docs(skills): add night-ops-directive and coderabbit-pr-gate agent skills - #22

Intent

Salvage the valuable content from stale/broken PR #5 (which bypassed no-mistakes and had failing checks) and reship it clean through no-mistakes. Recover from PR #5's branch: a new agent-only skill night-ops-directive (standing autonomous-dispatch directive covering delegator discipline during unattended work, firstmate's own watcher-heartbeat wake cadence since ScheduleWakeup/CronCreate are blocked for the primary session, human-only task triage via bd transfer, and a pointer to fork-first pushes), a new agent-only skill coderabbit-pr-gate (CodeRabbit-clean-plus-green merge gating: react to actionable review comments beyond bare pass/fail, treat rate limits as a paused external wait not a failure), and an AGENTS.md clarification that the qualifying test for a routine yolo-covered gate, including a PR merge, is reversibility. Both new skills carry metadata.internal=true like other .agents/skills/ entries, matching existing skill conventions (frontmatter, one-sentence-per-line prose). Do NOT reopen or reuse PR #5 itself; this is a clean new branch. Redo the changes cleanly rather than copying PR #5's diff verbatim: fix whatever caused PR #5's failing checks (lint, non-no-mistakes PR raise, portable serial behavior). Before writing, this task cross-checked against work already landed the same day: fork-first-brief-08 (fm-brief.sh already auto-injects a fork-first push rule into every push-mode ship brief, documented in AGENTS.md section 11) already fully covers PR #5's general fork-first-push AGENTS.md line, so that line was intentionally dropped (not carried over) to avoid duplicating it. PR #16 (pr-review-comment-wake) only added the wake mechanism for a new bot PR review comment (state/.pr-review-seen); it added no reaction policy, so coderabbit-pr-gate was kept as a complementary policy layer, not a duplicate, and now cross-references that wake mechanism instead of re-describing it. This task's diff is scoped to AGENTS.md and the two new .agents/skills/ directories only; no other files should change.

What Changed

• Added .agents/skills/night-ops-directive/SKILL.md, a standing autonomous-dispatch directive covering delegator discipline during unattended work, the primary session's watcher-heartbeat wake cadence (since ScheduleWakeup/CronCreate are blocked for it), human-only task triage via bd transfer, and a pointer to fork-first pushes.
• Added .agents/skills/coderabbit-pr-gate/SKILL.md, a CodeRabbit-clean-plus-green merge gating policy: react to actionable review comments beyond a bare pass/fail signal, and treat CodeRabbit rate limits as a paused external wait rather than a failure.
• Updated AGENTS.md to clarify that the qualifying test for a routine yolo-covered gate, including a PR merge, is reversibility.
• Registered both new SKILL.md files in docs/documentation-audiences.json so the tracked-documentation audience inventory stays in sync.

Risk Assessment

✅ Low: Purely additive change (two new agent-only skill docs plus a one-clause AGENTS.md clarification) confined to documentation/policy text; verified factual claims (script names, heartbeat backoff behavior, PreToolUse blocking of ScheduleWakeup/CronCreate, pr-review-seen wake mechanism) against the actual scripts and AGENTS.md, and diff scope matches the stated intent exactly with no unrelated files touched.

Testing

Captain, the documentation-audiences fix from round 1 already landed and the whole change checks out: the structural inventory test suite and the direct audience-check script both pass clean against the real repo tree, both new agent-only skills exist with correct frontmatter and content matching the salvaged PR #5 intent, every mechanism they reference (watcher heartbeat vars, pr-review-seen wake, bd transfer) is real, and the diff stays scoped to AGENTS.md plus the two new skill directories plus the required inventory registration — no other files touched, working tree clean.

ok - documentation inventory classifies every maintained prose surface exactly once
ok - classification, setup routing, and maintained-prose scope fail safely
ok - required documentation owner pointers cannot silently disappear
ok - local links resolve while dates, versions, commands, and incident prose remain semantically reviewed

fm-doc-audience-check: ok surfaces=63 local_links=168

Pipeline

Updates from git push no-mistakes