Skip to content

Actions: unifyai/unity

Actions

All workflows

Actions

Loading...
Loading

Showing runs from all workflows
1,040 workflow runs
1,040 workflow runs

Filter by Workflow

⌥ + click/return to exclude

Filter by Event

⌥ + click/return to exclude

Filter by Status

⌥ + click/return to exclude

Filter by Branch

⌥ + click/return to exclude

Filter by Actor

⌥ + click/return to exclude
Flow smoke (commit) Flow Smoke #96: Commit fb2e792 pushed by djl11
1s staging
staging
1s
Cancel the slow brain by turn id, not by queue slot Lint #276: Commit fb2e792 pushed by djl11
40s staging
staging
40s
Cancel the slow brain by turn id, not by queue slot When the fast brain resolves a turn itself (continuation or small talk) it must drop exactly the slow-brain run that turn spawned - not whatever happens to occupy the debouncer's two slots. The previous cancel nuked both slots, which could drop a prior still-thinking substantive answer or an unrelated act/SMS run. - Stamp a per-turn id (the existing _user_turn_seq) onto the inbound voice utterance events; thread it through interject_or_run -> reques... Tests #327: Commit fb2e792 pushed by djl11
7s staging
staging
7s
Flow smoke (commit) Flow Smoke #95: Commit 806aed5 pushed by djl11
1s staging
staging
1s
Add SMALLTALK: let the fast brain hold basic conversation Lint #275: Commit 806aed5 pushed by djl11
38s staging
staging
38s
Add SMALLTALK: let the fast brain hold basic conversation A greeting, "tell me about yourself", "what time is it where you are?", or "what did you just say?" no longer gets a "give me a moment" deferral. The fast brain answers pure social / biographical / simple self-context / repeat turns itself from its persona and recent history, and cancels the eagerly-started slow-brain run so it does not also answer. Everything substantive still defers. - prompt_builders: add build_smalltalk_messages + _SMALLTALK... Tests #326: Commit 806aed5 pushed by djl11
7s staging
staging
7s
Make the fast brain the front door for interrupted speech Every user interjection now gets an immediate fast-brain CONTINUE/DEFER decision; continuations are recursively interruptible; the slow brain is triggered eagerly but cancelled the instant the fast brain resumes a line. - Delete the fixed 0.5s handoff timer: an interrupted line's remainder waits for the next fast-brain turn to decide, so the fast brain can never lose the race. The slow brain only receives the remainder on an explicit DEFER. ... Tests #325: Commit 6e4e781 pushed by djl11
7s staging
staging
7s
Make the fast brain the front door for interrupted speech Lint #274: Commit 6e4e781 pushed by djl11
42s staging
staging
42s
Flow smoke (commit) Flow Smoke #94: Commit 6e4e781 pushed by djl11
2s staging
staging
2s
Flow smoke (commit) Flow Smoke #93: Commit 4d64677 pushed by djl11
1s staging
staging
1s
Recover a wedged voice worker so call-start tools can't strand forever The call-starting tools (make_call / make_whatsapp_call) are masked while is_ready_for_new_call is False, which requires a freshly prewarmed idle worker process (WORKER_READY marker). If a prewarmed process is consumed by a dispatch that never becomes a live job (e.g. a spurious re-dispatch into a tearing-down Unify Meet), the marker is cleared and never re-created. The watchdog only restarted a worker that had *exited*, so an alive-... Tests #324: Commit 4d64677 pushed by djl11
10s staging
staging
10s
Recover a wedged voice worker so call-start tools can't strand forever Lint #273: Commit 4d64677 pushed by djl11
45s staging
staging
45s
Constrain reference-quiz clues to strictly sci-fi The narration rule allowed "sci-fi or pop-culture" clues, which let the assistant pick general/fantasy references like Lord of the Rings. Require science-fiction only, with examples and an explicit fantasy/general-trivia exclusion. Tests #323: Commit 1d52baa pushed by djl11
7s staging
staging
7s
Constrain reference-quiz clues to strictly sci-fi Lint #272: Commit 1d52baa pushed by djl11
47s staging
staging
47s
Flow smoke (commit) Flow Smoke #92: Commit 1d52baa pushed by djl11
1s staging
staging
1s
Flow smoke (commit) Flow Smoke #91: Commit 13e4534 pushed by djl11
2s staging
staging
2s
Gracefully end Unify Meet so the Console leaves before room deletion When the assistant ended a meet, the agent's shutdown deleted the LiveKit room while the browser was still connected, force-evicting it. The abrupt WebRTC teardown fired the client's RTCDataChannel.onerror, surfacing as benign but noisy "Unknown DataChannel error on reliable/lossy" logs in the Next.js console. Now on an explicit meet stop the agent first publishes a call_ended status so the Console disconnects itself (a clean, clien... Tests #322: Commit 13e4534 pushed by djl11
8s staging
staging
8s
Gracefully end Unify Meet so the Console leaves before room deletion Lint #271: Commit 13e4534 pushed by djl11
45s staging
staging
45s
Soften fourth-wall policy to allow honest glitch acknowledgement Lint #270: Commit 9ef09e7 pushed by djl11
45s staging
staging
45s
Flow smoke (commit) Flow Smoke #90: Commit 9ef09e7 pushed by djl11
10s staging
staging
10s
Soften fourth-wall policy to allow honest glitch acknowledgement The strict "never break the fourth wall" rule made the brains fabricate excuses to "own" genuinely garbled output rather than surface the truth. Relax both the slow-brain and fast-brain prompts so the hard line is now single identity (never disown a word, never fragment into separate agents), while permitting a brief honest acknowledgement of a real glitch (e.g. "sorry, that came out garbled") instead of inventing a fake reason. Tests #321: Commit 9ef09e7 pushed by djl11
9s staging
staging
9s
Use fixed resume lead-ins to stop interrupted-line duplication Lint #269: Commit b60d67d pushed by djl11
46s staging
staging
46s
Use fixed resume lead-ins to stop interrupted-line duplication select_continuation asked the weak fast-brain model for a bridge lead-in, and it copied the remainder's first sentence verbatim; speaking lead_in + resume_text then said that sentence twice back-to-back. Now the model only classifies CONTINUE vs DEFER, and the lead-in is drawn from a fixed bank of bridges that can never overlap the resumed content (rotated for variety). The model's text is never used as the lead-in, so the duplication is eli... Tests #320: Commit b60d67d pushed by djl11
8s staging
staging
8s
Flow smoke (commit) Flow Smoke #89: Commit b60d67d pushed by djl11
8s staging
staging
8s
Re-add bundled fast-brain guidance (general; quiz only in gated context) Restores instant fast-brain confirm without the freeform-delegation footguns: guide_voice_agent gains an optional fast_brain_guidance note that is only honored when bundled with a spoken message (never guidance-alone, enforced in _run_llm), rides on FastBrainNotification, is stashed per slow-brain spoken turn (an empty note clears it), and reaches select_fast_reply via a general scoped block (use it only to directly answer what the... Tests #319: Commit e30327b pushed by djl11
10s staging
staging
10s