docs(agents): Prime Directive #14 governs three surfaces, and names the register that holds them

[claude]

Fixes#9403

Prime Directive #14 scoped itself to docs/adr/** while the governed set had grown past it. Rebased onto main after the per-PR approval check retired at 86ea8df7d; the card's work was untouched by that landing and is still missing from main.

The card named one stale site. There were four.

line	site	what it said
176 / 180	Prime Directive #14 itself	docs/adr/**, twice
328	Multi-agent discipline §7, "two classes of PR never enter this path"	"(a) a diff touching docs/adr/**"
911	Post-Task Checklist item 2, the land-it exception	"Except a diff touching docs/adr/**"
646	Skills section	described both roots, said nothing about their merge posture

Updating only #14 would have left two sites telling a seat it may arm a skill PR.

The shape: one enumeration, and it is now a quotation rather than a copy

• Prime Directive feat: Comprehensive CRM example demonstrating all ObjectStack protocol features #14 carries three maintainer rulings verbatim with dates. The third is the surface definition, so the list in the directive is the ruling itself rather than a hand-made paraphrase of it. It then names the register — GOVERNED_SURFACES in scripts/pm/check-governed-merges.mjs — and gives a one-liner that prints today's set.
• §7 and the Post-Task Checklist stop enumerating and defer to feat: Comprehensive CRM example demonstrating all ObjectStack protocol features #14 as a class. They cannot go stale as the set grows.
• The Skills section gains one class pointer, no list.

Four hand-maintained copies to one. Deleting the last one was considered and rejected: a seat must be able to decide from AGENTS.md alone whether the PR in front of it is governed.

What the rebase changed, beyond conflict resolution

86ea8df7d retired the per-PR check, deleted scripts/check-adr-merge-approval.mjs, moved the register, and widened the surface. Three assertions in the earlier draft of this PR were false against main and are rewritten:

1. The one-liner pointed at a deleted file. It now resolves against scripts/pm/check-governed-merges.mjs and was run, not transcribed:

   $ node -e "import('./scripts/pm/check-governed-merges.mjs').then(m=>console.log(m.GOVERNED_SURFACES.map(s=>s.glob).join(' · ')))"
   docs/adr/** · .claude/** · skills/** · AGENTS.md · CLAUDE.md
   

2. The required check ADR maintainer approval no longer exists. The enforcement half now states the real regime: a human merge IS the review record, discipline is the only pre-merge barrier, and detection is the report-only post-merge audit whose list every entry should match a merge the maintainer made in person. The directive says out loud that the retirement is not a relaxation.
3. The set is five surfaces, not three — and AGENTS.md and CLAUDE.md are on it, so this PR is itself a governed-surface PR. §7 now says so explicitly, because that is where a seat decides whether to arm.

The .claude/ widening was intended, not silent. The maintainer's approval quotes the unified list verbatim, including the parenthetical: 含 agents/hooks/settings,不只 skills. Confirmed in the audit script's header and reproduced in the directive as the ruling's own words.

CODEOWNERS re-read rather than recalled. It is now three rows: /docs/adr/, /.github/CODEOWNERS, /scripts/pm/check-governed-merges.mjs. Of the five governed surfaces exactly one is routed, so on the other four nothing requests the maintainer's review automatically. Stated as fact; no fix proposed here — #9513 stays open and undecided.

Verification

Gate union re-derived from the actual changed path and run after the final commit, at 428306907, clean tree, one file changed:

check:nul-bytes PASS check:adr-anchors PASS
check:pm-skill-id-lint PASS check:adr-links PASS
check:required-contexts PASS check:pm-half-states PASS
check:docs-audit-scope PASS check:pm-skill-ratchet PASS
check:doc-authoring PASS check:skill-frame-sync PASS
check:skill-compatibility PASS check:skill-frame-freshness PASS
check:pm-governed-merges --self-test PASS (25 assertions)

check:pm-skill-id-lint re-run after the rewrite as asked: green. The directive quotes three rulings and cites no issue number.

Reverse verification, both directions, from the committed state:

• Positive control — the file is scanned. Appending an issue-ID citation turned check:pm-skill-id-lint red naming the line. Restored byte-identical.
• The finding's measurement, re-run against the new register. Dropping CLAUDE.md back out of the directive's quoted list left the union green. Nothing ties the prose to GOVERNED_SURFACES; the register moving files did not change that. Filed as finding: nothing ties AGENTS.md's governed-surface enumeration to GOVERNED_SURFACES — #9491's pin covers check-run names, not path prefixes #9525.

No changeset: AGENTS.md publishes nothing. Labelled skip-changeset; Check Changeset reports skipped.

Findings filed alongside, unassigned, not touched here

• finding: nothing ties AGENTS.md's governed-surface enumeration to GOVERNED_SURFACES — #9491's pin covers check-run names, not path prefixes #9525 — nothing ties this prose to GOVERNED_SURFACES; the finding: required check-run names are pinned in the workflows but unpinned in every agent instruction file that names them, so a rename leaves the instructions stale with nothing red #9491 pin covers check-run names, not path prefixes.
• finding: AGENTS.md's Skills section lists 2 of the 5 internal playbooks and 9 of the 11 published skills #9526 — the Skills section's two inventories are stale (2 of 5 internal playbooks, 9 of 11 published skills).
• finding: CODEOWNERS routes docs/adr/ and the gate's enforcement chain, but neither governed skill surface #9513 stays open: CODEOWNERS routes one of five governed surfaces.

Generated by Claude Code

---

Generated by Claude Code

[claude]

Rebased onto main at 428306907. The per-PR approval check retired in 86ea8df7d while this PR sat at gates; the register moved, the surface widened to five, and the body above is rewritten against what is actually on main now. The card's own work was untouched by that landing — main's Prime Directive #14 is still a path rule naming one prefix.

One state change on this PR I did not make, recorded here rather than quietly fixed. From the timeline:

ready_for_review by os-steve 2026-08-18T07:51:53Z
added_to_merge_queue by os-steve 2026-08-18T07:52:06Z
removed_from_merge_queue by github-merge-queue[bot] 2026-08-18T08:08:35Z

The queue ejected it because the branch had gone dirty against 86ea8df7d — the merge conflict is the only reason it did not land. Since AGENTS.md joined the governed surface in that same commit, this is a governed PR, and it reached the queue about twenty minutes after the check that used to sit in front of that path retired.

I have converted it back to draft. Nothing was armed (auto_merge: null) and the queue had already dropped it, so the usual reason not to touch another actor's ready-flip — that flipping back destroys arming and queue membership in one step — did not apply, while my push had just made the PR mergeable again and left it sitting ready. Current state: draft: true, auto_merge: null, not in the queue, not on main. Adjudication is the PM's; the full timeline and my reasoning are in the round report on #9403.

Generated by Claude Code

---

Generated by Claude Code