Repository navigation
skills(objectstack-ui): state the one-app-per-package cap in App Navigation (GOVERNED — token ceiling raised to the landed count under a maintainer ruling) - #17310
Conversation
…gation
The App Navigation section tells an author to `register under
defineStack({ apps: [...] })` and stops there. Nothing in the published
bundle says the array holds one entry: measured across `skills/`, the
one-app rule appears 0 times in any spelling, while `defineStack` appears
54 times and `App.create` 3 — the zeros are readings, not an empty corpus.
An agent authoring `*.app.ts` from this skill therefore has no reason not
to write one app per audience, and finds out at `defineStack` load time,
after the design is written.
All three `ADR-0019` citations in the published bundle point at
`0019-approval-as-flow-node`, so this one names its record explicitly.
⚠️ This overruns the published-skills token ratchet by 59 tokens
(3874 against a ceiling of 3815, headroom measured at 0 before writing).
The ceiling is deliberately NOT touched and nothing else is trimmed to
pay for it — both are maintainer decisions. See the PR body.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012GKcPZbMoGq7WPzKLfRBTU
…e-app-rule-ui-skill
…landed count The one-app-per-package cap added to `skills/objectstack-ui/SKILL.md` sits 41 tokens over the shrink-only ratchet. The maintainer ruled on 2026-09-13 that this PR may grow the file; the ruling is quoted verbatim beside the row and in the PR body, which is the evidence bar the gate's header sets. Re-measured on the merged head rather than copied from the pre-merge CI run: `origin/main` deleted the `page` view-type row from the same file after that run, so the landed count is 3856 (15424 bytes), not the 3874 CI reported. The row is pinned AT that measurement, so it carries zero headroom. No other ceiling moves, and no text was re-wrapped or reflowed to buy tokens. Claude-Session: https://claude.ai/code/session_01NFSv55L8jzmE9yvi9UwZug Co-authored-by: Claude <noreply@anthropic.com>
维护者速读改了什么 —— 两个文件,共 +22 / −1。 ① ② 为什么改 —— ① 是因为这条规则在已发布技能包里一次都没出现过。实测( ② 是因为您已就这 41 个 token 给了裁决,而按门禁自己 header 的规定,「正文引用维护者裁决」是抬上限唯一合法的落地方式。 风险与代价(含回滚) —— 代价是 41 个 token,不是最初报的 59。 席位意见 —— 建议合并。三条理由,都是本席自己对着 diff 和树核验过的,不取报告自述:
你要做的 —— 人工合并(受管面 CI 读数,取于 ⛔ 本席未入队、未挂 auto-merge、未提交任何批准 review。 Generated by Claude Code |
…rding it (objectstack-ai#17766) Fixes objectstack-ai#17149 A `Claim:` comment that parses to ZERO branches is now reported as a MALFORMED claim instead of being discarded. The deliverable is triage 5620657752's, verbatim: 「⛔ do not fix the third spelling. **A parser that silently yields zero must say so** — an unparsed claim is an *unclassified* result, ⛔ never a *no*.」 Option 1 on the card (widening `claimedBranches` for the inline spelling) is **not** done here, and the branch reader's accept set is byte-identical — pinned as a case, because that is what keeps the NEXT unrecognised spelling loud on its first occurrence. ## The mechanism `governingClaim` picked the newest claim comment from which at least one branch parses, and `if (branches.length === 0) return;` threw the rest away. A discard is indistinguishable from an absence, so governance fell back to an older claim and said nothing. `claimGovernance(commentRows)` is now the two-part reading: - `governing` — what `governingClaim` has always returned. `governingClaim` is a one-line wrapper over it, so the ~18 readers that legitimately need only branches are untouched (proved below: the self-test's 3616 pre-existing cases pass unchanged on the refactor alone). - `malformed` — the NEWEST claim-shaped comment when it parses to zero branches, carrying its comment id, its timestamp, and the `created_at` of whatever governance fell back to. Recency is written once and both halves call it, so the two readings can never disagree about which comment is current. ## Every reader says so | reader | before | after | |---|---|---| | H20 / H27 dispatch liveness | probed the superseded branch, or (nothing parses) went silent as "no claim" | **H60**, a new `state`-band row, fires for every open `pm:dispatched` card, naming the comment id, what governance did instead, and the remedy | | `check-clause2-carriers` declaration limb | read the superseded claim's `Clause-②` line as if current | `cardDeclaration` returns `claim-branch-unparsed` **before** any line is read, from any comment (including the objectstack-ai#17366 correction path) | | `--pair N` | exit 4 with a verdict taken from the wrong comment | **exit 2 (UNJUDGED)**, with the whole reading printed | | the fold/lane rosters (H37/H38) | built on `governing` | unchanged — they need branches only, and H60 is the row that says the roster may rest on a superseded claim | **The exit code is the file's own, not a new one.** Its table already reads 2 as UNJUDGED — "an unread carrier is NOT a bare carrier and an unread thread is NOT an absent declaration (objectstack-ai#4690)" — and an unresolvable carrier is exactly that. Rendering it as 4 would make an unclassified result an adverse verdict, which is the reading the card refuses by name. Rendering it as 0-with-a-message is what the whole file exists against. ## The three questions the card asked ### 1. Where does the guard belong — carriers, half-states, or both? **Both, from one source.** The state is produced once in `check-half-states.mjs` (which owns `CLAIM_COMMENT_MARKER` and `claimedBranches`) and consumed in both files. A second detector in the carriers file is the drift that file's own docblocks refuse by name ("imported rather than restated ... so the two readers cannot drift"). ### 2. Which end? Option 2 only. See the four-axis reading below. ### 3. Is the fleet's claim template the cause? — MEASURED on the live board Read 2026-09-12T03:4xZ, REST, repo-scoped, over every open `pm:dispatched` card: - **28** open `pm:dispatched` cards. - **5** of them (18%) have a newest claim-shaped comment that parses to ZERO branches: **objectstack-ai#16310, objectstack-ai#16268, objectstack-ai#16251, objectstack-ai#16175, objectstack-ai#15234**. - 22 parse; 1 (objectstack-ai#13597) carries no claim-shaped comment at all. Four of the five claims were posted by ONE session (`session_012GKcPZbMoGq7WPzKLfRBTU`, the `domain:devx` execution PM seat) within **three seconds** of each other, 02:53:05Z–02:53:08Z, all in the same shape: ``` Claim: session_012GKcPZbMoGq7WPzKLfRBTU · claude/issue-16310-orphan-locale-key-gateable Clause-②: yes ``` That is not a run of typos; it is a template emitting a carrier no reader accepts. **objectstack-ai#16175 is the silent-fallback shape, live today**: its newest claim (5642984850, 2026-09-12) parses to zero, so governance falls back to its 2026-09-06 claim (5557414924) — which names a **different** branch, `claude/issue-16175-regen-sibling-stale-rules` against the current `claude/issue-16175-staleness-mtime-false-refusal`. Every downstream reader is probing the wrong ref, and the two `Clause-②` values happen to AGREE, which is the objectstack-ai#16589 near-miss recorded on objectstack-ai#16322 reproducing itself. ⛔ No `.claude/**` edit is made from this card, per the dispatch. The template finding is handed to the skills-lane seat; after this PR lands, those five cards are visible rather than silent, which is the point. ## Four-axis reading of option 1 vs option 2 - **实际业务需求** — measured, not speculative: 5 live carriers today, plus the objectstack-ai#16322 cost (two rounds, a director re-review, a re-issued claim) and its near-miss sibling. Option 1 addresses the one spelling in front of us; option 2 addresses the population that produces them. - **项目长远合理性** — option 1 is the treadmill the card names and objectstack-ai#16170 already bought once; each widening buys one spelling and leaves the next silent. Contract-first says the defect is at the reader's CONTRACT (a two-valued answer where three states exist), not in its accept set. - **防 AI 写代码犯错** — decisive here. Option 1 is consumer-side tolerance — the exact shape the axis forbids, and its failure mode is the silent one. Option 2 is a loud refusal at read time: a claim written in a shape the protocol does not accept is refused with the remedy named, and the seat cannot declare a claim the tooling does not honour. - **创业阶段不扩散需求** — option 2 adds no capability surface: one new reading of data already in hand, no new request, no new exit code, no new label written. Option 1 would grow the accept set permanently for a spelling the standing rule already forbids. ⇒ Option 2, on all four. Option 1 is deliberately NOT also done: doing both would let the accept set absorb the measured spelling and leave the new state unexercised on the live board, which is the one way to ship this fix and still not know whether it works. ## Verification All exit codes captured before any pipe (`cmd > log 2>&1; EXIT=$?`). | command | verdict line | exit | |---|---|---| | `pnpm check:pm-half-states` | `✓ check-half-states self-test: 3656 cases pass.` (was 3616) | 0 | | `pnpm check:pm-clause2-carriers` | `✓ check-clause2-carriers self-test: 493 cases pass (...)` (was 465) | 0 | | `pnpm check:pm-dispatch-gates` | `✓ dispatch-gates self-test: 1678 cases pass.` | 0 | | the derived union, 41 commands | `✓ dispatch-gates --ran: 41 derived famil(ies) accounted for — 41 run, 0 NOT-MEASURED (a DERIVED zero — all 41 recorded an exit code and none of them is 3).` | all 0 | The union was re-derived and re-run **after** the final commit, on head `b2b55c2b4` (`git rev-parse --short HEAD`), with `--repo objectstack-ai/objectstack` asserted. The first derivation printed a STALE TREE clause naming `check-skill-line-ratchet.mjs` and `check-widening-tells.mjs` (PR objectstack-ai#17760 had landed); `origin/main` was merged in, the list re-derived byte-identically at 41, and every command re-run on the merged head. Reconciliation was fed `command :: exit N` lines so the zero is derived, not claimed.⚠️ NOT MEASURED, and named rather than implied: the 47 artifact-roster families, the 11 declared-wide-population families, the 4 families taking a value from the workflow, and the 1 path-scheduled CI job are each outside the derived total — CI's, not this run's. ### Ablation — the new state can actually fail Committed first, then mutated on disk, then restored; the mutation and the restore are both proved by `git hash-object` against the HEAD blob rather than by an exit code. - Mutation: the pre-fix silent discard restored (`malformed` never populated). Anchor occurrences 1 → 0, file hash `f9f869cd` → `013b485b`. - Result: **20 of 3656** half-states cases and **14** carriers cases turn RED, every one of them in the new batteries. Direction: turns red, as pre-registered. - Restore: `git checkout HEAD -- scripts/pm/check-half-states.mjs`; `git diff HEAD` empty, hash back to `f9f869cd`. The script carries `trap ... EXIT INT TERM`. ## Live verification — the fix, run against the real board Both sweeps were run on the merged head, 2026-09-12T04:0xZ. **`check-half-states.mjs` (sweep, exit 0)** emits exactly **5** H60 rows — the same five cards the independent REST census found, arrived at through the fix's own code path: objectstack-ai#15234, objectstack-ai#16175, objectstack-ai#16251, objectstack-ai#16268, objectstack-ai#16310. Both sentence variants fire live: objectstack-ai#15234 gets the "NOTHING governs this card" reading, objectstack-ai#16175 gets "governance SILENTLY FELL BACK to an OLDER claim (2026-09-06...)". **`check-clause2-carriers.mjs` (sweep, exit 2)** reports **5 of 18 card/PR pairs UNJUDGED** — and two of them (objectstack-ai#15627 via PR objectstack-ai#17776, objectstack-ai#16565 via PR objectstack-ai#17310) are cards the `pm:dispatched` census could not see, because they are not in that population. Before this change the sweep exited 0 on all of them. **The before/after, measured rather than reasoned** — the base tree at `813f8e9f` materialised with `git archive` and its `cardDeclaration` run against the same live comment rows: | card | BEFORE | AFTER | |---|---|---| | objectstack-ai#16565 | `declared` `no` (`"Clause-②: no"`) | `claim-branch-unparsed` | | objectstack-ai#16175 | `declared` `no` (`"Clause-②: no"`) | `claim-branch-unparsed` | | objectstack-ai#15627 | `missing` | `claim-branch-unparsed` | Two of the three read as a confident `Clause-②: no` taken off a comment whose governance was never established — the objectstack-ai#16589 near-miss shape, twice, on today's board. The third read as `missing`, which sends the seat looking for a declaration line that is already written. Neither was a reading anybody could have found without opening the card by hand. ## Fixtures The measured bodies are quoted, never paraphrased, so a future widening cannot make these batteries pass by accident: - objectstack-ai#16322's two claims (5593513389 `Clause-②: no`, 5594909614 `Clause-②: yes`) — the inline spelling, and the two values DISAGREE, which is why reading the wrong one was a wrong answer and not merely an unlucky one.⚠️ Re-measured: the OLDER claim is branchless too, so on that card nothing parsed at all and `cardDeclaration`'s `pool = claimRows` fallback read the FIRST claim in thread order — a second silent fallback the card did not name, and the same fix reaches it. - The live 2026-09-12 specimen on objectstack-ai#16175, with its different-branch fallback. - objectstack-ai#16170's bulleted directive still parsing (its pin stays green, untouched at the H20 branch battery). - Controls: a well-formed newest claim still governs; an older branchless claim beside a well-formed newest one raises nothing; no claim comment at all still reads `absent`; an unreadable thread still reads `unreadable`. - The negative pins that keep this a STATE: the inline spelling still parses to ZERO, and `CLAIM_COMMENT_MARKER` still matches it. ## Acceptance notes - `skip-changeset` applies: `scripts/pm/**` ships in no package (fast-track path, no measurement owed). - `Clause-②: no`, as the claim declares — nothing published moves. - Scope held: no H22 docblock prose (objectstack-ai#17626 owns it), no widening tells (objectstack-ai#17618), no `.md`, no `.claude/**`. - H60 takes the `state` band, beside H34 — H34 reads a claim whose SEPARATOR the marker refuses, this one reads a claim whose BRANCH line the directive reader refuses; both are a live card contradicting itself, repaired on the board. ⛔ Not `stall`: the row does not claim the card is stopped, and the dev may be working perfectly well — what is broken is the READING. - ⛔ No age gate on H60, unlike H20's 60 minutes: a branchless claim is wrong at the instant it is posted and no later sweep frees it, because the protocol forbids a second `Claim:`. Pinned in both directions. - noted, not filed: `cardDeclaration`'s `pool = claimRows` path (used when nothing on the thread parses) picks the FIRST claim comment in thread order rather than the newest. This PR makes that path unreachable for the branchless case, so the residue is a recency question on a path no live shape now reaches. Successor: whoever next touches `cardDeclaration` — objectstack-ai#17098 is open against that function's neighbourhood. --- 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01MCLBsUgfykL74aU716rzVK --- _Generated by [Claude Code](https://claude.ai/code)_ --------- Co-authored-by: Claude <noreply@anthropic.com>
Item 2 of #16565. This PR does not close that card — item 1/3 ship in the companion PR #17309 and item 4 remains open. No closing keyword is used here.
Authorization for the ratchet raise
check-skills-token-ratchetis shrink-only, and its header sets exactly one evidence bar for the other direction: "The other direction lands only in a PR whose body quotes a maintainer ruling authorizing it." That ruling was given by the maintainer in PM chat on 2026-09-13, authorizing this PR specifically. Reproduced verbatim and untranslated, as the Communication rule requires:This supersedes the earlier revision of this PR, whose title and body said the change was "over the token ratchet by 59" and "needs a maintainer decision", and which deliberately left the ceiling untouched. That decision has now been made. The ceiling is raised in this revision, the PR is no longer red by design, and nothing was re-wrapped, reflowed or trimmed to buy tokens — the raise is paid for with the ruling, which is what the ruling is for.
The re-measurement — the raise is pinned AT the landed count, not at the number CI reported
The
3874in the earlier revision was measured ondc077e17, beforeorigin/mainwas merged in. It is stale, and copying it forward would have granted 18 tokens of headroom this file has not earned.origin/maindeleted thepagerow from this same file's view-type table while this PR sat open:So the arithmetic, re-derived on the merged head
4f22acc38d, in bytes because bytes are what this gate'sceil(utf8 bytes / 4)convention divides:The ceiling moves 3815 to 3856, by 41 rather than 59, and is pinned at the measurement — zero headroom, so the next token added to this file is paid for by deleting one from it, exactly as before. Independently reproducible without running the script, which is the point of the convention:
Exactly one ceiling entry moves. No other row in
CEILINGSis touched, and the raise carries its ruling in a comment on the row itself, in the shape the surrounding rows already use.维护者速读(草稿)
改了什么 —— 两处。①
skills/objectstack-ui/SKILL.md的 App Navigation 一节加三行(与上一版完全相同,未改一字):一个type: 'app'包只暴露一个 App,多受众是它内部用requiredPermissions门控的分组,并把ADR-0019写成指名到app-as-consumer-unit记录。②scripts/check-skills-token-ratchet.mjs里skills/objectstack-ui/SKILL.md这一行的上限由 3815 抬到 3856,并在该行上方按本文件既有格式记下裁决原话与算术。为什么改 —— ① 的理由没变:这是 AI 作者真正会读的地方。发布包里的实测:一个 app 的规则出现 0 次(所有拼写),而
defineStack出现 54 次、App.create3 次 —— 零是读数,不是空语料。② 是因为你已经就这 41 个 token 给了裁决(见上方 Authorization 节),按门禁自己的规则,这正是抬上限唯一合法的落地方式。风险与代价(含回滚) —— 代价是 41 个 token,不是上一版说的 59:
origin/main在本 PR 挂起期间删掉了同文件里的page视图类型行,少了 70 字节。抬到的 3856 是合入后实测值,余量仍为 0 —— 下一个作者往这个文件加一个 token,仍然只能靠在同文件删一个来付账,棘轮没有被放松,只是按裁决上移了一格。回滚成本仍为零:两个文件、一个常量、无代码、无生成物。⛔ 本 PR 没有跳过、禁用或隔离任何测试或门禁。席位意见 —— (留空,待席位定稿)
你要做的 —— 人工合并(治理面,
skills/**)。本 PR 仍不因合并关闭 #16565。The measurement behind the teaching, unchanged
Taken on
origin/mainatb90aff81f2, over the whole published bundle. Every zero carries a same-corpus positive control from the same pass.So the rule is absent in every spelling from a bundle that names
defineStack54 times, and the number an author would follow resolves, in this bundle's own usage, to the wrong record.Changeset: still the label, not a file
Not a changeset — the
skip-changesetlabel, already applied. Measured rather than assumed: no package manifest'sfiles[]containsskills,skills/is not a workspace member, and the root manifest is private. The second file this revision touches,scripts/check-skills-token-ratchet.mjs, is a repo gate script that ships in nothing.pr-automation.ymlspells this exact case out, including that an empty-frontmatter changeset is rejected and that the label is what "names no package" is for.Verification, re-run on the merged head
Gate family derived from this diff by the repo's own deriver, not recalled:
The gate this PR is about, and its self-test, on the merged head:
check-ratchet-remedy-authoritywas run deliberately, because this gate's own source carries a placement warning about prose nearCEILINGSflipping its classification frommarkedtoexcluded. The new comment sits inside the map, far from the declaration, and the classification holds:pnpm lintis a whole-repo scan CI owns; the local reading is a declared narrowing, with the three facts that make it a measurement rather than a skipped step: (a) the linted population is read fromeslint.config.mjsitself —files: ['**/*.{ts,tsx,mts,cts,js,jsx,mjs,cjs}']— which does not include.md, and eslint confirms it in its own words forSKILL.md("File ignored because no matching configuration was supplied"); (b) the file count is read from--format json: 2 paths handed in, 1 in population, 0 errors, 0 warnings on it; (c) the invariance claim — this repo runs oneeslint.config.mjswhich never enables type-aware linting for any file (noparserOptions.project, no typed rules; the config says so at itsQUERY_OPTIONS_TEST_GLOBScomment, with its own positive control), so a comment block added to one.mjscannot move the verdict on any untouched file.Control-character self-scan on both touched files, outside
check:nul-bytes:grep -naPover the C0 set plus DEL, zero hits.🤖 Generated with Claude Code
https://claude.ai/code/session_01NFSv55L8jzmE9yvi9UwZug
Generated by Claude Code