claude-code-config: Skill for Claude Code

Thread-level PR fleet manager. This skill turns the current thread into a dedicated monitor that watches every open PR you own and drives each one to merge-ready (or a named hard block) by dispatching the per-PR decision tree on a recurring cadence. Fix work (has-recoverable-blockers / verdict fixpr) is handled by parallel phase-a-fixer subagents (default cap 3, --max-parallel N). Merge-ready PRs get sequential /wrap dispatch only.

Auto-merge via /wrap. PMM never merges directly; merge-ready PRs dispatch the full /wrap workflow inline once gate + AC pass (CLAUDE.md "PR MERGE AUTHORIZATION"). Scope: only gh pr merge --squash via /wrap — never branch-protection changes, never dismissing human reviews, never bypassing gate or AC failures. With --confirm-merges off (default), PMM dispatches /wrap immediately on merge-ready PRs with no per-PR "merge now?" prompt.

Per-PR dispatch is inlined below. TODO: refactor to call /babysit-pr per discovered PR after #456 lands. Until #456 merges, Step 3's decision tree is the single owner of per-PR logic. When /babysit-pr exists, replace Step 3's inline branches with one /babysit-pr <PR> dispatch per discovered PR — the table, discovery, idempotency, and backoff scaffolding here stay unchanged.

Parent/subagent scope, prohibited actions, refusal template, and common misreads: references/pmm-scope.md.

Step 00: Resolve shared tooling (MANDATORY, before the tick gate)

/pr-monitor-and-manage is symlinked into every repo, but its helper scripts and reference docs are not — most repos carry no .claude/ directory. Resolve them once per tick, before anything reads state; never invoke a bare .claude/scripts/… path. Full contract and the classified dependency inventory: .claude/reference/portable-skill-resolution.md (issue #1189).

resolve_script() {
  local name="$1" candidate
  for candidate in \
    "$HOME/.claude/skills-worktree/.claude/scripts/$name" \
    "$HOME/.claude/scripts/$name" \
    ".claude/scripts/$name"; do
    if [[ -x "$candidate" ]]; then echo "$candidate"; return 0; fi
  done
  return 1
}
SESSION_STATE_SH=$(resolve_script session-state.sh || true)
MERGE_GATE_SH=$(resolve_script merge-gate.sh || true)
MERGE_SEQUENCE_SH=$(resolve_script merge-sequence.sh || true)
PR_ISSUE_REF_SH=$(resolve_script pr-issue-ref.sh || true)
REVIEWER_OF_SH=$(resolve_script reviewer-of.sh || true)
CR_HOURLY_SH=$(resolve_script cr-review-hourly.sh || true)
REPO_ROOT_SH=$(resolve_script repo-root.sh || true)
PR_STATE_SH=$(resolve_script pr-state.sh || true)
USAGE_HORIZON_SH=$(resolve_script usage-horizon.sh || true)

These bindings are in scope for every step of the tick, including the references/pmm-*.md procedures. Read reference docs (merge-sequencing.md, release-cadence.md) through the matching .claude/reference/ candidate order.

When something does not resolve, say so in one line; never skip the contract silently.

• SESSION_STATE_SH or MERGE_GATE_SH empty → required, and fatal for the tick. Print ERROR: <name> not found (checked all three paths) — PR fleet management unavailable and exit without acting. Every verdict this skill reaches is a merge decision; a fleet manager that cannot read the gate or persist a hard block must not guess, and must not act on a guess.
• MERGE_SEQUENCE_SH empty → optional. Print DEGRADED: merge-sequence.sh not found (checked all three paths) — overlap-aware sequencing unavailable, merging one PR per tick and serialize merges rather than sequencing them.
• USAGE_HORIZON_SH empty → optional, and conservative. Print DEGRADED: usage-horizon.sh not found (checked all three paths) — horizon consult unavailable, holding "unknown" (no new subagent dispatch) and run Step 3.7 with the verdict it already produces for an unresolved helper. unknown is the fail-closed posture, never clear: it suppresses new fixer dispatch while leaving merge-ready /wrap and rebases alone, so the fleet keeps landing work instead of freezing (subagent-thread-limit-park.md §8.1).
• REVIEWER_OF_SH, PR_ISSUE_REF_SH, CR_HOURLY_SH, REPO_ROOT_SH, PR_STATE_SH empty → optional. Print one DEGRADED: line naming the script and what is lost (reviewer re-trigger routing, issue back-reference, CR budget accounting, root-repo resolution, reviewer-engagement matrix), then continue with that one capability off. An empty PR_STATE_SH renders every engagement cell as ? — the tick still runs.