feat(loop): merge-poll Done-edge fallback + parallel drives (max_concurrent>1)

[mabry1985]

What

Two autonomy improvements to the orchestration loop — the genuine feature gaps left after the test/cleanup PRs.

1. Merge-poll fallback for the Done edge

The merge webhook (/webhook/pr) stays the primary path, but many deployments — local agents, NAT'd, tailnet-only — have no public URL GitHub can post to, so a merged feature would sit in in_review forever. The loop now polls each in_review PR via gh pr view … --json state and runs the same idempotent record_merge for any that merged.

• On by default; rate-limited to merge_poll_interval_s (60s); only probes in_review PRs → cheap.
• worktree.pr_is_merged() — the probe (a gh failure returns False, so the next poll just retries; never raises into the loop).
• worktree.reap_feature_worktree() — a shared id → worktree/branch reap; the webhook reap in api.py now uses it too (DRY).

2. Lift the max_concurrent cap

The loop drove exactly one feature at a time (the config knob existed but was ignored). It now drives up to max_concurrent features concurrently, each in its own worktree — the puller tops up to the cap each tick and re-checks promptly while busy. Per-feature in-flight worktrees are tracked (self._inflight) so shutdown cancels every drive and reaps every orphaned worktree. Default stays 1 (serial) — fully backward compatible.

Config (manifest + README updated)

max_concurrent: 1          # >1 builds features in parallel (each its own worktree)
merge_poll: true           # poll merged PRs as a fallback to the webhook Done edge
merge_poll_interval_s: 60  # how often the loop polls in_review PRs

Version → 0.3.0 (manifest + pyproject, kept in lockstep).

Tests

+11 (72 total), all green: pr_is_merged (merged/open/closed/gh-fail), reap_feature_worktree, _spawn_ready concurrency cap + claim accounting, _poll_merges (merged-only + reap), poll rate-limit + merge_poll: false, and the _active → _inflight tracking. ruff check . + ruff format --check . clean.

🤖 Generated with Claude Code

[protoquinn]

👀 Quinn is reviewing — verdict (PASS / WARN / FAIL) + findings to follow.

[protoquinn]

QA Audit — PR #6 | feat(loop): merge-poll Done-edge fallback + parallel drives (max_concurrent>1)

VERDICT: WARN

---

CI Status

• test: queued ⏳

Diff Review

• Clean DRY refactor: api.py webhook reap delegates to shared worktree.reap_feature_worktree()
• Concurrency model (_active → _drives set + _inflight dict) is well-structured; add_done_callback(self._drives.discard) is an elegant auto-clean pattern
• stop() properly cancels all drives, gathers with return_exceptions=True, then reaps every inflight worktree — no orphan risk
• Config parsing is defensive: max(1, int(...)), bool(merge_poll), adaptive sleep (idle=interval / busy=3s)
• _spawn_ready() is sync (correct — claim_next_ready is not async) and tops up to cap cleanly

Observations

• LOW: clawpatch structural review unavailable — repo not in project registry; 547-line diff reviewed via diff_summary only
• LOW: diff truncated at ~200/547 lines — _maybe_poll_merges() body, worktree.pr_is_merged(), worktree.reap_feature_worktree(), and all 11 new tests are not visible to this review pass
• LOW: CI still queued — tests unverified; formal PASS/FAIL deferred to a re-review once CI is terminal

— Quinn, QA Engineer

[protoquinn]

Submitted COMMENT review on protoLabsAI/projectBoard-plugin#6.