Codex review fix. The earlier approach synthesized a 'suppressed' beat from every
dismissed OUTCOME, which is wrong: a dismissal is not a later suppressed repeat,
and the old dismissal timestamp sank below the 12-row activity cap (invisible).
Now the negative-feedback loop is recorded WHEN IT HAPPENS. When shouldIntervene()
returns false specifically because a signature was dismissed before
(agent/podman.ts), the agent writes a durable SuppressionDoc to a new
`suppressions` collection, timestamped at the repeat. graph/live.ts materializes
those into kind:'suppressed' activity (recent -> surfaces at the top). The
suppressed collision is never written to `collisions` (agent returns before
recordCollision), so this is its own record.
Verified end-to-end: a transient demo-pod suppression renders at the TOP of the
activity stream; cleaned up after. shared build + backend/frontend typecheck +
eslint pass.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Addresses the P1 brittleness in the Step 3 verifier so learned_from edges
(graph/live.ts:413) can't be silently zeroed on stage.
- Capture overlap AT detection time as Collision.gitOverlap (podman.ts), while
engineer_states are still fresh, instead of re-deriving from possibly-stale
state when the user clicks. deriveWasRealCollision() now prefers this stored
evidence and only falls back to a live re-derivation for older collisions.
- Canonicalize engineer names (trim + lowercase) on both the capture and the
fallback path, so "Karti" vs "karti" no longer misses the git state.
- conflictKey now reuses the shared comparableBasename() helper (dedupe).
shared/src/collision.ts gains optional `gitOverlap?: boolean` (additive).
shared rebuilt; backend+frontend typecheck + eslint pass. PLAN.md rung 3 updated.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Step 1 — policy.ts shouldIntervene: suppress on a prior dismissal alone.
The former `&& !priorOutcome.wasRealCollision` term was dead code (outcomes
record wasRealCollision hardcoded true), so the 85 real dismissals in Atlas
were never used. Now a prior accepted===false suppresses the next identical
nudge. Spec: continual-learning/policy.md:41, spec.md:163.
Step 2 — podman.ts handle: only escalate severity to 'critical' (the spoken
alert trigger) when the recalled prior was an accepted *real* collision,
instead of blanket-escalating every recall. Surfaces learned routing in
preferredAction; stops dismissed/false priors over-escalating to voice.
Spec: continual-learning/policy.md:62-63, plan.md:66.
Documentation-first: adds PLAN.md section 8 "P0.5 - RSI negative-feedback
activation" with both rungs + follow-ups. No schema change. backend typecheck
passes. Independent of the MongoDB-cleanup handoff (Codex).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Both dedup gates (shouldIntervene per-pod cooldown + hasRecentInterventionForCollision per-file window) were purely time-based, so an unresolved conflict re-alerted every NUDGE_COOLDOWN_MS (30s on prod). Track active conflicts in-memory in PodMan and alert only on the transition into conflict; re-arm when a detection cycle no longer sees it.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Voice generation, critical escalation, and audio publish all worked, but
participants heard nothing: browsers block autoplay of the agent's audio
track until a user gesture. Add the canonical LiveKit unlock — listen for
AudioPlaybackStatusChanged and show an "Enable sound" button that calls
room.startAudio() so PodMan's voice cues are actually heard.
Also make collision alerts short and direct instead of chatty AI prose:
card: "Conflict: ram + yahya both on README.md (unpushed). Seen before."
voice: "Conflict. ram + yahya, both on README.md."
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AaCFWMkYQmTcuPsxaaACft
Live observations showed two engineers both dirty on README.md but zero
collisions firing. Two root causes:
1. normalize() only lowercased and prepended src/, so the same file at
different path depths ("agent.ts" vs "backend/src/agent.ts") never
matched. Replace with lowercased-basename matching.
2. Git ground truth (engineer_states.changedFiles) was only used to set
the unpushed flag, never to match the file. The strongest signal was
wasted.
Detector now fuses vision currentFile AND git changedFiles into one
file->engineers map, so a collision fires when two people share a dirty
file even if both screens aren't on it at the same instant.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AaCFWMkYQmTcuPsxaaACft
Reverse the best-effort error swallowing. MongoDB is core to PodMan's
continual-learning story and must always be used, so a broken memory
layer must surface immediately rather than silently masquerade as
working (which is how observations stayed at 0 unnoticed).
- agent verifies Mongo via initMemory() at boot; bad creds / unreachable
Atlas now fail loudly before joining the room, not mid-demo
- server exits on Mongo init failure instead of warning and limping on
- getGitStates and onScreenFrame no longer swallow Mongo errors
- memory persist() logs the failure and rethrows instead of warning
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AaCFWMkYQmTcuPsxaaACft
A Mongo auth/connection failure in getGitStates escaped uncaught and
killed the agent process on the first screen frame, so collision
detection never ran. Make git-state fusion best-effort (degrade to
vision-only) and wrap the whole onScreenFrame loop so no per-frame
Gemini/GitHub/Mongo error can crash the long-running agent.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AaCFWMkYQmTcuPsxaaACft
Add getGitStates(podId) to db.ts — queries the engineer_states collection
written by scripts/podman-agent.mjs and returns a map keyed by engineer name.
In PodMan.onScreenFrame(), fuse before detectCollisions(): if an engineer has
changedFiles in engineer_states, force hasUnpushedChanges=true on their context.
This overrides Gemini vision's unreliable pixel-based guess with deterministic
local git state — preventing the money-moment collision from silently not firing.
Key assumption: LiveKit participant identity matches the --name arg used when
running the git watcher (e.g. identity "alice" → --name alice).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AaCFWMkYQmTcuPsxaaACft
sb-iam