docs(agents): AGENTS.md counts the invocation guard's refusals as 「两条」 and enumerates exactly two — objectui#7814 adds a third, and the fix is a POINTER, not a new number

[os-try-charles]

Surfaced by the objectui#7814 dev as to file and ⛔ deliberately not edited by it — binding ruling 2 of that dispatch forbids a dev touching AGENTS.md (governed surface). Adopted and filed here by the domain:devx @ objectui seat.

The sentence

AGENTS.md:204 on main today, verbatim:

• 两条现在都会直接失败,由 scripts/vitest-invocation-guard.mjs 拦下:vitest root 不是仓根 → 拒绝;-- 后面还有参数 → 拒绝。

It states the count 两条 and enumerates exactly those two.

Control on the same read: AGENTS.md carries 34 ⛔ markers and 0 occurrences of a nonsense token ⇒ this is a reading of a file that exists and parses, ⛔ not a path that silently missed.

Why it goes stale

PR objectui#9504 (card objectui#7814) adds a third refusal — subsumed-positional-filter: when one positional filter contains another as a substring, Vitest unions them and the contained one adds zero files, so an appended path silently widens the run instead of narrowing it.

⇒ the moment that PR lands, 两条 is an undercount, and the enumeration beside it is incomplete.

⭐ The repair is a POINTER, ⛔ not 两条 → 三条

⭐ The objectui#7814 dev already solved this correctly in the file it was allowed to touch: the guard's own docstring header carried the same the two count, and the PR replaced it with a pointer to the pin test rather than with a new number. AGENTS.md wants the same treatment.

⛔ Changing the number is the wrong fix. This repo's own commandment #9 argues against writing the count down at all: a count in prose is a fact derived once and never re-derived, and the next refusal added will falsify it again. The durable form names the instrument:

…由 scripts/vitest-invocation-guard.mjs 拦下;它拒绝哪些形态,以 scripts/__tests__/vitest-invocation-guard.test.ts 为准。

⚠️ Third instance of this exact class filed by this seat today, which is the real signal: objectui#9493 (three "its merge" rows naming shas that are not on main) and objectui#9502 (two files asserting the shard jobs are not required checks, both false since 2026-08-24). ⭐ Same shape every time — a count or a list, written once, never re-derived.

⛔ Governed surface — fences for whoever takes it

1. ⛔ AGENTS.md is governed. A PR touching it is draft only, needs GOVERNED_APPROVERS (os-zhuang / hotlong) approval and a human merge. ⛔ This seat will never flip it ready, enqueue it, arm auto-merge, approve it or merge it.
2. ⚠️ It queues. AGENTS.md is a hot-file serial surface in this lane — objectui#9466 is already open against it awaiting an authorized approval, and the seat's queue has objectui#7833 and objectui#8875's second PR behind it. ⛔ Do not open a competing PR; take a slot.
3. ⚠️ Ordering: this card's premise only becomes true when objectui#9504 lands. ⛔ Do not edit the sentence before then, or the file will describe a refusal the tree does not have — the same defect pointing the other way.

Acceptance

1. The sentence no longer carries a refusal count, and points at the instrument that enumerates them.
2. ⛔ No new number is introduced anywhere in the paragraph.
3. The enumeration that remains (if any) is shown to match the guard's verdicts on the merged tree, with a control proving the reader could have found a mismatch.
4. gates_weakened: NONE — this is prose; ⛔ nothing about what the guard refuses may change.

Blocked-by: #7814
Restart-when: `grep -c 'subsumed-positional-filter' scripts/vitest-invocation-guard.mjs` reads ≥1 on `main`

[os-try-charles]

UNBLOCKED — pm:blocked → pm:queue. The Restart-when: fired on its own, 25 minutes after it was written.

domain:devx @ objectui seat, session_013VGeMu3p6qEFWR6K6GGLaW, R58, 2026-09-14T13:2xZ.

Blocked-by:   #7814   → closed/completed, PR objectui#9504 merged as 8fa7d69af2394b (13:27:24Z)
Restart-when: `grep -c 'subsumed-positional-filter' scripts/vitest-invocation-guard.mjs` reads ≥1 on `main`

Measured on main: 2. Control on the same file: a nonsense token reads 0, so the reader can return a zero and this one is a reading.

⇒ this card's premise is now true rather than anticipated: the guard really does carry a third refusal, so AGENTS.md:204's 「两条现在都会直接失败」 — which states the count and enumerates exactly two — is an undercount today, ⛔ not "will be".

⭐ The ordering fence in this card's body did its job: it said ⛔ do not edit the sentence before objectui#7814 lands, or the file would describe a refusal the tree does not have. That window is now closed in the right direction.

⛔ Nothing else about this card has changed

• ⛔ The repair is still a POINTER, ⛔ not 两条 → 三条. Commandment 完善设计器的每一个细节 #9 is the reason, and objectui#9504 already set the precedent in the file it was allowed to touch: the guard's own docstring carried the same the two count and the PR replaced it with a pointer to the pin test. ⭐ Do the same thing here.
• ⛔ AGENTS.md is governed surface — draft only, GOVERNED_APPROVERS approval, human merge. ⛔ This seat will never flip it ready, enqueue it, arm auto-merge, approve it or merge it.
• ⚠️ It queues. objectui#9466 is already open against AGENTS.md awaiting an authorized approval, with objectui#7833 and objectui#8875's second PR behind it. ⛔ Do not open a competing PR — take a slot.

⛔ Unassigned, and ⛔ not claimed by this comment.

---

Generated by Claude Code

[os-try-charles]

Claim: PM loop round R60
Session: session_015h79niBMyoB1xcaQje3uiz
Branch: claude/issue-9505-agents-md-invocation-guard-pointer
Worktree: objectui-issue-9505
Domain: domain:devx
File surface: AGENTS.md — the single sentence at/near :204 counting the invocation guard's refusals (stop on breach; explain in the report)
Container & model: S (mechanical-ish, but with judgement about the pointer's wording), mode:subagent, model: default judgement tier (opus) — quoting this fire's node scripts/pm/dispatch-gates.mjs --tier AGENTS.md: "no path-derived mandate: the surface hits none of the 3 declared glob(s)". ⚠️ AGENTS.md is governed but is ⛔ NOT on the 协议语义面 list (SKILL.md · the decision-framework copy · .claude/agents/os-dev.md · skills/**), so 强制条款① does not force the contract-review tier; this is a line-level prose change on the governance surface ⇒ default tier per 「治理面上一行级机械文本改动的卡，降为默认判断档施工」.
Clause-②: no
Thread-read: 5664750513
Serial constraints cleared: AGENTS.md — 0 of 10 open PRs hold it, re-measured this fire over all open PRs with 1,241 filenames examined (control); positive control, same instrument: scripts/pm/check-half-states.mjs → PR #9391, so the zero discriminates. ⭐ This card's own fence was 「behind objectui#9466」, and #9466 MERGED at 01:45Z — that is what released it. ⛔ objectui#9584 does not hold AGENTS.md (checked).

---

Seat: domain:devx @ objectui, session_015h79niBMyoB1xcaQje3uiz, R60, 2026-09-16T02:29Z.

Premise re-verified on origin/main at claim time, ⛔ not inherited from the R58 unblock comment:

reading	value
git grep -c 'subsumed-positional-filter' origin/main -- scripts/vitest-invocation-guard.mjs	2 ⇒ the third refusal really is on main
CONTROL (nonsense token, same file)	0 ⇒ the reader can return zero, so the 2 is a reading
git grep -c '两条现在都会直接失败' origin/main -- AGENTS.md	1 ⇒ the undercount is still live today

⇒ the defect is present, not anticipated. ⛔ Governed surface: this is a terminal four-piece, ⛔ not a landing.

---

Generated by Claude Code

[os-try-charles]

os-dev-report

{
  "issue": 9505,
  "status": "done",
  "branch": "claude/issue-9505-agents-md-invocation-guard-pointer",
  "pr": "https://github.com/objectstack-ai/objectui/pull/9587",
  "session": "session_015h79niBMyoB1xcaQje3uiz",
  "premise_still_valid": true,
  "refusal_set_carriers_found_in_AGENTS_md": {
    "count_or_enumeration_carriers": 3,
    "boundary_cases": 1,
    "detail": [
      "C1 section heading — 怎么跑测试（有两种写法会静默假绿 —— 现已机械拦截）: a count. REPAIRED (declared deviation).",
      "C2 陷阱一 / 陷阱二 — an ordinal enumeration of exactly two. LEFT as narrative, now explicitly flagged non-exhaustive by C3's replacement; matches what objectui#9504 did in the guard docstring, which kept its Trap 1/2/3 narratives and removed only the count.",
      "C3 the sentence the card names — 两条现在都会直接失败…: count plus inline enumeration. REPAIRED — now points at scripts/__tests__/vitest-invocation-guard.test.ts.",
      "BOUNDARY: the 路径过滤零匹配也不再是绿的 bullet describes what is in fact the missing-path-filter verdict, attributing the non-zero exit to passWithNoTests alone. Not a count, so not repaired; noted in the PR Acceptance notes."
    ],
    "control": "git grep -c on a nonsense token in AGENTS.md ⇒ 0 hits, exit 1, so the enumeration could have returned zero"
  },
  "summary": "Replaced the invocation guard's refusal COUNT in AGENTS.md with a pointer at scripts/__tests__/vitest-invocation-guard.test.ts, matching objectui#9504's precedent inside the guard's own docstring, per commandment #9. The measurement made the card's argument stronger than the card knew: evaluateVitestInvocation returns FOUR verdict codes today (double-dash-args, package-cwd, missing-path-filter, subsumed-positional-filter), not three — missing-path-filter was already outside 两条 before objectui#7814 existed. So incrementing to 三条 would have written a NEW false number. Two of three count carriers repaired; the ordinal trap narratives were kept and are now explicitly declared non-exhaustive. Governed surface: draft PR only, never flipped ready, never enqueued, no auto-merge, no approval, no labels written.",
  "gates": {
    "node scripts/check-governed-queue-guard.mjs --test AGENTS.md": 3,
    "node scripts/check-changeset-presence.mjs": 0,
    "node scripts/check-doc-links.mjs": 0,
    "node scripts/check-new-cross-file-line-citations.mjs": 0,
    "node scripts/check-doc-expression-carriage.mjs": 0,
    "node scripts/check-doc-fence-languages.mjs": 0,
    "node scripts/check-control-bytes.mjs": 0,
    "pnpm exec vitest run scripts/__tests__/vitest-invocation-guard.test.ts": 0
  },
  "gate_verdict_lines": {
    "governed-queue-guard": "⛔ GOVERNED — 1 of 1 path(s) are on a governed surface: AGENTS.md x1 — the repo-root agent instruction file. Exit 3 is the CORRECT answer here, not a failure.",
    "changeset-presence": "✅  No source or published contract of a released package changed in this range, so no changeset is owed. (authoritative 'nothing owed' statement; no skip-changeset label applied — it is a phantom in this repo)",
    "doc-links": "Links are valid across 17 scan roots.",
    "new-cross-file-line-citations": "VERDICT new-cross-file-line-citations: 0 new citation(s), enforcement report-only -> exit 0",
    "control-bytes": "✅  check-control-bytes: OK (scanned 7737 tracked text file(s); skipped 85 binary).",
    "doc-fences": "✅  check:doc-fences — every TypeScript block in 227 document(s) is fenced ts/tsx/typescript …",
    "invocation-guard pin test": "Test Files 1 passed (1) · Tests 44 passed (44) — the pointer target exists, is green, and is unchanged by this branch"
  },
  "gates_weakened": "NONE — prose only; nothing about what the guard refuses changed",
  "line_budget": {
    "file": "AGENTS.md",
    "lines_before": 590,
    "lines_after": 593,
    "net": "+3",
    "diff_stat": "1 file changed, 7 insertions(+), 4 deletions(-)"
  },
  "files_changed": ["AGENTS.md"],
  "deviations": [
    "DECLARED SCOPE DEVIATION: the claim's file surface named only the single sentence at/near :204. I also repaired carrier C1 (the section heading's 有两种 count), one line of the same paragraph-block, as a bounded in-place fix. All four conditions hold: same defect class; mechanical, shape pinned by the objectui#9504 precedent; no other claim holds the file (re-measured this fire: 0 of 10 open PRs touch AGENTS.md, 194 filenames examined, positive control scripts/pm/check-half-states.mjs ⇒ PR #9391 so the zero discriminates); same gate family, no new validation surface. Rationale: leaving it would have shipped a file whose own section title still said 两种 — falsified by this PR's own body. It is one word and trivially revertable. The claim's file-surface supplement is owed to the seat this round and this report is it.",
    "The claim comment quotes `node scripts/pm/dispatch-gates.mjs --tier AGENTS.md`, but that script does NOT exist in objectui (only scripts/pm/check-half-states.mjs does). The tier conclusion is independently correct — AGENTS.md is governed but not on the 协议语义面 list — so nothing here changes; flagged for the seat's own tooling.",
    "objectui has no scripts/pm/os-verify-lock.sh; the pin test was run through the sibling checkout's copy at /home/user/objectstack/scripts/pm/os-verify-lock.sh, since the lock is per-container, with OS_VERIFY_LOCK_SLOT=objectui-issue-9505. VERDICT command-exit 0, held 2s, waited 0s."
  ],
  "tests": "pnpm exec vitest run scripts/__tests__/vitest-invocation-guard.test.ts (repo root, no --, under the shared verify lock) ⇒ Test Files 1 passed (1), Tests 44 passed (44). No ablation is owed: this branch adds no guard and changes no behaviour — it is prose. On-disk mutation proof for the edit itself: grep -c '两条现在都会直接失败' AGENTS.md ⇒ 0 and grep -c '有两种写法会静默假绿' AGENTS.md ⇒ 0 (both were 1 before), grep -c 'scripts/__tests__/vitest-invocation-guard.test.ts' AGENTS.md ⇒ 2. Acceptance-3 control: git grep -o \"code: '…'\" on origin/main:scripts/vitest-invocation-guard.mjs ⇒ 4 distinct codes; control code: 'no-such-verdict' ⇒ 0 hits, exit 1.",
  "mcp_calls": "0 — no MCP GitHub tool of any kind was called, read or write",
  "api_writes": "4 — POST /repos/objectstack-ai/objectui/pulls (1, draft); git push to refs/heads/claude/issue-9505-agents-md-invocation-guard-pointer (2: empty-branch route probe, then the commit); POST /repos/objectstack-ai/objectui/issues/9505/comments (1, this report). ZERO label writes, zero PATCH.",
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: the 路径过滤零匹配也不再是绿的 bullet attributes missing-path-filter's refusal to passWithNoTests alone; its real guard message is 'vitest 调用被拒绝：路径过滤指向不存在的文件'. A missing/imprecise member, not a copy-and-it-fails example ⇒ below the filing bar. 承接者: whoever next expands this section.",
    "noted, not filed: the section expands two of the guard's four verdicts; no narrative exists for missing-path-filter or subsumed-positional-filter. Deliberately not added — acceptance item 2 forbids a new number in this paragraph. 承接者: the seat, if it wants the doc expanded.",
    "noted, not filed: QUICK_REFERENCE.md:57 points at AGENTS.md and the guard script and carries NO count — checked, it is clean, so this defect class has exactly one document carrier. 承接者: none needed."
  ]
}

---

Generated by Claude Code