Fix Pool — findings→fixes pooler

One coordinator session + N human testing slots produce a stream of freshly-filed issues. This skill turns the coordinator into the drain: poll → triage → dispatch → report, on a loop, with hard throttles. Capture stays with the testers (they have the context); pickup is centralized here (dedup + sequencing live in one place).

Invocation

/fix-pool arm [--interval 10m] [--max-agents 3] [--since <ISO|now>]
/fix-pool status
/fix-pool drain     # stop polling; let in-flight fix agents finish

Resolve config first

Read the task-manager adapter per _shared/adapters/{tracker.kind}.md (resolved from .claude/project.yml / project-context). All issue reads/links go through the adapter — never a hardcoded API. State file: .claude/.fix-pool-state.json (last_poll_iso, seen_ids[], dispatched{id→agent,pr}, held[]). Not committed.

Arm (the loop)

On arm, record the baseline timestamp (--since, default now) and start a polling cycle every --interval (default 10m) using the session's wake/scheduling mechanism (background monitor or ScheduleWakeup — whatever the harness provides; never a foreground sleep).

Each cycle:

1. Poll — adapter: issues created since last_poll_iso in the configured team/project (include tester-session-filed and manually-filed; exclude issues this session created itself unless labeled for the pool). Self-filter is mandatory in the poll query or first triage step — the hub files tickets too, and without it the pool re-triages its own output (observed live, 2026-07-17). Track hub-created ids in the state file and drop them on sight. Also drop issues whose domain an ACTIVE sibling workstream owns (e.g. a migration terminal's own ledger ticket) — that's a HOLD with a named owner, not a dispatch.

2. Triage each new issue (read title+body, check seen_ids):

   Verdict	Criteria	Action
   AUTO-FIX	Clear repro + scope is test-only, docs, or a small UI/logic fix; no schema, no auth/security semantics, no design decision	Dispatch a fix agent (worktree, own branch, PR to the integration branch, never merges)
   INVESTIGATE	Real defect, root cause unclear	Dispatch an investigate agent: repro + root cause + proposed fix posted as an issue comment; fix only if the cause turns out AUTO-FIX-class
   HOLD	Schema/migration, security posture, product/design decision, or anything one-way-door	Add to held[]; surface in the next report; never auto-built
   DUP	Matches an open issue/known finding	Link via adapter comment; no work

3. Throttle — at most --max-agents fix/investigate agents in flight (check TaskList); excess stays queued in order. New dispatches only as slots free.
4. Report — one short message per cycle only when something changed: new findings by verdict, dispatches started, PRs ready, held items. Silent cycles stay silent.

Hard rules

• PR-only. Pool agents never merge. Merges/deploys happen at operator break points — a deploy mid-test invalidates the testers' ground truth.
• No schema pushes, ever. Migration-class findings are always HOLD.
• Surface mutex. Never dispatch a fix touching a surface an active testing slot is walking (the operator names the active surfaces at arm time, or the walkthrough epics define them); queue until that slot completes.
• Dedup before dispatch. Two findings, one root cause → one agent, both issues linked.
• Honest triage beats fast triage. If AUTO-FIX vs HOLD is unclear, it's HOLD. A wrong auto-fix during a test session poisons two things at once.

Drain

drain: stop the poll loop, let in-flight agents finish and open their PRs, emit a final consolidated report: all findings routed, PRs awaiting merge, held items needing operator rulings. The pool is not drained until every finding has a disposition.