Build-ready spec for adding a "rework saved" ROI band to the member
Work History dialog: transparent heuristic over caught collisions,
additive shared field + backend queries + presentational component.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LuV8W8oNYRsDWKoqK8Mkqc
- docs/plans/research-overlap.md: self-contained build spec for code-edit
vs research cross-channel overlap nudge (vision classifies editing vs
research, semantic embedding match, collaboration nudge). Ready for an
implementing agent.
- CLAUDE.md: replace branch-first habit with push-directly-to-main policy
plus mandatory pull --rebase before push for concurrent teammates.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LuV8W8oNYRsDWKoqK8Mkqc
Reframe README around the real bottleneck (human coordination, not
engineering ability) with the "five-minute meeting" cost story; rewrite the
demo script to match. Update gemini/livekit/mongodb/digitalocean specs to
reflect shipped code (Gemini Live agent, TTS, embeddings, Lyria via
Interactions API; current API routes; hermes_jobs collections). Add hermes.md.
Rename graph.md -> cont_learning.md. Remove outdated/dead docs (agent-learning
scaffolding, graph-discovery, handoffs, superpowers, idea/plan/demo-setup) and
the bundled LiveKit starter under examples/. Rewire CLAUDE.md doc-first gate
off the deleted PLAN.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LuV8W8oNYRsDWKoqK8Mkqc
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>
Backend becomes authoritative for the verifier signal. recordOutcome now
overrides the client-supplied wasRealCollision with deriveWasRealCollision():
a flagged collision counts as REAL only if BOTH named engineers currently have
the collided file in their git changedFiles (getGitStates, 120s freshness TTL).
Conservative false when the collision is orphaned/missing or git state is
stale. Restores the (accepted x wasReal) 2x2 the spec assumes instead of the
hardcoded 107/107 true.
frontend/useInterventions.ts stops sending a hardcoded `true` (now a
backend-overridden placeholder). Spec: continual-learning/spec.md:98-108,
policy.md:35-42. backend+frontend typecheck + eslint pass. PLAN.md P0.5 rung 3
added.
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>
Fresh-session handoff covering the dynamic force-directed graph, honest de-noised
metrics, the click-to-explain Flow pane, the divergence vs main's parallel impl,
the best-of-both reconciliation (PR #38), build/verify quirks, and gotchas.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The 'Test audio' button becomes 'Background Music': each pod gets a calm looping track from Gemini Lyria 3 that sings the pod name once up front then stays instrumental. Backend GET /api/pods/:id/music generates via the Gemini interactions endpoint and caches the MP3 per pod in Mongo (pod_music); the frontend fetches and loops it via Web Audio, published pod-wide on the existing podman-beat track.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Lands the dynamic Team-memory redesign on top of main's continual-learning
pipeline. main already computes loop/activity in the API but its GraphView never
rendered them and kept a static-column graph; this swaps in the dynamic graph and
surfaces the rails, reusing main's richer PodLearningLoop / PodGraphActivity types.
Frontend (new frontend/src/components/graph/*, composed into GraphView.tsx):
- forceSim.ts: dependency-free force layout (charge, link springs, centroid
recenter, 2-pass collision, bounds clamp, alpha anneal) — no new deps/lockfile churn.
- GraphCanvas.tsx: SVG render from the sim — draggable + pinnable nodes, curved
edges, weight-sized shapes, fade-in, animated learned_from dash, risk-path
lighting / rest dimmed, label collision-avoidance.
- MetricsRail / LearningLoop / ActivityStream / SelectedNodePanel / encoding.ts:
the mock's rails + stream + detail panel in light shadcn. LearningLoop consumes
main's PodLearningLoop (steps + activeStep); ActivityStream consumes
PodGraphActivity (title + detail). SelectedNodePanel adds a Flow section that
narrates the path through a clicked node (flowNarrative); default copy is
mode-aware. Edge legend rounded out (editing/touches).
- GraphView polls every 5s and diffs (positions preserved), + ws /api/events nudge.
Stale selection (node gone across a poll) dropped so the canvas can't dim entirely.
Backend (surgical — main's materializer + buildLoop kept):
- live.ts: headline metric cards derived from the FINAL de-noised graph (Open risk
paths = distinct collision files; Learned owners = distinct owner engineers)
instead of raw collision-signature / accepted-outcome counts that inflate with
test churn (50 -> 4 risk files, 16 -> 1 owner on live). buildLoop untouched.
- demo.ts: metrics realigned to the demo graph (3 owners / 1 risk path / 100%).
Verified: lint + -r typecheck + -r build pass; Playwright confirmed the dynamic
graph (ticks on load, draggable), the loop rail (5 steps, ADAPT active) and
activity stream rendering main's shapes, honest metrics (3/1/100%), and the flow
narrative per node kind — zero page errors.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds repo_recent_commits (history/authorship, optional path or author) and
repo_find_commits (by commit message or pickaxe code search) so the voice
agent can answer who-changed-what and which-commit-introduced-X over the
local karti-ai/podman checkout.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Voice agent can now search the repo (code, symbols, files) via ripgrep over
the live local checkout of main. GitHub code-search index is empty for this
repo, so back it with the always-current clone instead.
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>
Surface audio-playback unlock and mic publish as persistent buttons in the pod control row (alongside Test audio / Test PodMan voice / Share screen), replacing the conditional 'Sound is off' banner that only appeared after a refresh. Enable mic calls setMicrophoneEnabled, which triggers the browser permission prompt so users can verify their mic is set up right.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Make the "Test audio" connectivity check pod-wide instead of local to the
publisher. Shared on/off state is derived from the presence of the podman-beat
track across participants (self-syncing on join/leave). Any participant can
stop it: the owner unpublishes directly; non-owners send a new additive
BEAT_STOP data message that the owner honors. Drives the button label, status
line, and status waveform for everyone, and guards against rapid-double-click
double-publish and mid-start unmount leaks.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Drop vite-plugin-pwa and add an index.html snippet that unregisters any leftover service worker and clears its caches on load. Continuous deploys during the event plus a self-destroying SW were forcing open tabs to reload mid-demo. Browsers self-heal on next load; re-add a precaching PWA post-event.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the confusing Join / Add-and-join / +Add trio on PodCard with one flow: type your name, hit Join → added to the roster (deduped server-side) and connected in one step. Removes the join-as-members[0] impersonation footgun where a second person could kick the first.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The deployed graph was an unreadable mess: full file paths centered on nodes
bled across columns, and collision labels showed garbage vision symbols
("infra/README.md### Running the git watcher").
- node labels are now the last two path segments (full path in summary)
- collision labels are just the file (drop the junk #symbol)
- cap file nodes to 9 (collision files prioritized)
demo-pod: 22 -> 20 clean nodes; no label bleed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
My/Team streams were one flat lane; source provenance was buried as a
gray badge by the filename. Group each lane into Signals (vision/git
inputs) and Reasoning & decisions (collision/intervention/outcome), and
promote source to a color-coded provenance chip with a readable kind
label. Presentation-only; no shared-type or backend changes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VpySDjxpM1usXMEngJRWqG
Materializer (live.ts) — cut demo-pod from 110 -> 22 nodes:
- cap to 8 recent collisions; collapse repeats by memorySignature
- collapse interventions to one (most recent) per collision
- filter junk 'files' (URLs, env vars, browser/app names, scratch/test) to
real source paths only
- prune test-artifact engineers (a/b/verify/codex-check/-testrepo) + any node
orphaned by that
- collisions referenced by accepted outcomes bypass the cap so the learned_from
money path never drops; risk-paths metric counts distinct signatures
UI: per-pod 'Team memory' action on each PodCard's menu (replaces the header
button that always opened pods[0]).
Verified against live demo-pod data (joins connect); typechecks clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Agent-published LiveKit audio is unreliable for voice: the track is
short-lived (publish, speak ~3s, unpublish) and blocked by browser
autoplay, so participants heard nothing even though the server published
fine. The VOICE_CUE text already arrives over the data channel, so speak
it in the browser via speechSynthesis instead — instant and reliable.
Prime speechSynthesis from user gestures (Enable sound, Share screen) so
later cues are allowed to play.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AaCFWMkYQmTcuPsxaaACft
Document the box topology so future sessions (and teammates' agents) don't
rediscover it the hard way: systemd-managed services (never launch the agent
manually — duplicate podman-hermes identities evict each other and silently
drop interventions), Caddy serving the static frontend from /var/www/podman,
the deploy procedure, and the shared-package build order.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AaCFWMkYQmTcuPsxaaACft
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
Verified materializePodGraph against real demo-pod data (joins all connect:
21/21 interventions->collisions, learned_from produced). Tuning:
- engineer nodes are case-insensitive (merges Shakthi/shakthi)
- git (engineer_states) now only CONFIRMS editing on files vision/collisions
already surfaced, instead of adding the whole repo diff (killed a 29-file
node explosion from a watcher running against the full podman repo)
Cut demo-pod from 110 -> 77 nodes; git file-edges 39 -> 6.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ramis