Clawdy v0.0.4

Added

• A built-in voice. Clawdy now talks back with its own bundled voice (Kokoro, running on your Mac — nothing is sent anywhere) instead of the robotic system voice. Six voices to pick from in the panel; "Heart" is the default. ElevenLabs is still there if you'd rather use your own key. Teach it names and product words in ~/.clawdy/pronunciations.txt. The download is about 170 MB bigger for it.
• Japanese names are pronounced properly by the built-in voice ("Aomori", "Shibuya", "Nebuta"): romanized Japanese is read syllable by syllable instead of being guessed as English spelling.
• The "thinking…" pill by the claw now appears the moment the model actually starts reasoning (or after 2.5 s of silence), not after 10 s — a quiet visual sign the turn is alive, with no extra spoken cue.
• The claw's pointing is timed to the built-in voice too: it lands on each element as it is named, sentence by sentence (previously only ElevenLabs had timing; with the built-in voice the cursor walked ahead of the speech).
• Pointing is more precise. Every screenshot the model sees now carries a pixel ruler along its edges, so a [POINT] coordinate comes from a printed scale instead of a guess (the model's estimates were landing a couple hundred pixels off on wide screens).
• Draw, then ask about it. When you draw on the screen while asking, Clawdy is now told the red strokes are your annotation, so "this road" means the one you traced (and it points at the marked thing, not a look-alike elsewhere).
• Research asks out loud. When a research run needs a quick answer first, Clawdy now asks the question in its voice (at the next quiet moment) and the pill turns into the thing you're talking to — the question, a mic, and "⌃⌥ to answer". Hold the keys and answer; the run continues. The typed box is still there if you click the pill.
• The research pill shows the brand claw and a quiet elapsed-time clock, so a minutes-long run never looks stuck.
• Speed settings. Pick the model and effort for quick answers in the menu bar panel; the control is the same for both engines. Sonnet + low effort is now the default for Claude (about a second faster to the first spoken word than Opus at half the cost, and no multi-second silent think before a longer answer: low effort halved time-to-first-text on substantive questions in testing). For Codex, low effort is about 2× faster than medium.
• Instant feedback, all in Clawdy's voice. About a second after you release the keys Clawdy says "let me check" / "let me take a look" (a natural beat, not the instant you stop), and if the answer is taking a while it says "still working on it, hang tight" at the twenty-second mark (the cursor's "thinking…" pill covers the stretch in between) — the register of a voice assistant, not a status line. Research runs now say "sure, I'll put a page together", "your page is ready", and "sorry, that one didn't work out" instead of playing system sounds. There are no sound effects, and no two voice outputs ever overlap: a research announcement waits until you and Clawdy are both quiet.
• Fast mode switch. Turns on the engine's own fast tier for quick answers: Claude Code fast mode (Opus only) or Codex service_tier=fast (about a second faster per reply). Off by default; both cost more.
• Per-turn latency log (log show --predicate 'subsystem == "com.clawdy.Clawdy" AND category == "latency"') so slow turns can be attributed to transcription, the model, or speech.

Changed

• Research pages and History are real windows now. While one is open Clawdy shows in the Dock and Cmd-Tab, so a page can't get lost behind other apps.
• Codex model picker. The panel shows your Codex default model and lets you pick any model Codex lists.
• Quieter recents list and History. The claw's recents list is a compact "Recent" list sized to its rows; History rows show the skill and time on a second line with a status dot only for running/failed runs, and the detail header shows skill · engine · time.
• Simpler menu bar panel. One right edge for every control, quieter hierarchy, the hotkey shown as keycaps, History and Quit on one row, no close button (click outside or press Escape). The "Use my Claude Code setup" toggle is gone: your setup always loads.

Removed

• The macOS system voice. The built-in Kokoro voice replaces it entirely (ElevenLabs remains as the bring-your-own-key option).

Fixed

• Stop stays available for the whole spoken answer. The panel dropped it (and the companion went "idle") the moment the reply's text was complete, even though a long answer keeps speaking for many seconds after that.
• The instant "let me check" was being cut off (and the "let me look" fillers and a queued "your page is ready" were being cancelled) by the request's own teardown right after key release. Cues now survive the request start; only a re-press or Stop cancels them. A research hand-off says one thing, not "okay" and then "on it".
• Opening or closing a research results page no longer forces a cold restart of the warm claude process on the next question.