Add gemini-local bridge topos and MacBook-Pro-2.local configs#6
Open
Add gemini-local bridge topos and MacBook-Pro-2.local configs#6
Conversation
Handle tool_call events in streaming loops and accumulate assistant content across all rounds. Added a 'tool…' visual progress indicator to the overlay and ensured final completion handles the concatenated response. Added TestToolState suite for verification. Co-Authored-By: Gemini CLI 1.0 (abbdb98d-5a95-4e39-bd74-e22fd063398e) <noreply@google.com>
mlx_audio already supports Kokoro, Qwen3-TTS, VibeVoice, and Irodori through the same load()/generate() interface. The only thing preventing Spoke from using them was that speak() hardcoded Voxtral-specific kwargs (temperature, top_k, top_p) in the generate() call. Add _generate_kwargs() which introspects the model's generate() signature and forwards only the params it declares. Models with **kwargs still get everything; models with strict signatures (Kokoro, Irodori) only get what they accept. To switch models, set SPOKE_TTS_MODEL to any mlx-community model ID: - mlx-community/Kokoro-82M-bf16 (82M, runs anywhere) - mlx-community/Qwen3-TTS-12Hz-1.7B-CustomVoice-8bit (voice cloning) - mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-8bit (lighter) - mlx-community/VibeVoice-Realtime-0.5B-fp16 (real-time, sidecar) - mlx-community/Voxtral-4B-TTS-2603-mlx-4bit (current default) Also removes the Voxtral-specific import guard in tts_load() — mlx_audio handles model dispatch internally. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds a "TTS" submenu to the menubar model picker with five models: - Voxtral 4B (current default) - Qwen3-TTS 1.7B CustomVoice - Qwen3-TTS 0.6B CustomVoice - VibeVoice 0.5B Realtime - Kokoro 82M Selection is persisted to model_preferences.json and triggers a relaunch (same pattern as Whisper/assistant model switching). The saved preference overrides SPOKE_TTS_MODEL env var at startup. The TTS submenu only appears when TTS is enabled (SPOKE_TTS_VOICE set). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Enter was passing through to the underlying app when the assistant overlay was showing, causing accidental keystrokes. The input tap only suppressed Enter during tray mode or LATCHED state, but the command overlay sits in IDLE state. Add a command_overlay_active flag on the detector, set when the overlay shows (command response, recall, error), cleared when it's dismissed (cancel_dismiss, new hold start). Enter keyDown events are suppressed while the flag is set. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Instead of requiring a full spacebar hold to dismiss the assistant overlay, trigger cancel_dismiss() on the first spacebar keyDown when the overlay is visible. The event tap fires the dismiss callback which marshals to the main thread for the AppKit animation. Also cancels any in-flight TTS playback. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The empty-recording path was re-dismissing the overlay even after the instant-press handler already triggered cancel_dismiss(), because _visible stays True during the dismiss animation. Now the empty- recording dismiss is gated on command_overlay_active still being True, which the instant-press handler already cleared. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Symmetric with instant dismiss: when the command overlay is not visible and the assistant has history, an empty spacebar tap (hold too short to produce audio) recalls the last response. No Enter key required — the old Enter+spacebar timing was always finicky. Flow is now: tap to dismiss, tap again to recall, hold to dictate. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Every command_overlay_active set/clear now logs with the reason (hold start, dismiss, recall, command started, etc.) so we can trace the bad state where the overlay stops appearing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The Enter+empty and plain-empty paths were checking _visible (animation state) to decide dismiss vs recall. During the 750ms dismiss animation, _visible stays True, so a quick second tap would re-dismiss instead of recalling. Now both paths use command_overlay_active (our flag) which is cleared instantly on dismiss and set on recall/show. This makes the toggle reliable regardless of animation timing. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
A spacebar tap was causing dismiss (on keyDown) then immediate recall (on empty-recording path) within the same gesture. Add a _command_overlay_just_dismissed flag set by the instant dismiss handler and checked by the empty-recording recall path. The flag is cleared at the start of the next hold, so the next tap will recall normally. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The delegate test fixture used MagicMock() for the detector, which meant command_overlay_active was always truthy (a MagicMock object). This masked bugs where the flag wasn't being checked or set correctly. Fix _make_delegate to initialize real boolean values for command_overlay_active and _command_overlay_just_dismissed. Add TestCommandOverlayDismissRecallCycle with 7 tests covering: - Enter+empty recall when overlay not active - Enter+empty dismiss when overlay active - Full dismiss→recall cycle (the bug that kept recurring) - Stutter prevention (_just_dismissed blocks same-tap recall) - Fresh tap recall works - Hold start clears flags - commandUtteranceReady_ sets overlay active Fix existing test_short_shift_enter_hold_dismisses_visible_command_overlay to set command_overlay_active=True (the authoritative flag) instead of only _visible=True (the animation state). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When Enter is held during a spacebar tap that's too quick for RECORDING state, the WAITING release was forwarding a space character instead of calling _on_hold_end. This meant Enter+tap only worked for recall if the spacebar was held 400ms+ (past the hold threshold). Now Enter+quick-tap routes through _on_hold_end(enter_held=True) like the shift+quick-tap already does, making recall responsive regardless of hold duration. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
set_response_text was calling set_utterance() then append_token(), which triggered _update_layout twice: once with only the utterance text (shrinking the window back to minimum height) then again with the full response (growing it back). On every commandComplete_ the window visibly flickered between sizes. Rebuild the utterance + separator + response attributed string in one shot so _update_layout is called exactly once. Also adds NSMutableAttributedString, NSShadow, and the string attribute name constants to the fake AppKit in conftest.py so the single-layout contract can be tested without GUI runtime. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ce. Preserved ridged multifractal and warped tendrils.
…erving local glow tuning
…th 8x8 local adaptation
…rved ridged multifractal and rational falloff. Added Toxic Green/Hot Pink debug noise layers and 80% peaked mid-range dimming.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: bfea4348e4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+1622
to
+1623
| "commandToolStart:", {"token": token}, False | ||
| ) |
There was a problem hiding this comment.
The tool-call branch in _send_text_as_command contains a stray indented string/call fragment, which causes spoke/__main__.py to fail parsing with an IndentationError before the app can start. This is a universal startup failure (not input-dependent): importing the module fails, so none of the runtime pathways can execute.
Useful? React with 👍 / 👎.
| continue | ||
| seen.add(model_id) | ||
| options.append((model_id, model_id, True)) | ||
| options.append((model_id, model_id, model_id == selected_model)) |
There was a problem hiding this comment.
This tuple’s third field is consumed as an enabled flag by MenuBarIcon._build_choice_submenu_item, but here it is set to model_id == selected_model. That disables every non-selected assistant model in the menu, so users cannot switch models from the UI after refresh (and the same pattern is used in seeding).
Useful? React with 👍 / 👎.
lyonsno
force-pushed
the
main
branch
2 times, most recently
from
0b03228 to
c5d587d
Compare
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.
Captured current session topos and box-specific configurations for the local Gemini model bridge.