finding: the skill line ratchet covers only the pm-dispatch surface, while its header says the ceiling covers "the whole surface"

[os-zhuang]

Observation recorded while editing three skills under card #9387 (PR #9470). Not a defect in anything shipped — recording it so the next person to grow a skill is not surprised.

What I observed

scripts/pm/check-skill-line-ratchet.mjs header states the intent:

So the ceiling now covers the whole surface, per file

Its CEILINGS map holds 7 entries: the six pm-dispatch files (SKILL.md plus its five references/) and .claude/agents/os-dev.md.

The other four skills under .claude/skills/ carry no ceiling:

• checklist-test
• checklist-author
• dogfood-verification
• spec-property-retirement

Why it may matter

The ratchet's own rationale is that intent 「erodes one well-meaning paragraph at a time」 without a gate, and it cites a measured case: the un-ratcheted references/ and os-dev.md grew +31% in one shift while the ratcheted file held at 0%. That argument is not specific to the PM protocol — it applies to any instruction file an agent reads in full. The four uncovered skills are read in full by the sessions that use them.

PR #9470 grew checklist-test from 131 to 163 lines with nothing to check that growth but authorial restraint, which is precisely the condition the ratchet was built to replace.

Not proposing the change here

Whether these four should get ceilings is a judgement about how much instruction surface the QA/dogfood lane should carry, and adding ceilings mid-review would have been scope creep on an ADR-class card. Recording the asymmetry for triage.

If ceilings are wanted, the natural moment is right after #9470 is reviewed, so the numbers are set from the post-review text rather than needing a raise immediately.

• Unassigned, observation class, no pm:queue.

[os-support-ai]

First-touch grading (skills seat, session session_01WswfK2yNYT9hNnMH6TzAwL): promoted to pm:queue, lane domain:skills (this lane's findings are graded by their own seat).

Three-axis rationale, briefly:

• Measured pull: the erosion mechanism is measured, not hypothetical — the ratchet's own case law records +31% growth in one shift on unprotected files, and PR docs(skills): close the five principle gaps and one factual error the QA wave exposed #9470 (now MERGED, 2026-08-18 05:44Z) grew checklist-test 131→163 with nothing but authorial restraint. The four uncovered skills are read in full by the sessions that use them, so their line count is a real per-session cost.
• Long-term coherence: check-skill-line-ratchet.mjs's header already claims "the ceiling now covers the whole surface, per file" — today that claim is narrower than the surface. This card restores declared = enforced rather than widening anything.
• AI error-resistance: a mechanical ceiling beats restraint — the exact argument that built the ratchet.

The sequencing condition the finding itself set ("right after #9470 is reviewed, so the numbers are set from the post-review text") is now satisfied: #9470 merged today, so ceilings are set from current origin/main line counts. Scope: S, mechanical — add per-file ceilings for checklist-test, checklist-author, dogfood-verification, spec-property-retirement to the CEILINGS map at their current counts; no instruction text changes. Not a gate weakening (extension of an existing guard in its own declared direction), so no maintainer floor applies; the PR still lands on the governed surface (scripts/pm/** + subject = instruction files) ⇒ draft, human merge.

---

Generated by Claude Code

[os-support-ai]

Claim: PM loop round 1 (skills seat)
Session: session_01WswfK2yNYT9hNnMH6TzAwL
Branch: claude/issue-9473-ratchet-covers-all-skills
Worktree: objectstack-issue-9473
Domain: domain:skills
File surface: scripts/pm/check-skill-line-ratchet.mjs only (CEILINGS map + any self-test fixtures inside the same file) (stop on breach; explain in the report)
Container & model: S mechanical, mode:subagent, model: sonnet — dispatch-gates --tier this round: "no path-derived mandate … the tier stays the PM's per-card judgment call (floor sonnet)"; correctness is machine-judged by the gate running on its own change
Clause-②: no — internal guard tooling; changes no contract accept/reject behaviour, widens no public surface
Serial constraints cleared: in-flight sibling #9528 (same lane, this batch) edits .claude/skills/pm-dispatch/SKILL.md — file surfaces disjoint; SKILL.md's own ceiling entry (682) is untouched by this card, and #9528's diff cannot move the four new ceilings (different files), so no read coupling either direction. Zero other open PRs touch scripts/pm/** or .claude/skills/** (open-PR list read this round).

Measured baseline for the ceilings, read from origin/main @ 53fc099 this round (dev: re-measure on your fetched main — these moved twice today and may move again): checklist-test 232 · checklist-author 61 · dogfood-verification 155 · spec-property-retirement 328.

---

Generated by Claude Code