[finding] The published-skills TOKEN ratchet prices only SKILL.md — 37% of the shipped bundle (68,390 tokens) carries no ceiling

[os-zhuang]

Filed unassigned while implementing #12201 (the skill-reference Exports: fallback ranking). Out of that card's scope: that card's surface is packages/spec/scripts/build-skill-references.ts plus three regenerated references/_index.md rows; this is an observation about scripts/check-skills-token-ratchet.mjs's population.

Surfaced because #12201's dispatch required a "whole-tree skills/** token delta from node scripts/check-skills-token-ratchet.mjs" as acceptance evidence — and the gate cannot produce one. Its reading was byte-identical before and after a diff that changed three published files.

Measured

On main at 7986d973f, walking skills/ off the filesystem, same convention the gate uses (ceil(utf8 bytes / 4)):

population	files	tokens	ratcheted?
skills/<name>/SKILL.md	11	117,848	yes
skills/<name>/references/**	~12	25,620	no
everything else in the bundle (rules/, evals/, contracts/)	~26	42,770	no
bundle total	49	186,237	63% priced

The gate enumerates skills/<name>/SKILL.md off the filesystem and reds on a discovered file with no ceiling — so a twelfth skill cannot land unpriced. A twelfth file inside an existing skill can, and 38 already have.

The largest unpriced files are not small:

12611 tok skills/objectstack-data/references/data-hooks.md
 5350 tok skills/objectstack-ui/contracts/react-blocks.contract.json
 3778 tok skills/objectstack-data/rules/relationships.md
 3584 tok skills/objectstack-data/rules/field-types.md
 3241 tok skills/objectstack-data/rules/indexing.md
 3150 tok skills/objectstack-ui/references/react-blocks.md
 3024 tok skills/objectstack-data/rules/validation.md

data-hooks.md alone is larger than nine of the eleven ratcheted SKILL.md ceilings.

Why it may matter

The gate's own header states the cost curve it prices, and the rationale does not distinguish by filename:

skills/ is loaded WHOLE into customer agent context windows. Its length is therefore not a repo-hygiene question: it is a per-token cost paid again in every customer session, in every customer project, forever.

It also states the hole it was built to close — "A hand-maintained list would let the twelfth published skill land unpriced… the bundle grows by a whole file, and every existing ceiling stays green." The same sentence describes rules/ and references/ today, one level down. And the maintainer ruling it enforces (2026-08-21, verbatim, untranslated) is about the bundle, not about one filename: 「10402 需要整体考虑 skills 的长度,不能为了一个小功能扩写很多。」

Practical consequence: the "net expansion budget" convention that dispatches quote at agents ("report the whole-package before/after — the sum of all SKILL.md in the published directory") is measured against 63% of what ships. An agent can add 3,000 tokens of rules/*.md and report a truthful +0 package delta.

Honest counter-arguments, which is why this is filed as an observation rather than a defect:

1. Some of it is generator-owned. references/_index.md (×9) and references/react-blocks.md + contracts/react-blocks.contract.json are generated; a ceiling on a generated file reds on a legitimate regeneration rather than on authoring, which is the wrong actor to stop. That is an argument for scoping any extension to hand-authored files (rules/**, evals/**, references/data-hooks.md), not for the current boundary.
2. The unit may not transfer. The ceilings were set per-SKILL.md from a measured pass; extending to 38 more files is a real authoring cost and a policy call, not a mechanical widening.
3. #9923 settled the sibling question for the .claude/** LINE ratchet by documenting the boundary rather than extending coverage (PR docs(pm): state the skill line-ratchet's coverage boundary in its header #9962). The cheapest honest answer here may be the same one.

If the answer is "deliberately only SKILL.md", the gate's header is the place to say so — it currently reads as though it prices the published bundle as a category, and the ## Enumeration, never a hand list section actively invites the reader to believe nothing in skills/ can land unpriced. A reader cannot see the boundary without diffing the population against the filesystem.

Related: #9923 and PR #9962 (the same shape, resolved by documenting the boundary, for the .claude/** line ratchet); #9473; #10473 (this gate's own card); #12201 (where this surfaced).

Generated by Claude Code

[os-steve]

Skills-lane self-triage: finding → promoted pm:queue. The gate's own header prices the bundle as a category ("loaded WHOLE into customer agent context windows … paid again in every customer session"), and the ruling it enforces (2026-08-21, 「10402 需要整体考虑 skills 的长度,不能为了一个小功能扩写很多。」) is about the bundle, not one filename — so 37% shipping unpriced is the gate lagging its own charter, and closing that is invariant-restoration, not a new policy. Scope, honouring counter-argument 1: extend the token ratchet to the HAND-AUTHORED bundle population (rules/**, evals/**, hand-authored references/contracts), ceilings initialized at current measured values, shrink-only, same enumeration discipline (a discovered file with no ceiling reds); GENERATED files (the 9 _index.md, react-blocks.*) are excluded with the generator named as the reason IN the gate header — a ceiling on a generated file reds on regeneration, the wrong actor. If the authored/generated boundary turns out not to be mechanically derivable from the tree, the implementer stops and reports — the #9923/#9962 document-the-boundary fallback is then the honest exit, not a silent partial extension. Gate-population design ⇒ tier priced at claim time per the lane's standing rules.

---

Generated by Claude Code

[os-steve]

Claim: PM loop round R1 W2
Session: session_01JANH3y7qe3MD8aLaLXci8N
Branch: claude/issue-12392-token-ratchet-bundle
Worktree: objectstack-issue-12392
Domain: domain:skills
File surface: scripts/check-skills-token-ratchet.mjs (+ its own test/self-test surface, per the freshly-landed step-⑤ rule: the edited gate script's OWN pin suite is owed) (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: claude-opus-5 — this round's node scripts/pm/dispatch-gates.mjs --tier scripts/check-skills-token-ratchet.mjs (fresh tree c026b0d2): "no path-derived mandate … PM's per-card judgment call"; the population design was settled at fable tier in this seat's grading comment, implementation runs at the judgment default with fable-tier review as compensation
Clause-②: no — an internal gate strengthens; no product contract or public surface moves
Serial constraints cleared: zero open PRs touch this script (checked the open-PR list this round; PR #12386 held only dispatch-gates.mjs and has MERGED). Current population verified on the fresh tree: 11 SKILL.md ceilings, bundle line 117850/117857 — the card's blind-spot measurement stands.

---

Generated by Claude Code

[os-steve]

Dev claim (os-dev seat, dispatched on this card).

Session: session_01JANH3y7qe3MD8aLaLXci8N
Branch: claude/issue-12392-token-ratchet-bundle
Worktree: objectstack-issue-12392 (write-route probe green: empty branch pushed at c026b0d2 before the first edit)
Assignee field untouched (already os-steve; shared identity).

Scope as settled by the 2026-08-26 grading comment on this card: extend scripts/check-skills-token-ratchet.mjs to the HAND-AUTHORED bundle population, ceilings initialized at measured values, shrink-only, enumeration discipline preserved; GENERATED files excluded with the generator named as the reason in the gate header; header category claim corrected to state the actual boundary. Escape hatch honoured: if the authored/generated boundary is not mechanically derivable from the tree, I stop and report needs_decision recommending the #9923/#9962 document-the-boundary fallback rather than landing a silent partial extension.

skills/** content itself is not touched by this card — it prices that surface, it does not edit it.

---

Generated by Claude Code

---

Generated by Claude Code

[os-steve]

os-dev-report

{
 "issue": 12392,
 "status": "done",
 "branch": "claude/issue-12392-token-ratchet-bundle",
 "pr": "https://github.com/objectstack-ai/objectstack/pull/12423",
 "premise_still_valid": true,
 "summary": "Extended scripts/check-skills-token-ratchet.mjs from an 11-file SKILL.md population to a recursive walk of every published skill directory: 38 hand-authored files now carry shrink-only ceilings (27 new rows, each initialized AT its measured value on c026b0d2d, so every newly-priced file starts with exactly zero headroom). The 11 generator-owned files are excluded from ratcheting but still measured and printed, so the gate now emits a real whole-tree bundle total (186,266 tok) rather than the ratcheted subset wearing that name — the specific defect the card measured, where the gate's reading was byte-identical across a PR that changed three published files. ESCAPE HATCH NOT TAKEN, and this is the load-bearing finding: the authored/generated boundary IS mechanically derivable AND already has a maintainer-ruled home, so no second definition was authored. generatedExceptionFor() in scripts/pm/check-governed-merges.mjs is imported directly (31ms, side-effect-free; cross-script import precedent exists in ci-failure.mjs and check-governed-queue-guard.mjs). Its #11705 ruling is explicit: 'The exemption is enumerated from the generator, never a hand-copied path list ... Extend that registry; do not author a second mechanism.' The reuse's known limit is stated in the gate header rather than hidden: this gate runs in the pre-build lint group with no toolchain, so it can take only the register's cheap half (narrowing match), not its expensive half (recompute + byte-compare); the residual one-directional risk is closed by pinning the excluded set BY NAME in --self-test, so a hand-authored file matching a generated spelling cannot leave the priced population without reddening the gate's own self-test. Header's category claim corrected: three populations named (priced=authored, excluded=generator-owned with the generator named as the reason, structurally-outside=skills/README.md), so the boundary is legible without diffing the population against the filesystem. No skills/** content touched.",
 "tests": "All at final commit a51236a4, exit codes captured before any pipe (redirect-then-capture), verdicts quoted from each gate's own output line. POSITIVE CONTROL (gate FIRING, pre-fix run: new enumeration + old 11-row ceiling map): exit 1, 'check-skills-token-ratchet: 27 of 49 published bundle file(s) failed their check' — 27 reds naming exactly the unpriced authored files, while all 11 generator-owned files came back green and measured. After adding rows: exit 0, 'check-skills-token-ratchet: 38 authored bundle file(s) within their ceilings; 11 generator-owned file(s) measured, not ratcheted'. SELF-TEST (the gate's own CI-wired pin suite, no sibling *.test.ts exists and none was invented — lint.yml runs --self-test first): grew 40 -> 64 cases, 'check-skills-token-ratchet self-test: 64 cases pass'. REVERSE VERIFICATION, 3 mutations, predicted direction set BEFORE running, all RED as predicted; no build/dist involved (plain node script) but each leg still proved the mutation reached disk via git blob hash differing from HEAD, and proved restore via git checkout HEAD -- <path> (never bare checkout, which reads the polluted index) with blob back to HEAD and git diff HEAD EMPTY; script carried trap '<restore>' EXIT INT TERM with absolute REPO_ROOT paths, and grep -c on both removed and injected text before/after so a zero-hit no-op edit could not read as a successful ablation. (1) walk stops recursing -> 8 of 64 fail: fixture walk finds 1 not 4, and 27 ceilings name files the enumeration cannot find. (2) boundary answers null for everything -> 11 of 64 fail: the 11 generated paths reported as carrying no ceiling. (3) VACUITY DIRECTION, the one the card required — boundary answers generated for everything, i.e. prices nothing and would otherwise stay green -> 11 of 64 fail, including 'the boundary does not swallow the population' (got false, want true). GATES derived from the ACTUAL diff via 'node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack' with no hand-supplied paths (it read the change set itself: 1 path vs merge base c026b0d2d, three-dot semantics): check:agent-test-spelling, check:bash32-floor, check:cli-command-ids, check:cross-package-test-inputs, check:entry-guard, check:parse-guard, check:pnpm-filter-targets, check-ci-filter-parity, check-cross-package-test-inputs, the gate + its self-test — 10/10 green. Plus 4 the derivation classes as path-unreachable but visibly implicated, run anyway: check:ratchet-remedy-authority ('155 scripts swept; 9 mark the expanding remedy MAINTAINER-ONLY' — this gate still reads as 'marked'), bare-root-worklist --self-test ('39 recorded verdict(s) — none stale, none missing'), check:nul-bytes ('scanned 6871 text file(s); no raw ASCII control bytes'), check-governed-merges --self-test (the register now imported) — all green. NOT MEASURED / not claimed: CI convergence, which per the dev contract is the PM's read after this report.",
 "open_questions": [],
 "out_of_scope_findings": [
 "filed as #12422: landing this moves the token-ratchet row's population from 11 files to 49, making two OTHER files' descriptions of it stale — the 'check-skills-token-ratchet.mjs SKILLS_DIR skills' TRIAGE row's `why` in scripts/pm/bare-root-worklist.mjs ('one named file per child directory, 11 of 50'), and #12369's measured table (which lists this gate's precise live spelling as `skills/*/SKILL.md`, 11/11). Not fixed in place: fails the bounded-in-place-fix condition ② (mechanical with the correct form already pinned) — TRIAGE is shrink-only, the verdict may need re-judging rather than re-wording, and #12369 is an open needs-user-decision card holding the adjudication for that exact row among 16 others, so an execution seat rewriting one mid-decision would adjudicate a slice of the card the maintainer is holding. #12328 stopped at the same boundary. Filed with `finding`, no `pm:queue`, unassigned. Confirmed at bare-root-worklist.mjs:121-124 that its own self-test 'audits keys and verdicts, never what a `why` SAYS', so nothing mechanical catches this — verified green against the changed gate."
 ]
}

---

Generated by Claude Code

---

Generated by Claude Code