Merge pull request #11 from yusufipk/local-models

Transcribe and clean up on this machine, fetching both programs

README.md

@@ -1,9 +1,9 @@
 # Dikte
 
-Press `Ctrl+Space`, talk, press again. The recording goes to OpenAI or OpenRouter
+Press `Ctrl+Space`, talk, press again. The recording is transcribed on this
-for transcription, a model cleans it up (dropping the *uh*s, the
+machine by default, a model cleans it up (dropping the *uh*s, the restarts, the
-restarts, the missing punctuation), and the result lands in your clipboard and
+missing punctuation), and the result lands in your clipboard and is pasted into
-is pasted into whatever window you were typing in.
+whatever window you were typing in.
 
 Built for KDE Plasma 6 on Wayland. No dependencies beyond system packages:
 just the Python standard library and PyQt6.
@@ -41,10 +41,13 @@ two global shortcuts, whose keys are its two arguments. `./update.sh` pulls and
 puts all of that back, keeping the keys you chose; `./uninstall.sh` takes it away
 again and leaves your settings and dictations alone unless you pass `--purge`.
 
-Three keys go in the settings window: **OpenAI**, **Groq** and **OpenRouter**.
+Speech to text and cleanup each pick a provider in the settings window, and both
-Speech to text runs on any of them (`gpt-4o-transcribe` by default), cleanup
+can run here, on whisper.cpp and llama.cpp: the program and the model are
-on OpenRouter (`google/gemini-3.5-flash-lite`) or, when either is installed, on
+downloaded from that window, checksummed, so nothing has to be installed first
-Claude Code or Codex instead, so a single OpenRouter key can cover both. They fall back to `OPENAI_API_KEY`, `GROQ_API_KEY` and
+and nothing leaves the machine. Otherwise speech to text runs on **OpenAI**,
+**Groq** or **OpenRouter** (`gpt-4o-transcribe` by default) and 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 `~/.config/dikte/config.json`, mode 600. Cleanup can be switched off,
 in which case the raw transcript is pasted, and a thinking model's effort can be
@@ -147,8 +150,10 @@ 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
-api.py            transcription on either provider, OpenRouter cleanup (stdlib only)
+api.py            transcription and cleanup requests (stdlib only)
-cleanup.py        who rewrites the transcript: OpenRouter, Claude Code or Codex
+cleanup.py        who rewrites the transcript: OpenRouter, here, Claude or Codex
+ggml.py           whisper.cpp and llama.cpp here: fetch, verify, keep serving
+hub.py            what GitHub and Hugging Face have on offer today
 worker.py         transcribe → clean up → clipboard → paste
 vad.py            deciding whether a recording holds speech at all
 filetranscribe.py file transcription: ffmpeg, chunking, timestamps

README.tr.md

@@ -1,9 +1,8 @@
 # Dikte
 
-`Ctrl+Space`'e bas, konuş, tekrar bas. Ses OpenAI'ye ya da OpenRouter'a gidip
+`Ctrl+Space`'e bas, konuş, tekrar bas. Ses varsayılan olarak bu makinede yazıya
-yazıya çevrilir, bir model transkripti temizler (ıı'lar,
+çevrilir, bir model transkripti temizler (ıı'lar, tekrarlar, eksik noktalama),
-tekrarlar, eksik noktalama), sonuç panoya kopyalanır ve o an yazdığın pencereye
+sonuç panoya kopyalanır ve o an yazdığın pencereye yapıştırılır.
-yapıştırılır.
 
 KDE Plasma 6 / Wayland için yazıldı. Sistem paketleri dışında bağımlılığı yok:
 sadece Python standart kütüphanesi ve PyQt6.
@@ -42,12 +41,15 @@ son sürümü çeker ve bunları senin seçtiğin tuşlarla yerine koyar;
 `./uninstall.sh` hepsini geri alır, `--purge` demedikçe ayarlarına ve
 diktelerine dokunmaz.
 
-Ayarlar penceresinde üç anahtar istenir: **OpenAI**, **Groq** ve **OpenRouter**.
+Sesi yazıya çevirme ve temizleme, ayarlar penceresinde ayrı ayrı sağlayıcı
-Sesi yazıya çevirme üçünden birinde çalışır (varsayılan `gpt-4o-transcribe`),
+seçer; ikisi de burada çalışabilir, whisper.cpp ve llama.cpp üzerinde: program da
-temizleme OpenRouter'da (`google/gemini-3.5-flash-lite`) ya da kuruluysa Claude
+model de o pencereden, sha256 doğrulamasıyla indirilir, yani önceden hiçbir şey
-Code veya Codex'te, yani tek bir OpenRouter anahtarı ikisine de yeter. Boş bırakırsan `OPENAI_API_KEY`,
+kurman gerekmez ve makineden hiçbir şey çıkmaz. Bulutu seçersen sesi yazıya
-`GROQ_API_KEY` ve `OPENROUTER_API_KEY` kullanılır;
+çevirme **OpenAI**, **Groq** ya da **OpenRouter**'da (varsayılan
-anahtarlar `~/.config/dikte/config.json`
+`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. Temizlemeyi tamamen kapatabilirsin, o zaman ham transkript
 yapıştırılır; modelin yanındaki kutudan düşünme seviyesini de seçebilirsin.
 
@@ -146,8 +148,10 @@ 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
-api.py            iki sağlayıcıda transkript + OpenRouter temizleme (yalnız stdlib)
+api.py            transkript ve temizleme istekleri (yalnız stdlib)
-cleanup.py        transkripti kim temizler: OpenRouter, Claude Code ya da Codex
+cleanup.py        transkripti kim temizler: OpenRouter, burası, Claude ya da Codex
+ggml.py           whisper.cpp ve llama.cpp'yi indirip burada çalıştırma
+hub.py            GitHub ve Hugging Face'te bugün ne olduğu
 worker.py         transkript → temizleme → pano → yapıştırma
 vad.py            kayıtta gerçekten konuşma var mı kararı
 filetranscribe.py dosyadan transkript: ffmpeg, parçalama, zaman damgaları

api.py

@@ -1,9 +1,13 @@
-"""OpenAI, Groq and OpenRouter calls, stdlib only.
+"""OpenAI, Groq, OpenRouter and this machine, stdlib only.
 
-Transcription runs on any of the three: Groq and OpenRouter both mirror OpenAI's
+Transcription runs on any of the four: Groq and OpenRouter both mirror OpenAI's
-/audio/transcriptions endpoint field for field, so one multipart request serves
+/audio/transcriptions endpoint field for field, and ggml.py starts whisper.cpp
-all of them and only the key, the base URL and the model id change. Cleanup is
+on that same path, so one multipart request serves all of them and only the key,
-always OpenRouter.
+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.
+
+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.
 """
 
 import collections
@@ -14,6 +18,7 @@ import secrets
 import urllib.error
 import urllib.request
 
+import ggml
 from i18n import t
 
 APP_URL = "https://github.com/yusufipk/dikte"
@@ -22,6 +27,12 @@ OPENAI_URL = "https://api.openai.com/v1"
 GROQ_URL = "https://api.groq.com/openai/v1"
 OPENROUTER_URL = "https://openrouter.ai/api/v1"
 
+# 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
+# time, and a long recording on a machine without a graphics card takes a good
+# deal of it. Cutting that off would throw the work away for nothing.
+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.
@@ -33,9 +44,11 @@ def timestamp_model(provider, selected=""):
 
     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.
+    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.
     """
-    if provider == "groq":
+    if provider in ("groq", "local"):
         return selected or "whisper-large-v3-turbo"
     return "openai/whisper-1" if provider == "openrouter" else "whisper-1"
 
@@ -114,7 +127,11 @@ def _multipart(fields, file_field, file_path):
 
 
 def _headers(provider, api_key, content_type=None):
-    headers = {"Authorization": f"Bearer {api_key}", "User-Agent": USER_AGENT}
+    headers = {"User-Agent": USER_AGENT}
+    # A server on this machine has nothing to authorise, and sending it a
+    # bearer token would only be a made-up one.
+    if api_key:
+        headers["Authorization"] = f"Bearer {api_key}"
     if content_type:
         headers["Content-Type"] = content_type
     if provider == "openrouter":
@@ -124,16 +141,45 @@ def _headers(provider, api_key, content_type=None):
     return headers
 
 
+def serving(server):
+    """The base URL of a local server, started if it is not up yet.
+
+    It picks its own port, so this is the first moment its address exists.
+    serve() is idempotent: once it is running this costs nothing.
+    """
+    try:
+        return server.serve()
+    except ggml.LocalError as exc:
+        raise ApiError(str(exc)) from None
+
+
+def local_failure(service, server, exc):
+    """A server that died mid-request, explained by its own output.
+
+    Without this the message is that the connection dropped, when the reason for
+    it was printed by the process at the other end.
+    """
+    detail = server.error()
+    return ApiError(f"{service}: {exc}" + (f" ({detail})" if detail else ""),
+                    exc.status)
+
+
 def _transcribe_request(target, wav_path, language, prompt, response_format,
                         granularity=None, timeout=300):
-    if not target.api_key:
+    if target.provider == "local":
+        # The timeouts here are sized for a hosted API, where a slow answer is a
+        # bill running. Locally the only thing being spent is time.
+        target = target._replace(base_url=serving(ggml.whisper))
+        timeout = max(timeout, LOCAL_TIMEOUT)
+    elif not target.api_key:
         raise ApiError(t("{service} API key is empty. Add it in Settings.",
                          service=target.service))
     fields = [("model", target.model), ("response_format", response_format)]
     if language and language != "auto":
         fields.append(("language", language))
     # OpenRouter takes the hint field and throws it away, so spare it the bytes.
-    # The same words still reach the cleanup model as a glossary.
+    # The same words still reach the cleanup model as a glossary. whisper.cpp
+    # takes it as the initial prompt, the way OpenAI does.
     if prompt and target.provider != "openrouter":
         fields.append(("prompt", prompt))
     if granularity:
@@ -145,14 +191,58 @@ def _transcribe_request(target, wav_path, language, prompt, response_format,
             _headers(target.provider, target.api_key, ctype), timeout=timeout,
         )
     except ApiError as exc:
+        if target.provider == "local":
+            raise local_failure(target.service, ggml.whisper, exc) from None
         raise explain(exc, target.service) from None
 
 
+# Whisper marks the start of a word with a leading space, so a piece of text
+# that does not begin with one continues the word before it rather than starting
+# a new one. Both helpers below turn on that.
+def _continues_a_word(previous, following):
+    return bool(previous) and not previous[-1:].isspace() and not following[:1].isspace()
+
+
+def _local_text(text):
+    """whisper.cpp's segments, joined back into the flowing line OpenAI returns.
+
+    Its plain text puts one segment per line, and a segment boundary falls
+    wherever the tokens fell, which in Turkish lands inside a word about as
+    often as between two. Nothing takes the line break's place: whisper's own
+    leading spaces are what separate the words, and a break inside "değ|iller"
+    has nothing on either side of it worth keeping.
+    """
+    return "".join(text.split("\n"))
+
+
+def _merge_word_splits(segments):
+    """Fold a segment that begins mid-word into the one it continues.
+
+    The hosted whisper-1 hands back segments cut on sentences; whisper.cpp cuts
+    them on tokens, and a subtitle cue reading "değ" is not a cue. The times are
+    joined along with the text, so the merged segment still covers the whole
+    word.
+    """
+    merged = []
+    for seg in segments:
+        text = seg.get("text") or ""
+        if merged and _continues_a_word(merged[-1]["text"], text):
+            merged[-1]["text"] += text
+            merged[-1]["end"] = seg.get("end") or merged[-1]["end"]
+            continue
+        merged.append({"text": text, "start": seg.get("start") or 0.0,
+                       "end": seg.get("end") or 0.0})
+    return merged
+
+
 def transcribe(target, wav_path, language="", prompt="", timeout=300):
     data = _transcribe_request(
         target, wav_path, language, prompt, "json", timeout=timeout
     )
-    text = (data.get("text") or "").strip()
+    text = data.get("text") or ""
+    if target.provider == "local":
+        text = _local_text(text)
+    text = text.strip()
     if not text:
         raise ApiError(t("Transcript came back empty."))
     return text
@@ -166,6 +256,8 @@ def transcribe_segments(target, wav_path, language="", prompt="", timeout=300):
         granularity="segment", timeout=timeout,
     )
     segments = data.get("segments") or []
+    if target.provider == "local":
+        segments = _merge_word_splits(segments)
     out = []
     for seg in segments:
         text = (seg.get("text") or "").strip()
@@ -174,18 +266,55 @@ def transcribe_segments(target, wav_path, language="", prompt="", timeout=300):
             end = float(seg.get("end") or 0.0)
             out.append((start, max(end, start), text))
     if not out:
-        text = (data.get("text") or "").strip()
+        text = data.get("text") or ""
+        if target.provider == "local":
+            text = _local_text(text)
+        text = text.strip()
         if not text:
             raise ApiError(t("Transcript came back empty."))
         out = [(0.0, 0.0, text)]
     return out
 
 
+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.
+    """
+    if not reasoning:
+        return
+    if provider == "local-llm":
+        # 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 reasoning != "none":
+        # The thinking itself is never shown, so ask for it to be left out.
+        payload["reasoning"] = {"effort": reasoning, "exclude": True}
+
+
+def local_ceiling(text):
+    """How much of a reply is worth waiting for from a model on this machine.
+
+    Cleanup gives back what it was given, near enough, so a reply several times
+    the length of the transcript is a model that has lost the thread rather than
+    one doing the job. A small one will happily repeat the transcript until the
+    context is full, and every one of those tokens is a second of somebody
+    waiting. A hosted model is left alone: there the same runaway is rare, and a
+    ceiling would cut the minutes short instead.
+    """
+    return max(512, len(text))
+
+
 def cleanup(text, api_key, model, system_prompt, reasoning="",
-            base_url=OPENROUTER_URL, timeout=180):
+            base_url=OPENROUTER_URL, timeout=180, provider="openrouter",
-    if not api_key:
+            service="OpenRouter"):
+    if not api_key and provider != "local-llm":
         raise ApiError(t("{service} API key is empty. Add it in Settings.",
-                         service="OpenRouter"))
+                         service=service))
     payload = {
         "model": model,
         "temperature": 0,
@@ -194,25 +323,30 @@ def cleanup(text, api_key, model, system_prompt, reasoning="",
             {"role": "user", "content": f"<transcript>\n{text}\n</transcript>"},
         ],
     }
-    # An empty level means "whatever the model does on its own"; anything else is
+    if provider == "local-llm":
-    # one of OpenRouter's efforts. The thinking itself is never shown, so ask for
+        payload["max_tokens"] = local_ceiling(text)
-    # it to be left out of the reply.
+    _thinking(payload, provider, reasoning)
-    if reasoning:
-        payload["reasoning"] = {"effort": reasoning, "exclude": True}
     try:
         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)))
-    content = ((choices[0].get("message") or {}).get("content") or "").strip()
+    message = choices[0].get("message") or {}
+    content = (message.get("content") or "").strip()
     if not content:
+        # A thinking model can spend the whole reply on the thinking and leave
+        # nothing to paste. Worth naming, because the fix is a setting rather
+        # than a retry: cleanup is not a job that wants thinking.
+        if message.get("reasoning_content") or message.get("reasoning"):
+            raise ApiError(t("The cleanup model spent its whole reply on "
+                             "thinking. Set Thinking to \u201cOff\u201d."))
         raise ApiError(t("The cleanup model returned an empty reply."))
     return content
 

cleanup.py

@@ -20,9 +20,10 @@ import tempfile
 
 import api
 import assistant
+import ggml
 from i18n import t
 
-PROVIDERS = ("openrouter", "claude", "codex")
+PROVIDERS = ("openrouter", "local", "claude", "codex")
 
 
 class CleanupError(api.ApiError):
@@ -47,6 +48,8 @@ def executable(name):
 def model(conf):
     """Which model does the cleaning, for the history and the settings window."""
     name = provider(conf)
+    if name == "local":
+        return conf["local_llm_model"]
     if name == "claude":
         return conf["cleanup_claude_model"].strip() or "haiku"
     if name == "codex":
@@ -65,10 +68,34 @@ def run(text, conf, system_prompt, timeout=180):
             reasoning=conf["cleanup_reasoning"],
             base_url=conf["openrouter_base_url"], timeout=timeout,
         )
+    if name == "local":
+        return _local(text, conf, system_prompt, timeout)
     runner = _claude if name == "claude" else _codex
     return runner(text, conf, system_prompt, timeout)
 
 
+def _local(text, conf, system_prompt, timeout):
+    """llama.cpp, on this machine, answering the request OpenRouter answers.
+
+    No key and no bill, and the address does not exist until the server is up,
+    which is what starting it here is for. The timeout is the hosted one raised:
+    the only thing being spent is time.
+    """
+    service = t("Local model")
+    try:
+        return api.cleanup(
+            text, "", conf["local_llm_model"], system_prompt,
+            reasoning=conf["local_llm_reasoning"],
+            base_url=api.serving(ggml.llm),
+            timeout=max(timeout, api.LOCAL_TIMEOUT),
+            provider="local-llm", service=service,
+        )
+    except api.ApiError as exc:
+        # A server that died mid-request would otherwise report only that the
+        # connection dropped, when the reason is in its own output.
+        raise api.local_failure(service, ggml.llm, exc) from None
+
+
 def _wrap(text):
     """The same fence the OpenRouter call puts around it: this is the material,
     not the instruction, however much of it reads like one."""

config.py

@@ -7,7 +7,9 @@ import os
 import pathlib
 
 import api
+import ggml
 import i18n
+from i18n import t
 
 
 def _xdg(var, default):
@@ -368,18 +370,49 @@ DEFAULTS = {
     "groq_base_url": "https://api.groq.com/openai/v1",
     "openrouter_api_key": "",
     "openrouter_base_url": "https://openrouter.ai/api/v1",
-    "transcribe_provider": "openai",  # a key of TRANSCRIBERS
+    "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",
     "language": "tr",
     "transcribe_prompt": "",
+
+    # --- whisper.cpp, on this machine ---------------------------------------
+    # The program and the model are both fetched from Settings; empty means
+    # nothing has been downloaded yet, which is what opens Settings on a first
+    # run.
+    # Pointed at the suggestion rather than at nothing, so the settings window
+    # opens with the Download button already on the right model.
+    "local_model": ggml.SUGGESTED_WHISPER,
+    "local_threads": 0,             # 0 -> whisper.cpp picks
+    "local_gpu": True,
+    "local_preload": True,          # load the model while Dikte starts, rather
+                                    # than on the first dictation
+    "local_binary": "",             # empty -> whichever copy ggml.py finds
+
     "cleanup_enabled": True,
-    "cleanup_provider": "openrouter",  # openrouter | claude | codex
+    "cleanup_provider": "openrouter",  # a name in cleanup.PROVIDERS
     "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_reasoning": "",        # empty -> whatever the model does by default
+
+    # --- llama.cpp, on this machine -----------------------------------------
+    # Kept apart from the meeting settings on purpose. Cleanup is punctuation
+    # and filler words, which a small model does in a moment; the minutes are a
+    # summary of an hour, which it does not.
+    "local_llm_model": "",          # a file name, e.g. gemma-3-4b-it-Q4_K_M.gguf
+    # Where the model list is read from; the settings window offers the
+    # publishers ggml.py knows of and takes any other one that is typed in.
+    "local_llm_repo": ggml.SUGGESTED_LLM[0],
+    "local_llm_threads": 0,
+    "local_llm_gpu": True,
+    "local_llm_context": 8192,
+    "local_llm_binary": "",
+    "local_llm_preload": False,     # heavier than whisper, so only when asked
+    # Off rather than empty: a model trained to think will, and 300 tokens of
+    # reasoning about a comma is 300 tokens of waiting.
+    "local_llm_reasoning": "none",
     "cleanup_prompt": "",           # empty -> language-specific default
     "auto_paste": True,
     "paste_shortcut": "ctrl+v",
@@ -528,14 +561,63 @@ class Config:
         return self.api_key("openrouter_api_key")
 
     def transcribe_target(self):
-        """Key, endpoint and model for whichever provider does speech to text."""
+        """Key, endpoint and model for whichever provider does speech to text.
+
+        The local one is not in the table and leaves its base URL empty on
+        purpose: the server picks a port when it starts, and reading a setting
+        must not be what launches a process. api.py fills the address in when it
+        is about to send the request, which is the moment the server is needed
+        anyway.
+        """
         name = self["transcribe_provider"]
+        if name == "local":
+            return api.Target("local", t("Local whisper"), "", "",
+                              self["local_model"])
         if name not in TRANSCRIBERS:
-            name = DEFAULTS["transcribe_provider"]
+            # A config written by a fork, or by a version that dropped one. The
+            # shipped default is not in the table, so this names the hosted one
+            # to land on rather than reading it from there.
+            name = "openai"
         who = TRANSCRIBERS[name]
         return api.Target(name, who.service, self.api_key(who.key),
                           self[who.url], self[who.model])
 
+    def transcribe_ready(self):
+        """Whether speech to text could run right now, without opening Settings."""
+        if self["transcribe_provider"] == "local":
+            return self.local_whisper_ready()
+        return bool(self.transcribe_target().api_key)
+
+    def local_whisper_ready(self):
+        return bool(ggml.program_path(ggml.WHISPER, self["local_binary"])
+                    and self["local_model"]
+                    and ggml.have_model(ggml.whisper_model_path(self["local_model"])))
+
+    def local_llm_ready(self):
+        return bool(ggml.program_path(ggml.LLAMA, self["local_llm_binary"])
+                    and self["local_llm_model"]
+                    and ggml.have_model(ggml.llm_model_path(self["local_llm_model"])))
+
+    def apply_local(self):
+        """Hand the local settings to the servers, restarting what they change."""
+        ggml.whisper.configure(
+            model=self["local_model"],
+            threads=int(self["local_threads"]),
+            gpu=bool(self["local_gpu"]),
+            binary=self["local_binary"],
+        )
+        ggml.llm.configure(
+            model=self["local_llm_model"],
+            threads=int(self["local_llm_threads"]),
+            gpu=bool(self["local_llm_gpu"]),
+            binary=self["local_llm_binary"],
+            context=int(self["local_llm_context"]),
+        )
+
+    def uses_local_llm(self):
+        """Whether anything is set to run the local cleanup model."""
+        return self["cleanup_provider"] == "local"
+
     def cleanup_prompt(self, with_timestamps=False, with_speakers=False,
                        subtitles=False):
         turkish = i18n.language() == "tr"

dikte.py

@@ -7,16 +7,20 @@ terminal talks to. Every verb it answers is in cli.py, which is also what runs
 command line says "there is no instance to talk to, so be one".
 """
 
+import contextlib
 import json
 import os
+import signal
+import socket
 import sys
+import threading
 
 # A Wayland client cannot place a window in a screen corner, so the indicator
 # is drawn through XWayland.
 if os.environ.get("XDG_SESSION_TYPE") == "wayland" and os.environ.get("DISPLAY"):
     os.environ.setdefault("QT_QPA_PLATFORM", "xcb")
 
-from PyQt6.QtCore import QTimer, QElapsedTimer  # noqa: E402
+from PyQt6.QtCore import QTimer, QElapsedTimer, QSocketNotifier  # noqa: E402
 from PyQt6.QtGui import QAction, QIcon  # noqa: E402
 from PyQt6.QtNetwork import QLocalServer, QLocalSocket  # noqa: E402
 from PyQt6.QtWidgets import QApplication, QMenu, QSystemTrayIcon  # noqa: E402
@@ -25,6 +29,7 @@ import assistant  # noqa: E402
 import audio  # noqa: E402
 import cli  # noqa: E402
 import config as cfg  # noqa: E402
+import ggml  # noqa: E402
 import hotkey  # noqa: E402
 import i18n  # noqa: E402
 import ipc  # noqa: E402
@@ -92,6 +97,9 @@ class Dikte:
         self.meeting_recorder = audio.MeetingRecorder()
         self.meetings = MeetingPipeline(self.conf)
         self.evdev = hotkey.EvdevHotkey()
+        # Before anything of ours is started: a server from a Dikte that was
+        # killed outright is still holding a model in memory.
+        ggml.sweep()
 
         self.recorder.level.connect(self._on_level)
         self.recorder.stopped.connect(self._on_recorded)
@@ -805,9 +813,44 @@ class Dikte:
         # Don't drop the object while its own signal is still being delivered.
         QTimer.singleShot(0, lambda: setattr(self, "settings_window", None))
 
+    def _apply_local(self):
+        """Pass the local settings on, and hold the models ready if asked to.
+
+        Loading a model takes a second or two for whisper and longer for an LLM.
+        Doing it while Dikte starts rather than on the first dictation is the
+        whole reason a server is kept alive instead of running the program once
+        per recording; the checkboxes are there for the machine whose memory is
+        wanted elsewhere.
+        """
+        self.conf.apply_local()
+        wanted = []
+        if self.conf["transcribe_provider"] == "local":
+            if self.conf["local_preload"] and self.conf.local_whisper_ready():
+                wanted.append((ggml.whisper, "whisper"))
+        else:
+            ggml.whisper.stop()      # give the memory back when it is not in use
+        if self.conf.uses_local_llm():
+            if self.conf["local_llm_preload"] and self.conf.local_llm_ready():
+                wanted.append((ggml.llm, "llama"))
+        else:
+            ggml.llm.stop()
+
+        def warm():
+            for server, name in wanted:
+                try:
+                    server.serve()
+                except ggml.LocalError as exc:
+                    # Not worth an indicator: the first dictation raises the
+                    # same thing where the user can act on it.
+                    print(f"dikte: {name}: {exc}", file=sys.stderr)
+
+        if wanted:
+            threading.Thread(target=warm, daemon=True).start()
+
     def _apply_settings(self):
         self.overlay.corner = self.conf["overlay_corner"]
         self.ask_overlay.corner = self.conf["overlay_corner"]
+        self._apply_local()
         self._build_tray()
         self._refresh_tray()
         if self.conf["evdev_hotkey"]:
@@ -836,6 +879,9 @@ class Dikte:
             self.meeting_recorder.stop()
         self.overlay.dismiss()
         self.ask_overlay.dismiss()
+        # Also on the restart path, which replaces the process without ever
+        # reaching atexit and would otherwise leave the models in memory.
+        ggml.stop_all()
         self.tray.hide()
 
 
@@ -861,6 +907,41 @@ def main():
     return run_app([arg for arg in argv if arg != "--gui"])
 
 
+def install_signal_handlers(app):
+    """Quit properly on the signals a session sends, rather than dying where we stand.
+
+    Qt spends its time blocked inside C, and a Python signal handler only runs
+    between bytecodes, so on its own it would not run until the next event
+    arrived, which for an idle tray icon may be never. set_wakeup_fd writes the
+    signal number to a socket instead, and a notifier turns that into an event
+    Qt does deliver.
+
+    Worth the trouble because of what shutdown() does: a logout sends SIGTERM,
+    and without this a whisper.cpp or llama.cpp server outlives the session
+    holding its model in memory. SIGKILL cannot be caught at all, which is what
+    ggml.sweep() is for.
+
+    Returns the objects it made; they have to stay alive to keep working.
+    """
+    reader, writer = socket.socketpair()
+    reader.setblocking(False)
+    writer.setblocking(False)
+    signal.set_wakeup_fd(writer.fileno())
+    notifier = QSocketNotifier(reader.fileno(), QSocketNotifier.Type.Read)
+
+    def woken():
+        with contextlib.suppress(OSError):
+            reader.recv(64)
+        app.quit()          # aboutToQuit runs shutdown()
+
+    notifier.activated.connect(woken)
+    for sig in (signal.SIGINT, signal.SIGTERM, signal.SIGHUP):
+        # A handler that does nothing, so that the default action, stopping the
+        # process where it stands, is replaced by the wakeup above.
+        signal.signal(sig, lambda *_: None)
+    return reader, writer, notifier
+
+
 def run_app(args):
     command = args[0] if args else ""
 
@@ -868,6 +949,12 @@ def run_app(args):
     app.setApplicationName("Dikte")
     app.setDesktopFileName("dikte")
     app.setQuitOnLastWindowClosed(False)
+    # Before Dikte is built, because building it is what may start a server, and
+    # a signal arriving in the middle of that would otherwise take the default
+    # action and leave the server behind. A signal this early lands in the
+    # socket and is delivered as soon as the event loop starts. Held in a name
+    # so that the notifier and its socket outlive this function.
+    signal_plumbing = install_signal_handlers(app)  # noqa: F841
 
     if not QSystemTrayIcon.isSystemTrayAvailable():
         print("dikte: no system tray found, running anyway")
@@ -916,7 +1003,10 @@ def run_app(args):
 
     # No key for the chosen transcription provider means nothing can work yet,
     # so the settings window is the only useful thing to open.
-    if command == "settings" or not dikte.conf.transcribe_target().api_key:
+    # A transcription provider that cannot run yet, whether that is a missing
+    # API key or a model nobody has downloaded, means nothing can work, so the
+    # settings window is the only useful thing to open.
+    if command == "settings" or not dikte.conf.transcribe_ready():
         dikte.open_settings()
     elif command == "toggle":
         QTimer.singleShot(0, dikte.toggle)

ggml.py

@@ -0,0 +1,792 @@
+"""Speech to text and cleanup on this machine: whisper.cpp and llama.cpp.
+
+Two programs, one treatment. Fetch a release from GitHub, unpack it under the
+data directory, fetch a model from Hugging Face, then keep one server alive on a
+port of its own. Both of them speak the shape api.py already sends to the hosted
+providers, so what the rest of Dikte sees is a base URL and nothing else:
+whisper-server is started on `--inference-path /v1/audio/transcriptions`, the
+exact path api.py builds, and llama-server answers /v1/chat/completions the way
+OpenRouter does.
+
+A server rather than a one-shot run, because the model is the slow part. Loading
+a large whisper model takes a second or two while transcribing a few seconds of
+speech takes a fraction of one, and an LLM is worse: a server pays that once and
+a run per dictation pays it every time.
+
+Nothing downloaded is trusted for having arrived. Every file is checked against
+the sha256 its index published, and the bytes go to a `.part` that is only
+renamed once the whole thing is there, so an interrupted download can never be
+mistaken for a working one.
+
+This module imports hub and the string table, and nothing else of Dikte's: it
+knows how to fetch a file and how to run a process, and nothing about dictation.
+Its errors leave as LocalError and api.py turns them into the ApiError the
+interface already knows how to show.
+"""
+
+import atexit
+import collections
+import ctypes.util
+import hashlib
+import http.client
+import json
+import os
+import pathlib
+import platform
+import shutil
+import signal
+import socket
+import subprocess
+import tarfile
+import threading
+import time
+import urllib.error
+import urllib.request
+
+import hub
+from i18n import t
+
+HOST = "127.0.0.1"
+# The path api.py asks for, so its URL and the server's line up.
+INFERENCE_PATH = "/v1/audio/transcriptions"
+
+DATA_DIR = (pathlib.Path(os.environ.get("XDG_DATA_HOME")
+                         or os.path.expanduser("~/.local/share")) / "dikte")
+BIN_DIR = DATA_DIR / "bin"
+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
+DOWNLOAD_CHUNK = 1 << 20
+
+# `health` is the path that answers only once the model is in memory. whisper
+# does not have one and does not need one: it binds its port after the model is
+# loaded, so the port opening is the signal.
+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")
+
+# 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
+# publishes.
+WHISPER_MODELS_REPO = "ggerganov/whisper.cpp"
+LLM_AUTHOR = "ggml-org"
+
+# What the whisper repository holds besides models: Core ML encoders for Apple
+# hardware and the odd loose file.
+WHISPER_PREFIX = "ggml-"
+WHISPER_SUFFIX = ".bin"
+
+# 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-")
+# 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
+
+# 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.
+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/SmolLM3-3B-GGUF",
+)
+# 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".
+SUGGESTED_WHISPER = "ggml-large-v3-turbo-q5_0.bin"
+
+
+class LocalError(Exception):
+    pass
+
+
+def human_size(count):
+    for unit in ("B", "KB", "MB", "GB"):
+        if count < 1024 or unit == "GB":
+            return f"{count:.0f} {unit}" if unit == "B" else f"{count:.1f} {unit}"
+        count /= 1024.0
+    return f"{count:.1f} GB"
+
+
+# --- fetching -------------------------------------------------------------
+
+
+def download(item, target, on_progress=None, should_stop=None, require_hash=True):
+    """Fetch one hub.Item to `target`. True when it landed, False when stopped.
+
+    The bytes go to a `.part` that is renamed only after both the length and the
+    hash agree with what the index said. A truncated file would otherwise sit
+    there looking installed and fail much later, inside a server, as a corrupt
+    model; a file that is the right length but the wrong content is worse, and
+    this is a program as often as it is a model.
+
+    A file whose index published no hash is refused rather than taken on trust.
+    Everything fetched here is either run or parsed by something written in C++,
+    and GitHub did not always publish a digest: a release old enough to predate
+    that would otherwise install unchecked, which is the one case where this
+    would matter most and say least.
+    """
+    target = pathlib.Path(target)
+    if require_hash and not item.sha256:
+        raise LocalError(t("{name} is published without a checksum, so there is "
+                           "no way to tell what arrived. Nothing was installed.",
+                           name=item.name))
+    part = target.with_name(target.name + ".part")
+    try:
+        target.parent.mkdir(parents=True, exist_ok=True)
+    except OSError as exc:
+        raise LocalError(t("Could not create {path}: {error}",
+                           path=target.parent, error=exc)) from exc
+
+    request = urllib.request.Request(item.url, headers={"User-Agent": hub.USER_AGENT})
+    digest = hashlib.sha256()
+    done = 0
+    try:
+        with urllib.request.urlopen(request, timeout=60) as response:
+            total = int(response.headers.get("Content-Length") or item.size or 0)
+            with open(part, "wb") as out:
+                while True:
+                    if should_stop is not None and should_stop():
+                        part.unlink(missing_ok=True)
+                        return False
+                    block = response.read(DOWNLOAD_CHUNK)
+                    if not block:
+                        break
+                    out.write(block)
+                    digest.update(block)
+                    done += len(block)
+                    # More than was announced: a body that does not end is the
+                    # one way this loop could run until the disk is full.
+                    if total and done > total:
+                        part.unlink(missing_ok=True)
+                        raise LocalError(t("{name} is longer than it said it "
+                                           "would be.", name=item.name))
+                    if on_progress is not None:
+                        on_progress(done, total)
+        # A proxy notice or an error page that came back as 200 would otherwise
+        # be renamed into place and only fail when something tries to read it.
+        if total and done != total:
+            part.unlink(missing_ok=True)
+            raise LocalError(t("The download stopped early ({done} of {total}).",
+                               done=human_size(done), total=human_size(total)))
+        if item.sha256 and digest.hexdigest() != item.sha256:
+            part.unlink(missing_ok=True)
+            raise LocalError(t("{name} does not match its published checksum. "
+                               "Nothing was installed.", name=item.name))
+        part.replace(target)
+        return True
+    except urllib.error.HTTPError as exc:
+        part.unlink(missing_ok=True)
+        exc.close()   # it holds the response body open until it is collected
+        raise LocalError(t("Could not download {name}: HTTP {code}",
+                           name=item.name, code=exc.code)) from exc
+    except urllib.error.URLError as exc:
+        part.unlink(missing_ok=True)
+        raise LocalError(t("Could not download {name}: {error}",
+                           name=item.name, error=exc.reason)) from exc
+    except OSError as exc:
+        # A connection cut mid-body arrives here too, and gigabytes in is
+        # exactly where that happens.
+        part.unlink(missing_ok=True)
+        raise LocalError(t("Could not write {name}: {error}",
+                           name=item.name, error=exc)) from exc
+
+
+# --- the programs ---------------------------------------------------------
+
+
+def _arch():
+    machine = platform.machine().lower()
+    if machine in ("aarch64", "arm64"):
+        return "arm64"
+    return "x64"
+
+
+def _has_vulkan():
+    """Whether a Vulkan loader is installed, which decides which build to fetch.
+
+    llama.cpp publishes no CUDA build for Linux, so Vulkan is what a graphics
+    card gets here. The build without it is smaller and runs on the CPU, and
+    fetching the Vulkan one for a machine that cannot load it would only make
+    the download bigger.
+    """
+    return bool(ctypes.util.find_library("vulkan"))
+
+
+def _wanted_assets(program):
+    """Asset name endings to accept, best first."""
+    arch = _arch()
+    if program is LLAMA and _has_vulkan():
+        return (f"bin-ubuntu-vulkan-{arch}.tar.gz", f"bin-ubuntu-{arch}.tar.gz")
+    return (f"bin-ubuntu-{arch}.tar.gz",)
+
+
+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."""
+    try:
+        record = json.loads(_install_record(program).read_text(encoding="utf-8"))
+        path = record.get("binary") or ""
+    except (OSError, ValueError):
+        return ""
+    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 ""
+
+
+def program_path(program, custom=""):
+    """Which copy of the program to run, or "" when there is none.
+
+    A system one wins over a downloaded one. The distribution package is built
+    against whatever the machine has, which on this platform means it may reach
+    the graphics card, while the release binaries carry CPU backends only.
+    """
+    custom = (custom or "").strip()
+    if custom:
+        return custom if os.path.isfile(custom) and os.access(custom, os.X_OK) else ""
+    return shutil.which(program.binary) or installed_program(program)
+
+
+def system_program(program):
+    """Whether the program came from the system rather than from Dikte."""
+    return bool(shutil.which(program.binary))
+
+
+def _find_binary(root, name):
+    for path in sorted(pathlib.Path(root).rglob(name)):
+        if path.is_file():
+            return path
+    return None
+
+
+def _extract(archive, into):
+    """Unpack a release tarball, refusing anything that reaches outside `into`.
+
+    The archives lay their libraries next to their binaries and are linked with
+    an $ORIGIN runpath, so a whole directory is what has to survive the trip and
+    the binary cannot be lifted out of it.
+    """
+    try:
+        with tarfile.open(archive, "r:gz") as tar:
+            try:
+                tar.extractall(into, filter="data")
+            except TypeError:      # Python without the extraction filters
+                tar.extractall(into)
+    except (tarfile.TarError, OSError) as exc:
+        raise LocalError(t("Could not unpack {name}: {error}",
+                           name=os.path.basename(str(archive)), error=exc)) from exc
+
+
+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.
+    """
+    try:
+        tag, assets = hub.release(program.repo, tag or "latest", 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:
+        raise LocalError(t("{repo} {tag} has no build for this machine.",
+                           repo=program.repo, tag=tag))
+
+    into = BIN_DIR / program.name / tag
+    shutil.rmtree(into, ignore_errors=True)
+    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, program.binary)
+        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")
+    except OSError as exc:
+        raise LocalError(t("Could not install {name}: {error}",
+                           name=program.name, error=exc)) from exc
+    finally:
+        try:
+            archive.unlink(missing_ok=True)
+        except OSError:
+            pass
+    _drop_old_versions(program, keep=tag)
+    return str(binary)
+
+
+def _drop_old_versions(program, keep):
+    """Leave one unpacked release behind, not one per update."""
+    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)
+    except OSError:
+        pass
+
+
+# --- the models -----------------------------------------------------------
+
+
+def whisper_models(refresh=False):
+    """[hub.Item] for every whisper model on offer, smallest first."""
+    try:
+        files = hub.files(WHISPER_MODELS_REPO, refresh=refresh)
+    except hub.HubError as exc:
+        raise LocalError(str(exc)) from exc
+    models = [f for f in files
+              if f.name.startswith(WHISPER_PREFIX) and f.name.endswith(WHISPER_SUFFIX)
+              and f.size > 0]
+    return sorted(models, key=lambda f: f.size)
+
+
+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)]
+    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
+        # say so where it matters, when a download is asked for.
+        found = []
+    if not found:
+        return list(SUGGESTED_LLM)
+    first = [r for r in SUGGESTED_LLM if r in found]
+    return first + [r for r in found if r not in first]
+
+
+def llm_quants(repo, refresh=False):
+    """[hub.Item] for the model files in one GGUF repository, smallest first."""
+    try:
+        files = hub.files(repo, refresh=refresh)
+    except hub.HubError as exc:
+        raise LocalError(str(exc)) from exc
+    out = []
+    for item in files:
+        name = item.name.rsplit("/", 1)[-1]
+        if not name.endswith(".gguf") or name.startswith(GGUF_SKIP):
+            continue
+        # A model split across files needs all of them and a different command
+        # line; anything cleanup wants fits in one.
+        if "-of-000" in name or not 0 < item.size <= GGUF_MAX_BYTES:
+            continue
+        out.append(item)
+    return sorted(out, key=lambda f: f.size)
+
+
+def whisper_model_path(name):
+    return MODELS_DIR / "whisper" / name
+
+
+def llm_model_path(name):
+    return MODELS_DIR / "llm" / name.rsplit("/", 1)[-1]
+
+
+def have_model(path):
+    path = pathlib.Path(path)
+    return path.is_file() and path.stat().st_size > 0
+
+
+def installed_whisper_models():
+    return sorted(p.name for p in (MODELS_DIR / "whisper").glob("*.bin"))
+
+
+def installed_llm_models():
+    return sorted(p.name for p in (MODELS_DIR / "llm").glob("*.gguf"))
+
+
+def delete_model(path):
+    try:
+        pathlib.Path(path).unlink()
+    except FileNotFoundError:
+        pass
+    except OSError as exc:
+        raise LocalError(t("Could not delete the model: {error}", error=exc)) from exc
+
+
+# --- one server -----------------------------------------------------------
+
+
+def _free_port():
+    """A port nothing is listening on, handed straight to the server.
+
+    Between closing this socket and the server binding it, something else could
+    take it; that is why a start retries rather than trusting the number.
+    """
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
+        sock.bind((HOST, 0))
+        return sock.getsockname()[1]
+
+
+def _listening(port):
+    try:
+        with socket.create_connection((HOST, port), timeout=0.5):
+            return True
+    except OSError:
+        return False
+
+
+def _healthy(port, path):
+    """Whether the model is in memory, for a server that says so.
+
+    Spoken over http.client rather than urllib because this never leaves the
+    machine: it is the same question as _listening, one layer up.
+    """
+    connection = http.client.HTTPConnection(HOST, port, timeout=2)
+    try:
+        connection.request("GET", path)
+        # 503 for as long as the model is still being read in.
+        return connection.getresponse().status == 200
+    except (http.client.HTTPException, OSError):
+        return False
+    finally:
+        connection.close()
+
+
+def _tail(path, lines=3):
+    try:
+        with open(path, encoding="utf-8", errors="replace") as fh:
+            found = [line.strip() for line in fh if line.strip()]
+    except OSError:
+        return ""
+    return " | ".join(found[-lines:])
+
+
+class Server:
+    """One process, started when something needs it and stopped when nothing does.
+
+    `build` turns the settings into a command line; everything else about
+    running a server is the same for both programs.
+    """
+
+    def __init__(self, program, build, defaults):
+        self.program = program
+        self._build = build
+        self._settings = dict(defaults)
+        # Two locks on purpose. `_lock` is held for the length of a dictionary
+        # lookup, so the interface can ask what is running while a model is
+        # being loaded; `_starting` is held across the start itself, which can
+        # take a minute and which two threads must not both do.
+        self._lock = threading.Lock()
+        self._starting = threading.Lock()
+        self._proc = None
+        self._port = 0
+        self._log = ""
+        self._key = None
+
+    # ---- settings --------------------------------------------------------
+
+    def configure(self, **changes):
+        """Apply settings. A server started on the old ones is stopped."""
+        with self._lock:
+            for key, value in changes.items():
+                if value is not None and key in self._settings:
+                    self._settings[key] = value
+            stale = self._proc is not None and self._key != self._settings_key()
+        if stale:
+            self.stop()
+
+    def settings(self):
+        with self._lock:
+            return dict(self._settings)
+
+    def _settings_key(self):
+        """What a running server would have to be restarted for."""
+        return json.dumps(self._settings, sort_keys=True, default=str)
+
+    # ---- process ---------------------------------------------------------
+
+    @property
+    def running(self):
+        with self._lock:
+            return self._proc is not None and self._proc.poll() is None
+
+    def base_url(self):
+        with self._lock:
+            return f"http://{HOST}:{self._port}/v1" if self._port else ""
+
+    def error(self):
+        """The last thing the server printed, for a failure after it started."""
+        with self._lock:
+            log = self._log
+        return _tail(log) if log else ""
+
+    def serve(self):
+        """The base URL of a server that is up and running the current settings."""
+        ready = self._current_url()
+        if ready:
+            return ready
+        with self._starting:
+            # Somebody may have started it while this thread waited its turn.
+            ready = self._current_url()
+            if ready:
+                return ready
+            self.stop()
+            with self._lock:
+                settings, key = dict(self._settings), self._settings_key()
+            proc, port, log = self._launch(settings)
+            with self._lock:
+                self._proc, self._port, self._log, self._key = proc, port, log, key
+            return self.base_url()
+
+    def _current_url(self):
+        with self._lock:
+            up = self._proc is not None and self._proc.poll() is None
+            return (f"http://{HOST}:{self._port}/v1"
+                    if up and self._key == self._settings_key() else "")
+
+    def _launch(self, settings):
+        args = self._build(settings)        # raises LocalError when unusable
+        last = ""
+        for _ in range(3):
+            port = _free_port()
+            log = DATA_DIR / f"{self.program.name}-server.log"
+            try:
+                log.parent.mkdir(parents=True, exist_ok=True)
+                sink = open(log, "wb")
+            except OSError as exc:
+                raise LocalError(t("Could not start {name}: {error}",
+                                   name=self.program.name, error=exc)) from exc
+            try:
+                with sink:
+                    proc = subprocess.Popen(
+                        args + ["--host", HOST, "--port", str(port)],
+                        stdout=sink, stderr=subprocess.STDOUT,
+                        stdin=subprocess.DEVNULL,
+                    )
+            except OSError as exc:
+                raise LocalError(t("Could not start {name}: {error}",
+                                   name=self.program.name, error=exc)) from exc
+
+            # 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)
+            try:
+                ready = 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
+                # running would leak a loaded model with nobody to ask it
+                # anything, which is the whole failure this class is careful
+                # about elsewhere.
+                self._kill(proc)
+                self._forget()
+                raise
+            if 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():
+                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):
+        deadline = time.monotonic() + STARTUP_TIMEOUT
+        while time.monotonic() < deadline:
+            if proc.poll() is not None:
+                return False
+            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
+            time.sleep(0.1)
+        self._kill(proc)
+        return False
+
+    @staticmethod
+    def _kill(proc, gently=False):
+        """Stop a process of ours, and wait for it rather than assume."""
+        if proc is None or proc.poll() is not None:
+            return
+        if gently:
+            proc.terminate()
+            try:
+                proc.wait(timeout=5)
+                return
+            except subprocess.TimeoutExpired:
+                pass
+        proc.kill()
+        try:
+            proc.wait(timeout=5)
+        except subprocess.TimeoutExpired:
+            pass
+
+    def stop(self):
+        with self._lock:
+            proc, self._proc = self._proc, None
+            self._port, self._log, self._key = 0, "", None
+        self._kill(proc, gently=True)
+        if proc is not None:
+            self._forget()
+
+    # ---- servers a killed Dikte left behind -------------------------------
+
+    def _pid_file(self):
+        return DATA_DIR / f"{self.program.name}-server.pid"
+
+    def _remember(self, pid):
+        try:
+            path = self._pid_file()
+            path.parent.mkdir(parents=True, exist_ok=True)
+            path.write_text(str(pid))
+        except OSError:
+            pass      # the sweep is a safety net, not something to fail a run over
+
+    def _forget(self):
+        try:
+            self._pid_file().unlink()
+        except OSError:
+            pass
+
+    def _is_ours(self, pid):
+        """Whether that pid is still the server this Dikte started.
+
+        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.
+        """
+        try:
+            blob = pathlib.Path(f"/proc/{pid}/cmdline").read_bytes()
+        except OSError:
+            return False
+        return (self.program.binary.encode() in blob
+                and str(DATA_DIR).encode() in blob)
+
+    def sweep(self):
+        """Kill a server a previous Dikte left behind. True when one was found.
+
+        stop() and atexit cover every exit that gets to run code. A SIGKILL does
+        not, and neither does a session torn down from under it, and the server
+        would then sit there holding the model with nothing left alive to ask it
+        anything.
+        """
+        try:
+            pid = int(self._pid_file().read_text().strip())
+        except (OSError, ValueError):
+            return False
+        self._forget()
+        if not self._is_ours(pid):
+            return False
+        try:
+            os.kill(pid, signal.SIGTERM)
+        except OSError:
+            return False
+        return True
+
+
+# --- the two of them ------------------------------------------------------
+
+
+def _whisper_args(settings):
+    binary = program_path(WHISPER, settings["binary"])
+    if not binary:
+        raise LocalError(t("whisper.cpp is not installed. Settings → API and "
+                           "models → Download."))
+    model = whisper_model_path(settings["model"])
+    if not settings["model"] or not have_model(model):
+        raise LocalError(t("No whisper model has been downloaded yet. "
+                           "Settings → API and models → Download."))
+    args = [
+        binary, "-m", str(model),
+        "--inference-path", INFERENCE_PATH,
+        # Whatever language the request does not name. api.py leaves the field
+        # out when the language is "auto", and the server's own default is
+        # English rather than detection.
+        "-l", "auto",
+        # Stock phrases invented for near-silence come from non-speech tokens,
+        # and verbose_json otherwise pays for a language probability sweep
+        # nothing here reads.
+        "-sns", "-nlp",
+    ]
+    if int(settings["threads"]) > 0:
+        args += ["-t", str(int(settings["threads"]))]
+    if not settings["gpu"]:
+        args.append("-ng")
+    return args
+
+
+def _llm_args(settings):
+    binary = program_path(LLAMA, settings["binary"])
+    if not binary:
+        raise LocalError(t("llama.cpp is not installed. Settings → API and "
+                           "models → Download."))
+    model = llm_model_path(settings["model"])
+    if not settings["model"] or not have_model(model):
+        raise LocalError(t("No local cleanup model has been downloaded yet. "
+                           "Settings → API and models → Download."))
+    args = [binary, "-m", str(model), "-c", str(int(settings["context"]))]
+    # All of them, or as many as fit: llama.cpp stops offloading when the card
+    # is full rather than failing, and a build with no GPU backend ignores it.
+    args += ["-ngl", "99" if settings["gpu"] else "0"]
+    if int(settings["threads"]) > 0:
+        args += ["-t", str(int(settings["threads"]))]
+    return args
+
+
+whisper = Server(WHISPER, _whisper_args, {
+    "model": "",
+    "threads": 0,
+    "gpu": True,
+    "binary": "",
+})
+
+llm = Server(LLAMA, _llm_args, {
+    "model": "",
+    "threads": 0,
+    "gpu": True,
+    "binary": "",
+    # A dictation and its prompt are short. This is sized for the longest
+    # cleanup block rather than for a conversation, and it is what the model
+    # costs in memory beyond its own weights.
+    "context": 8192,
+})
+
+SERVERS = (whisper, llm)
+
+
+def sweep():
+    """Clean up after a Dikte that was killed outright. True when one was found."""
+    return any([server.sweep() for server in SERVERS])
+
+
+def stop_all():
+    for server in SERVERS:
+        server.stop()
+
+
+# Dikte stops the servers itself on quit and on restart; this catches the paths
+# that skip that, such as an unhandled exception on the way out.
+atexit.register(stop_all)

hub.py

@@ -0,0 +1,186 @@
+"""Where the programs and the models come from: GitHub releases and Hugging Face.
+
+Both answer plain JSON over HTTPS without a key, and both publish a sha256 for
+every file they hand out: GitHub as the asset digest, Hugging Face as the LFS
+object id. Nothing that lands on disk is trusted for having arrived, which
+matters more here than it usually would, because half of what is fetched is a
+program Dikte then runs.
+
+The lists are read rather than kept. A model catalogue written into the source
+means a release of Dikte for every new model, and a pinned whisper.cpp version
+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.
+"""
+
+import collections
+import json
+import os
+import pathlib
+import time
+import urllib.error
+import urllib.parse
+import urllib.request
+
+from i18n import t
+
+GITHUB_API = "https://api.github.com"
+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")
+# 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
+
+# `sha256` is empty for the few files neither side stores in LFS; those are the
+# small ones, and a checksum is only worth having where there is something to
+# check.
+Item = collections.namedtuple("Item", "name url size sha256")
+Repo = collections.namedtuple("Repo", "id downloads updated")
+
+
+class HubError(Exception):
+    pass
+
+
+def _get(url, timeout=20):
+    request = urllib.request.Request(url, headers={
+        "User-Agent": USER_AGENT,
+        "Accept": "application/json",
+    })
+    try:
+        with urllib.request.urlopen(request, timeout=timeout) as response:
+            return json.loads(response.read().decode("utf-8"))
+    except urllib.error.HTTPError as exc:
+        exc.close()   # it holds the response body open until it is collected
+        raise HubError(t("{url} answered HTTP {code}.",
+                         url=urllib.parse.urlsplit(url).netloc, code=exc.code)) from exc
+    except urllib.error.URLError as exc:
+        raise HubError(t("Could not reach {url}: {error}",
+                         url=urllib.parse.urlsplit(url).netloc,
+                         error=exc.reason)) from exc
+    except (ValueError, OSError) as exc:
+        raise HubError(t("Could not read the answer from {url}: {error}",
+                         url=urllib.parse.urlsplit(url).netloc, error=exc)) from exc
+
+
+def _cache_file(key):
+    safe = "".join(c if c.isalnum() or c in "-._" else "-" for c in key)
+    return CACHE_DIR / f"{safe}.json"
+
+
+def _read_cache(key, ttl):
+    """What was stored under this key, or None. `ttl` of 0 ignores the age."""
+    path = _cache_file(key)
+    try:
+        age = time.time() - path.stat().st_mtime
+        if ttl and age > ttl:
+            return None
+        return json.loads(path.read_text(encoding="utf-8"))
+    except (OSError, ValueError):
+        return None
+
+
+def _write_cache(key, payload):
+    try:
+        CACHE_DIR.mkdir(parents=True, exist_ok=True)
+        _cache_file(key).write_text(json.dumps(payload), encoding="utf-8")
+    except OSError:
+        pass          # a cache that cannot be written is not a failed lookup
+
+
+def _fetch(key, url, ttl=CACHE_TTL, refresh=False):
+    """The JSON at `url`, from the cache when it is fresh enough.
+
+    A lookup that fails falls back to the cache however old it is: an offline
+    settings window that shows yesterday's list is worth a great deal more than
+    one that shows an error.
+    """
+    if not refresh:
+        cached = _read_cache(key, ttl)
+        if cached is not None:
+            return cached
+    try:
+        payload = _get(url)
+    except HubError:
+        stale = _read_cache(key, 0)
+        if stale is not None:
+            return stale
+        raise
+    _write_cache(key, payload)
+    return payload
+
+
+def _digest(value):
+    """GitHub writes its digests as "sha256:…"; Hugging Face writes the hash."""
+    value = (value or "").strip()
+    return value.split(":", 1)[1] if value.startswith("sha256:") else value
+
+
+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}"
+    data = _fetch(f"gh-{repo}-{tag or 'latest'}",
+                  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
+
+
+def files(repo, revision="main", refresh=False):
+    """[Item] for every file in a Hugging Face repository.
+
+    The size is there whether or not the file is in LFS; the hash is only there
+    when it is, which for anything worth downloading it always is.
+    """
+    data = _fetch(f"hf-tree-{repo}-{revision}",
+                  f"{HF_API}/models/{repo}/tree/{revision}?recursive=true",
+                  refresh=refresh)
+    if not isinstance(data, list):
+        raise HubError(t("{repo} did not return a file list.", repo=repo))
+    out = []
+    for entry in data:
+        if entry.get("type") != "file":
+            continue
+        path = entry.get("path") or ""
+        lfs = entry.get("lfs") or {}
+        out.append(Item(
+            path,
+            f"{HF_FILES}/{repo}/resolve/{revision}/{urllib.parse.quote(path)}",
+            int(lfs.get("size") or entry.get("size") or 0),
+            _digest(lfs.get("oid") or lfs.get("sha256")),
+        ))
+    return out
+
+
+def repos(author="", search="", limit=40, refresh=False):
+    """[Repo] of GGUF repositories, newest first.
+
+    Filtered by author on purpose. Hugging Face's own trending list is open to
+    everyone and reads like it: asking it for the popular GGUF today answers
+    with a wall of roleplay merges, which is not what a dictation transcript
+    wants cleaning up. An author is a small enough thing to trust and a large
+    enough one to keep the list current without Dikte being updated.
+    """
+    query = {"filter": "gguf", "sort": "lastModified", "direction": "-1",
+             "limit": str(limit)}
+    if author:
+        query["author"] = author
+    if search:
+        query["search"] = search
+    url = f"{HF_API}/models?{urllib.parse.urlencode(query)}"
+    data = _fetch(f"hf-models-{author}-{search}-{limit}", url, refresh=refresh)
+    if not isinstance(data, list):
+        raise HubError(t("Hugging Face did not return a model list."))
+    return [Repo(m.get("id") or "", int(m.get("downloads") or 0),
+                 m.get("lastModified") or "")
+            for m in data if m.get("id")]

install.sh

@@ -132,5 +132,6 @@ fi
 
 echo
 ok "Done. Start it with:  dikte"
-say "The settings window opens on first run; add an OpenAI, Groq or OpenRouter key."
+say "The settings window opens on first run: download a speech model, or add"
+say "an OpenAI or OpenRouter key instead."
 echo

settings_ui.py

@@ -19,6 +19,7 @@ import audio
 import cleanup
 import config as cfg
 import filetranscribe
+import ggml
 import hotkey
 import ipc
 import meeting
@@ -31,8 +32,10 @@ LANGUAGES = [
     ("German", "de"), ("French", "fr"), ("Spanish", "es"), ("Arabic", "ar"),
 ]
 CORNERS = ["bottom-left", "bottom-right", "top-left", "top-right"]
-# The provider box offers what config knows how to reach, in that order.
+# The provider box offers what config knows how to reach, this machine first.
-TRANSCRIBE_PROVIDERS = [(who.service, name) for name, who in cfg.TRANSCRIBERS.items()]
+TRANSCRIBE_PROVIDERS = ([("This machine (whisper.cpp)", "local")]
+                        + [(who.service, name)
+                           for name, who in cfg.TRANSCRIBERS.items()])
 # Starting points for the model box; "Fetch model list" replaces them with
 # whatever the provider offers today.
 TRANSCRIBE_MODELS = {
@@ -50,11 +53,12 @@ CLEANUP_MODELS = [
     "google/gemini-2.5-flash-lite", "anthropic/claude-haiku-4.5",
     "openai/gpt-5-mini", "meta-llama/llama-3.3-70b-instruct",
 ]
-# The same two CLIs the agent can run on, doing the smaller job instead. They
+# In the order they answer in. A request to OpenRouter is over in a second, a
-# are offered second: a request to OpenRouter is over in a second, and a CLI
+# model here takes a little longer and costs nothing, and the two CLIs the agent
-# opens a session first.
+# can run on open a whole session to do the smaller job.
 CLEANUP_PROVIDERS = [
-    ("OpenRouter", "openrouter"), ("Claude Code", "claude"), ("Codex", "codex"),
+    ("OpenRouter", "openrouter"), ("This machine (llama.cpp)", "local"),
+    ("Claude Code", "claude"), ("Codex", "codex"),
 ]
 # Cleaning up a sentence is the lightest thing either of them will ever be
 # asked, so the small model comes first.
@@ -116,6 +120,338 @@ AUDIO_FILTER = ("*.mp3 *.wav *.m4a *.ogg *.opus *.flac *.aac *.wma "
                 "*.mp4 *.mkv *.webm *.mov *.avi")
 
 
+class LocalModelBox(QGroupBox):
+    """The program, the model, and the two downloads that put them there.
+
+    One class for whisper.cpp and llama.cpp, because the job is the same one
+    twice: say whether the program is here, offer the models somebody publishes,
+    fetch the chosen one, and stay usable while a gigabyte arrives. Nothing is
+    listed in the source; `repos` and `models` are asked at the moment the box is
+    opened, so a model published this morning is in the list this afternoon.
+    """
+
+    _listed = pyqtSignal(list, str)
+    _quants = pyqtSignal(list, str)
+    # qint64 rather than int, which is C++'s 32-bit one: a 2.3 GB model is more
+    # than fits in it, and the count comes out the far side negative.
+    _progress = pyqtSignal("qint64", "qint64")
+    _finished = pyqtSignal(str, str)
+    _installed = pyqtSignal(str, str)
+
+    changed = pyqtSignal()
+
+    def __init__(self, program, title, models, model_path, repos=None, parent=None):
+        super().__init__(title, parent)
+        self.program = program
+        self._models = models          # () -> [hub.Item], or (repo) -> [hub.Item]
+        self._model_path = model_path  # (name) -> Path
+        self._repos = repos            # None, or () -> [repo id]
+        self._downloading = False
+        self._pending = False
+        self._stop = False
+        self._wanted = ""              # the model to select once a list arrives
+
+        form = QFormLayout(self)
+
+        self.program_label = QLabel("")
+        self.program_label.setWordWrap(True)
+        self.install_button = QPushButton(t("Download"))
+        self.install_button.clicked.connect(self._install_program)
+        form.addRow(t("Program"), self._side_by_side(self.program_label,
+                                                     self.install_button))
+
+        if self._repos is not None:
+            self.repo = QComboBox()
+            self.repo.setEditable(True)
+            self.repo.setToolTip(t("A Hugging Face repository of GGUF files. The "
+                                   "list is fetched; any other one can be typed in."))
+            self.repo.currentTextChanged.connect(self._repo_changed)
+            form.addRow(t("Publisher"), self.repo)
+
+        self.model = QComboBox()
+        self.download_button = QPushButton(t("Download"))
+        self.download_button.clicked.connect(self._download)
+        self.delete_button = QPushButton(t("Delete"))
+        self.delete_button.clicked.connect(self._delete)
+        form.addRow(t("Model"), self._side_by_side(self.model,
+                                                   self.download_button,
+                                                   self.delete_button))
+        self.model.currentIndexChanged.connect(self._model_changed)
+
+        self.status = QLabel("")
+        self.status.setWordWrap(True)
+        form.addRow(self.status)
+
+        self._listed.connect(self._on_listed)
+        self._quants.connect(self._on_listed)
+        self._progress.connect(self._on_progress)
+        self._finished.connect(self._on_finished)
+        self._installed.connect(self._on_installed)
+
+    @staticmethod
+    def _fit_popup(combo):
+        """Let the list that drops down be as wide as its longest row.
+
+        A combo box hands its own width to the list under it and elides
+        whatever does not fit, which lands in the middle of the name:
+        `ggml-org/Qwen....7B-Base-GGUF` is not a model anybody can choose
+        between. The box itself stays the width the form gave it.
+        """
+        view = combo.view()
+        view.setTextElideMode(Qt.TextElideMode.ElideNone)
+        metrics = combo.fontMetrics()
+        widest = max((metrics.horizontalAdvance(combo.itemText(row))
+                      for row in range(combo.count())), default=0)
+        # Room for the frame and for a scroll bar, which a long list will have.
+        view.setMinimumWidth(widest + view.verticalScrollBar().sizeHint().width() + 24)
+
+    @staticmethod
+    def _side_by_side(*widgets):
+        layout = QHBoxLayout()
+        layout.setContentsMargins(0, 0, 0, 0)
+        for index, widget in enumerate(widgets):
+            layout.addWidget(widget, 1 if index == 0 else 0)
+        holder = QWidget()
+        holder.setLayout(layout)
+        return holder
+
+    # ---- what is here ----------------------------------------------------
+
+    def selected(self):
+        return self.model.currentData() or ""
+
+    def repository(self):
+        return self.repo.currentText().strip() if self._repos is not None else ""
+
+    def load(self, model, repo=""):
+        """Show what is stored. What else is on offer is asked for on the way up.
+
+        Nothing is fetched here: building the settings window is not the same as
+        opening it, and a list nobody is looking at is not worth a request. What
+        is already on this disk is shown straight away either way.
+        """
+        self._wanted = model
+        self._pending = True
+        self._show_program()
+        if self._repos is not None:
+            self.repo.blockSignals(True)
+            self.repo.clear()
+            self.repo.addItems(list(ggml.SUGGESTED_LLM))
+            self.repo.setCurrentText(repo or ggml.SUGGESTED_LLM[0])
+            self.repo.blockSignals(False)
+            self._fit_popup(self.repo)
+        self._fill_models([])
+
+    def showEvent(self, event):
+        super().showEvent(event)
+        if self._pending:
+            self._pending = False
+            if self._repos is not None:
+                self._fill_repos(self.repository())
+            self._fetch_models(self.repository())
+
+    def _show_program(self):
+        path = ggml.program_path(self.program)
+        if not path:
+            self.program_label.setText(t("Not installed."))
+            self.install_button.setVisible(True)
+            return
+        self.install_button.setVisible(not ggml.installed_program(self.program)
+                                       and not ggml.system_program(self.program))
+        if ggml.system_program(self.program):
+            # Worth saying which one is running: a distribution package is built
+            # for this machine and may reach the graphics card, while the
+            # released binaries carry processor backends only.
+            self.program_label.setText(t("Installed on the system: {path}", path=path))
+        else:
+            self.program_label.setText(
+                t("Downloaded, version {version}.",
+                  version=ggml.installed_version(self.program) or "?"))
+
+    # ---- the lists -------------------------------------------------------
+
+    def _fill_repos(self, current):
+        def work():
+            self._listed.emit([("repos", ggml.llm_repos())], "")
+
+        threading.Thread(target=work, daemon=True).start()
+
+    def _repo_changed(self):
+        if not self._downloading:
+            self._fetch_models(self.repository())
+
+    def _fetch_models(self, repo=""):
+        self.status.setText(t("Fetching the model list…"))
+
+        def work():
+            try:
+                found = self._models(repo) if self._repos is not None else self._models()
+                self._quants.emit([("models", found)], "")
+            except ggml.LocalError as exc:
+                self._quants.emit([], str(exc))
+
+        threading.Thread(target=work, daemon=True).start()
+
+    def _on_listed(self, payload, error):
+        if error:
+            self.status.setText(error)
+            self._refresh_buttons()
+            return
+        kind, found = payload[0]
+        if kind == "repos":
+            current = self.repo.currentText()
+            self.repo.blockSignals(True)
+            self.repo.clear()
+            self.repo.addItems(found)
+            self.repo.setCurrentText(current)
+            self.repo.blockSignals(False)
+            self._fit_popup(self.repo)
+            return
+        self._fill_models(found)
+
+    def _fill_models(self, items):
+        """One row per model, saying what it weighs and whether it is here."""
+        wanted = self._wanted or self.selected()
+        here = [name for name in (self._model_path(i.name).name for i in items)]
+        self.model.blockSignals(True)
+        self.model.clear()
+        for item, name in zip(items, here):
+            mark = (t("downloaded") if ggml.have_model(self._model_path(item.name))
+                    else ggml.human_size(item.size))
+            self.model.addItem(f"{name}  ({mark})", name)
+            self.model.setItemData(self.model.count() - 1, item, Qt.ItemDataRole.UserRole + 1)
+        # A model that was downloaded and then dropped from the list upstream is
+        # still on this disk and still works, so it stays on offer.
+        for name in self._on_disk():
+            if self.model.findData(name) < 0:
+                self.model.addItem(f"{name}  ({t('downloaded')})", name)
+        # And one that is chosen but not here, because the file was deleted from
+        # underneath or the settings came from another machine, stays chosen:
+        # Save reads this box, and a row missing here would quietly empty the
+        # setting rather than showing that the model needs downloading again.
+        if wanted and self.model.findData(wanted) < 0:
+            self.model.addItem(f"{wanted}  ({t('not downloaded')})", wanted)
+        index = self.model.findData(wanted)
+        self.model.setCurrentIndex(max(index, 0))
+        self.model.blockSignals(False)
+        self._fit_popup(self.model)
+        self._wanted = ""
+        self._model_changed()
+
+    def _on_disk(self):
+        return (ggml.installed_whisper_models() if self.program is ggml.WHISPER
+                else ggml.installed_llm_models())
+
+    # ---- fetching --------------------------------------------------------
+
+    def _install_program(self):
+        self.install_button.setEnabled(False)
+        self.program_label.setText(t("Downloading…"))
+
+        def work():
+            try:
+                ggml.install_program(self.program, on_progress=self._report)
+                self._installed.emit("", "")
+            except ggml.LocalError as exc:
+                self._installed.emit("", str(exc))
+
+        threading.Thread(target=work, daemon=True).start()
+
+    def _on_installed(self, _, error):
+        self.install_button.setEnabled(True)
+        self._show_program()
+        if error:
+            self.program_label.setText(error)
+        self.changed.emit()
+
+    def _current_item(self):
+        return self.model.currentData(Qt.ItemDataRole.UserRole + 1)
+
+    def _download(self):
+        if self._downloading:
+            self._stop = True
+            return
+        item = self._current_item()
+        if item is None:
+            return
+        self._downloading, self._stop = True, False
+        self._refresh_buttons()
+
+        def work():
+            try:
+                landed = ggml.download(item, self._model_path(item.name),
+                                       on_progress=self._report,
+                                       should_stop=lambda: self._stop)
+                self._finished.emit(item.name if landed else "", "")
+            except ggml.LocalError as exc:
+                self._finished.emit("", str(exc))
+
+        threading.Thread(target=work, daemon=True).start()
+
+    def _report(self, done, total):
+        self._progress.emit(done, total)
+
+    def _on_progress(self, done, total):
+        share = f" ({done * 100 // total}%)" if total else ""
+        text = t("Downloading: {done} of {total}{share}",
+                 done=ggml.human_size(done), total=ggml.human_size(total or done),
+                 share=share)
+        if self._downloading:
+            self.status.setText(text)
+        else:
+            self.program_label.setText(text)
+
+    def _on_finished(self, name, error):
+        self._downloading = False
+        if error:
+            self.status.setText(error)
+        elif not name:
+            self.status.setText(t("Download stopped."))
+        self._fill_models_from_current()
+        self.changed.emit()
+
+    def _fill_models_from_current(self):
+        """Redraw the rows without asking anybody anything again."""
+        items = [self.model.itemData(i, Qt.ItemDataRole.UserRole + 1)
+                 for i in range(self.model.count())]
+        self._wanted = self.selected()
+        self._fill_models([i for i in items if i is not None])
+
+    def _delete(self):
+        name = self.selected()
+        if not name or not ggml.have_model(self._model_path(name)):
+            return
+        if QMessageBox.question(self, t("Delete model"),
+                                t("Delete {name} from this machine?", name=name)) \
+                != QMessageBox.StandardButton.Yes:
+            return
+        try:
+            ggml.delete_model(self._model_path(name))
+        except ggml.LocalError as exc:
+            self.status.setText(str(exc))
+        self._fill_models_from_current()
+        self.changed.emit()
+
+    def _model_changed(self):
+        self._refresh_buttons()
+        self.changed.emit()
+
+    def _refresh_buttons(self):
+        name = self.selected()
+        here = bool(name) and ggml.have_model(self._model_path(name))
+        self.delete_button.setEnabled(here and not self._downloading)
+        self.download_button.setText(t("Stop") if self._downloading else t("Download"))
+        self.download_button.setEnabled(self._downloading or (bool(name) and not here))
+        if self._downloading:
+            return
+        if not name:
+            self.status.setText(t("Nothing downloaded yet."))
+        elif here:
+            self.status.setText(t("Ready: {name}.", name=name))
+        else:
+            self.status.setText(t("{name} has not been downloaded yet.", name=name))
+
+
 class SettingsWindow(QDialog):
     applied = pyqtSignal()
 
@@ -143,9 +479,9 @@ class SettingsWindow(QDialog):
         self.setWindowTitle(t("Dikte Settings"))
         self.resize(680, 640)
 
-        tabs = QTabWidget(self)
+        tabs = self.tabs = QTabWidget(self)
         tabs.addTab(self._general_tab(), t("General"))
-        tabs.addTab(self._api_tab(), t("API and models"))
+        self.api_tab_index = tabs.addTab(self._api_tab(), t("API and models"))
         tabs.addTab(self._prompt_tab(), t("Cleanup rules"))
         tabs.addTab(self._assistant_tab(), t("Agent"))
         tabs.addTab(self._meeting_tab(), t("Meeting"))
@@ -176,6 +512,10 @@ class SettingsWindow(QDialog):
             self.meetings.finished.connect(self._on_minutes_finished)
             self.meetings.failed.connect(self._on_minutes_failed)
         self._load()
+        # On a machine where nothing can transcribe yet, this window was opened
+        # because of that, so open it on the tab that fixes it.
+        if not conf.transcribe_ready():
+            self.tabs.setCurrentIndex(self.api_tab_index)
 
     # ---- tabs ----------------------------------------------------------
 
@@ -274,20 +614,52 @@ class SettingsWindow(QDialog):
         stt_form = QFormLayout(stt)
         self.transcribe_provider = QComboBox()
         for label, value in TRANSCRIBE_PROVIDERS:
-            self.transcribe_provider.addItem(label, value)
+            self.transcribe_provider.addItem(t(label), value)
         stt_form.addRow(t("Provider"), self.transcribe_provider)
 
+        # A hosted provider takes any model id that is typed at it; the local
+        # one offers what has been published. One row each, and only the rows of
+        # whoever is chosen are on screen.
+        self.stt_form = stt_form
         self.transcribe_model = QComboBox()
         self.transcribe_model.setEditable(True)
         self.refresh_transcribe_models = QPushButton(t("Fetch model list"))
         self.refresh_transcribe_models.clicked.connect(self._load_transcribe_models)
-        stt_form.addRow(t("Model"),
+        self.transcribe_model_row = self._row(self.transcribe_model,
-                        self._row(self.transcribe_model, self.refresh_transcribe_models))
+                                              self.refresh_transcribe_models)
+        stt_form.addRow(t("Model"), self.transcribe_model_row)
         # A spanning row: in the narrow field column a wrapped label gets a
         # height that fits one line, and the rest of the text is cut off.
         self.transcribe_status = QLabel("")
         self.transcribe_status.setWordWrap(True)
         stt_form.addRow(self.transcribe_status)
+
+        self.local_whisper = LocalModelBox(
+            ggml.WHISPER, t("On this machine"),
+            ggml.whisper_models, ggml.whisper_model_path)
+        stt_form.addRow(self.local_whisper)
+
+        self.local_gpu = QCheckBox(t("Use the graphics card"))
+        self.local_gpu.setToolTip(
+            t("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."))
+        self.local_preload = QCheckBox(t("Load the model when Dikte starts"))
+        self.local_preload.setToolTip(
+            t("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."))
+        self.local_threads = QSpinBox()
+        self.local_threads.setRange(0, 64)
+        self.local_threads.setSpecialValueText(t("Automatic"))
+        self.local_options = QWidget()
+        options_form = QFormLayout(self.local_options)
+        options_form.setContentsMargins(0, 0, 0, 0)
+        options_form.addRow("", self.local_gpu)
+        options_form.addRow("", self.local_preload)
+        options_form.addRow(t("Threads"), self.local_threads)
+        stt_form.addRow(self.local_options)
+
         self.transcribe_provider.currentIndexChanged.connect(self._provider_changed)
         outer.addWidget(stt)
 
@@ -301,9 +673,10 @@ class SettingsWindow(QDialog):
             self.cleanup_provider.addItem(t(label), value)
         self.cleanup_provider.setToolTip(t(
             "OpenRouter is the quickest and the only one that needs nothing "
-            "installed. Claude Code and Codex clean up on the subscription you "
+            "installed. llama.cpp runs here, on a model downloaded below. Claude "
-            "already have, without a second key, and take a few seconds longer "
+            "Code and Codex clean up on the subscription you already have, "
-            "because each one opens a session to do it."
+            "without a second key, and take a few seconds longer because each "
+            "one opens a session to do it."
         ))
         self.cleanup_provider.currentIndexChanged.connect(self._cleanup_provider_changed)
         orr_form.addRow(t("Runs on"), self.cleanup_provider)
@@ -342,6 +715,33 @@ class SettingsWindow(QDialog):
         self.models_label = QLabel(t("Runs on OpenRouter."))
         self.models_label.setWordWrap(True)
         orr_form.addRow(self.models_label)
+
+        self.local_llm = LocalModelBox(
+            ggml.LLAMA, t("On this machine"),
+            ggml.llm_quants, ggml.llm_model_path, repos=ggml.llm_repos)
+        orr_form.addRow(self.local_llm)
+
+        self.local_llm_gpu = QCheckBox(t("Use the graphics card"))
+        self.local_llm_preload = QCheckBox(t("Load the model when Dikte starts"))
+        self.local_llm_preload.setToolTip(
+            t("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."))
+        self.local_llm_reasoning = QComboBox()
+        for label, value in REASONING_LEVELS:
+            self.local_llm_reasoning.addItem(t(label), value)
+        self.local_llm_reasoning.setToolTip(
+            t("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."))
+        self.local_llm_options = QWidget()
+        llm_form = QFormLayout(self.local_llm_options)
+
+        llm_form.setContentsMargins(0, 0, 0, 0)
+        llm_form.addRow("", self.local_llm_gpu)
+        llm_form.addRow("", self.local_llm_preload)
+        llm_form.addRow(t("Thinking"), self.local_llm_reasoning)
+        orr_form.addRow(self.local_llm_options)
+
         outer.addWidget(orr)
         outer.addStretch(1)
         return page
@@ -985,6 +1385,11 @@ class SettingsWindow(QDialog):
         self._shown_provider = ""
         self._select_data(self.transcribe_provider, conf["transcribe_provider"])
         self._provider_changed()  # selecting index 0 fires no signal
+        self.local_gpu.setChecked(conf["local_gpu"])
+        self.local_preload.setChecked(conf["local_preload"])
+        self.local_threads.setValue(int(conf["local_threads"]))
+        self.local_whisper.load(conf["local_model"])
+
         self.cleanup_enabled.setChecked(conf["cleanup_enabled"])
         self.cleanup_model.setCurrentText(conf["cleanup_model"])
         self.cleanup_claude_model.setCurrentText(conf["cleanup_claude_model"])
@@ -994,6 +1399,10 @@ class SettingsWindow(QDialog):
         self._select_data(self.cleanup_provider, conf["cleanup_provider"])
         self._cleanup_provider_changed()  # selecting index 0 fires no signal
         self._select_data(self.cleanup_reasoning, conf["cleanup_reasoning"])
+        self.local_llm_gpu.setChecked(conf["local_llm_gpu"])
+        self.local_llm_preload.setChecked(conf["local_llm_preload"])
+        self._select_data(self.local_llm_reasoning, conf["local_llm_reasoning"])
+        self.local_llm.load(conf["local_llm_model"], conf["local_llm_repo"])
         self.cleanup_prompt.setPlainText(conf["cleanup_prompt"] or cfg.default_cleanup_prompt())
         self.file_cleanup_prompt.setPlainText(
             conf["file_cleanup_prompt"] or cfg.default_file_cleanup_prompt()
@@ -1063,12 +1472,17 @@ class SettingsWindow(QDialog):
         conf["filter_hallucinations"] = self.filter_hallucinations.isChecked()
         conf["keep_audio"] = self.keep_audio.isChecked()
 
-        provider = self.transcribe_provider.currentData() or "openai"
+        provider = self.transcribe_provider.currentData() or "local"
-        self._models[provider] = self.transcribe_model.currentText().strip()
+        if provider in self._models:
+            self._models[provider] = self.transcribe_model.currentText().strip()
         conf["transcribe_provider"] = provider
         for name, who in cfg.TRANSCRIBERS.items():
             conf[who.key] = self._key_fields[name].text().strip()
             conf[who.model] = self._models[name].strip() or cfg.DEFAULTS[who.model]
+        conf["local_model"] = self.local_whisper.selected()
+        conf["local_gpu"] = self.local_gpu.isChecked()
+        conf["local_preload"] = self.local_preload.isChecked()
+        conf["local_threads"] = self.local_threads.value()
 
         conf["cleanup_enabled"] = self.cleanup_enabled.isChecked()
         conf["cleanup_provider"] = self.cleanup_provider.currentData() or "openrouter"
@@ -1080,6 +1494,11 @@ class SettingsWindow(QDialog):
             "" if codex_cleanup_model == t("Codex's own default") else codex_cleanup_model
         )
         conf["cleanup_reasoning"] = self.cleanup_reasoning.currentData() or ""
+        conf["local_llm_model"] = self.local_llm.selected()
+        conf["local_llm_repo"] = self.local_llm.repository()
+        conf["local_llm_gpu"] = self.local_llm_gpu.isChecked()
+        conf["local_llm_preload"] = self.local_llm_preload.isChecked()
+        conf["local_llm_reasoning"] = self.local_llm_reasoning.currentData() or ""
 
         # Store an empty prompt when it matches the default, so switching the
         # interface language also switches the prompt language.
@@ -1163,10 +1582,17 @@ class SettingsWindow(QDialog):
 
     def _provider_changed(self):
         """Swap the model box over to the newly chosen provider's own model."""
-        if self._shown_provider:
+        if self._shown_provider in TRANSCRIBE_MODELS:
             self._models[self._shown_provider] = self.transcribe_model.currentText().strip()
-        provider = self.transcribe_provider.currentData() or "openai"
+        provider = self.transcribe_provider.currentData() or "local"
         self._shown_provider = provider
+        local = provider == "local"
+        self.stt_form.setRowVisible(self.transcribe_model_row, not local)
+        self.stt_form.setRowVisible(self.transcribe_status, not local)
+        self.stt_form.setRowVisible(self.local_whisper, local)
+        self.stt_form.setRowVisible(self.local_options, local)
+        if local:
+            return
         self.transcribe_model.clear()
         self.transcribe_model.addItems(TRANSCRIBE_MODELS[provider])
         self.transcribe_model.setCurrentText(self._models[provider])
@@ -1404,9 +1830,15 @@ class SettingsWindow(QDialog):
                                         provider == "claude")
         self.cleanup_form.setRowVisible(self.cleanup_codex_model,
                                         provider == "codex")
+        self.cleanup_form.setRowVisible(self.cleanup_reasoning,
+                                        provider != "local")
+        self.cleanup_form.setRowVisible(self.local_llm, provider == "local")
+        self.cleanup_form.setRowVisible(self.local_llm_options, provider == "local")
         binary = cleanup.executable(provider)
         found = shutil.which(binary) if binary else ""
-        if not binary:
+        if provider == "local":
+            self.models_label.setText(t("Runs on this machine, on llama.cpp."))
+        elif not binary:
             self.models_label.setText(t("Runs on OpenRouter."))
         elif found:
             self.models_label.setText(t("Found: {path}", path=found))

tests/support.py

@@ -18,6 +18,7 @@ import sys
 import tempfile
 import unittest
 import urllib.error
+import urllib.request
 import wave
 from unittest import mock
 
@@ -40,6 +41,12 @@ linux_only = unittest.skipUnless(
 )
 
 
+def _no_network(*args, **kwargs):
+    raise AssertionError(
+        "a test reached the network; wrap the call in support.fake_urlopen"
+    )
+
+
 def _no_exec(*args, **kwargs):
     raise AssertionError(
         "a test reached os.execv, which would replace the test process with the "
@@ -78,6 +85,11 @@ class DikteTest(unittest.TestCase):
         # with it and hang, so it fails loudly here instead.
         self.patch_attr(os, "execv", _no_exec)
 
+        # Every way out of here goes through urllib, so closing it is enough to
+        # keep the suite offline. A test that means to answer a request patches
+        # this again through fake_urlopen.
+        self.patch_attr(urllib.request, "urlopen", _no_network)
+
     # ---- helpers ---------------------------------------------------------
 
     def path(self, *parts):

tests/test_api.py

@@ -10,6 +10,7 @@ import os
 import unittest
 
 import api
+import ggml
 from tests.support import (
     DikteTest,
     fake_urlopen,
@@ -481,3 +482,99 @@ class ModelLists(DikteTest):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+class FakeServer:
+    """A ggml.Server as far as api.py is concerned."""
+
+    def __init__(self, url="http://127.0.0.1:9999/v1", fails="", log=""):
+        self.url = url
+        self.fails = fails
+        self.log = log
+        self.starts = 0
+
+    def serve(self):
+        self.starts += 1
+        if self.fails:
+            raise ggml.LocalError(self.fails)
+        return self.url
+
+    def error(self):
+        return self.log
+
+
+LOCAL = api.Target("local", "Local whisper", "", "", "ggml-base.bin")
+
+
+class TranscribeHere(DikteTest):
+    def setUp(self):
+        super().setUp()
+        self.wav = str(self.path("clip.wav"))
+        os.makedirs(self.root, exist_ok=True)
+        with open(self.wav, "wb") as fh:
+            fh.write(b"RIFFfake")
+        self.server = FakeServer()
+        self.patch_attr(ggml, "whisper", self.server)
+
+    def test_the_address_comes_from_the_server_it_starts(self):
+        with fake_urlopen({"text": "hello"}) as calls:
+            api.transcribe(LOCAL, self.wav)
+        self.assertEqual(self.server.starts, 1)
+        self.assertEqual(calls[0].full_url,
+                         "http://127.0.0.1:9999/v1/audio/transcriptions")
+
+    def test_nothing_local_is_authorised(self):
+        with fake_urlopen({"text": "hello"}) as calls:
+            api.transcribe(LOCAL, self.wav)
+        self.assertNotIn("Authorization", calls[0].headers)
+
+    def test_a_server_that_will_not_start_is_the_error_shown(self):
+        self.patch_attr(ggml, "whisper", FakeServer(fails="no model downloaded"))
+        with self.assertRaises(api.ApiError) as caught:
+            api.transcribe(LOCAL, self.wav)
+        self.assertIn("no model downloaded", str(caught.exception))
+
+    def test_a_server_that_dies_mid_request_says_what_it_printed(self):
+        self.patch_attr(ggml, "whisper", FakeServer(log="out of memory"))
+        with fake_urlopen(url_error("connection reset")):
+            with self.assertRaises(api.ApiError) as caught:
+                api.transcribe(LOCAL, self.wav)
+        self.assertIn("out of memory", str(caught.exception))
+
+    def test_the_hint_reaches_whisper_as_its_initial_prompt(self):
+        with fake_urlopen({"text": "hi"}) as calls:
+            api.transcribe(LOCAL, self.wav, prompt="Dikte, Paraşüt")
+        self.assertEqual(multipart_fields(calls[0])["prompt"], "Dikte, Paraşüt")
+
+    def test_a_word_broken_over_two_lines_is_put_back_together(self):
+        # whisper.cpp cuts on tokens and writes one segment per line, which in
+        # Turkish lands inside a word about as often as between two.
+        with fake_urlopen({"text": "Onlar akraba değ\niller. Ve\n devamı."}):
+            # The line break inside a word leaves nothing in its place; the
+            # one between two words is where whisper's own leading space is.
+            self.assertEqual(api.transcribe(LOCAL, self.wav),
+                             "Onlar akraba değiller. Ve devamı.")
+
+    def test_a_local_timeout_is_not_a_hosted_one(self):
+        # Nothing is being spent but time, and a long file on a machine without
+        # a graphics card takes a good deal of it.
+        with fake_urlopen({"text": "hi"}):
+            api.transcribe(LOCAL, self.wav, timeout=300)
+        self.assertGreaterEqual(api.LOCAL_TIMEOUT, 600)
+
+    def test_segments_that_continue_a_word_are_merged(self):
+        reply = {"segments": [
+            {"start": 0.0, "end": 1.0, "text": " Onlar akraba değ"},
+            {"start": 1.0, "end": 1.4, "text": "iller."},
+            {"start": 2.0, "end": 3.0, "text": " Başka bir cümle."},
+        ]}
+        with fake_urlopen(reply):
+            out = api.transcribe_segments(LOCAL, self.wav)
+        self.assertEqual([text for _, _, text in out],
+                         ["Onlar akraba değiller.", "Başka bir cümle."])
+        self.assertEqual(out[0][1], 1.4)     # the merged cue covers the whole word
+
+    def test_the_loaded_model_is_the_one_asked_for_again(self):
+        with fake_urlopen({"segments": [{"start": 0, "end": 1, "text": " hi"}]}) as calls:
+            api.transcribe_segments(LOCAL, self.wav)
+        self.assertEqual(multipart_fields(calls[0])["model"], "ggml-base.bin")

tests/test_cleanup.py

@@ -13,7 +13,9 @@ from unittest import mock
 
 import api
 import cleanup
-from tests.support import DikteTest
+import ggml
+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=""):
@@ -209,3 +211,76 @@ class Codex(DikteTest):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+class Here(DikteTest):
+    """llama.cpp, answering the request OpenRouter answers."""
+
+    def setUp(self):
+        super().setUp()
+        self.conf = self.config(cleanup_provider="local",
+                                local_llm_model="gemma.gguf")
+        self.server = FakeServer()
+        self.patch_attr(ggml, "llm", self.server)
+
+    def test_the_address_comes_from_the_server_it_starts(self):
+        with fake_urlopen(chat_reply("Done.")) as calls:
+            self.assertEqual(cleanup.run("uh, done", self.conf, "the rules"),
+                             "Done.")
+        self.assertEqual(self.server.starts, 1)
+        self.assertEqual(calls[0].full_url,
+                         "http://127.0.0.1:9999/v1/chat/completions")
+
+    def test_no_key_is_wanted_and_none_is_sent(self):
+        with fake_urlopen(chat_reply("Done.")) as calls:
+            cleanup.run("uh, done", self.conf, "the rules")
+        self.assertNotIn("Authorization", calls[0].headers)
+
+    def test_thinking_is_turned_off_in_the_words_llama_cpp_uses(self):
+        with fake_urlopen(chat_reply("Done.")) as calls:
+            cleanup.run("uh, done", self.conf, "the rules")
+        self.assertEqual(sent_json(calls[0])["chat_template_kwargs"],
+                         {"enable_thinking": False})
+
+    def test_the_models_own_default_asks_for_nothing(self):
+        self.conf["local_llm_reasoning"] = ""
+        with fake_urlopen(chat_reply("Done.")) as calls:
+            cleanup.run("uh, done", self.conf, "the rules")
+        self.assertNotIn("chat_template_kwargs", sent_json(calls[0]))
+
+    def test_a_reply_longer_than_the_transcript_is_cut_off(self):
+        # A small model will repeat the transcript until the context is full,
+        # and every one of those tokens is a second of somebody waiting.
+        with fake_urlopen(chat_reply("Done.")) as calls:
+            cleanup.run("x" * 4000, self.conf, "the rules")
+        self.assertEqual(sent_json(calls[0])["max_tokens"], 4000)
+
+    def test_a_short_dictation_still_gets_room_to_answer(self):
+        with fake_urlopen(chat_reply("Done.")) as calls:
+            cleanup.run("uh, done", self.conf, "the rules")
+        self.assertEqual(sent_json(calls[0])["max_tokens"], 512)
+
+    def test_a_reply_that_was_all_thinking_names_the_setting_that_fixes_it(self):
+        reply = {"choices": [{"message": {"content": "", "reasoning": "hmm"}}]}
+        with fake_urlopen(reply), self.assertRaises(api.ApiError) as caught:
+            cleanup.run("uh, done", self.conf, "the rules")
+        self.assertIn("Thinking", str(caught.exception))
+
+    def test_a_server_that_will_not_start_is_the_error_shown(self):
+        self.patch_attr(ggml, "llm", FakeServer(fails="llama.cpp is not installed"))
+        with self.assertRaises(api.ApiError) as caught:
+            cleanup.run("uh, done", self.conf, "the rules")
+        self.assertIn("llama.cpp", str(caught.exception))
+
+    def test_a_server_that_dies_mid_request_says_what_it_printed(self):
+        self.patch_attr(ggml, "llm", FakeServer(log="out of memory"))
+        with fake_urlopen(url_error("connection reset")):
+            with self.assertRaises(api.ApiError) as caught:
+                cleanup.run("uh, done", self.conf, "the rules")
+        self.assertIn("out of memory", str(caught.exception))
+
+    def test_no_cli_is_started_for_it(self):
+        patcher, calls = fake_run(stdout="never")
+        with patcher, fake_urlopen(chat_reply("Done.")):
+            cleanup.run("uh, done", self.conf, "the rules")
+        self.assertEqual(calls, [])

tests/test_config.py

@@ -12,7 +12,9 @@ import unittest
 from unittest import mock
 
 import api
+import cleanup
 import config as cfg
+import ggml
 import i18n
 from tests.support import DikteTest
 
@@ -131,8 +133,17 @@ class Keys(DikteTest):
 
 
 class TranscribeTarget(DikteTest):
-    def test_openai_by_default(self):
+    def test_this_machine_by_default(self):
-        target = self.config(openai_api_key="sk-test").transcribe_target()
+        target = cfg.Config().transcribe_target()
+        self.assertEqual(target.provider, "local")
+        self.assertEqual(target.api_key, "")
+        # Empty on purpose: the server picks a port when it starts, and reading
+        # a setting must not be what starts it.
+        self.assertEqual(target.base_url, "")
+
+    def test_openai_when_it_is_picked(self):
+        target = self.config(transcribe_provider="openai",
+                             openai_api_key="sk-test").transcribe_target()
         self.assertEqual(target.provider, "openai")
         self.assertEqual(target.service, "OpenAI")
         self.assertEqual(target.api_key, "sk-test")
@@ -165,7 +176,8 @@ class TranscribeTarget(DikteTest):
         self.assertEqual(target.provider, "openai")
 
     def test_a_self_hosted_endpoint(self):
-        conf = self.config(openai_base_url="http://localhost:8080/v1")
+        conf = self.config(transcribe_provider="openai",
+                           openai_base_url="http://localhost:8080/v1")
         self.assertEqual(conf.transcribe_target().base_url, "http://localhost:8080/v1")
 
 
@@ -444,3 +456,69 @@ class Defaults(unittest.TestCase):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+class LocalCleanup(DikteTest):
+    def test_the_local_model_is_what_the_history_records(self):
+        conf = self.config(cleanup_provider="local",
+                           local_llm_model="gemma-3-4b-it-Q4_K_M.gguf")
+        self.assertEqual(cleanup.provider(conf), "local")
+        self.assertEqual(cleanup.model(conf), "gemma-3-4b-it-Q4_K_M.gguf")
+
+    def test_it_needs_no_program_on_the_path(self):
+        # whisper.cpp and llama.cpp are fetched rather than installed, so unlike
+        # Claude Code and Codex there is no executable to look for.
+        self.assertEqual(cleanup.executable("local"), "")
+
+    def test_the_minutes_do_not_follow_the_cleanup_provider(self):
+        # A 4B model here will strip the filler words out of a dictation and
+        # will not write up an hour long meeting.
+        conf = self.config(cleanup_provider="local")
+        self.assertEqual(conf["meeting_model"], cfg.DEFAULTS["meeting_model"])
+
+    def test_only_the_cleanup_setting_asks_for_the_local_model(self):
+        self.assertFalse(cfg.Config().uses_local_llm())
+        self.assertTrue(self.config(cleanup_provider="local").uses_local_llm())
+
+
+class ReadyToRun(DikteTest):
+    def setUp(self):
+        super().setUp()
+        self.patch_attr(ggml, "MODELS_DIR", self.path("models"))
+
+    def install(self, name):
+        path = ggml.whisper_model_path(name)
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_bytes(b"model")
+
+    def test_a_missing_program_is_not_ready(self):
+        with mock.patch("shutil.which", return_value=None):
+            self.install("ggml-base.bin")
+            conf = self.config(local_model="ggml-base.bin")
+            self.assertFalse(conf.transcribe_ready())
+
+    def test_a_missing_model_is_not_ready_either(self):
+        with mock.patch("shutil.which", return_value="/usr/bin/whisper-server"):
+            conf = self.config(local_model="ggml-base.bin")
+            self.assertFalse(conf.transcribe_ready())
+
+    def test_both_halves_in_place(self):
+        with mock.patch("shutil.which", return_value="/usr/bin/whisper-server"):
+            self.install("ggml-base.bin")
+            conf = self.config(local_model="ggml-base.bin")
+            self.assertTrue(conf.transcribe_ready())
+
+    def test_a_hosted_provider_is_ready_when_it_has_a_key(self):
+        conf = self.config(transcribe_provider="openai", openai_api_key="sk-test")
+        self.assertTrue(conf.transcribe_ready())
+
+    def test_the_settings_reach_the_servers(self):
+        conf = self.config(local_model="ggml-base.bin", local_threads=4,
+                           local_gpu=False, local_llm_model="gemma.gguf",
+                           local_llm_context=4096)
+        conf.apply_local()
+        self.addCleanup(ggml.whisper.configure, model="", threads=0, gpu=True)
+        self.assertEqual(ggml.whisper.settings()["model"], "ggml-base.bin")
+        self.assertEqual(ggml.whisper.settings()["threads"], 4)
+        self.assertFalse(ggml.whisper.settings()["gpu"])
+        self.assertEqual(ggml.llm.settings()["context"], 4096)

tests/test_ggml.py

@@ -0,0 +1,664 @@
+"""Fetching a program and a model, and keeping a server alive on them.
+
+No network and no whisper.cpp: the downloads are answered from memory, and the
+servers are stand-in scripts that take the same arguments and open their port
+when they are told to, which is the only thing the code waits on.
+"""
+
+import contextlib
+import hashlib
+import io
+import os
+import signal
+import sys
+import tarfile
+import textwrap
+import threading
+import time
+from unittest import mock
+
+import ggml
+import hub
+from tests.support import (DikteTest, fake_urlopen, http_error, json_body,
+                           linux_only, url_error)
+
+
+def body(data, length=None):
+    """What urlopen hands back for a download: a reader with a length header."""
+    class Body:
+        def __init__(self):
+            self._buf = io.BytesIO(data)
+            self.headers = {"Content-Length":
+                            str(len(data) if length is None else length)}
+
+        def read(self, count=-1):
+            return self._buf.read(count)
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *_):
+            return False
+    return Body()
+
+
+def item(name, data, url="https://example.invalid/f", sha=True):
+    return hub.Item(name, url, len(data),
+                    hashlib.sha256(data).hexdigest() if sha else "")
+
+
+@contextlib.contextmanager
+def serving(release, archive):
+    """Answer by what is being asked for rather than by what came before.
+
+    An install asks GitHub what the release is and then asks for one file out of
+    it, and the first of those two comes from the cache the second time around.
+    Answering in order would then hand the archive request the release listing.
+    """
+    def opener(request, timeout=None):
+        url = request.full_url
+        if "api.github.com" in url:
+            return json_body(release)
+        return body(archive)
+
+    with mock.patch("urllib.request.urlopen", side_effect=opener) as calls:
+        yield calls
+
+
+def tarball(entries):
+    """A .tar.gz laid out the way the releases are: one directory of files."""
+    buf = io.BytesIO()
+    with tarfile.open(fileobj=buf, mode="w:gz") as tar:
+        for name, content in entries.items():
+            info = tarfile.TarInfo(name)
+            info.size = len(content)
+            info.mode = 0o755
+            tar.addfile(info, io.BytesIO(content))
+    return buf.getvalue()
+
+
+class Local(DikteTest):
+    """A test with its own bin, models and cache directories."""
+
+    def setUp(self):
+        super().setUp()
+        self.patch_attr(ggml, "DATA_DIR", self.path("data"))
+        self.patch_attr(ggml, "BIN_DIR", self.path("data", "bin"))
+        self.patch_attr(ggml, "MODELS_DIR", self.path("data", "models"))
+        self.patch_attr(hub, "CACHE_DIR", self.path("cache"))
+
+
+# --- downloading ----------------------------------------------------------
+
+
+class Download(Local):
+    def test_it_lands_and_the_part_file_is_gone(self):
+        data = b"a model, more or less" * 100
+        target = self.path("data", "models", "m.bin")
+        with fake_urlopen(body(data)):
+            self.assertTrue(ggml.download(item("m.bin", data), target))
+        self.assertEqual(target.read_bytes(), data)
+        self.assertFalse(target.with_name("m.bin.part").exists())
+
+    def test_a_wrong_checksum_installs_nothing(self):
+        data = b"the bytes that arrived"
+        wrong = hub.Item("m.bin", "https://example.invalid/f", len(data), "f" * 64)
+        target = self.path("data", "models", "m.bin")
+        with fake_urlopen(body(data)):
+            with self.assertRaises(ggml.LocalError) as caught:
+                ggml.download(wrong, target)
+        self.assertIn("checksum", str(caught.exception))
+        self.assertFalse(target.exists())
+        self.assertFalse(target.with_name("m.bin.part").exists())
+
+    def test_a_body_shorter_than_its_header_installs_nothing(self):
+        data = b"half of it"
+        target = self.path("data", "models", "m.bin")
+        with fake_urlopen(body(data, length=len(data) * 2)):
+            with self.assertRaises(ggml.LocalError):
+                ggml.download(item("m.bin", data), target)
+        self.assertFalse(target.exists())
+
+    def test_a_file_with_no_published_checksum_is_refused(self):
+        # Everything fetched here is run or parsed by something written in C++,
+        # and GitHub did not always publish a digest.
+        data = b"a program, say"
+        target = self.path("data", "models", "m.bin")
+        with fake_urlopen(body(data)):
+            with self.assertRaises(ggml.LocalError) as caught:
+                ggml.download(item("m.bin", data, sha=False), target)
+        self.assertIn("checksum", str(caught.exception))
+        self.assertFalse(target.exists())
+
+    def test_nothing_is_asked_for_before_it_is_refused(self):
+        # The refusal is not worth a gigabyte of somebody's bandwidth first.
+        with fake_urlopen(body(b"never read")) as calls:
+            with self.assertRaises(ggml.LocalError):
+                ggml.download(item("m.bin", b"x", sha=False), self.path("m.bin"))
+        self.assertEqual(calls, [])
+
+    def test_stopping_leaves_nothing_behind(self):
+        data = b"x" * (ggml.DOWNLOAD_CHUNK * 3)
+        target = self.path("data", "models", "m.bin")
+        with fake_urlopen(body(data)):
+            landed = ggml.download(item("m.bin", data), target,
+                                   should_stop=lambda: True)
+        self.assertFalse(landed)
+        self.assertFalse(target.exists())
+        self.assertFalse(target.with_name("m.bin.part").exists())
+
+    def test_progress_is_reported_against_the_total(self):
+        data = b"y" * (ggml.DOWNLOAD_CHUNK + 5)
+        seen = []
+        with fake_urlopen(body(data)):
+            ggml.download(item("m.bin", data), self.path("data", "m.bin"),
+                          on_progress=lambda done, total: seen.append((done, total)))
+        self.assertEqual(seen[-1], (len(data), len(data)))
+        self.assertGreater(len(seen), 1)
+
+    def test_a_refused_connection_says_which_file(self):
+        with fake_urlopen(url_error("no route to host")):
+            with self.assertRaises(ggml.LocalError) as caught:
+                ggml.download(item("m.bin", b"x"), self.path("data", "m.bin"))
+        self.assertIn("m.bin", str(caught.exception))
+
+    def test_an_http_error_is_not_written_to_disk(self):
+        target = self.path("data", "m.bin")
+        with fake_urlopen(http_error(404)):
+            with self.assertRaises(ggml.LocalError):
+                ggml.download(item("m.bin", b"x"), target)
+        self.assertFalse(target.exists())
+
+
+# --- installing a program -------------------------------------------------
+
+
+class InstallProgram(Local):
+    def setUp(self):
+        super().setUp()
+        # Built once, because the release listing has to publish its checksum
+        # and a tarball is not the same bytes twice.
+        self.archive = tarball({
+            "whisper-bin-ubuntu-x64/whisper-server": b"#!/bin/sh\nexit 0\n",
+            "whisper-bin-ubuntu-x64/libwhisper.so": b"not really a library",
+        })
+
+    def release(self, *names, archive=None):
+        digest = hashlib.sha256(self.archive if archive is None else archive)
+        return {"tag_name": "v1.9.1", "assets": [
+            {"name": name, "browser_download_url": f"https://example.invalid/{name}",
+             "size": 10, "digest": "sha256:" + digest.hexdigest()}
+            for name in names]}
+
+    def install(self, *names, archive=None):
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        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)
+        return path, [call.args[0].full_url for call in calls.call_args_list]
+
+    def test_the_binary_and_its_libraries_land_together(self):
+        path, _ = self.install("whisper-bin-ubuntu-x64.tar.gz")
+        self.assertTrue(os.path.isfile(path))
+        self.assertTrue(os.access(path, os.X_OK))
+        self.assertTrue(os.path.isfile(os.path.join(os.path.dirname(path),
+                                                    "libwhisper.so")))
+
+    def test_the_build_for_this_machine_is_the_one_fetched(self):
+        _, urls = self.install("whisper-bin-x64.zip", "whisper-bin-ubuntu-arm64.tar.gz",
+                               "whisper-bin-ubuntu-x64.tar.gz")
+        self.assertTrue(urls[1].endswith("whisper-bin-ubuntu-x64.tar.gz"))
+
+    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")):
+            with self.assertRaises(ggml.LocalError) as caught:
+                ggml.install_program(ggml.WHISPER)
+        self.assertIn("this machine", str(caught.exception))
+
+    def test_what_was_installed_is_remembered(self):
+        path, _ = self.install("whisper-bin-ubuntu-x64.tar.gz")
+        self.assertEqual(ggml.installed_program(ggml.WHISPER), path)
+        self.assertEqual(ggml.installed_version(ggml.WHISPER), "v1.9.1")
+
+    def test_a_record_pointing_at_a_deleted_binary_counts_for_nothing(self):
+        path, _ = self.install("whisper-bin-ubuntu-x64.tar.gz")
+        os.unlink(path)
+        self.assertEqual(ggml.installed_program(ggml.WHISPER), "")
+
+    def test_the_archive_is_not_kept(self):
+        self.install("whisper-bin-ubuntu-x64.tar.gz")
+        left = list((self.path("data", "bin", "whisper")).glob("*.tar.gz"))
+        self.assertEqual(left, [])
+
+    def test_the_previous_version_is_swept_up(self):
+        self.install("whisper-bin-ubuntu-x64.tar.gz")
+        old = self.path("data", "bin", "whisper", "v1.9.0")
+        old.mkdir(parents=True)
+        (old / "whisper-server").write_bytes(b"older")
+        self.install("whisper-bin-ubuntu-x64.tar.gz")
+        self.assertFalse(old.exists())
+
+    def test_an_archive_without_the_binary_is_refused(self):
+        empty = tarball({"whisper-bin-ubuntu-x64/README": b"nothing here"})
+        with self.assertRaises(ggml.LocalError) as caught:
+            self.install("whisper-bin-ubuntu-x64.tar.gz", archive=empty)
+        self.assertIn("whisper-server", str(caught.exception))
+
+
+    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
+        # still have none.
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        listing = {"tag_name": "v1.8.0", "assets": [
+            {"name": "whisper-bin-ubuntu-x64.tar.gz",
+             "browser_download_url": "https://example.invalid/w.tar.gz",
+             "size": 10}]}
+        with serving(listing, self.archive):
+            with self.assertRaises(ggml.LocalError) as caught:
+                ggml.install_program(ggml.WHISPER)
+        self.assertIn("checksum", str(caught.exception))
+        self.assertEqual(ggml.installed_program(ggml.WHISPER), "")
+
+    def test_an_archive_that_is_not_what_was_promised_installs_nothing(self):
+        listing = self.release("whisper-bin-ubuntu-x64.tar.gz")
+        other = tarball({"whisper-bin-ubuntu-x64/whisper-server": b"#!/bin/sh\nrm -rf\n"})
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        with serving(listing, other):
+            with self.assertRaises(ggml.LocalError) as caught:
+                ggml.install_program(ggml.WHISPER)
+        self.assertIn("checksum", str(caught.exception))
+        self.assertEqual(ggml.installed_program(ggml.WHISPER), "")
+
+    def test_an_archive_cannot_write_outside_the_directory_it_is_opened_in(self):
+        # An archive is not a trusted thing to unpack: a member named ../../ is
+        # how one writes over a file it was never given.
+        escape = tarball({"../../../escaped": b"should not land"})
+        path = self.path("data", "bin", "whisper", "v1.9.1")
+        with self.assertRaises(ggml.LocalError):
+            self.install("whisper-bin-ubuntu-x64.tar.gz", archive=escape)
+        self.assertFalse(self.path("escaped").exists())
+        self.assertFalse((path.parent.parent / "escaped").exists())
+
+    def test_a_symlink_out_of_the_directory_does_not_survive_either(self):
+        buf = io.BytesIO()
+        with tarfile.open(fileobj=buf, mode="w:gz") as tar:
+            info = tarfile.TarInfo("whisper-bin-ubuntu-x64/whisper-server")
+            info.type, info.linkname = tarfile.SYMTYPE, "/etc/passwd"
+            tar.addfile(info)
+        with self.assertRaises(ggml.LocalError):
+            self.install("whisper-bin-ubuntu-x64.tar.gz", archive=buf.getvalue())
+
+    def test_everything_is_asked_for_over_tls(self):
+        for url in (hub.GITHUB_API, hub.HF_API, hub.HF_FILES):
+            with self.subTest(url=url):
+                self.assertTrue(url.startswith("https://"))
+
+    def test_llama_takes_the_vulkan_build_when_there_is_a_loader(self):
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: True)
+        self.assertEqual(ggml._wanted_assets(ggml.LLAMA)[0],
+                         "bin-ubuntu-vulkan-x64.tar.gz")
+
+    def test_llama_falls_back_to_the_plain_build_without_one(self):
+        self.patch_attr(ggml, "_arch", lambda: "x64")
+        self.patch_attr(ggml, "_has_vulkan", lambda: False)
+        self.assertEqual(ggml._wanted_assets(ggml.LLAMA), ("bin-ubuntu-x64.tar.gz",))
+
+
+class WhichCopyRuns(Local):
+    def test_a_system_build_wins_over_a_downloaded_one(self):
+        self.patch_attr(ggml, "installed_program", lambda program: "/data/whisper-server")
+        with mock.patch("shutil.which", return_value="/usr/bin/whisper-server"):
+            self.assertEqual(ggml.program_path(ggml.WHISPER), "/usr/bin/whisper-server")
+
+    def test_the_downloaded_one_is_used_when_there_is_no_system_build(self):
+        self.patch_attr(ggml, "installed_program", lambda program: "/data/whisper-server")
+        with mock.patch("shutil.which", return_value=None):
+            self.assertEqual(ggml.program_path(ggml.WHISPER), "/data/whisper-server")
+
+    def test_a_setting_pointing_at_nothing_is_no_program(self):
+        self.assertEqual(ggml.program_path(ggml.WHISPER, "/nowhere/whisper-server"), "")
+
+    def test_a_setting_pointing_at_a_program_wins(self):
+        mine = self.path("mine")
+        mine.write_text("#!/bin/sh\n")
+        mine.chmod(0o755)
+        with mock.patch("shutil.which", return_value="/usr/bin/whisper-server"):
+            self.assertEqual(ggml.program_path(ggml.WHISPER, str(mine)), str(mine))
+
+
+# --- the lists ------------------------------------------------------------
+
+
+WHISPER_TREE = [
+    {"type": "file", "path": "ggml-base.bin", "size": 147951465,
+     "lfs": {"oid": "a" * 64}},
+    {"type": "file", "path": "ggml-large-v3-turbo-q5_0.bin", "size": 574041195,
+     "lfs": {"oid": "b" * 64}},
+    {"type": "file", "path": "ggml-base-encoder.mlmodelc.zip", "size": 37922638,
+     "lfs": {"oid": "c" * 64}},
+    {"type": "file", "path": "README.md", "size": 3196},
+]
+
+GGUF_TREE = [
+    {"type": "file", "path": "gemma-3-4b-it-Q4_K_M.gguf", "size": 2489000000,
+     "lfs": {"oid": "a" * 64}},
+    {"type": "file", "path": "gemma-3-4b-it-Q8_0.gguf", "size": 4130000000,
+     "lfs": {"oid": "b" * 64}},
+    {"type": "file", "path": "mmproj-model-f16.gguf", "size": 851000000,
+     "lfs": {"oid": "c" * 64}},
+    {"type": "file", "path": "mtp-gemma-4-E4B-it-Q4_0.gguf", "size": 59000000,
+     "lfs": {"oid": "d" * 64}},
+    {"type": "file", "path": "huge-00001-of-00009.gguf", "size": 40000000000,
+     "lfs": {"oid": "e" * 64}},
+    {"type": "file", "path": "README.md", "size": 100},
+]
+
+
+class Catalogue(Local):
+    def test_only_models_are_offered_and_the_small_ones_first(self):
+        with fake_urlopen(WHISPER_TREE):
+            models = ggml.whisper_models()
+        self.assertEqual([m.name for m in models],
+                         ["ggml-base.bin", "ggml-large-v3-turbo-q5_0.bin"])
+
+    def test_the_core_ml_encoders_are_not_models(self):
+        with fake_urlopen(WHISPER_TREE):
+            names = [m.name for m in ggml.whisper_models()]
+        self.assertNotIn("ggml-base-encoder.mlmodelc.zip", names)
+
+    def test_the_projector_and_the_draft_head_are_not_models(self):
+        with fake_urlopen(GGUF_TREE):
+            names = [q.name for q in ggml.llm_quants("ggml-org/gemma-3-4b-it-GGUF")]
+        self.assertEqual(names,
+                         ["gemma-3-4b-it-Q4_K_M.gguf", "gemma-3-4b-it-Q8_0.gguf"])
+
+    def test_a_model_split_across_files_is_left_out(self):
+        with fake_urlopen(GGUF_TREE):
+            names = [q.name for q in ggml.llm_quants("ggml-org/gemma-3-4b-it-GGUF")]
+        self.assertNotIn("huge-00001-of-00009.gguf", names)
+
+    def test_the_suggestions_come_first_and_the_rest_follow(self):
+        listing = [{"id": "ggml-org/something-new-GGUF"},
+                   {"id": ggml.SUGGESTED_LLM[0]}]
+        with fake_urlopen(listing):
+            found = ggml.llm_repos()
+        self.assertEqual(found[0], ggml.SUGGESTED_LLM[0])
+        self.assertIn("ggml-org/something-new-GGUF", found)
+
+    def test_an_unreachable_list_still_offers_the_suggestions(self):
+        with fake_urlopen(url_error()):
+            self.assertEqual(ggml.llm_repos(), list(ggml.SUGGESTED_LLM))
+
+    def test_an_unreachable_whisper_list_is_an_error_worth_showing(self):
+        with fake_urlopen(url_error()):
+            with self.assertRaises(ggml.LocalError):
+                ggml.whisper_models()
+
+    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")
+        path.parent.mkdir(parents=True)
+        path.write_bytes(b"model")
+        self.assertEqual(ggml.installed_whisper_models(), ["ggml-base.bin"])
+        self.assertTrue(ggml.have_model(path))
+
+    def test_an_empty_file_is_not_a_model(self):
+        path = ggml.llm_model_path("ggml-org/x-GGUF/model.gguf")
+        path.parent.mkdir(parents=True)
+        path.write_bytes(b"")
+        self.assertFalse(ggml.have_model(path))
+
+    def test_a_model_is_named_by_its_file_not_its_repository(self):
+        self.assertEqual(ggml.llm_model_path("ggml-org/x-GGUF/model.gguf").name,
+                         "model.gguf")
+
+
+# --- keeping a server alive -----------------------------------------------
+
+
+STAND_IN = textwrap.dedent("""
+    import http.server, sys, threading, time
+
+    args = sys.argv[1:]
+
+    def opt(name, default=""):
+        return args[args.index(name) + 1] if name in args else default
+
+    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"))
+
+    class Handler(http.server.BaseHTTPRequestHandler):
+        def do_GET(self):
+            ok = time.monotonic() - started >= healthy_after
+            self.send_response(200 if ok else 503)
+            self.end_headers()
+            self.wfile.write(b"{}")
+
+        def log_message(self, *a):
+            pass
+
+    server = http.server.HTTPServer((opt("--host"), int(opt("--port"))), Handler)
+    print("listening on " + opt("--port"), flush=True)
+    server.serve_forever()
+""")
+
+
+class Servers(Local):
+    def setUp(self):
+        super().setUp()
+        self.path("data").mkdir(parents=True, exist_ok=True)
+        # Named for the program and kept inside the data directory, because that
+        # is what the sweep looks for on a command line.
+        self.script = self.path("data", "whisper-server.py")
+        self.script.write_text(STAND_IN)
+        self.addCleanup(ggml.stop_all)
+        self.servers = []
+
+    def server(self, program=ggml.WHISPER, **settings):
+        defaults = {"extra": []}
+        defaults.update(settings)
+        made = ggml.Server(
+            program,
+            lambda values: [sys.executable, str(self.script)] + list(values["extra"]),
+            defaults,
+        )
+        self.servers.append(made)
+        self.addCleanup(made.stop)
+        return made
+
+    def test_a_started_server_hands_back_its_address(self):
+        server = self.server()
+        url = server.serve()
+        self.assertRegex(url, r"^http://127\.0\.0\.1:\d+/v1$")
+        self.assertTrue(server.running)
+
+    def test_the_second_call_does_not_start_a_second_one(self):
+        server = self.server()
+        first = server.serve()
+        self.assertEqual(server.serve(), first)
+
+    def test_a_settings_change_stops_what_was_running(self):
+        server = self.server()
+        server.serve()
+        server.configure(extra=["--wait", "0"])
+        self.assertFalse(server.running)
+
+    def test_the_new_settings_are_what_the_next_start_uses(self):
+        server = self.server()
+        server.serve()
+        server.configure(extra=["--healthy-after", "0"])
+        second = server.serve()
+        self.assertTrue(server.running)
+        self.assertTrue(second)
+
+    def test_a_program_that_dies_reports_what_it_printed(self):
+        server = self.server(extra=["--die"])
+        with self.assertRaises(ggml.LocalError) as caught:
+            server.serve()
+        self.assertIn("no such file", str(caught.exception))
+        self.assertFalse(server.running)
+
+    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.
+        server = self.server(program=ggml.LLAMA, extra=["--healthy-after", "0.4"])
+        started = time.monotonic()
+        server.serve()
+        self.assertGreaterEqual(time.monotonic() - started, 0.4)
+
+    def test_a_start_that_never_becomes_ready_gives_up(self):
+        self.patch_attr(ggml, "STARTUP_TIMEOUT", 0.5)
+        server = self.server(program=ggml.LLAMA, extra=["--healthy-after", "30"])
+        with self.assertRaises(ggml.LocalError):
+            server.serve()
+
+    def test_stopping_leaves_nothing_running(self):
+        server = self.server()
+        server.serve()
+        server.stop()
+        self.assertFalse(server.running)
+        self.assertEqual(server.base_url(), "")
+
+    def test_the_last_thing_it_printed_is_available(self):
+        server = self.server()
+        server.serve()
+        self.assertIn("listening", server.error())
+
+    def test_asking_what_is_running_does_not_wait_for_a_start(self):
+        """A model being loaded must not freeze the settings window.
+
+        The interface asks a running server what it is doing while a start is in
+        flight, and a lock held across the whole start would stop it dead.
+        """
+        server = self.server(extra=["--wait", "0.6"])
+        answers = []
+
+        def start():
+            server.serve()
+
+        thread = __import__("threading").Thread(target=start)
+        thread.start()
+        try:
+            time.sleep(0.15)
+            began = time.monotonic()
+            answers.append(server.settings())
+            answers.append(server.running)
+            self.assertLess(time.monotonic() - began, 0.2)
+        finally:
+            thread.join(timeout=10)
+
+    @linux_only
+    def test_a_server_a_killed_dikte_left_behind_is_swept_up(self):
+        server = self.server()
+        server.serve()
+        # What a SIGKILL of Dikte leaves: the child still running, the pid file
+        # still on disk, and nothing left that knows about either.
+        proc, server._proc = server._proc, None
+        self.assertTrue(server.sweep())
+        self.assertEqual(proc.wait(timeout=5), -signal.SIGTERM)
+
+    @linux_only
+    def test_a_pid_that_belongs_to_something_else_is_left_alone(self):
+        server = self.server()
+        server._remember(os.getpid())     # this test runner, not a server
+        self.assertFalse(server.sweep())
+
+    def test_no_pid_file_is_nothing_to_sweep(self):
+        self.assertFalse(self.server().sweep())
+
+    def test_a_start_that_goes_wrong_takes_its_process_with_it(self):
+        started = []
+
+        def explode(inner, proc, port):
+            started.append(proc)
+            raise RuntimeError("something in the wait went wrong")
+
+        self.patch_attr(ggml.Server, "_wait_ready", explode)
+        server = self.server()
+        with self.assertRaises(RuntimeError):
+            server.serve()
+        # Nothing else holds a reference to it, so leaving it running would leak
+        # a loaded model with nobody left to ask it anything.
+        self.assertIsNotNone(started[0].poll())
+        self.assertFalse(server.sweep())      # and the pid file went with it
+
+
+class Arguments(Local):
+    """What the two command lines say, since neither program is here to say it."""
+
+    def setUp(self):
+        super().setUp()
+        self.binary = self.path("whisper-server")
+        self.binary.write_text("#!/bin/sh\n")
+        self.binary.chmod(0o755)
+
+    def whisper_model(self, name="ggml-base.bin"):
+        path = ggml.whisper_model_path(name)
+        path.parent.mkdir(parents=True, exist_ok=True)
+        path.write_bytes(b"model")
+        return name
+
+    def test_the_inference_path_is_the_one_api_py_builds(self):
+        args = ggml._whisper_args({"binary": str(self.binary), "gpu": True,
+                                   "threads": 0, "model": self.whisper_model()})
+        self.assertIn("--inference-path", args)
+        self.assertEqual(args[args.index("--inference-path") + 1],
+                         "/v1/audio/transcriptions")
+
+    def test_detection_rather_than_english_when_nothing_is_asked_for(self):
+        args = ggml._whisper_args({"binary": str(self.binary), "gpu": True,
+                                   "threads": 0, "model": self.whisper_model()})
+        self.assertEqual(args[args.index("-l") + 1], "auto")
+
+    def test_the_graphics_card_is_turned_off_rather_than_asked_for(self):
+        settings = {"binary": str(self.binary), "gpu": False, "threads": 2,
+                    "model": self.whisper_model()}
+        args = ggml._whisper_args(settings)
+        self.assertIn("-ng", args)
+        self.assertEqual(args[args.index("-t") + 1], "2")
+
+    def test_a_missing_model_is_a_message_about_settings(self):
+        with self.assertRaises(ggml.LocalError) as caught:
+            ggml._whisper_args({"binary": str(self.binary), "gpu": True,
+                                "threads": 0, "model": "ggml-nothing.bin"})
+        self.assertIn("Settings", str(caught.exception))
+
+    def test_a_missing_program_says_so_before_a_missing_model(self):
+        with mock.patch("shutil.which", return_value=None):
+            with self.assertRaises(ggml.LocalError) as caught:
+                ggml._whisper_args({"binary": "", "gpu": True, "threads": 0,
+                                    "model": self.whisper_model()})
+        self.assertIn("whisper.cpp", str(caught.exception))
+
+    def test_the_layers_go_to_the_card_when_there_is_one(self):
+        model = ggml.llm_model_path("m.gguf")
+        model.parent.mkdir(parents=True, exist_ok=True)
+        model.write_bytes(b"gguf")
+        args = ggml._llm_args({"binary": str(self.binary), "gpu": True,
+                               "threads": 0, "model": "m.gguf", "context": 4096})
+        self.assertEqual(args[args.index("-ngl") + 1], "99")
+        self.assertEqual(args[args.index("-c") + 1], "4096")
+
+    def test_no_card_means_no_layers_offloaded(self):
+        model = ggml.llm_model_path("m.gguf")
+        model.parent.mkdir(parents=True, exist_ok=True)
+        model.write_bytes(b"gguf")
+        args = ggml._llm_args({"binary": str(self.binary), "gpu": False,
+                               "threads": 0, "model": "m.gguf", "context": 4096})
+        self.assertEqual(args[args.index("-ngl") + 1], "0")
+
+
+class Sizes(DikteTest):
+    def test_bytes_are_written_the_way_a_download_is_talked_about(self):
+        self.assertEqual(ggml.human_size(512), "512 B")
+        self.assertEqual(ggml.human_size(574041195), "547.4 MB")
+        self.assertEqual(ggml.human_size(3_095_033_483), "2.9 GB")
+

tests/test_hub.py

@@ -0,0 +1,184 @@
+"""What GitHub and Hugging Face are asked, and what is believed of the answer."""
+
+import json
+
+import hub
+from tests.support import DikteTest, fake_urlopen, http_error, url_error
+
+RELEASE = {
+    "tag_name": "v1.9.1",
+    "assets": [
+        {"name": "whisper-bin-ubuntu-x64.tar.gz",
+         "browser_download_url": "https://example.invalid/ubuntu-x64.tar.gz",
+         "size": 9379235, "digest": "sha256:" + "a" * 64},
+        {"name": "whisper-bin-x64.zip",
+         "browser_download_url": "https://example.invalid/win.zip",
+         "size": 100, "digest": "sha256:" + "b" * 64},
+        {"name": "no-url-here.zip", "size": 1},
+    ],
+}
+
+TREE = [
+    {"type": "file", "path": ".gitattributes", "size": 1477},
+    {"type": "file", "path": "ggml-base.bin", "size": 147951465,
+     "lfs": {"oid": "c" * 64, "size": 147951465}},
+    {"type": "directory", "path": "extra"},
+    {"type": "file", "path": "extra/ggml-tiny.bin", "size": 77691713,
+     "lfs": {"oid": "d" * 64, "size": 77691713}},
+]
+
+MODELS = [
+    {"id": "ggml-org/gemma-3-4b-it-GGUF", "downloads": 44606,
+     "lastModified": "2026-07-01T00:00:00.000Z"},
+    {"id": "ggml-org/gpt-oss-20b-GGUF", "downloads": 47975},
+    {"noid": True},
+]
+
+
+class Releases(DikteTest):
+    def setUp(self):
+        super().setUp()
+        self.patch_attr(hub, "CACHE_DIR", self.path("cache"))
+
+    def test_the_tag_and_the_assets_come_back(self):
+        with fake_urlopen(RELEASE) as calls:
+            tag, assets = hub.release("ggml-org/whisper.cpp")
+        self.assertEqual(tag, "v1.9.1")
+        self.assertEqual([a.name for a in assets],
+                         ["whisper-bin-ubuntu-x64.tar.gz", "whisper-bin-x64.zip"])
+        self.assertEqual(calls[0].full_url,
+                         "https://api.github.com/repos/ggml-org/whisper.cpp/"
+                         "releases/latest")
+
+    def test_the_sha256_prefix_is_dropped(self):
+        with fake_urlopen(RELEASE):
+            _, assets = hub.release("ggml-org/whisper.cpp")
+        self.assertEqual(assets[0].sha256, "a" * 64)
+
+    def test_a_tag_asks_for_that_tag(self):
+        with fake_urlopen(RELEASE) as calls:
+            hub.release("ggml-org/whisper.cpp", "v1.9.1")
+        self.assertTrue(calls[0].full_url.endswith("/releases/tags/v1.9.1"))
+
+    def test_a_release_with_no_assets_is_an_error(self):
+        with fake_urlopen({"tag_name": "v1", "assets": []}):
+            with self.assertRaises(hub.HubError):
+                hub.release("ggml-org/whisper.cpp")
+
+    def test_the_second_call_asks_nobody(self):
+        with fake_urlopen(RELEASE) as calls:
+            hub.release("ggml-org/whisper.cpp")
+            hub.release("ggml-org/whisper.cpp")
+        self.assertEqual(len(calls), 1)
+
+    def test_a_refresh_asks_again(self):
+        with fake_urlopen(RELEASE) as calls:
+            hub.release("ggml-org/whisper.cpp")
+            hub.release("ggml-org/whisper.cpp", refresh=True)
+        self.assertEqual(len(calls), 2)
+
+    def test_an_old_cache_beats_no_answer(self):
+        with fake_urlopen(RELEASE):
+            hub.release("ggml-org/whisper.cpp")
+        # Old enough that it would normally be fetched again, and no network
+        # to fetch it with.
+        for path in self.path("cache").iterdir():
+            os_utime(path)
+        with fake_urlopen(url_error()):
+            tag, assets = hub.release("ggml-org/whisper.cpp")
+        self.assertEqual(tag, "v1.9.1")
+        self.assertEqual(len(assets), 2)
+
+    def test_no_cache_and_no_network_says_so(self):
+        with fake_urlopen(url_error("no route to host")):
+            with self.assertRaises(hub.HubError) as caught:
+                hub.release("ggml-org/whisper.cpp")
+        self.assertIn("api.github.com", str(caught.exception))
+
+    def test_an_http_error_names_the_host_and_the_code(self):
+        with fake_urlopen(http_error(404, "nope")):
+            with self.assertRaises(hub.HubError) as caught:
+                hub.release("ggml-org/nothing")
+        self.assertIn("404", str(caught.exception))
+
+
+class Files(DikteTest):
+    def setUp(self):
+        super().setUp()
+        self.patch_attr(hub, "CACHE_DIR", self.path("cache"))
+
+    def test_directories_are_left_out(self):
+        with fake_urlopen(TREE):
+            files = hub.files("ggerganov/whisper.cpp")
+        self.assertEqual([f.name for f in files],
+                         [".gitattributes", "ggml-base.bin", "extra/ggml-tiny.bin"])
+
+    def test_the_url_is_the_one_that_serves_the_bytes(self):
+        with fake_urlopen(TREE):
+            files = hub.files("ggerganov/whisper.cpp")
+        self.assertEqual(
+            files[1].url,
+            "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin")
+
+    def test_the_lfs_object_id_is_the_checksum(self):
+        with fake_urlopen(TREE):
+            files = hub.files("ggerganov/whisper.cpp")
+        self.assertEqual(files[1].sha256, "c" * 64)
+        self.assertEqual(files[1].size, 147951465)
+
+    def test_a_file_outside_lfs_has_no_checksum(self):
+        with fake_urlopen(TREE):
+            files = hub.files("ggerganov/whisper.cpp")
+        self.assertEqual(files[0].sha256, "")
+
+    def test_an_answer_that_is_not_a_list_is_an_error(self):
+        with fake_urlopen({"error": "Invalid username or password."}):
+            with self.assertRaises(hub.HubError):
+                hub.files("ggml-org/whisper.cpp")
+
+
+class Repos(DikteTest):
+    def setUp(self):
+        super().setUp()
+        self.patch_attr(hub, "CACHE_DIR", self.path("cache"))
+
+    def test_it_asks_for_one_author_and_for_gguf(self):
+        with fake_urlopen(MODELS) as calls:
+            found = hub.repos(author="ggml-org")
+        self.assertIn("author=ggml-org", calls[0].full_url)
+        self.assertIn("filter=gguf", calls[0].full_url)
+        self.assertEqual([r.id for r in found],
+                         ["ggml-org/gemma-3-4b-it-GGUF", "ggml-org/gpt-oss-20b-GGUF"])
+
+    def test_a_missing_download_count_is_zero(self):
+        with fake_urlopen(MODELS):
+            found = hub.repos(author="ggml-org")
+        self.assertEqual(found[0].downloads, 44606)
+        self.assertEqual(found[1].updated, "")
+
+
+def os_utime(path):
+    """Backdate a cache file past its time to live."""
+    import os
+    import time
+    old = time.time() - hub.CACHE_TTL - 60
+    os.utime(path, (old, old))
+
+
+class CacheOnDisk(DikteTest):
+    def setUp(self):
+        super().setUp()
+        self.patch_attr(hub, "CACHE_DIR", self.path("cache"))
+
+    def test_what_is_stored_is_what_came_back(self):
+        with fake_urlopen(RELEASE):
+            hub.release("ggml-org/whisper.cpp")
+        stored = [json.loads(p.read_text()) for p in self.path("cache").iterdir()]
+        self.assertEqual(stored[0]["tag_name"], "v1.9.1")
+
+    def test_a_cache_that_cannot_be_written_is_not_a_failure(self):
+        self.patch_attr(hub, "CACHE_DIR", self.path("nope", "deeper"))
+        self.path("nope").write_text("a file where a directory would go")
+        with fake_urlopen(RELEASE):
+            tag, _ = hub.release("ggml-org/whisper.cpp")
+        self.assertEqual(tag, "v1.9.1")

tests/test_ui.py

@@ -11,6 +11,7 @@ from unittest import mock
 
 from PyQt6.QtWidgets import QApplication, QMessageBox
 
+import cleanup
 import config as cfg
 import hotkey
 import overlay as overlay_module
@@ -44,11 +45,20 @@ CHANGED = {
     "groq_transcribe_model": "whisper-large-v3",
     "openrouter_transcribe_model": "openai/whisper-1",
     "cleanup_enabled": False,
-    "cleanup_provider": "claude",
+    "cleanup_provider": "local",
     "cleanup_model": "some/other-model",
     "cleanup_claude_model": "opus",
     "cleanup_codex_model": "gpt-5",
     "cleanup_reasoning": "high",
+    "local_model": "ggml-small.bin",
+    "local_gpu": False,
+    "local_preload": False,
+    "local_threads": 6,
+    "local_llm_model": "gemma-3-4b-it-Q4_K_M.gguf",
+    "local_llm_repo": "ggml-org/gemma-4-E2B-it-GGUF",
+    "local_llm_gpu": False,
+    "local_llm_preload": True,
+    "local_llm_reasoning": "low",
     "cleanup_prompt": "Only fix the punctuation.",
     "file_cleanup_prompt": "Keep the stamps where they are.",
     "transcribe_prompt": "Paraşüt, OpenFrame",
@@ -213,7 +223,13 @@ class Settings(DikteTest):
         window = self.window(cfg.Config())
         offered = [window.transcribe_provider.itemData(i)
                    for i in range(window.transcribe_provider.count())]
-        self.assertEqual(offered, list(cfg.TRANSCRIBERS))
+        self.assertEqual(offered, ["local"] + list(cfg.TRANSCRIBERS))
+
+    def test_the_cleanup_box_offers_everyone_cleanup_py_dispatches_to(self):
+        window = self.window(cfg.Config())
+        offered = [window.cleanup_provider.itemData(i)
+                   for i in range(window.cleanup_provider.count())]
+        self.assertEqual(sorted(offered), sorted(cleanup.PROVIDERS))
 
     def test_the_answer_to_a_test_lands_under_the_key_it_was_asked_about(self):
         """One signal serves all three buttons, so it carries which one asked."""
@@ -335,3 +351,78 @@ class Overlay(DikteTest):
 
 if __name__ == "__main__":
     unittest.main()
+
+
+class LocalModels(DikteTest):
+    """The download boxes, without a network and without either program."""
+
+    def window(self, conf):
+        window = settings_ui.SettingsWindow(conf)
+        self.addCleanup(window.deleteLater)
+        self.addCleanup(window.close)
+        return window
+
+    def test_it_opens_where_the_missing_model_is_fixed(self):
+        # Nothing can transcribe on a fresh install, which is why this window
+        # was opened at all.
+        window = self.window(cfg.Config())
+        self.assertEqual(window.tabs.currentIndex(), window.api_tab_index)
+
+    def test_it_opens_where_it_was_left_when_everything_works(self):
+        conf = self.config(transcribe_provider="openai", openai_api_key="sk-test")
+        self.assertEqual(self.window(conf).tabs.currentIndex(), 0)
+
+    def test_a_model_that_is_not_here_yet_survives_a_save(self):
+        # The box is filled from what is on this disk, so a model that was
+        # deleted from underneath is not in the list. Dropping it on save would
+        # quietly empty the setting instead of asking for the download again.
+        conf = self.config(local_model="ggml-large-v3-turbo-q5_0.bin")
+        with mock.patch.object(QMessageBox, "information"):
+            self.window(conf)._save()
+        self.assertEqual(conf["local_model"], "ggml-large-v3-turbo-q5_0.bin")
+
+    def test_nothing_is_fetched_for_a_window_nobody_opened(self):
+        # DikteTest closes the network, so a request would fail the test. The
+        # lists are asked for when the box is shown, not when it is built.
+        window = self.window(cfg.Config())
+        self.assertTrue(window.local_whisper._pending)
+
+    def test_a_model_bigger_than_two_gigabytes_counts_up_rather_than_down(self):
+        # 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)
+        _app.processEvents()
+        self.assertIn("2.3 GB", box.status.text())
+        self.assertNotIn("-", 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".
+        box = self.window(cfg.Config()).local_llm
+        box.repo.addItem("ggml-org/a-model-with-a-name-that-runs-on-and-on-GGUF")
+        box._fit_popup(box.repo)
+        view = box.repo.view()
+        self.assertEqual(view.textElideMode(), settings_ui.Qt.TextElideMode.ElideNone)
+        widest = max(box.repo.fontMetrics().horizontalAdvance(box.repo.itemText(row))
+                     for row in range(box.repo.count()))
+        self.assertGreaterEqual(view.minimumWidth(), widest)
+
+    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))
+        self.assertFalse(window.stt_form.isRowVisible(window.local_whisper))
+        window._select_data(window.transcribe_provider, "local")
+        self.assertFalse(window.stt_form.isRowVisible(window.transcribe_model_row))
+        self.assertTrue(window.stt_form.isRowVisible(window.local_whisper))
+
+    def test_only_the_chosen_cleaner_is_on_screen(self):
+        window = self.window(cfg.Config())
+        self.assertTrue(window.cleanup_form.isRowVisible(window.cleanup_model_row))
+        self.assertFalse(window.cleanup_form.isRowVisible(window.local_llm))
+        window._select_data(window.cleanup_provider, "local")
+        self.assertTrue(window.cleanup_form.isRowVisible(window.local_llm))
+        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))

tests/test_worker.py

@@ -234,7 +234,7 @@ class Chain(DikteTest):
         self.assertEqual(row["raw"], "uh, book it for Thursday")
         self.assertEqual(row["text"], "Book it for Thursday.")
         self.assertEqual(row["duration"], 2.0)
-        self.assertEqual(row["model"], self.conf["transcribe_model"])
+        self.assertEqual(row["model"], self.conf.transcribe_target().model)
         self.assertEqual(row["mode"], "")
 
     def test_a_command_is_recorded_as_one(self):