[coordination] Ticket 3 — deferred: pipeline cards, live progress line, loop-run orphan sweeper

[serge-ivo]

Filed, deliberately not built. Follow-ons to #205 / #206.

A. Pipeline runs on the board

A pipeline run writes pipeline_runs but no board card, so a pipeline subordinate is invisible to a supervisor. Fix: mirrorRuntimeTask a pipeline.run card in lib/pipeline-run-start.ts and flip its status in workflows/pipeline-run.ts.

Deferred because it has no Coder-Lead relevance and would add a card to every pipeline agent's board — a wider blast radius than #206 for zero benefit to the stated goal.

Worth noting as the architectural test: when this lands, subordinate_status picks it up with no supervision-side change at all. That is the evidence the boundary drawn in #205 is right.

B. A plain-language progress line — the honest equivalent of the terminal tail

Even with #205 and #206, the Lead sees card titles, outcomes and step counts — not the terminal text the hardcoded Overseer reads. There is no clean way for supervision to reach that, and it must not try (see the reverted 3f14bd3).

The architecturally correct close is another domain write: have the Pilot upsert a short human summary into its own delegation card's description, from the "thought" branch of the same onEvent hook #206 uses, throttled (e.g. every 5th action). The Lead then reads live progress through a generic record and still knows nothing about tmux.

This is the deliberate finish line for Overseer parity, not an oversight. Deferred pending a decision that outcomes + progress is insufficient in practice.

C. Orphan sweeper for agent_loop_runs

expireOrphanedRuntimeTasks reaps stranded instance_runtime_tasks. Nothing does the equivalent for agent_loop_runs or pipeline_runs, so a run whose Workflow died mid-step sits running forever with a stale iteration.

Both workflows have code defenses (agent-loop.ts force-closes in a catch, coding-session.ts closes on the no-runner path — "a running row nobody will ever close is worse than a failed one, because the supervisor keeps waiting"), but neither is a queryable staleness signal and neither survives an isolate death.

#205 mitigates the symptom by deriving "stalled" from last_progress_at. The structural fix is a sweeper. Note this also weakens the case for ever unifying the run tables: more write paths into one table means more ways to strand a row, and a stranded row is permanent data.

[serge-ivo]

Shipped in b9d38ab (deployed, CI green).

A — pipeline cards: verified live. Ran pump_test on the Lead Finder; a card appeared on its board: Pipeline: pump_test (chat) → Done, detail 1 step(s), 1 → leads — 1 seen · 1 added. The architectural claim this ticket was filed to test holds: subordinate_status picks these up with no supervision-side change, and there is now a test asserting exactly that — summarizeSubordinates buckets a pipeline card correctly while knowing nothing about pipelines.

B — progress line: half verified live, half review + unit tests. Found and fixed a real gap en route: a delegation's outcome text only ever went to reasoning, which is in NEITHER generic reader's fallback chain, so a finished delegation showed the console what happened and showed a supervisor a blank. Confirmed live — the Lead now quotes outcome: failed — run error: No runner connected — run \pags up`` verbatim off the card, where before it had nothing to read.

The mid-run "thought" line needs a Pilot that actually executes steps, and the runner is currently offline, so that half is unit-tested + review-only for now. Its load-bearing guard IS pinned: the write is a partial UPDATE gated on status = 'running', so a progress write racing the terminal write cannot flip a completed card back to running (mutation-tested).

C — sweeper: shipped. Runs on the per-minute cron, 3h of silence, measured with the same COALESCE(last_progress_at, started_at) rule the supervisor uses so the two cannot disagree about what "quiet" means. The UPDATE re-checks status = 'running', so a run that closes itself between the SELECT and the UPDATE is not overwritten with a failure. Migration 0068 adds partial indexes on the open set, which stay the size of the open set rather than of all history.

Bonus confirmation of #206 from the same live run: the workflow's no-runner early-return path closed its coding-session card to failed rather than stranding it at running — the fourth of the four transition points, working in production.