fix(tests): the suite typed on the real keyboard, and the second model is retired - #60
Merged
Merged
Conversation
…l is retired
THE SUITE TYPED INTO HIS SCREEN, ALL DAY, AND THAT IS THE HEADLINE. Two tests
drive `_stream_loop` with a fixed `Heard` and monkeypatch `dictate._inject`,
believing the clipboard was the way out to the machine. It is not the only one:
`dictate.place` tries `platforms.type_text` FIRST, and that is a real
`CGEventKeyboardSetUnicodeString` with nothing in front of it. So every run of
the suite typed both fixtures into whatever window had focus, back to back:
hello there my friend and also yougokigen you desu ne totemo ii tenki
Reported as "I keep getting this same random paste everywhere, even tho I'm not
using murmurflow", which is exactly right — it was not MurmurFlow, it was
MurmurFlow's tests. The giveaway was that it was byte-identical every time: a
hallucination differs every time, a fixture does not. `tests/conftest.py` shuts
that ONE door for every test, autouse, and a test asserts it stays shut. Only
that door: the clipboard and the recorder have tests that drive them on
purpose, and stubbing those would replace the thing under test with the stub.
AND SILENCE IS CUT BEFORE ANYTHING TRANSCRIBES IT. Whisper invents words when
handed audio with nothing in it — the same 12s of speech, three ways:
speech alone "...but just in this text box,"
+ 20s digital silence "...but just in this text box, Thank you."
+ 20s faint room noise "...but just in this text box.."
A word list cannot catch that (this file already said so) and neither can
whisper's own per-segment `no_speech_prob`: the invented " Thank you." came
back at 0.000 among real speech at -0.05. Both were built and both were
measured useless before `trim_trailing_quiet`, which was measured to work.
THE SECOND MODEL AND ITS SERVER ARE GONE. `partial_model`, `PARTIAL_PREFERENCE`,
`DEFAULT_PARTIAL_MODEL`, `partial_port`, `start_partial_server`, `partial_at`,
`livePass` and the whole `at`/`model` parameter family are deleted; one model
answers the live pass and writes the final, on one server, on one queue.
`stop_server` still sweeps the port above ours, because a version that no
longer starts a small server must still reap the one an older one left.
THE LANGUAGE PIN IS GONE with it: whisper-server reports back whatever language
it was TOLD to decode, so pinning made the "is that one of yours" gate blind
for the whole clip after its first second — and a partial is pasted.
AND `whole_clip_read` IS DELETED, four hours after it shipped. It never fired:
a big-model pass costs 1.5-2.4s, so several seconds of speech are always unread
when the key comes up, and the shortcut refused every time (measured on his own
log: 4830ms, 8929ms, 5528ms, all after the change). It also reused a PINNED
read as the final transcript, which is the guarantee the pin's own docstring
promised never to break.
STREAM_TIMEOUT 10s -> 30s: a pass on a two-minute clip costs more than ten.
README and the website carry the measured numbers now: one model, 1.5s a pass
under 30s of speech and 2.4s past it, and the microphone that closes itself.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P7oJz8M9QzsdJimoM318cj
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Advanced Run ID: 📒 Files selected for processing (8)
📝 WalkthroughWalkthroughChangesSingle-model transcription
3D star-field rendering
Estimated code review effort: 4 (Complex) | ~45 minutes Sequence Diagram(s)sequenceDiagram
participant RecordingLoop
participant trim_trailing_quiet
participant WarmServer
participant Finalizer
RecordingLoop->>trim_trailing_quiet: trim trailing quiet audio
RecordingLoop->>WarmServer: request language-gated live pass
Finalizer->>trim_trailing_quiet: trim recorded audio
Finalizer->>WarmServer: request final transcription
WarmServer-->>Finalizer: return final transcript
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The headline: the test suite typed into his screen, all day
It was not MurmurFlow. It was MurmurFlow's tests.
Two tests drive
_stream_loopwith a fixedHeardand monkeypatchdictate._inject, believing the clipboard was the way out to the machine. It is not the only one:dictate.placetriesplatforms.type_textfirst, and that is a realCGEventKeyboardSetUnicodeStringwith nothing in front of it. So every run of the suite typed both fixtures into whatever window had focus, back to back and with no space between them.The giveaway was that it was byte-identical every time. A hallucination differs every time; a fixture does not.
tests/conftest.pyshuts that one door for every test, autouse, and a test asserts it stays shut. Only that door: the clipboard and the recorder have tests that drive them on purpose, and stubbing those would replace the thing under test with the stub — a suite reporting green over a function it never ran is worse than the bug.And silence is cut before anything transcribes it
Whisper invents words when handed audio with nothing in it. The same 12 seconds of speech, three ways:
...but just in this text box,...but just in this text box, **Thank you.**...but just in this text box..A word list cannot catch that — this file already said so, whisper answers silence in a different invented language each time — and neither can whisper's own per-segment
no_speech_prob: the invented " Thank you." came back at 0.000, sitting among real speech at -0.05. Both were built and both were measured useless beforetrim_trailing_quiet, which was measured to work.This matters more now that a forgotten microphone closes itself after 15 seconds: that clip ends with 15 seconds of exactly the audio whisper invents into.
The bridges, burned
partial_model,PARTIAL_PREFERENCE,DEFAULT_PARTIAL_MODEL,partial_port,start_partial_server,partial_at,livePass, and the wholeat/modelparameter family — deleted. One model answers the live pass and writes the final, on one server, on one queue.stop_serverstill sweeps the port above ours, because a version that no longer starts a small server must still reap the one an older one left.The language pin went with it. whisper-server reports back whatever language it was told to decode, so pinning made the "is that one of yours" gate blind for the whole clip after its first second — and a partial is pasted.
And
whole_clip_readis deleted, four hours after it shipped. It never fired: a big-model pass costs 1.5–2.4s, so several seconds of speech are always unread when the key comes up. Measured on his own log after the change: 4830ms, 8929ms, 5528ms — the shortcut refused every time. It also reused a pinned read as the final transcript, which is the guarantee the pin's own docstring promised never to break.STREAM_TIMEOUT10s → 30s: a pass on a two-minute clip costs more than ten.Docs and website
Both carry the measured numbers now: one model, 1.5s a pass under 30s of speech and 2.4s past it, the punctuation trade with the
livePassnote removed, and the microphone that closes itself.Gate
206 passed, ruff clean, mypy clean. 483 lines deleted, 412 added.🤖 Generated with Claude Code
https://claude.ai/code/session_01P7oJz8M9QzsdJimoM318cj
Summary by CodeRabbit
New Features
large-v3-turbomodel.Improvements
Documentation