[finding] the retirement playbook's walk-radius mandate names `<pkg>#test`, but since #16466 the 6 split packages hash the radius on `#test:repo` — and the gate reds if it is on `#test`

[claude]

The retirement playbook's walk-radius mandate names turbo.json's <pkg>#test inputs. Since #16466 split the test task, that is the wrong task for 6 of this repo's packages, and the gate enforces the opposite in both directions — so an executor following the sentence literally reddens the build twice and has to work out why from the gate's message rather than the instruction that sent them there.

Surfaced while dispatching #15567 (closed as discharged). ⛔ Every reading below was taken by the filing seat on origin/main 2bdabe60e, not transcribed from a report. Filed bare and unassigned — ⛔ no domain:*, no grade, no type; all three are triage's.

⚠️ Routing input only. The subject is .claude/skills/spec-property-retirement/SKILL.md, a governed surface — ⛔ not this seat's to edit, and ⛔ this seat did not touch it.

The sentence, verbatim

.claude/skills/spec-property-retirement/SKILL.md, the tree-scoped-pin checklist item:

⭐ 半径按包申报一次:写进 scripts/cross-package-test-inputs.mjs 的 CROSS_PACKAGE_TEST_INPUTS 并配齐 turbo.json 的 <pkg>#test inputs,turbo 才 hash 得到它、CI 分片并集才看得见;… ⛔ 半径未申报的 tree-scoped pin 不是完成的退役。

What the tree actually requires

Measured on origin/main 2bdabe60e, from turbo.json:

task	inputs	$TURBO_ROOT$ entries
@objectstack/spec#test	4	0
@objectstack/spec#test:repo	34	30

⇒ On a split package the radius hashes on #test:repo, and #test must carry none. scripts/check-cross-package-test-inputs.mjs enforces both directions, and its own self-test pins the failing case: "a wide input LEFT on PKG#test reds even when test:repo is complete."

⇒ Following the sentence literally puts the inputs on the task the gate requires to be empty. Two reds: the missing declaration on test:repo, and the forbidden one on test.

The population is 6, measured — ⛔ not "spec only"

Packages declaring a test:repo script (git grep -l '"test:repo"' -- 'packages/**/package.json'):

packages/core   packages/objectql   packages/rest
packages/runtime   packages/spec     packages/types

The playbook never mentions the split

git grep -c "test:repo" -- .claude/skills/spec-property-retirement/SKILL.md   →  0
control, same file: CROSS_PACKAGE_TEST_INPUTS                                 →  1

⇒ The zero is a reading, not a broken pattern: the file is readable and the neighbouring symbol is found. The playbook has no notion that #test may be the wrong task.

⭐ Why this is worth a card rather than a footnote

The mandate's own closing clause is ⛔ 半径未申报的 tree-scoped pin 不是完成的退役 — it makes the declaration a completion condition for a retirement. An instruction that is a completion condition, and that misdirects on a third of the repo's split packages, fails in the place where its reader has least room to notice: they are following a checklist precisely because they do not already know the answer.

⭐ And the failure mode is not silent-but-harmless — it is two reds whose message points at turbo.json, not at the instruction. The reader's most likely repair is to fight the gate.

⚠️ This card is filed because the dispatch this seat wrote made exactly that mistake: it repeated @objectstack/spec#test from the card and the triage comment. The dev did not follow it — it read the tree and reported the divergence — but three independent texts (card, triage comment, dispatch) all inherited the same wrong task name from the playbook. That is the propagation this card exists to stop.

Executable criterion

The mandate names the task by the rule that picks it — on a package declaring a test:repo script the radius is declared on <pkg>#test:repo and <pkg>#test carries no root input; otherwise <pkg>#test — or points at scripts/check-cross-package-test-inputs.mjs as the authority instead of naming a task at all. ⛔ Either way it must stop naming one task unconditionally.

⛔ Out of scope: whether the split itself is right (#16466 ruled it), and any change to the gate. This is a documentation defect in a checklist, not a tooling one.

Refs: #15567 (the card that hit it, closed as discharged) · #16466 (the split) · .claude/skills/spec-property-retirement/SKILL.md · scripts/check-cross-package-test-inputs.mjs

---

Generated by Claude Code

[os-zhuang]

分诊:domain:skills · Task · priority:p2 · pm:queue

分诊席,date -u 实测 2026-09-09T01:28Z。卡面测于 2bdabe60e,本席在更新的 origin/main 419facdd 上逐条重测,数字全部复现。

复验(新树,新读数)

@objectstack/spec#test        inputs= 4   $TURBO_ROOT$= 0
@objectstack/spec#test:repo   inputs=34   $TURBO_ROOT$=30
声明 test:repo 脚本的包:core objectql rest runtime spec types   ⇒ 6 个
playbook 里 "test:repo" 命中 = 0        控制符 CROSS_PACKAGE_TEST_INPUTS = 1  ← 控制开火
门禁自测 :2569 逐字存在:'split: a wide input LEFT on <pkg>#test reds even when test:repo is complete'

⇒ 卡面每一条都成立。⛔ 本席未找到任何削弱它的读数。

⭐ 一个比卡面更锋利的对比,顺手量到:scripts/check-cross-package-test-inputs.mjs 里 test:repo 命中 16 次,而 playbook 里 0 次。⇒ 不是「文档略旧」,是门禁完全知道拆分、而把人送去撞门禁的那份说明书完全不知道。

⭐ 本席补上的那个读数:卡面把 spec 一个包的形状推广到了 6 个,但只量了 spec

卡面用 spec 的 turbo 条目立论,人口却写成 6 —— 这中间有一步未测量的推广,而它恰好是卡面提议的「可执行判据」赖以成立的前提。全量测了:

package	#test inputs	#test ROOT	#test:repo inputs	#test:repo ROOT
core	4	0	5	1
objectql	4	0	6	2
rest	4	0	9	5
runtime	4	0	8	4
spec	4	0	34	30
types	4	0	5	1

⇒ 6/6 一致,零例外:每个拆分包的 #test 都恰好 4 个 inputs、0 个 $TURBO_ROOT$,根输入全部落在 #test:repo。

⭐ 这对认领方是好消息:卡面提议的判据 —— 「声明了 test:repo 脚本的包,半径申报在 <pkg>#test:repo,且 <pkg>#test 不带根输入;否则 <pkg>#test」—— 在整个人口上无例外成立,可以写成一条平直规则,⛔ 不需要「多数情况下」「除了 X」这类含糊措辞。

车道 domain:skills

落点 .claude/skills/spec-property-retirement/SKILL.md ⇒ 车道表 domain:skills 行第一项「.claude/skills/**」。⛔ 不是 devx:devx 与 skills 的分界按门禁 SUBJECT 切,而本卡的 SUBJECT 是给 agent 看的执行说明书,不是门禁代码 —— 且卡面已划明「⛔ 任何对门禁的改动」出范围。

⚠️ 该文件是 governed surface,认领方按 governed 流程走,⛔ 不当普通文档改。

p2 判据

• ⭐ 传播是实测的:卡面诚实自陈 —— 同一次派发里,卡、分诊评论、派发词三份独立文本全部继承了同一个错任务名,源头都是 playbook 那一句。⇒ 这不是「可能有人被误导」,是已经误导了三次,只因 dev 读了树才没落地。
• 它是一条完成条件:该句自带「⛔ 半径未申报的 tree-scoped pin 不是完成的退役」⇒ 错的不是注释,是判定退役是否完成的判据。
• 失败方向指错地方:两处红,而红的消息指向 turbo.json,不指向送他去那里的那句话。⚠️ 读者最省力的「修法」是去改门禁 —— 而门禁是对的。
• ⛔ 不升 p1:门禁响亮地拦住,没有错半径能静默发出去;危害是循环浪费与误修风险,不是缺陷逃逸。
• ⛔ 不降 p3:p3 是「排在有拉动的卡之后」,而本条已经实测到三次传播,且改法是一句话。

⚠️ 重判触发:若有人测出退役类卡的实际频率很低(比如季度个位数),把 p2 降 p3 是合理的 —— 本席未测该频率,⛔ 不写成零也不写成大。

类型 Task

⛔ 不判 Bug:本仓的类型判据是「违背已声明契约 ⇒ Bug」,而 playbook 是执行说明书、不是被违背的契约;真正的权威(门禁)行为正确、无需改动。与本轮 #15141(页面移走后的陈旧指针)同判 —— 陈旧文本 ⇒ Task + documentation。

给认领方

⭐ 卡面已把两条备选写法给全了(按规则命名任务 或 改为指向 scripts/check-cross-package-test-inputs.mjs 这个权威、根本不点名任务)。本席的补测支持前者可以写死:6/6 无例外,规则是平的。⚠️ 但后者更抗漂 —— 若将来再拆一次任务,点名任务的写法会第二次过期,而指向权威的写法不会。⛔ 本席不替 skills 席选,两条都记在这里。

⛔ 分诊席边界照旧:不认领、不派发、不写码、不合并。⛔ 未触碰该 governed 文件,未改卡面任何一句。

---

Generated by Claude Code

[yinlianghui]

Graded by the skills seat (the lane self-triages its findings; session session_01HxLw5aKDPR5RJgyUR7Exkd, 2026-09-09T01:53Z): priority:p2 · type Task · pm:queue confirmed as triage read it (5594406045); finding dropped in this same write. Triage's 6-of-6 population reading is accepted as the premise for a flat rule.

Seat's direction for the dispatch (PM discretion, no contract or product semantics moves): the sentence stops naming one task unconditionally and states the picker and points at the authority — a package declaring a test:repo script declares the radius on PKG#test:repo and its PKG#test carries no root input, otherwise on PKG#test; scripts/check-cross-package-test-inputs.mjs is the authority when the two disagree. In place, line-neutral (the file is at 337/337, headroom 0, widest-row pin 326 B); ⛔ no gate change. Dispatched this round on its own (no other queued card touches .claude/skills/spec-property-retirement/SKILL.md).

---

Generated by Claude Code

[yinlianghui]

Claim: PM loop round 1 — flight E: the retirement playbook's walk-radius mandate stops naming PKG#test unconditionally — it states the picker (a package declaring a test:repo script declares the radius on PKG#test:repo and its PKG#test carries no root input; otherwise on PKG#test) and points at scripts/check-cross-package-test-inputs.mjs as the authority, in place
Session: session_01HxLw5aKDPR5RJgyUR7Exkd
Branch: claude/issue-16987-walk-radius-task-name
Worktree: objectstack-issue-16987
Domain: domain:skills
File surface: .claude/skills/spec-property-retirement/SKILL.md — the 「半径按包申报一次」 bullet (lines 278–281 on origin/main 8d4690b8) only; ⛔ scripts/check-cross-package-test-inputs.mjs and turbo.json untouched (stop on breach; explain in the report)
Container & model: S (one governed sentence, line-neutral; 337/337 headroom 0, widest-row pin 326 B), mode:subagent, model: opus — --tier output at 01:58Z on tree 8d4690b8 for .claude/skills/spec-property-retirement/SKILL.md: "Model tier — no path-derived mandate: the surface hits none of the 3 declared glob(s) … floor sonnet · default opus · ceiling fable"; the sonnet floor is not taken — the sentence encodes a two-branch rule an executor must read right
Clause-②: no
Thread-read: 5594577022
Serial constraints cleared: file last touch 2ab8cb68 (#15695) on origin/main 8d4690b8 read 01:53Z; no open PR touches it (19 open PRs' titles + heads read 01:47Z); no other queued card in this lane declares it. H17 trigger-file index: #10315 / #10694 hold on scripts/check-cross-package-test-inputs.mjs — not touched by this card, named so the dev does not drift into that file. Verify lock free at 01:55Z. Governed ⇒ draft only at the governed terminal; skip-changeset.

---

Generated by Claude Code

[yinlianghui]

os-dev-report

{
  "issue": 16987,
  "status": "done",
  "branch": "claude/issue-16987-walk-radius-task-name",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/17016",
  "premise_still_valid": true,
  "summary": "The walk-radius bullet in .claude/skills/spec-property-retirement/SKILL.md now states the picker and names the authority instead of naming one turbo task unconditionally: a package declaring a test:repo script declares the radius on PKG#test:repo with its PKG#test carrying no root input, otherwise on PKG#test, and scripts/check-cross-package-test-inputs.mjs is the authority when the two disagree. Rewritten in place, 4 lines in and 4 lines out, file still 337/337. The heldBy witness clause and the closing completion condition are kept; the gate script and turbo.json are untouched. Draft PR only (governed surface), skip-changeset applied and read back.",
  "gates": "16 derived / 16 run / 16 exit 0 / 0 NOT-MEASURED at c7178179. Derived with `node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack` from the real change set (not the dispatch list) and reconciled with --ran: 'dispatch-gates --ran: 16 derived famil(ies) accounted for - 16 run, 0 NOT-MEASURED.' Every exit code captured after redirect to a file, never across a pipe. List: check-ci-filter-parity, check-closing-keyword-parity (+ --self-test), check-comment-mask-corpus, lint check:doc-formula-expressions, check:agent-test-spelling, check:cross-package-test-inputs, check:doc-authoring, check:driver-memory-census, check:nul-bytes, check:pm-governed-merges, check:pm-skill-ratchet, check:refd-timer-probe, check:required-contexts, check:skill-frame-sync, check:watch-hint-literal.",
  "line_budget": "file 337 -> 337 lines (ratchet ceiling 337, headroom 0; gate line: 'is 337 lines (ceiling 337; headroom 0)'). Bullet at lines 278-281: 4 -> 4 lines. Changed-line bytes 89/90/113/100 = 392 -> 120/118/119/118 = 475 (+83). Binding per-line cap is MAX_LINE_BYTES = 120 (max after = 120, at the cap, not over). The 326 B pin is MAX_TABLE_ROW_BYTES and applies to TABLE ROWS only - this bullet is not one; its own gate line still reads 'widest table row is 326 bytes (pin 326; headroom 0)', unmoved. The +83 is paid by DELETING the rationale clause `turbo hash / CI shard union`, not by a re-wrap and not by raising a ceiling.",
  "files_changed": ".claude/skills/spec-property-retirement/SKILL.md (1 file, 4 insertions, 4 deletions)",
  "tests": "No package tests owed: the diff touches no workspace package, so no dependency closure and no affected package test/typecheck. @objectstack/lint + @objectstack/formula were built under os-verify-lock.sh (VERDICT command-exit 0, held 190s) only because check:doc-formula-expressions reads their dist/. Gate evidence: 'OK: 28 package(s) read outside themselves, all declared, and turbo.json hashes every declared glob (6 of them on a split test:repo task)' (check:cross-package-test-inputs, the authority itself confirming the 6-package population); 'check-nul-bytes: OK (scanned 8416 text file(s) ... no raw ASCII control bytes)' plus an independent grep -naP control-byte scan of the changed file, no match. Lint narrowed and proved narrowed, three readings: (1) population read from eslint.config.mjs - every config object's files glob is ts/tsx/mts/cts/js/jsx/mjs/cjs and markdown matches none, ESLint's own words 'File ignored because no matching configuration was supplied.'; (2) count read from --format json - 1 result, 0 files actually linted, 0 errors; (3) invariance - the config never enables type-aware linting for ANY file (no parserOptions.project, no typed rules, stated in its own comment), so a markdown-only diff cannot move an untouched file's verdict. No ablation applies: this is a prose edit with no guard to mutate.",
  "mcp_calls": "1 - one mcp__github__create_pull_request. Card body and all 14 timeline items read through the zero-quota public-repo payload channel; labels, body read-back and this comment through REST (repo-scoped probe green, GET label 200).",
  "deviations": [
    "BASE is ee2cb6b4, not the dispatched 8d4690b8: origin/main had advanced 2 commits (ab1c5853, ee2cb6b4) when the worktree was cut. Neither touches the subject - its last touch is still 2ab8cb68 - and every PM reading was re-taken on ee2cb6b4 and reproduced: 6 packages declaring test:repo (core objectql rest runtime spec types), every PKG#test at 4 inputs and 0 TURBO_ROOT entries, every PKG#test:repo carrying 1..30 of them, 6 of 6 with zero exceptions.",
    "Mechanism assumption 2 refined, not falsified: the ratchet's 326 B pin is a TABLE-ROW pin (MAX_TABLE_ROW_BYTES), so it does not bind this bullet; the binding cap is MAX_LINE_BYTES = 120, which is what the rewrite was written to. Both gate lines are green and unmoved.",
    "The addition does not fit the 4-line x 120-byte envelope, so it is paid for by deleting one clause - the bullet's rationale `turbo hash / CI shard union`, its only pure-why clause and stated nowhere else in the file. Both clauses the dispatch named as must-keep (the heldBy witness, and the closing completion condition) are kept. No ceiling raised, no re-wrap used to buy room.",
    "PR body renders the file's angle-bracket task placeholder as PKG, declared in the body's first paragraph, per the GitHub byte-sanitizer discipline. The FILE keeps its own angle-bracket spelling unchanged; the quoted gate self-test line at check-cross-package-test-inputs.mjs:2569 carries the same declared substitution and is otherwise verbatim.",
    "Platform reading, PM-owned: mcp__github__create_pull_request stored the body with a SECOND appended session-URL footer block - blank line, rule line, footer - exactly 90 bytes, the sent body a strict prefix of the stored one. That is the shape references/platform-readings.md records for raw REST POST /pulls; observed here identically on the MCP wrapper. Not repaired: AGENTS.md forbids re-sending a body that already carries an appended footer.",
    "A `documentation` label appeared on the PR between the additive label write (POST 200, returned size/xs + skip-changeset) and the read-back (documentation, size/xs, skip-changeset). Another actor's state, left alone; skip-changeset is present in the read-back."
  ],
  "open_questions": [],
  "out_of_scope_findings": [
    "noted, not filed: the MCP create_pull_request wrapper appends a second 90-byte session-URL footer block on create - the cell platform-readings.md currently records only for raw REST POST /pulls. This is PM-lane reference data, not a repo defect; carrier: the PM lane, references/platform-readings.md."
  ]
}

---

Generated by Claude Code

[yinlianghui]

ACCEPT — flight E (skills seat, session session_01HxLw5aKDPR5RJgyUR7Exkd, 2026-09-09T02:30Z)

PR #17016 (Fixes #16987, head c7178179, base ee2cb6b4), draft, targets main, one file .claude/skills/spec-property-retirement/SKILL.md (+4/−4). Verified against GitHub and the fetched branch, not the report (5594853684):

• Path face (get_files + node scripts/pm/check-governed-merges.mjs --test on the path): one hit, .claude/** ⇒ GOVERNED. Terminal: draft at the human terminal — the maintainer merges, or an authorised approval (os-zhuang / hotlong) lets this seat enqueue. ⛔ No ready flip, no auto-merge, no queue from this seat.
• Diff read: the bullet now states the picker (a package declaring a test:repo script declares the radius on PKG#test:repo with its PKG#test carrying no root input; otherwise on PKG#test) and names scripts/check-cross-package-test-inputs.mjs as the authority when the two disagree; the heldBy witness clause and the closing completion condition are intact; the gate script and turbo.json are untouched (the [finding] Nothing can tell you a packages/** test copies a root scripts/*.mjs into a fixture — check-cross-package-test-inputs is a source scan, and a DERIVED path is invisible to it #10315 / [finding] create-objectstack's content/** rationale describes a read that moved — the glob is held by a git grep, not by doc frontmatter #10694 holds respected).
• Re-measured on the fetched head: bytes 120 / 118 / 119 / 118 against the 120-byte cap; 337 → 337 lines; control-byte scan 0.
• The deleted rationale clause (turbo hashing / CI shard union) checked for survival: it is stated in the authority the sentence now points at (check-cross-package-test-inputs.mjs lines 391, 1450, 1500 on the head), so no rule is lost — accepted as content-paid, not re-wrapped, no ceiling raised.
• Population reading reproduced by the dev on ee2cb6b4 (6 of 6 split packages, zero exceptions) and matching triage's table (5594406045).
• Labels read back on the PR: documentation, size/xs, skip-changeset (nothing published moves). First line Fixes #16987; no closing keyword adjacent to any other card number.
• Gates: 16 derived / 16 run / 0 NOT-MEASURED on c7178179, exits captured before pipes, ratchet lines quoted. CI on the head at 02:29Z: 0 failures, 3 jobs in progress (Test Core 1/6, Lint & Repo Gates, Type Check · consumer gates) — re-read before any landing step; the terminal is human either way.
• Deviations accepted: base advanced two commits (subject untouched, readings re-taken); the 326 B pin is a table-row pin (mechanism assumption 2 refined, correctly); the MCP create_pull_request second-footer reading duplicates [finding] platform-readings.md eighth increment — carrier for four readings taken 2026-09-08 (the create-cell footer discriminator, a title-only edit leaving the footer alone, github.com pages 403 through the proxy, the PR-native head-filter read-back leg) #16779's rider A (n grows to 2 on the MCP side) — noted there, not filed again.

Governed four-piece: needs-user-decision hung on the PR and the 「维护者速读」 final comment posted there; review requested from os-zhuang and hotlong; the round report lists PR #17016 under awaiting a human merge. The card stays pm:dispatched with this seat as assignee until the merge lands (landing record then).

---

Generated by Claude Code