finding(docs): ci-cd-pipeline.md:551 still says check-skills-paths.mjs reads markdown under skills/ only, missing the .claude/skills/ root #7358 added

Filed out-of-scope from #7362 (that card's file surface was content/docs/guide/agent-skills.md and, conditionally, ci-cd-pipeline.md:581 only — this is a different line, a different sentence, unassigned).

What is stale

content/docs/guide/ci-cd-pipeline.md:551, in the "Skill Guide Paths (skills-paths.yml)" section:

Runs scripts/check-skills-paths.mjs, which reads every markdown file under skills/ and asks, of each in-repo path the prose states inside a backtick code span, whether it exists on disk.

This is evergreen present-tense prose describing the script's current scan surface (not a "measured at commit X" statement like the main@6422aa891 sentence a few lines below, which #7965 already reviewed and ruled correct because it declares its measurement point). #7358 widened SCAN_ROOTS to ['skills', '.claude/skills'], so the sentence has been incomplete since then — today's real reading (node scripts/check-skills-paths.mjs) is:

✅  check-skills-paths: OK (88/89 stated path(s) resolve across 20 guide file(s); 1 baselined).
    skills/ — 28/28 resolve across 16 file(s)
    .claude/skills/ — 60/61 resolve across 4 file(s)

A reader who takes "reads every markdown file under skills/" literally would conclude .claude/skills/** guides (including the very console-development.md / no-touch-zones.md pair #7251 moved there) are outside this gate's reach — the exact wrong conclusion #7358 exists to prevent, and the same failure mode #7404 fixed for .github/workflows/skills-paths.yml's header comment. This is the same defect class, one file over, still open.

Why not fixed on the spot

#7362's dispatch boundary was content/docs/guide/agent-skills.md plus ci-cd-pipeline.md:581 "ONLY as the re-measured baseline truth requires" — this is a different sentence at a different line, so it's out of that claim's file surface.

Suggested fix

Name both roots, e.g. "reads every markdown file under skills/ and .claude/skills/" — two words, no numeral to go stale. scripts/__tests__/ci-cd-pipeline-doc.test.ts does not appear to pin this sentence (grepped for the surrounding tokens with no hit), so a prose edit shouldn't move any pin.

finding, no pm:queue — this is an observation for PM triage, not a defect I'm asking to be actioned on a timeline.

Triage: lands in domain:devx (content/docs/guide/ci-cd-pipeline.md:551); rationale: class (a) — wrong, not merely incomplete, which is the boundary that decides it. Bug.

⭐ The sentence is evergreen present tense ("reads every markdown file under skills/"), and the live reading shows the omitted root holds most of the population:

skills/          — 28/28 resolve across 16 file(s)
.claude/skills/  — 60/61 resolve across  4 file(s)

⇒ A reader taking it literally concludes .claude/skills/** guides are outside this gate's reach — including the very console-development.md / no-touch-zones.md pair objectui#7251 moved there — the exact wrong conclusion objectui#7358 exists to prevent. ⛔ That is a false belief with a named consequence, not an omitted member.

⭐ And it is the N−1 shape again: objectui#7404 fixed this same defect class one file over, in .github/workflows/skills-paths.yml's header comment. ⇒ One carrier of the fact was corrected after #7358 widened SCAN_ROOTS to ['skills', '.claude/skills']; this one was not. Same reasoning that admitted objectstack#16198 and objectui#8031 today.

⭐ The filer also drew the distinction that matters on this page and got it right: this is not the main@6422aa891 sentence a few lines below, which objectui#7965 already reviewed and ruled correct because it declares its measurement point. ⇒ A dated claim and an evergreen claim rot differently, and only the evergreen one is wrong here.

priority:p3: one sentence, prose only, no behaviour, no gate moves.

⭐ Fix is two words — "reads every markdown file under skills/ and .claude/skills/" — and the filer applies the right lesson unprompted: no numeral, so nothing to go stale. ⇒ That is the same discipline this seat ratified on objectstack#16200, #16007 and objectui#8122 today; ⛔ do not "improve" it by adding today's counts.

⭐ Pin check already done and worth trusting: scripts/__tests__/ci-cd-pipeline-doc.test.ts does not appear to pin this sentence (grepped for the surrounding tokens, no hit), so a prose edit should move no pin. ⚠️ Re-confirm at PR time rather than inheriting it.

⇒ Three live findings now sit on this one page, and they split cleanly by kind: this card and objectui#7965 are prose edits and could ride together in one PR; objectui#8015 (graded p3 today) is a pin extension to ci-cd-pipeline-doc.test.ts and is a different verification surface. ⛔ None is a duplicate of the others; ⚠️ but whoever takes any of them should check the other two first, since a fourth PR on this page in a week is waste.

⛔ Correctly out of objectui#7362's scope — its dispatch boundary was agent-skills.md plus ci-cd-pipeline.md:581 "ONLY as the re-measured baseline truth requires", and this is a different sentence at a different line.

⛔ This seat grades and routes only: not claimed, not dispatched, no code.

Generated by Claude Code

Claim: PM loop round R46 — domain:devx @ objectui execution seat.
Session: session_01FhBNJcLRZLe8M87VcUgpKr
Branch: claude/issue-8006-ci-cd-pipeline-skills-paths-roots
Worktree: /home/user/objectui-issue-8006
Domain: domain:devx
File surface: the one prose sentence at content/docs/guide/ci-cd-pipeline.md:551 describing check-skills-paths.mjs's scan surface. ⛔ Not ci-cd-pipeline.md:581 (the main@6422aa891 measured-baseline sentence #7965 already ruled correct — it declares its measurement point and is a different kind of claim); ⛔ not the Lint section, which is #8015's.
Dispatch: one os-dev subagent of this session (mode:subagent, model: opus); the dispatch note follows. Labels pm:queue → pm:dispatched and assignee set in one write, read back at 2026-09-07T21:05Z.

fold-or-serial, answered explicitly because #8006 and #8015 share content/docs/guide/ci-cd-pipeline.md: SERIAL, ⛔ not folded.

Gate ① of the fold admission test (same defect form and same fix) fails, and the titles are exactly the trap the gate exists to catch — both read as "a stale claim in ci-cd-pipeline.md", which is a shared keyword, not a shared defect:

finding(docs): ci-cd-pipeline.md:551 still says check-skills-paths.mjs reads markdown under skills/ only, missing the .claude/skills/ root #7358 added #8006 is a stale sentence: check-skills-paths.mjs's SCAN_ROOTS widened to ['skills', '.claude/skills'] in check-skills-paths lost 55 of its 93 assertions when the contributor guides moved out of skills/ — widen SCAN_ROOT to cover .claude/skills #7358 and the prose still names one root. The fix is two words of prose, and the card records that no pin covers that sentence.
finding(docs): ci-cd-pipeline.md's Lint section names 1 of the 4 scripts/check-*.mjs gates that job runs, and nothing pins it the way ci.yml's job table is pinned #8015 is a coverage gap in a pin: scripts/__tests__/ci-cd-pipeline-doc.test.ts scopes its "names every first-party command the job actually runs" assertion to ci.yml alone, so lint.yml's step list drifts unwatched — three blocking gates undescribed. Its fix reaches into the test file, not only the doc.

⇒ Different defect forms, different fixes, different file surfaces (#8015 also touches scripts/__tests__/), each independently checkable. Gates ②–⑤ are moot once ① fails. #8006 is the chain head; #8015 is dispatched only after it lands and rebases rather than racing it.

⛔ Excluded and named so they are not mistaken for members: #8122 (drifted counts, a different document) and #8102 (the i18n gate's class-8 header) — same "stale prose" keyword, different subjects.

Generated by Claude Code

Dispatched — domain:devx @ objectui execution seat, PM session session_01FhBNJcLRZLe8M87VcUgpKr, R46, 2026-09-07T21:05Z. Three of three slots in flight (with #8029 and #8062; disjoint surfaces).

Dev. One os-dev subagent on branch claude/issue-8006-ci-cd-pipeline-skills-paths-roots, worktree /home/user/objectui-issue-8006, from current origin/main (4b5e07a53 or later). Race re-read at 21:06Z: one Claim: on this card, 5575616484, naming that branch.

⛔ Re-derivation before any edit, and ⛔ no numeral in the prose. The brief has the dev run node scripts/check-skills-paths.mjs and read SCAN_ROOTS directly rather than reuse this card's quoted verdict — and it forbids writing any count into the sentence. The card's own suggested fix is numeral-free on purpose ("under skills/ and .claude/skills/") so it cannot go stale a second time; a fresh count in prose would just reset the same clock.

The exclusions are two sentences in the same document that are NOT this defect. ⛔ Line 581 is the main@6422aa891 measured-baseline sentence, which #7965 already ruled correct because it declares its measurement point — a different form of claim under a standing ruling, ⛔ not to be "fixed". ⛔ The Lint section is #8015's.

fold-or-serial: SERIAL, answered on the claim comment. #8006 and #8015 share this file and are exactly the trap gate ① exists to catch — both read as "a stale claim in ci-cd-pipeline.md", which is a shared keyword, not a shared defect. #8006 is a stale sentence fixed in prose with no pin covering it; #8015 is a coverage gap in a pin (ci-cd-pipeline-doc.test.ts scopes its command-parity assertion to ci.yml, so lint.yml drifts unwatched) whose fix reaches into scripts/__tests__/. Different forms, different fixes, different surfaces. ⇒ #8006 is the chain head; #8015 is dispatched only after it lands.

⚠️ One instruction that can flip the plan: the card claims no pin covers the sentence. The dev must verify that rather than inherit it — if a pin does cover it, both move together.

Generated by Claude Code

os-dev-report

{
  "issue": 8006,
  "status": "done",
  "branch": "claude/issue-8006-ci-cd-pipeline-skills-paths-roots",
  "worktree": "/home/user/objectui-issue-8006",
  "pr": "https://github.com/objectstack-ai/objectui/pull/8417",
  "pr_number": 8417,
  "head_sha": "76597948d4784dca44eba6145463d759936a1774",
  "base": "main @ 4b5e07a53c78ed4fcdabf4d92bbb813ac5f5e7ea",
  "premise_still_valid": true,
  "premise_notes": "Defect real and unchanged in substance. Two line numbers in the card have drifted: the target sentence is at line 574 today, not 551; the main@6422aa891 measured-baseline sentence the fence protects is at line 599, not 581. I honoured the fence by CONTENT (the sentence that declares its own measurement point), not by line number, and left it untouched. Attribution checked: the widening landed in 7c7729859, titled for objectui#7358 and merged as PR #7407 — the card's #7358 is correct; the script docblock's #7251 is the earlier MOVE that opened the gap, not the widening.",
  "re_derived_scan_roots": "export const SCAN_ROOTS = ['skills', '.claude/skills'];  (scripts/check-skills-paths.mjs:154)",
  "re_derived_gate_verdict": "node scripts/check-skills-paths.mjs -> exit 0 | check-skills-paths: OK (88/89 stated path(s) resolve across 20 guide file(s); 1 baselined). / skills/ - 28/28 resolve across 16 file(s) / .claude/skills/ - 60/61 resolve across 4 file(s)",
  "pin_covers_the_sentence": false,
  "pin_evidence": "Static: grep -in 'skill' scripts/__tests__/ci-cd-pipeline-doc.test.ts returns ZERO lines, with a positive control on the same file (grep -c 'ci-cd-pipeline' = 30) so the grep is live. Repo-wide: the only occurrence of the sentence text anywhere is the doc itself. Dynamic: with the fix committed, the sentence was replaced by an obvious falsehood and the ELEVEN test files that read content/docs/guide/ci-cd-pipeline.md still passed 11/11 files, 404/404 tests, and check-doc-links still exited 0. Observed ablation direction is STAYS GREEN — that is the finding, not a failed reverse validation, and it is exactly why the sentence could drift unnoticed. Restore leg proved by blob hash equality against HEAD (ad91e99c287f532abcbeee4faaa884baed13c18e) plus an empty git diff HEAD, never by an exit code; mutation leg proved on disk by injected-marker count 1 and deleted-text count 0 plus a differing blob hash.",
  "files_changed": ["content/docs/guide/ci-cd-pipeline.md (+4 / -4, one paragraph re-wrapped)"],
  "changeset": "none owed — node scripts/check-changeset-presence.mjs exit 0: 'Compared the working tree with 4b5e07a53 (merge-base with origin/main): 1 file(s) changed, 0 of them published source of a package the release covers ... No source or published contract of a released package changed in this range, so no changeset is owed.' No skip-changeset label applied: this repository has no such mechanism and the pin in this very document asserts it is never wired into a workflow or a gate.",
  "summary": "One sentence in the Skill Guide Paths section of content/docs/guide/ci-cd-pipeline.md now names both scan roots: 'reads every markdown file under `skills/` and `.claude/skills/`'. No numeral appears in the new prose, so it cannot go stale the way a count would; the phrasing matches what the neighbouring Skill Examples section already writes for the same pair of roots, so the two sections now agree. Scope fences honoured: the main@6422aa891 measured-baseline sentence (objectui#7965 ruled its form correct because it declares its measurement point) and the Lint section (objectui#8015, serialised behind this card on the same file) were left untouched. PR is DRAFT; the ready flip is the PM's read. The file is not governed surface — check-governed-queue-guard.mjs --test says NOT GOVERNED — but the dispatch's draft instruction governs regardless.",
  "tests": "ALL run in the worktree at repo root, exit codes captured by redirect-then-capture, never through a pipe. (1) node scripts/check-skills-paths.mjs -> 0, verdict quoted above. (2) node scripts/check-doc-links.mjs -> 0, 'Links are valid across 17 scan roots.' (3) node scripts/check-doc-fence-languages.mjs -> 0, 'every TypeScript block in 227 document(s) is fenced ts/tsx/typescript ...'. (4) node scripts/check-control-bytes.mjs -> 0, 'scanned 6661 tracked text file(s); skipped 85 binary'. Plus a direct control-byte sweep of the edited file: grep -naP '[\\x00-\\x08\\x0b\\x0c\\x0e-\\x1f\\x7f]' -> no match. (5) node scripts/check-changeset-presence.mjs -> 0, quoted above. (6) node scripts/check-governed-queue-guard.mjs --test content/docs/guide/ci-cd-pipeline.md -> 0, 'NOT GOVERNED - 1 path(s) checked against 5 governed surface(s); none matched.' (7) The eleven test files that actually READ this document (enumerated by grep -rln 'guide/ci-cd-pipeline.md'), run through the shared verify lock: OS_VERIFY_LOCK_SLOT=objectui-8006-doc-pins bash /home/user/objectstack/scripts/pm/os-verify-lock.sh -c 'pnpm exec vitest run scripts/__tests__/ci-cd-pipeline-doc.test.ts ... check-handler-key-read-sites.test.ts' -> 'os-verify-lock: VERDICT command-exit 0 - held the lock 17s - waited 0s'; vitest reported 'Test Files 11 passed (11) / Tests 404 passed (404)'. (8) ABLATION, run on the committed tree with a trap '<restore>' EXIT INT TERM using absolute paths from git rev-parse --show-toplevel: sentence replaced by a falsehood, on-disk mutation proved (marker count 1, original-text count 0, blob 84ea1fb9 differs from HEAD blob ad91e99c), same eleven files re-run -> 11 passed / 404 passed, check-doc-links -> 0; restore verified by blob hash equality and empty git diff HEAD. No build/dist leg applies: the subject under test is markdown read from the working tree, not a package entry resolved through exports.",
  "local_verification_scope": "Narrowed and declared: this diff touches one markdown file under content/docs, so the affected-package set is empty of source packages. I ran every gate whose scan surface includes content/docs prose (links, fences, control bytes, changeset presence) plus every test file that reads this document by path. I did NOT run pnpm lint or the full pnpm test — repo-wide sweeps are CI's run, and this narrowing is declared here rather than silently taken.",
  "mcp_calls": "4 — create_pull_request, pull_request_read (body read-back), search_issues (one targeted dedupe hit, channel switch declared below), add_issue_comment (this report). Card body, labels and comment read-back went through the zero-quota repo-scoped REST channel instead.",
  "channel_declaration": "Probed first as required: repo-scoped REST read GET /repos/objectstack-ai/objectui/issues/8006 -> HTTP 200, so REST is live in this container for repository-scoped paths. But GET /search/issues -> HTTP 403 with body 'This GitHub API path is not available: sessions are bound to their configured repositories. Use repository-scoped endpoints', on BOTH the control query and the target query — the proxy's path whitelist, exactly as this repo's AGENTS.md records. Per the dispatch rule I therefore switched to ONE targeted MCP search_issues for the dedupe read. It returned total_count 21 with items, so it is self-validating and needed no known-hit control.",
  "body_readback": "PR 8417 body re-read to its end after creation: intact, no tag-shaped fragment eaten, no angle-bracket shapes were written into it in the first place. The platform appended its own bare attribution footer block below the session-URL line, the documented normalisation. Closing-keyword scan of the stored body: the only close/fix/resolve keyword adjacent to a card number is the intended leading 'Fixes #8006'; #7358, #7407, #7965, #8015 all appear bare. The commit message carries NO card trailer at all (verified: zero close/fix/resolve matches), so a squash cannot double-declare the relationship.",
  "open_questions": [],
  "out_of_scope_findings": [],
  "out_of_scope_note": "Nothing new filed, and this is a decision rather than a silence. The one candidate observation — 'prose in this document asserts a script constant with nothing pinning it' — was dedupe-read before filing and the class is already owned: objectui#8015 is OPEN on exactly this file for exactly this shape (a doc claim about gate scripts that no pin watches), and objectui#4170 plus objectui#7901 are closed cards on the same class ('a second unpinned current-state claim in the same section', 'nothing watches the counts published in content/docs prose'). Filing another would be the duplicate the search step exists to prevent. Flagged here for PM triage rather than dropped.",
  "four_axis_analysis": {
    "options": [
      "A — name both roots in numeral-free prose: 'reads every markdown file under `skills/` and `.claude/skills/`' (the card's own suggestion, and the phrasing the neighbouring Skill Examples section already uses). IMPLEMENTED.",
      "B — restate the sentence with the current measurement, e.g. 'reads N markdown files across two roots', following the form of the `main@6422aa891` baseline sentence further down the page.",
      "C — stop naming roots in prose at all and point the reader at the constant: 'reads every markdown file under the roots listed in `SCAN_ROOTS`'.",
      "D — leave the prose and add a pin test that derives the sentence from `SCAN_ROOTS`, so the document is mechanically forced to follow the script."
    ],
    "axis_1_real_business_need": "实测而非推测:`node scripts/check-skills-paths.mjs` 今天答出两个 root,`.claude/skills/` 一侧是 60/61 across 4 file(s) —— 也就是这条句子漏掉的那一半,恰好是被检查断言数更多的那一半。谁在读这句话?本仓每一个 agent 席位:这份文档是 CI 行为的唯一权威叙述,而这个 gate 的整个存在理由(objectui#7358)就是「守卫停止看某棵树时没有任何东西变红」。方案 A 服务的是这个真实存在的读者场景。方案 B 服务的是「让读者知道当前规模」—— 这不是任何人实际需要的问题:读者需要知道的是「我的文件在不在门禁范围内」,一个计数答不了这个问题。方案 D 服务的是一个真实需求(防再次漂移),但它的成本落在别处,见轴二。⇒ A 与 D 有真实拉动,B 没有。",
    "axis_2_long_term_soundness": "北极星是 contract-first、no workarounds。这里的「契约」是 `SCAN_ROOTS` 这个常量,散文是它的下游消费者。A 让散文陈述契约的**内容**(两个 root),这是稳定的:root 集合变了,散文才需要动,而那正是应该有人重读这一段的时刻。B 让散文陈述契约的**当前测量结果**,那是本卡自己要修的那类缺陷的再生产 —— 页面上已经有一条 `main@6422aa891` 句子必须靠声明自己的测量点才活得下去(objectui#7965 的裁决),再造一条只会把那个例外变成惯例。C 表面上最「contract-first」,实际最差:它把读者从文档踢回源码,而这份文档存在的全部理由就是替读者读源码;而且它连**可证伪性**都丢了 —— 一句「见 SCAN_ROOTS」永远不会过时,也永远不会告诉任何人任何事。D 长期最正确,但它是一次门禁扩张,该由拥有该文件钉子缺口的那张卡(objectui#8015,已串行排在本卡之后、且正是「钉子看不见」这一类)统一做,而不是在一张 p3 散文卡里顺手加一条新验证面。⇒ A 现在最优,D 归 #8015 的射程。",
    "axis_3_prevent_ai_mistakes": "这条轴在本卡上权重最高,因为受害者就是 AI 席位本身。失败形态已实测:一个照字面读这句话的 agent 会得出「`.claude/skills/**` 的向导不在门禁内」,于是在那里写下一个死坐标而不去核对 —— 这正是 #7358 存在要防的那个结论,也是当初 55/93 条断言静默离开检查面的同一族错误。判据是「结构上更难写错」:A 让两个 root 都被逐字写出,一个 AI 无法从中读出「只有一个 root」;B 引入一个数字,而数字必然漂移,漂移后的文档比含糊的文档更危险 —— 它给出一个**自信的错误读数**,而 AI 对带数字的断言的信任度更高;C 制造的是宽容消费端的镜像:文档不再声明任何可被证伪的事实,AI 读到「见常量」要么去读源码(那这句话白写),要么按自己的猜测继续(而猜测正是要防的)。「声明即强制」在这里读作:文档声明的范围必须与运行时实际扫描的范围一致,A 是唯一让两者逐字对上的选项。⇒ A 明确胜出,B 反向,C 无效。",
    "axis_4_startup_focus_no_scope_creep": "创业阶段聚焦原则(维护者 2026-08-04)与「过渡也从紧、不渐进」(维护者 2026-08-27)。A 是 4 行 diff、零新增机制、零新增维护面,是本卡范围的**下限**实现,完全符合从紧。D 会新增一条门禁 + 一份钉子测试 + 一个未来要维护的推导,这是能力扩张;在没有拉动证据(本卡是 p3、类已被 #8015 持有)时按 implementation-first 处置 —— 不做。B 不但不省事,还会把这一段拖进「每次 gate 规模变化都要回来改文档」的循环,是最贵的选项。另外三条边界也按从紧处理并已落地:不碰 line 599 的已裁句子、不碰 Lint 一节(#8015 会 rebase 到本改动之上)、不新立重复卡。⇒ A。",
    "recommendation": "A,四条轴一致指向它,没有需要交维护者拍板的权衡。轴三是决定性的:唯一一个让 AI 席位在结构上读不出错误结论的选项;轴一与轴四把 B 和 C 排除(一个再生产本卡要修的缺陷类,一个把可证伪性丢掉);轴二承认 D 长期最正确,但它属于 objectui#8015 的射程,在这里做等于在一张 p3 散文卡上偷偷扩张验证面。已按 A 实施。"
  }
}

Verification artefacts for this run live in the session transcript and in PR 8417's body; the branch head is 76597948d4784dca44eba6145463d759936a1774.

Generated by Claude Code

Navigation Menu

Repository navigation

finding(docs): ci-cd-pipeline.md:551 still says check-skills-paths.mjs reads markdown under `skills/` only, missing the `.claude/skills/` root #7358 added #8006

What is stale

Why not fixed on the spot

Suggested fix

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

finding(docs): ci-cd-pipeline.md:551 still says check-skills-paths.mjs reads markdown under skills/ only, missing the .claude/skills/ root #7358 added #8006

Description

What is stale

Why not fixed on the spot

Suggested fix

Activity

os-zhuang commented on Sep 7, 2026

claude commented on Sep 7, 2026

claude commented on Sep 7, 2026

baozhoutao commented on Sep 7, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

finding(docs): ci-cd-pipeline.md:551 still says check-skills-paths.mjs reads markdown under `skills/` only, missing the `.claude/skills/` root #7358 added #8006