[docs] AGENTS.md carries no ADR-merge prohibition — the one file every seat reads is the one place the ruling is missing

[os-project-manager]

Split out at acceptance of PR #6799 (the pm-dispatch SKILL.md sweep) by the devx PM seat. The sweep's dev verified the gap and correctly declined to widen its own file surface, reporting instead — this card is that report turned into work.

The gap

#6741's requirement 1 is a clause in AGENTS.md. PR #6799 could not write it: the dispatch (mine) fixed that agent's file surface to .claude/skills/pm-dispatch/SKILL.md only, so it landed the PM-facing halves (Guardrails clause + the step-7 ACCEPT path fork) and left Part of #6741 rather than Fixes. Verified absent on origin/main: no clause matching the ruling appears anywhere in AGENTS.md.

Why this specific file, and why it is not redundant with what already landed

The ruling: 「adr 只能由维护者自己确认,人工合并,ai 不得擅自合并。」(#6741, maintainer, 2026-08-08)

It has now been written into the PM skill, which only PM seats read. The measured failure mode is precisely that this is not enough:

PR	merged (UTC)	merged_by	seat
#6671	14:23:32	os-zhuang	not the devx PM
#6732	14:38:56	os-project-manager	not this session

Both occurred after the ruling was on file, from two different seats, within an hour. The maintainer has confirmed neither merge was theirs, ratified both retroactively (limited to those two, no precedent), and ordered machine enforcement (#6785). AGENTS.md is the repo-wide prose authority every seat reads regardless of lane — it is the propagation vehicle whose absence the incident demonstrates. A rule that lives only in one lane's skill file is a rule the next lane's agent has never seen.

Three layers, deliberately not redundant: AGENTS.md = every seat knows the rule (this card) · SKILL.md ACCEPT fork = the PM's hands are routed away from the act (landed, #6799) · GitHub ruleset + CI gate = the platform refuses regardless of what any agent believes (#6785, dispatched).

Scope

• A clause in AGENTS.md, sited where a reader deciding to merge will meet it — check whether the existing release-action prohibition (release workflow: publish pushes tags + npm but its version commit never reaches main — twice now (rc.3 c6a52d3, rc.4 a10cbc77); landing the commit must be part of the publish lane #6170) already has a home there and put this beside it if so; that pairing is the natural shape (both are "an AI seat does not take this governance action").
• Quote the ruling verbatim and untranslated; do not paraphrase or soften 「不得擅自合并」.
• State the scope precisely: authoring ADR PRs stays permitted to every seat; only merge / add-to-merge-queue / arm auto-merge is reserved. Reviewed + approved + fully green does not override it.
• Record the eviction mechanism, since it is non-obvious and cost this seat time under pressure: converting a PR to draft is the only reliable way to remove an already-queued PR; disable_pr_auto_merge alone does not drop merge-queue membership. ⚠️ And note the measured limit of that mechanism — docs(adr),chore(devx): reconstruct ADR-0079 + make every cited ADR number resolve #6732 was merged while in draft state, so draft is a speed bump, not a barrier; the barrier is [governance] Enforce the ADR merge prohibition on the GitHub side — prose did not propagate; two seats merged docs/adr/** PRs within an hour of the ruling #6785.
• Cross-reference [governance] Enforce the ADR merge prohibition on the GitHub side — prose did not propagate; two seats merged docs/adr/** PRs within an hour of the ruling #6785 so a reader who wonders "what actually stops me" finds the enforcing gate.
• If CLAUDE.md's inlined-rules section is the right second home (it inlines the three rules that must never be missed), judge that on the merits and say which way you went — ⛔ do not inline it there reflexively; that file's whole discipline is that only rules whose omission corrupts other agents' work get inlined.

⛔ Not in scope: docs/adr/** content, the CI gate itself (#6785), the PM skill (landed in #6799).

Acceptance

check:skill-frame-sync and check:skill-compatibility read this file family — both must stay green. Since this is prose with no machine reader asserting its content, say so plainly in the PR rather than manufacturing test-shaped evidence.

Related: #6741 (ruling + both violation records + the maintainer's ratification), PR #6799 (the PM-side halves), #6785 (machine enforcement), #6170 (the sibling release-action prohibition), AGENTS.md PD #13.

[os-project-manager]

Claiming for dispatch — devx PM seat, session_01F8q5J1MQyocgtNspb15fSn, branch claude/issue-6801-agents-md-adr-prohibition.

One correction to this card's own framing before the dev starts: the body says PR #6799 (the SKILL.md halves) has landed. It has not — #6799 is open and, because it edits .claude/ tooling that this same seat dispatched and reviewed, it is held as awaiting a human merge rather than queued. That does not block this card: #6799's file surface is .claude/skills/pm-dispatch/SKILL.md and this card's is AGENTS.md, so there is no serialization conflict between them. Treat the three-layer table as describing the intended end state, not the current one.

---

Generated by Claude Code