Streams

/streams — the parallel-contribution janitor

You maintain state/active-streams.yaml as the live coordination substrate for every parallel stream, and you garbage-collect abandoned work. The operating model is adrs/drafts/parallel-contribution-operating-model.md; the protocol every stream follows is playbooks/stream-coordination.md; the resource inventory is adrs/drafts/parallel-contribution-infra.md (R1-R9 / O1-O10). This skill is the read + reconcile + GC view.

$ARGUMENTS: empty → the full dashboard. <stream-id> → focus one stream. gc → run the reclaim sweep.

1. Read + verify-live (never trust the record — O4/O8/O9)

Read state/active-streams.yaml. For every entry, re-derive the truth from the working trees and docker — a record is a claim, the tree is the truth:

• git -C ../odd-platform rev-parse HEAD + --abbrev-ref HEAD + git status --porcelain (vs the recorded head/branch).
• the same for ../documentation and every ../odd-platform-<id> worktree (git worktree list).
• git -C ../odd-platform status lineage/ + the probe_run_id in lineage/odd-platform/feature-flows.yaml (who actually holds the R9 lineage lock — never guess a stream; O8).
• docker ps (which <id>-* / probe-* stacks + ports + images are live) — a <id>-* or probe-* stack whose <id> has no LIVE entry is an ORPHAN stack (a crashed/manual run that skipped teardown) → GC it (§3).
• cat state/locks/heavy-e2e.holder — the flock holder of the serialized heavy-e2e lock (the authoritative "who is running the full regression right now"; absent = the lock is free; see run-regression.sh).

Reconcile the registry to what you observe: fix stale heads, wrong lock holders, and the unowned_dirty_state block. Stamp each touched entry with a verified-at. Commit the reconciliation (explicit path, atomic).

2. Classify each stream

• LIVE — updated within the heartbeat window (default 4h) and a non-terminal phase; its worktree + stack match the record. Leave it; just report.
• STALE — updated older than the window AND a non-terminal phase (intake…pr-pending). The session likely ended mid-flight. Candidate for resume/reclaim.
• ORPHANED — a worktree / contrib branch / held port / image tag with no registry entry, OR an entry whose isolation no longer exists on disk. Also: unowned_dirty_state (dirty paths no entry claims — e.g. a probe-run's lineage/** residue, a stray spec). Surface it; never sweep it (O10).
• COMPLETE — terminal (merged/done/blocked/review-complete). Should already be cleared; clear it.

3. Offer per STALE / ORPHANED stream (the GC) — gc mode

For each, present the options; never act destructively without the work captured or an explicit discard:

• resume — the worktree + branch + record are intact; print the exact commands to re-attach (the cd ../odd-platform-<id>, the SUT tag, the ports) and the next pending DoD gate. Nothing is reclaimed.
• reclaim — free the namespace only after the work is captured (the contrib branch is pushed to origin, or its uncommitted diff is saved to state/abandoned/<id>.patch with a one-line provenance header and surfaced). Then free, in order: tear down its stack — docker compose -p <id> -f lineage/_extractor/probe-stacks/odd-minimal.docker-compose.yml down -v (containers + network + volume; the backstop for a run that crashed before run-regression.sh's teardown trap fired) — drop the image tag (docker image rm odd-platform:odd-team-sut-<id>), remove the worktree, release the ports, clear the registry entry. Print what was captured + where before freeing anything.
• discard — maintainer-confirmed only. Even then, save the diff to state/abandoned/<id>.patch first (a discard is recoverable for one release cycle), then free the namespace.

The invariant (non-negotiable): abandoned work is never silently lost and never silently blocking. Every reclaim either pushes the branch or parks the patch; every discard parks the patch and is human-confirmed.

4. The dashboard (default, no args)

Print a compact table: id · role · work-item · phase · LIVE/STALE/ORPHANED · isolation (worktree/image/ports) · lineage-lock? · last-verified. Then:

• Next-free namespace — the lowest unused stream id pattern + the next free port pair + the next image tag, so a starting stream can copy them (the shared_resources.next_free values, re-derived live).
• Serialized-resource status — who holds lineage/** (R9); the heavy-e2e flock (cat state/locks/heavy-e2e.holder — the authoritative one-at-a-time holder that run-regression.sh acquires, G-C2); and the wants: e2e queue order.
• GC candidates — the STALE/ORPHANED list with the recommended action each.

Rules

• Verify live over record, always (O4/O8/O9). If the registry and the tree disagree, the tree wins; fix the registry.
• Never sweep what you do not own (O10). Unowned dirty state is reported for its owner to reconcile; this skill does not git checkout -- / git add / revert another stream's paths. The single exception is a human-confirmed discard, and even then the diff is parked first.
• Capture before reclaim. No worktree is removed and no branch is deleted until its work is pushed or parked.
• Atomic registry commits (explicit path; never git add -A — R5/O3).
• This skill is read-mostly + reconcile; it does not run builds, suites, /enrich, or pushes of code. It curates the coordination substrate so the other skills run safely in parallel.