+                        lambda pid: self.activated.append(pid) or False)
+        watch = self.watching(4242, dikte_in_front=True)
+        watch.tick()
+        watch.tick()
+        self.assertEqual(self.activated, [4242, 4242])
+        self.assertTrue(watch.running)
+
+    def test_a_front_that_was_never_taken_is_left_where_it_is(self):
+        """The microphone does not always take it, and pulling an application
+        forward that is already there is one flicker for nothing."""
+        watch = self.watching(4242, dikte_in_front=False)
+        watch.tick()
+        self.assertEqual(self.activated, [])
+        self.assertTrue(watch.running)    # still waiting for the moment
+
+    def test_it_gives_up_rather_than_watching_for_ever(self):
+        watch = self.watching(4242, dikte_in_front=False)
+        with mock.patch.object(self.dikte.time, "monotonic",
+                               return_value=self.dikte.time.monotonic() + 60):
+            watch.tick()
+        self.assertFalse(watch.running)
+        self.assertEqual(self.activated, [])
+
+    def test_a_dictation_started_in_dikte_itself_watches_nothing(self):
+        """Settings is a window of ours, and the front is already where it
+        belongs."""
+        self.assertIsNone(self.watching(os.getpid(), dikte_in_front=True))
+
+    def test_a_second_recording_calls_off_the_watch_the_first_one_left(self):
+        """The older watch remembers where the older recording started, and by
+        now that is the wrong window to be pulling forward."""
+        dikte = self.bare()
+        first = self.watching(4242, dikte_in_front=False, on=dikte)
+        second = self.watching(1111, dikte_in_front=False, on=dikte)
+        self.assertFalse(first.running)
+        self.assertTrue(second.running)
+        second.tick()                      # and the survivor is the new one
+        self.patch_attr(self.mac_window_module(), "is_frontmost", lambda: True)
+        second.tick()
+        self.assertEqual(self.activated, [1111])
+
+    def test_a_recording_nobody_needs_watching_for_still_calls_off_the_old_one(self):
+        """Starting the next one from Dikte's own window is not a reason to
+        leave the last one's watch running."""
+        dikte = self.bare()
+        first = self.watching(4242, dikte_in_front=False, on=dikte)
+        self.watching(os.getpid(), dikte_in_front=False, on=dikte)
+        self.assertFalse(first.running)
+        self.assertIsNone(dikte._front_watch)
+
+    def test_nobody_in_front_is_nobody_to_go_back_to(self):
+        self.assertIsNone(self.watching(None, dikte_in_front=True))
+
+    def mac_window_module(self):
+        from dikte import mac_window
+        return mac_window
+
+
+class EveryRecordingProtectsTheFront(DikteTest):
+    """Three ways in, dictation, agent and meeting, and all three open the same
+    avfoundation capture, so all three take the front the same way. What is
+    checked here is the order: the front has to be noted before the microphone
+    is opened, and the watch armed after, or there is nothing to go back to.
+    """
+
+    def setUp(self):
+        super().setUp()
+        from dikte import app as dikte_module
+        self.dikte = dikte_module
+        self.order = []
+
+    def app(self, **attributes):
+        """A Dikte that records the order it does things in, and nothing else.
+
+        The methods under test are called unbound against it, so the stand-ins
+        go on the object rather than on the class.
+        """
+        dikte = mock.Mock(**attributes)
+        dikte._run_id = 0
+        dikte.conf = {"mic_target": "", "max_seconds": 60,
+                      "meeting_mic_target": "", "meeting_system_target": "",
+                      "meeting_max_seconds": 60}
+        dikte._the_front.side_effect = lambda: self.order.append("noted") or 4242
+        dikte._give_the_front_back.side_effect = (
+            lambda pid: self.order.append(f"watching {pid}"))
+        dikte._begin_recording = (
+            lambda owner: self.dikte.Dikte._begin_recording(dikte, owner))
+        dikte.recorder.start.side_effect = (
+            lambda *_a: self.order.append("microphone"))
+        dikte.meeting_recorder.start.side_effect = (
+            lambda *_a: self.order.append("microphone"))
+        return dikte
+
+    def test_a_dictation_notes_the_front_before_the_indicator_is_even_shown(self):
+        dikte = self.app(state=self.dikte.IDLE, recording=False)
+        dikte.overlay.show_recording.side_effect = (
+            lambda *_a: self.order.append("indicator"))
+        self.dikte.Dikte.start(dikte)
+        self.assertEqual(self.order,
+                         ["noted", "indicator", "microphone", "watching 4242"])
+
+    def test_the_agent_does_the_same(self):
+        dikte = self.app(ask_state=self.dikte.IDLE, recording=False)
+        self.dikte.Dikte.start_ask(dikte)
+        self.assertEqual(self.order, ["noted", "microphone", "watching 4242"])
+
+    def test_a_meeting_does_the_same(self):
+        """The one most worth protecting: the user is in a call."""
+        dikte = self.app(meeting_state=self.dikte.M_IDLE)
+        dikte.meeting_recorder.active = True
+        self.dikte.Dikte.start_meeting(dikte)
+        self.assertEqual(self.order, ["noted", "microphone", "watching 4242"])
+
+    def test_a_meeting_whose_microphone_never_opened_watches_nothing(self):
+        dikte = self.app(meeting_state=self.dikte.M_IDLE)
+        dikte.meeting_recorder.active = False
+        self.dikte.Dikte.start_meeting(dikte)
+        self.assertEqual(self.order, ["noted", "microphone"])
+
 class Overlay(DikteTest):
     def overlay(self, **kwargs):
         widget = overlay_module.Overlay(**kwargs)
@@ -596,6 +1058,17 @@ class Overlay(DikteTest):
                 widget.show_recording()
                 widget._reposition()
 
+    def test_a_named_screen_is_used_instead_of_the_pointer_screen(self):
+        screen = mock.Mock()
+        screen.name.return_value = "DP-1"
+        screen.availableGeometry.return_value = settings_ui.QRect(1920, 0, 1920, 1080)
+        widget = self.overlay(screen_name="DP-1")
+        with mock.patch.object(QApplication, "screens", return_value=[screen]), \
+                mock.patch.object(QApplication, "screenAt") as screen_at:
+            widget._reposition()
+        screen_at.assert_not_called()
+        self.assertEqual(widget.pos(), QPoint(1948, 995))
+
     def test_a_warning_and_an_error_both_show(self):
         widget = self.overlay()
         widget.show_warning("cleanup failed")
@@ -654,7 +1127,13 @@ class MeetingSources(DikteTest):
                 only_these_tools(), \
                 mock.patch.object(settings_ui.SettingsWindow, "_load_models"), \
                 mock.patch.object(settings_ui.SettingsWindow,
-                                  "_load_transcribe_models"):
+                                  "_load_transcribe_models"), \
+                mock.patch.object(settings_ui.SettingsWindow,
+                                  "_load_codex_models"), \
+                mock.patch.object(settings_ui.SettingsWindow,
+                                  "_load_agy_models"), \
+                mock.patch.object(settings_ui.SettingsWindow,
+                                  "_load_hosted_models"):
             window = settings_ui.SettingsWindow(cfg.Config())
         self.addCleanup(window.deleteLater)
         self.addCleanup(window.close)
@@ -683,6 +1162,13 @@ class LocalModels(DikteTest):
         # "nothing can transcribe" question from its real binary and model.
         self.patch_attr(ggml, "BIN_DIR", self.path("bin"))
         self.patch_attr(ggml, "MODELS_DIR", self.path("models"))
+        # And one with Codex on it would ask it for its model list.
+        self.enterContext(mock.patch.object(settings_ui.SettingsWindow,
+                                            "_load_codex_models"))
+        self.enterContext(mock.patch.object(settings_ui.SettingsWindow,
+                                            "_load_agy_models"))
+        self.enterContext(mock.patch.object(settings_ui.SettingsWindow,
+                                            "_load_hosted_models"))
 
     def window(self, conf):
         window = settings_ui.SettingsWindow(conf)
@@ -719,12 +1205,21 @@ class LocalModels(DikteTest):
         # Qt's int is C++'s 32-bit one, and a 2.3 GB model is more than fits in
         # it: the count came out the far side negative, at "-1%".
         box = self.window(cfg.Config()).local_llm
-        box._downloading = True
-        box._report(1_048_576, 2_489_757_856)
+        box._report("model", 1_048_576, 2_489_757_856)
         _app.processEvents()
         self.assertIn("2.3 GB", box.status.text())
         self.assertNotIn("-", box.status.text())
 
+    def test_each_download_reports_into_its_own_label(self):
+        """The two can run at once; the tag, not a flag read later, says
+        which label the bytes belong to."""
+        box = self.window(cfg.Config()).local_llm
+        box._report("program", 10, 100)
+        box._report("model", 20, 100)
+        _app.processEvents()
+        self.assertIn("10", box.program_label.text())
+        self.assertIn("20", box.status.text())
+
     def test_a_long_model_name_is_not_cut_in_half(self):
         # The list under a combo box takes the box's width and elides what does
         # not fit, in the middle: "ggml-org/Qwen....7B-Base-GGUF".
@@ -737,6 +1232,319 @@ class LocalModels(DikteTest):
                      for row in range(box.repo.count()))
         self.assertGreaterEqual(view.minimumWidth(), widest)
 
+    @staticmethod
+    def _item(name, size=1 << 20):
+        return hub.Item(name, f"https://example.invalid/{name}", size, "")
+
+    @staticmethod
+    def _rows(box):
+        """Every row's text, headings included."""
+        return [box.model.itemText(row) for row in range(box.model.count())]
+
+    @staticmethod
+    def _repos(box):
+        return [box.repo.itemText(row) for row in range(box.repo.count())]
+
+    @staticmethod
+    def _roomy():
+        """Stand on a machine with room for every suggestion.
+
+        The order the publishers come in follows the memory, so a test that
+        reads it has to say which machine it is standing on. A build runner
+        with 7 GB in it puts the two Gemma 4 rows last and is right to.
+        """
+        return mock.patch.object(ggml, "total_memory", return_value=64 << 30)
+
+    @staticmethod
+    def _offered(box):
+        """The model names in the box, headings and duplicates left out."""
+        names = []
+        for row in range(box.model.count()):
+            name = box.model.itemData(row)
+            if name and name not in names:
+                names.append(name)
+        return names
+
+    def test_a_row_with_nothing_to_fetch_does_not_offer_a_download(self):
+        # The model the settings name is not in the list any more, so its row
+        # was rebuilt from the name alone and carries no file to fetch. The
+        # button stayed lit and the press did nothing at all.
+        box = self.window(self.config(local_llm_model="gone.gguf")).local_llm
+        box.load("gone.gguf", "ggml-org/SmolLM3-3B-GGUF")
+        self.assertEqual(box.selected(), "gone.gguf")
+        self.assertFalse(box.download_button.isEnabled())
+        self.assertIn("gone.gguf", box.status.text())
+        self.assertIn("publisher", box.status.text())
+
+    def test_a_model_without_its_program_does_not_say_it_is_ready(self):
+        # The model runs on the program above it, and "Ready" over a missing
+        # one is what had people asking why nothing transcribed.
+        box = self.window(cfg.Config()).local_whisper
+        path = ggml.whisper_model_path("ggml-small.bin")
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_bytes(b"not really a model")
+        box.load("ggml-small.bin")
+        self.assertFalse(ggml.program_path(ggml.WHISPER))
+        self.assertNotIn("Ready", box.status.text())
+        self.assertIn("program", box.status.text())
+
+    def test_changing_the_publisher_changes_the_model(self):
+        # The model chosen under the old publisher is not published by the new
+        # one. Carried over, it was added back as "not downloaded" and selected
+        # again, and the box looked as though the change had not taken.
+        box = self.window(self.config(local_llm_model="gemma-3-4b-it-Q4_K_M.gguf",
+                                      local_llm_repo="ggml-org/gemma-3-4b-it-GGUF")).local_llm
+        box.load("gemma-3-4b-it-Q4_K_M.gguf", "ggml-org/gemma-3-4b-it-GGUF")
+        box.repo.blockSignals(True)
+        box.repo.setCurrentText("ggml-org/SmolLM3-3B-GGUF")
+        box.repo.blockSignals(False)
+        box._on_listed([("models", [self._item("SmolLM3-Q4_K_M.gguf")],
+                         "ggml-org/SmolLM3-3B-GGUF")], "")
+        self.assertEqual(box.selected(), "SmolLM3-Q4_K_M.gguf")
+        self.assertEqual(self._offered(box), ["SmolLM3-Q4_K_M.gguf"])
+
+    def test_a_list_for_a_publisher_that_is_no_longer_chosen_is_dropped(self):
+        # Every change starts its own request, and they do not come back in the
+        # order they went out.
+        box = self.window(cfg.Config()).local_llm
+        box.load("", "ggml-org/SmolLM3-3B-GGUF")
+        box.repo.blockSignals(True)
+        box.repo.setCurrentText("ggml-org/SmolLM3-3B-GGUF")
+        box.repo.blockSignals(False)
+        box._on_listed([("models", [self._item("SmolLM3-Q4_K_M.gguf")],
+                         "ggml-org/SmolLM3-3B-GGUF")], "")
+        box._on_listed([("models", [self._item("gemma-3-4b-it-Q4_K_M.gguf")],
+                         "ggml-org/gemma-3-4b-it-GGUF")], "")
+        self.assertEqual(box.selected(), "SmolLM3-Q4_K_M.gguf")
+
+    def test_the_publisher_box_is_not_asked_on_every_keystroke(self):
+        box = self.window(cfg.Config()).local_llm
+        with mock.patch.object(box, "_fetch_models") as fetch:
+            for text in ("g", "gg", "ggm", "ggml-org/SmolLM3-3B-GGUF"):
+                box.repo.setCurrentText(text)
+            fetch.assert_not_called()
+            box._later.setInterval(0)
+            box._later.start()
+            _app.processEvents()
+            time.sleep(0.05)
+            _app.processEvents()
+        self.assertEqual(fetch.call_count, 1)
+    def test_the_models_are_grouped_by_the_model_rather_than_by_size(self):
+        # Sorted by size alone, the turbo files land between the two medium
+        # ones, half a screen from the model they are a copy of.
+        box = self.window(cfg.Config()).local_whisper
+        with mock.patch.object(ggml, "total_memory", return_value=8 << 30), \
+                mock.patch.object(ggml, "accelerator", return_value=""):
+            box._on_listed([("models", [
+                self._item("ggml-medium-q5_0.bin", 539 << 20),
+                self._item("ggml-large-v3-turbo-q5_0.bin", 574 << 20),
+                self._item("ggml-medium-q8_0.bin", 823 << 20),
+                self._item("ggml-large-v3-turbo.bin", 1624 << 20),
+            ], "")], "")
+        rows = self._rows(box)
+        # The two medium files under one heading, the two turbo ones under
+        # theirs, and the model rather than the file deciding the order.
+        self.assertEqual(rows[rows.index("medium"):],
+                         ["medium",
+                          "ggml-medium-q5_0.bin  (539.0 MB, 5-bit)",
+                          "ggml-medium-q8_0.bin  (823.0 MB, 8-bit)",
+                          "large-v3-turbo",
+                          "ggml-large-v3-turbo-q5_0.bin  "
+                          "(574.0 MB, 5-bit, recommended)",
+                          "ggml-large-v3-turbo.bin  (1.6 GB, 16-bit)"])
+        # A heading is not a model, and nothing can be saved from one.
+        self.assertIsNone(box.model.itemData(rows.index("medium")))
+
+    def test_the_row_for_this_machine_is_on_top_and_says_so(self):
+        box = self.window(cfg.Config()).local_whisper
+        with mock.patch.object(ggml, "total_memory", return_value=8 << 30), \
+                mock.patch.object(ggml, "accelerator", return_value=""):
+            box._on_listed([("models", [
+                self._item("ggml-tiny.bin", 77 << 20),
+                self._item("ggml-large-v3-turbo-q5_0.bin", 574 << 20),
+            ], "")], "")
+        self.assertEqual(box.selected(), "ggml-large-v3-turbo-q5_0.bin")
+        self.assertEqual(box.model.itemData(1), "ggml-large-v3-turbo-q5_0.bin")
+        self.assertIn(t("recommended"), box.model.itemText(1))
+
+    def test_a_model_the_memory_cannot_hold_says_so_on_its_row(self):
+        box = self.window(cfg.Config()).local_llm
+        box.repo.blockSignals(True)
+        box.repo.setCurrentText("ggml-org/x-GGUF")
+        box.repo.blockSignals(False)
+        with mock.patch.object(ggml, "total_memory", return_value=8 << 30):
+            box._on_listed([("models", [
+                self._item("small-Q4_0.gguf", 1 << 30),
+                self._item("huge-Q8_0.gguf", 12 << 30),
+            ], "ggml-org/x-GGUF")], "")
+        rows = {box.model.itemData(row): box.model.itemText(row)
+                for row in range(box.model.count())}
+        self.assertNotIn(t("too big for this machine"), rows["small-Q4_0.gguf"])
+        self.assertIn(t("too big for this machine"), rows["huge-Q8_0.gguf"])
+
+    def test_a_recommended_row_is_not_listed_twice_after_a_download(self):
+        # It has a row of its own on top as well as one in its group, and
+        # reading the rows back the way a finished download does was doubling
+        # it in the list every time.
+        box = self.window(cfg.Config()).local_whisper
+        with self._roomy():
+            box._on_listed([("models", [
+                self._item("ggml-tiny.bin", 77 << 20),
+                self._item("ggml-large-v3-turbo-q5_0.bin", 574 << 20),
+            ], "")], "")
+            before = self._offered(box)
+            box._fill_models_from_current()
+        self.assertEqual(self._offered(box), before)
+        names = [box.model.itemData(row) for row in range(box.model.count())]
+        self.assertEqual(len([n for n in names if n]), len(before) + 1)
+
+    def test_a_processor_build_is_not_recommended_the_accurate_model(self):
+        # The Vulkan loader is on the machine but what was installed is the
+        # processor build, so there is no card in play whatever the loader
+        # says, and a 1 GB model on a processor is a wait somebody is sitting
+        # through with a sentence half typed.
+        binary = self.path("bin/whisper/v1.9.3/whisper-server")
+        binary.parent.mkdir(parents=True)
+        binary.write_text("")
+        binary.chmod(0o755)
+        self.path("bin/whisper/installed.json").write_text(json.dumps(
+            {"tag": "v1.9.3", "binary": str(binary), "backend": "processor"}))
+        self.patch_attr(ggml.shutil, "which", lambda name: None)
+        box = self.window(cfg.Config()).local_whisper
+        with mock.patch.object(ggml, "total_memory", return_value=32 << 30), \
+                mock.patch.object(ggml, "accelerator", return_value="Vulkan"):
+            self.assertEqual(box._suggested(), ggml.SUGGESTED_WHISPER)
+
+    def test_a_publisher_with_nothing_to_offer_says_why(self):
+        # Half of what ggml-org publishes is split across files or past the
+        # size cap, and an empty box read as though the click had not landed.
+        box = self.window(cfg.Config()).local_llm
+        box.repo.blockSignals(True)
+        box.repo.setCurrentText("ggml-org/gpt-oss-120b-GGUF")
+        box.repo.blockSignals(False)
+        box._on_listed([("models", [], "ggml-org/gpt-oss-120b-GGUF")], "")
+        self.assertIn("ggml-org/gpt-oss-120b-GGUF", box.status.text())
+        self.assertIn("publisher", box.status.text())
+
+    def test_an_empty_box_nobody_has_asked_yet_is_not_a_publisher_fault(self):
+        box = self.window(cfg.Config()).local_llm
+        box.load("", "ggml-org/SmolLM3-3B-GGUF")
+        self.assertNotIn("publisher", box.status.text())
+
+    def test_only_the_suggested_publishers_are_offered_to_start_with(self):
+        # Forty repository ids is not a choice anybody can make.
+        box = self.window(cfg.Config()).local_llm
+        with self._roomy():
+            box._on_listed([("repos", [ggml.SUGGESTED_LLM[0],
+                                       "ggml-org/something-else-GGUF"], "")], "")
+            self.assertEqual(self._repos(box), list(ggml.SUGGESTED_LLM))
+
+    def test_a_suggestion_missing_from_the_listing_is_still_offered(self):
+        # The listing is the forty repositories touched most recently, and a
+        # publisher that has not been updated in a season falls off it while
+        # still being the one to point at.
+        box = self.window(cfg.Config()).local_llm
+        box._on_listed([("repos", ["ggml-org/something-else-GGUF"], "")], "")
+        self.assertIn(ggml.SUGGESTED_LLM[0], self._repos(box))
+
+    def test_the_switch_brings_the_rest_and_keeps_them_apart(self):
+        box = self.window(cfg.Config()).local_llm
+        with self._roomy():
+            box._on_listed([("repos", [ggml.SUGGESTED_LLM[0],
+                                       "ggml-org/something-else-GGUF"], "")], "")
+            box.every_repo.setChecked(True)
+            rows = self._repos(box)
+        self.assertEqual(rows[:len(ggml.SUGGESTED_LLM)],
+                         list(ggml.SUGGESTED_LLM))
+        # A separator rather than a heading: the box is typed into as well as
+        # chosen from, and a heading would land in the field as a repository.
+        self.assertEqual(rows[len(ggml.SUGGESTED_LLM)], "")
+        self.assertEqual(rows[-1], "ggml-org/something-else-GGUF")
+
+    def test_a_publisher_typed_in_is_not_dropped_by_the_next_fetch(self):
+        box = self.window(cfg.Config()).local_llm
+        box.repo.blockSignals(True)
+        box.repo.setCurrentText("ggml-org/something-else-GGUF")
+        box.repo.blockSignals(False)
+        box._on_listed([("repos", [ggml.SUGGESTED_LLM[0],
+                                   "ggml-org/something-else-GGUF"], "")], "")
+        self.assertFalse(box.every_repo.isChecked())
+        self.assertIn("ggml-org/something-else-GGUF", self._repos(box))
+        self.assertEqual(box.repository(), "ggml-org/something-else-GGUF")
+
+    def test_the_chosen_publisher_is_said_in_words(self):
+        # A repository id names the publisher, the parameter count and the
+        # shape of the weights, and none of that says whether to click it.
+        box = self.window(cfg.Config()).local_llm
+        box.repo.setCurrentText(ggml.SUGGESTED_LLM[0])
+        self.assertTrue(box.repo_note.text())
+        box.repo.setCurrentText("ggml-org/nobody-wrote-a-note-GGUF")
+        self.assertEqual(box.repo_note.text(), "")
+
+    def test_the_box_says_what_this_machine_will_run_on(self):
+        box = self.window(cfg.Config()).local_whisper
+        with mock.patch.object(ggml, "accelerator", return_value="Vulkan"), \
+                mock.patch.object(ggml, "total_memory", return_value=32 << 30):
+            box._show_machine()
+        self.assertIn("Vulkan", box.machine_label.text())
+        self.assertIn("32.0 GB", box.machine_label.text())
+
+    def test_a_machine_with_no_card_is_told_it_is_on_the_processor(self):
+        box = self.window(cfg.Config()).local_whisper
+        with mock.patch.object(ggml, "accelerator", return_value=""), \
+                mock.patch.object(ggml, "total_memory", return_value=8 << 30):
+            box._show_machine()
+        self.assertIn("processor", box.machine_label.text())
+
+    def test_a_processor_build_where_the_vulkan_one_belongs_says_so(self):
+        # The Vulkan whisper-server is published by hand, and until it is
+        # there the download lands upstream's processor build. Said nowhere,
+        # an idle graphics card looks exactly like one that is being used.
+        binary = self.path("bin/whisper/v1.9.3/whisper-server")
+        binary.parent.mkdir(parents=True)
+        binary.write_text("")
+        binary.chmod(0o755)
+        self.path("bin/whisper/installed.json").write_text(json.dumps(
+            {"tag": "v1.9.3", "binary": str(binary), "backend": "processor"}))
+        # A whisper-server on this machine's PATH would win over the download.
+        self.patch_attr(ggml.shutil, "which", lambda name: None)
+        label = self.window(cfg.Config()).local_whisper.program_label.text()
+        self.assertIn("v1.9.3", label)
+        self.assertIn("Vulkan", label)
+
+    def test_an_ordinary_install_is_reported_without_a_word_about_vulkan(self):
+        binary = self.path("bin/whisper/v1.9.3/whisper-server")
+        binary.parent.mkdir(parents=True)
+        binary.write_text("")
+        binary.chmod(0o755)
+        self.path("bin/whisper/installed.json").write_text(json.dumps(
+            {"tag": "v1.9.3", "binary": str(binary)}))
+        self.patch_attr(ggml.shutil, "which", lambda name: None)
+        label = self.window(cfg.Config()).local_whisper.program_label.text()
+        self.assertIn("v1.9.3", label)
+        self.assertNotIn("Vulkan", label)
+
+    def test_a_downloaded_program_can_still_be_asked_for_again(self):
+        # The button used to disappear the moment anything landed, which left
+        # no way to pick up a newer whisper.cpp, or the Vulkan build on a
+        # machine whose driver was installed after Dikte was.
+        binary = self.path("bin/whisper/v1.9.3/whisper-server")
+        binary.parent.mkdir(parents=True)
+        binary.write_text("")
+        binary.chmod(0o755)
+        self.path("bin/whisper/installed.json").write_text(json.dumps(
+            {"tag": "v1.9.3", "binary": str(binary)}))
+        self.patch_attr(ggml.shutil, "which", lambda name: None)
+        box = self.window(cfg.Config()).local_whisper
+        self.assertTrue(box.install_button.isVisibleTo(box))
+        self.assertEqual(box.install_button.text(), t("Download again"))
+
+    def test_a_system_copy_is_not_offered_for_download(self):
+        # Nothing Dikte downloads would be run while one is on the PATH.
+        self.patch_attr(ggml.shutil, "which", lambda name: "/usr/bin/" + name)
+        box = self.window(cfg.Config()).local_whisper
+        self.assertFalse(box.install_button.isVisibleTo(box))
+
     def test_only_the_chosen_transcriber_is_on_screen(self):
         window = self.window(self.config(transcribe_provider="openai"))
         self.assertTrue(window.stt_form.isRowVisible(window.transcribe_model_row))
@@ -754,3 +1562,29 @@ class LocalModels(DikteTest):
         self.assertFalse(window.cleanup_form.isRowVisible(window.cleanup_model_row))
         # Its own thinking box, because the two default to opposite things.
         self.assertFalse(window.cleanup_form.isRowVisible(window.cleanup_reasoning))
+
+    def test_each_cleaner_brings_its_own_model_row_and_no_other(self):
+        window = self.window(cfg.Config())
+        rows = {"openrouter": window.cleanup_model_row,
+                "gemini": window.cleanup_gemini_model_row,
+                "claude": window.cleanup_claude_model,
+                "codex": window.cleanup_codex_model,
+                "agy": window.cleanup_agy_model}
+        for chosen, row in rows.items():
+            with self.subTest(provider=chosen):
+                window._select_data(window.cleanup_provider, chosen)
+                for name, other in rows.items():
+                    self.assertEqual(window.cleanup_form.isRowVisible(other),
+                                     name == chosen)
+
+    def test_each_agent_brings_its_own_box_and_no_other(self):
+        window = self.window(cfg.Config())
+        boxes = {"claude": window.claude_box, "codex": window.codex_box,
+                 "agy": window.agy_box, "openrouter": window.openrouter_box}
+        for chosen, box in boxes.items():
+            with self.subTest(provider=chosen):
+                window._select_data(window.assistant_provider, chosen)
+                for name, other in boxes.items():
+                    # isHidden rather than isVisible: the window itself is never
+                    # shown in a test, so nothing in it is ever visible.
+                    self.assertEqual(other.isHidden(), name != chosen)

tests/test_vad.py

@@ -128,6 +128,12 @@ class Hallucinations(DikteTest):
         self.assertFalse(vad.looks_like_hallucination("Bugün toplantı var.", 2.0))
         self.assertFalse(vad.looks_like_hallucination("Send it on Thursday.", 2.0))
 
+    def test_a_one_word_answer_is_believed(self):
+        # Whisper invents both over silence, but people dictate both as whole
+        # answers, and losing a real answer costs more than passing a fake one.
+        self.assertFalse(vad.looks_like_hallucination("You.", 1.5))
+        self.assertFalse(vad.looks_like_hallucination("Bye.", 1.5))
+
     def test_an_empty_transcript_counts_as_invented(self):
         self.assertTrue(vad.looks_like_hallucination("   ", 2.0))
         self.assertTrue(vad.looks_like_hallucination("...", 2.0))
@@ -138,8 +144,8 @@ class Hallucinations(DikteTest):
                 self.assertTrue(vad.looks_like_hallucination(text, 2.0))
 
     def test_the_boundary_is_the_max_duration(self):
-        self.assertTrue(vad.looks_like_hallucination("you", 6.0))
-        self.assertFalse(vad.looks_like_hallucination("you", 6.1))
+        self.assertTrue(vad.looks_like_hallucination("thanks for watching", 6.0))
+        self.assertFalse(vad.looks_like_hallucination("thanks for watching", 6.1))
 
 
 if __name__ == "__main__":

tests/test_worker.py

@@ -8,6 +8,7 @@ afterwards. A pull request that reorders any of it shows up here.
 import contextlib
 import io
 import os
+import threading
 import unittest
 from unittest import mock
 
@@ -30,9 +31,11 @@ class Chain(DikteTest):
 
     def run_chain(self, ask=False, paste_override=None, duration=2.0,
                   transcript="uh, book it for Thursday",
+                  transcribe_error=None,
                   cleaned="Book it for Thursday.",
                   cleanup_error=None, answer=("Booked.", ""), rms=None,
-                  clipboard=b"what was there before", paste_error=None):
+                  clipboard=b"what was there before", paste_error=None,
+                  focus=None):
         pipeline = worker.Pipeline(self.conf)
         done, failures, stages, cancels = [], [], [], []
         pipeline.finished.connect(lambda *args: done.append(args))
@@ -46,7 +49,10 @@ class Chain(DikteTest):
         # The chain reports its own failures on stderr, which a test run has no
         # use for.
         with contextlib.redirect_stderr(io.StringIO()), \
-                mock.patch.object(api, "transcribe", return_value=transcript) as tr, \
+                mock.patch.object(
+                    api, "transcribe",
+                    **({"side_effect": transcribe_error} if transcribe_error
+                       else {"return_value": transcript})) as tr, \
                 mock.patch.object(api, "cleanup", cleanup), \
                 mock.patch.object(assistant, "ask", return_value=answer) as ask_call, \
                 mock.patch.object(paste, "copy") as copy, \
@@ -60,7 +66,8 @@ class Chain(DikteTest):
                      "copy": copy, "copy_bytes": copy_bytes, "press": press,
                      "read_clipboard": read_clipboard}
             pipeline._work(self.wav, duration,
-                           self.rms if rms is None else rms, ask, paste_override)
+                           self.rms if rms is None else rms, ask, paste_override,
+                           focus)
         return {"done": done, "failures": failures, "stages": stages,
                 "cancelled": cancels, **calls}
 
@@ -72,7 +79,15 @@ class Chain(DikteTest):
         self.assertEqual(run["done"][0],
                          ("uh, book it for Thursday", "Book it for Thursday.", ""))
         run["copy"].assert_called_once_with("Book it for Thursday.")
-        run["press"].assert_called_once_with(self.conf["paste_shortcut"])
+        run["press"].assert_called_once_with(self.conf["paste_shortcut"],
+                                             focus=None)
+
+    def test_the_paste_is_told_where_the_dictation_started(self):
+        """Whoever was in front when the recording began is where the keys are
+        meant to go, and the press is the only part that can act on it."""
+        run = self.run_chain(focus=4242)
+        run["press"].assert_called_once_with(self.conf["paste_shortcut"],
+                                             focus=4242)
 
     def test_the_stages_are_named_as_they_happen(self):
         run = self.run_chain()
@@ -108,11 +123,57 @@ class Chain(DikteTest):
         run = self.run_chain()
         run["copy_bytes"].assert_not_called()
 
-    def test_the_clipboard_is_put_back_when_the_keypress_fails(self):
+    def test_a_failed_keypress_leaves_the_transcript_on_the_clipboard(self):
+        """The press failing is a warning, not a lost dictation: restoring the
+        old clipboard over the text would leave nothing to paste by hand."""
         self.conf["restore_clipboard"] = True
         run = self.run_chain(paste_error=paste.PasteError("not trusted"))
-        self.assertIn("not trusted", run["failures"][0])
-        run["copy_bytes"].assert_called_once_with(b"what was there before")
+        self.assertEqual(run["failures"], [])
+        raw, text, warning = run["done"][0]
+        self.assertIn("not trusted", warning)
+        run["copy_bytes"].assert_not_called()
+
+    def test_a_failed_keypress_still_reaches_the_history(self):
+        self.run_chain(paste_error=paste.PasteError("not trusted"))
+        rows = cfg.read_history()
+        self.assertEqual(len(rows), 1)
+        self.assertEqual(rows[0]["text"], "Book it for Thursday.")
+        # The row goes in before the paste is attempted, so the paste failing
+        # has to be written back into it: the record tells the whole truth.
+        self.assertIn("not trusted", rows[0]["cleanup_error"])
+
+    def test_a_failed_transcription_keeps_the_audio(self):
+        """Speech the user cannot repeat from memory must survive the failure."""
+        run = self.run_chain(transcribe_error=api.ApiError("server down"))
+        self.assertIn("server down", run["failures"][0])
+        self.assertIn("kept", run["failures"][0])
+        kept = list(cfg.RECORDINGS_DIR.glob("*.wav"))
+        self.assertEqual(len(kept), 1)
+        self.assertFalse(os.path.exists(self.wav))
+
+    def test_two_failures_in_one_second_keep_both_recordings(self):
+        self.run_chain(transcribe_error=api.ApiError("down"))
+        self.wav = make_wav(self.path("clip2.wav"), speech(2.0))
+        with mock.patch.object(worker.time, "strftime",
+                               return_value="20260820-120000"):
+            self.run_chain(transcribe_error=api.ApiError("down"))
+            self.wav = make_wav(self.path("clip3.wav"), speech(2.0))
+            self.run_chain(transcribe_error=api.ApiError("down"))
+        self.assertEqual(len(list(cfg.RECORDINGS_DIR.glob("*.wav"))), 3)
+
+    def test_the_history_row_says_whether_cleanup_actually_ran(self):
+        """The ask path cleans under its own setting; the record follows the
+        run, not the dictation gate."""
+        self.conf["cleanup_enabled"] = False
+        self.conf["assistant_cleanup"] = True
+        self.run_chain(ask=True)
+        row = cfg.read_history()[0]
+        self.assertNotEqual(row["cleanup_model"], "")
+        cfg.clear_history()
+        self.conf["cleanup_enabled"] = True
+        self.conf["assistant_cleanup"] = False
+        self.run_chain(ask=True)
+        self.assertEqual(cfg.read_history()[0]["cleanup_model"], "")
 
     def test_the_transcription_is_told_the_language_and_the_glossary(self):
         self.conf["language"] = "tr"
@@ -278,13 +339,41 @@ class Chain(DikteTest):
 
 
 class Busy(DikteTest):
-    def test_a_second_run_while_one_is_going_is_ignored(self):
+    def test_a_second_run_while_one_is_going_waits_its_turn(self):
+        """The microphone is free while a transcript is being cleaned up, so
+        the next dictation can already have been spoken by then. It has to run
+        once the first is done, in the order they were spoken, on one thread."""
         pipeline = worker.Pipeline(self.config())
-        pipeline._thread = mock.Mock(is_alive=lambda: True)
-        self.assertTrue(pipeline.busy)
-        with mock.patch.object(worker.threading, "Thread") as thread:
-            pipeline.run("/tmp/nope.wav", 1.0)
-        thread.assert_not_called()
+        order = []
+        started, gate = threading.Event(), threading.Event()
+
+        def work(wav_path, *_rest):
+            order.append(wav_path)
+            started.set()
+            if wav_path == "first.wav":
+                gate.wait(5)
+
+        with mock.patch.object(pipeline, "_work", side_effect=work):
+            pipeline.run("first.wav", 1.0)
+            self.assertTrue(started.wait(5))
+            pipeline.run("second.wav", 1.0)
+            # Held, not dropped and not running beside the first.
+            self.assertEqual(order, ["first.wav"])
+            gate.set()
+            pipeline._thread.join(5)
+        self.assertEqual(order, ["first.wav", "second.wav"])
+
+    def test_a_run_arriving_after_the_queue_drained(self):
+        """The worker thread ends with the queue; the next run brings one."""
+        pipeline = worker.Pipeline(self.config())
+        order = []
+        with mock.patch.object(pipeline, "_work",
+                               side_effect=lambda wav, *rest: order.append(wav)):
+            pipeline.run("first.wav", 1.0)
+            pipeline._thread.join(5)
+            pipeline.run("second.wav", 1.0)
+            pipeline._thread.join(5)
+        self.assertEqual(order, ["first.wav", "second.wav"])
 
     def test_the_chunk_length_matches_the_level_meter(self):
         """The silence thresholds are read in seconds, so the two must agree."""