docs(pm-dispatch): keep two copies of the decision frame, not four

[os-litant]

Part of #14685
Part of #13597

Item 4 of decision batch 3, ruled B. From the ruling comment (5520452691), verbatim:

B — the four-axis decision frame keeps two copies (internal-pm and published-pm); the dev-side copies are replaced by a {decision_frame} placeholder the dispatching PM fills at dispatch time; check:skill-frame-sync watches two sites. Confidence gap carried: whoever executes verifies that the published-dev block has no reader other than the placeholder path before removing it.

---

1. The confidence gap, discharged BEFORE any edit

The ruling made removal conditional on a reader census. Run at the merge base 5ff5f9576, over .claude/ skills/ scripts/ .github/ packages/ docs/, for four things: the published-dev start marker (Analyze every option on ... fixed axes:), the literal published-dev, {decision_frame}, and the internal-dev start marker. Every hit, and what it is:

site	what it is	depends on the published-dev BLOCK?
skills/objectstack-pm-dispatch/SKILL.md:633-662	the block itself	it IS the block
skills/objectstack-pm-dispatch/SKILL.md:623	"with the block below in place of its {decision_frame} line"	YES — the placeholder path, and the only one
skills/objectstack-pm-dispatch/rules/dev-template.md:99	the {decision_frame} line the instruction above fills	via that instruction only
scripts/check-skill-frame-sync.mjs:136-143	the published-dev COPIES entry	the gate that watches it
scripts/check-skill-frame-sync.mjs:714,719,723,724,727	self-test fixtures addressing the published-dev id	the gate's own fixtures
scripts/check-skill-frame-freshness.mjs	imports COPIES / AXIS_MAP wholesale; five fixtures spelled .claude/agents/os-dev.md	reads the TABLE, never the block text
scripts/pm/dispatch-gates.mjs:15382-15397	spawned import reading the FILE list out of COPIES for the tier mandate	reads FILES, not copies
scripts/pm/check-dispatch-gates.mjs:202	FRAME_MODULE watch-hint constant	names the module path only
scripts/check-skills-token-ratchet.mjs:334,345,359	ceiling-ledger prose	comments, no behaviour
scripts/pm/check-skill-line-ratchet.mjs:455	ceiling-ledger prose	comments, no behaviour

No test, generator, workflow, doc or fixture quotes the block by fingerprint. The gap resolves in favour of removal: the placeholder path and the two frame gates are the whole reader set, and both gates read the COPIES table rather than the prose.

The same census over the internal-dev start marker found exactly one non-gate hit — the block itself at .claude/agents/os-dev.md:400.

2. What changed, site by site

skills/objectstack-pm-dispatch/SKILL.md — the embedded dev template loses the frame copy: the declaring sentence through the binding sentence (old :633-662), which is precisely the span the gate's start and binding anchors delimit.

The stop trigger above it (old :627-632, "When to STOP instead of coding …") stays, and that is a judgement call worth naming. It is not part of the frame — the gate's anchors do not reach it — and it is the escalation TRIGGER rather than the analysis. rules/dev-template.md carries it nowhere else (grepped: the template mentions needs_decision only at :34 inside the contract-first rule and at :112 in the report schema), so deleting it with the frame would have removed the dev's instruction to stop at all, which the ruling did not ask for.

The paste instruction now names the surviving copy by its heading:

Paste rules/dev-template.md verbatim into every dispatch prompt; the PM fills
its {…} placeholders. {decision_frame} takes the stop trigger below, followed
by the whole `#### The four-axis decision frame (binding)` section above: the
frame has ONE copy in this skill and the PM pastes it at dispatch time, so no
dev-side copy can drift from it.

.claude/agents/os-dev.md — the frame block (old :400-421, declaring sentence through the binding sentence, which ran into the same paragraph as the unrelated blocked rule) collapses to one sentence. The blocked rule is preserved byte-for-byte. Neither :232 nor :303 is touched (PR #14808 holds them); the diff is one hunk at :400.

Three judgement calls in that one sentence:

1. It says the frame comes from the PM's own copy, not specifically from rules/dev-template.md, because that template is the PUBLISHED skill's artifact. The internal PM composes dispatch prompts per .claude/skills/pm-dispatch/references/dispatch-runbook.md and has no {decision_frame} slot (grep: the token exists only under skills/objectstack-pm-dispatch/). The published slot is named parenthetically, as the brief asked, without claiming the internal lane uses it. See finding F5.
2. It adds "if the dispatch prompt did not carry it, stop and ask; do not invent your own axes." os-dev.md's own doctrine is that standard clauses live in that file and prompt silence is normal — so moving the frame OUT of it opens a hole that a silent prompt would otherwise walk through.
3. It is written in Chinese and deliberately avoids the phrases fixed axes, 固定评估轴, 条评估轴 and any WORD-axis decision frame form. Those are FINGERPRINTS in the sync gate's anti-dormancy scan, whose population is per file: os-dev.md stops being a declared file with this change, so any of those phrasings would have turned the gate red. Verified after the edit — grep -nP for all four patterns over os-dev.md returns nothing.

scripts/check-skill-frame-sync.mjs — COPIES 4 → 2.

- internal-pm .claude/skills/pm-dispatch/SKILL.md KEPT
- internal-dev .claude/agents/os-dev.md REMOVED
- published-pm skills/objectstack-pm-dispatch/SKILL.md KEPT
- published-dev skills/objectstack-pm-dispatch/SKILL.md REMOVED

Every spelled count is now derived: the header ("FOUR hand-written copies across three files"), the COPIES docblock, the axisEntryStarts docblock, the freshness-gate cross-reference, and the count-disagreement message (the four copies do not agree becomes the ${results.length} copies do not agree). report() and the success line already derived theirs from COPIES.length.

The same-file overlap guard is kept even though it is vacuous today — no two copies share a file any more. Re-declaring such a pair is one edit away, and that guard is what would make such a pair prove it compares two different sections rather than the same one twice.

scripts/check-skill-frame-freshness.mjs — it imported COPIES but its self-test SPELLED .claude/agents/os-dev.md in five fixtures (cases 1, 2, 6, 7 and the two withUnreadableCopy calls in 8 and 9). Those broke on the reduction, exactly as that file's own #8024 rule predicts. They now derive SAMPLE_COPY_ID / SAMPLE_FRAME_FILE from COPIES[0], so the next change to the table re-points the fixtures instead of stranding them.

3. The anti-dormancy scan still refuses a returning copy — measured, both directions

Committed first, then mutated, with an EXIT INT TERM trap restoring absolute paths; each mutation proved on disk by counting the injected marker before reading any verdict; each restore proved by git diff HEAD empty AND by comparing git hash-object against the path's HEAD blob.

Direction A — the removed text returns to the now-undeclared .claude/agents/os-dev.md. REFUSED. Marker occurrences after mutation: 1 (was 0). node scripts/check-skill-frame-sync.mjs exit 1:

✗ check-skill-frame-sync: 1 problem(s).
• .claude/agents/os-dev.md
looks like ANOTHER copy of the decision frame (matched /fixed axes/) but is
not declared in COPIES, so nothing keeps it in sync.

Restore: git diff HEAD empty; worktree blob 968d4fd10a653f013635a8a48bdee25a25efc648 equals the HEAD blob.

Direction B — the same text returns to the still-declared skills/objectstack-pm-dispatch/SKILL.md. NOT REFUSED (exit 0). This is a real, measured limit and it is not new: the scan skips any file present in declaredFiles, so a second undeclared copy inside an already-declared file has always been invisible to it. Restore proved the same way (blob 3cbbed0f501736a0db3a5ecdd9e6cda43d44b434).

What that limit means for this PR, stated rather than implied: what keeps a second published copy from being written is the paste instruction naming the surviving section by heading, plus review — not the scan. Closing it mechanically needs a per-occurrence criterion the gate does not have (the published file legitimately carries two more fixed axes fingerprints, at :534 and in the :539 heading). A comment next to the new self-test case says exactly this, so the next reader does not mistake direction B for a bug introduced here.

A self-test case pins direction A against the REAL path this change undeclared, rather than a synthetic one.

4. Counts (merge base 5ff5f9576 → head)

file	bytes	lines	tokens (ceil bytes/4)
skills/objectstack-pm-dispatch/SKILL.md	38807 → 36898 (−1909)	753 → 726 (−27)	9702 → 9225 (−477), ceiling 9708
.claude/agents/os-dev.md	42743 → 41196 (−1547)	469 → 450 (−19)	— (line-ratcheted), ceiling 469
scripts/check-skill-frame-sync.mjs	40961 → 43180 (+2219)	960 → 990 (+30)	not ratcheted
scripts/check-skill-frame-freshness.mjs	46344 → 47480 (+1136)	1065 → 1085 (+20)	not ratcheted

Whole published package skills/objectstack-pm-dispatch/** (all markdown): 11540 → 11063 tokens, −477.

Two ratcheted files, both shrink; no ceiling row is edited, in either ratchet. The measured token saving is −477, not the −630 the card estimated — the difference is the stop trigger, kept for the reason in section 2. The script growth (+50 lines across two gates) is entirely the derivation constants, the new regression case and the comments recording the measured limit; neither script is under a ratchet.

5. Gates

Union derived at the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands:

dispatch-gates: gate list derived from the tree of 'objectstack-ai/objectstack' at commit ae616fbc91 (/home/user/objectstack-14685-item4).
dispatch-gates: change set derived from git — 4 path(s) vs merge base 5ff5f9576 of 'origin/main' and HEAD

Every exit code captured by redirect before any pipe.

All 32 re-run at the final head ae616fbc91, after the last commit. 31 green, 1 NOT MEASURED, 0 red. The load-bearing ones, quoted from their own verdict lines:

command	exit	its own verdict line
pnpm check:skill-frame-sync	0	✓ check-skill-frame-sync self-test: 13 cases pass, plus 5 dispatch-gates declaration cases. / ✓ check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
node scripts/check-skill-frame-freshness.mjs	0	✓ check-skill-frame-freshness: the decision frame in this tree is current with origin/main (fetched just now). (12 self-test cases pass)
node scripts/check-skills-token-ratchet.mjs	0	✓ check-skills-token-ratchet: skills/objectstack-pm-dispatch/SKILL.md is 9225 tokens (ceiling 9708; headroom 483).
pnpm check:pm-skill-ratchet	0	✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 450 lines (ceiling 469; headroom 19).
node scripts/check-self-test-wired.mjs	0	✓ check-self-test-wired: every one of the 164 script(s) CI runs that ship a --self-test has that self-test run by CI.
pnpm check:pm-dispatch-gates	0	✓ dispatch-gates self-test: 1241 cases pass.
pnpm --filter @objectstack/spec run check:skill-docs	0	✓ skills/README.md / ✓ content/docs/ai/skills-reference.mdx — nothing to regenerate
pnpm check:ratchet-remedy-authority	0	green
pnpm check:pm-skill-id-lint	0	green
pnpm check:nul-bytes	0	green (plus a direct grep -naP for control bytes over both edited markdown files: no match)
pnpm check:doc-authoring, check:role-word, check:agent-model-declared, check:agent-test-spelling, check:corpus-claim-drift, check:skill-compatibility, check:skill-identifier-liveness, check:watch-hint-literal, check:entry-guard, check:parse-guard, check:cli-command-ids, check:bash32-floor, check:pnpm-filter-targets, check:pm-governed-merges, check:cross-package-test-inputs	0	green
node scripts/check-ci-filter-parity.mjs, check-cross-package-test-inputs.mjs, check-shard-attestation.mjs, check-whole-set-label-write.mjs, scripts/pm/bare-root-worklist.mjs --self-test	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	first run answered PREREQUISITE NOT MET (exit 3, @objectstack/formula then @objectstack/lint unbuilt). Prerequisite supplied via the shared verify lock (turbo run build --filter=@objectstack/formula, then --filter=@objectstack/lint) and re-run to a real verdict: ✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 426 files / 1365 TS blocks judged clean
node scripts/check-test-completeness.mjs	3	NOT MEASURED — see below

One red was found and fixed rather than reported: on the first pass check:pm-skill-ratchet said ✗ check-skill-line-ratchet: .claude/agents/os-dev.md has 1 line(s) over the 120-byte budget: L400 (121B). The ratchet prices a per-line byte budget as well as a line ceiling. The replacement sentence was rewrapped at legal CJK break points, content byte-identical, and the whole union was re-derived and re-run at the resulting head — which is the head every number above comes from.

NOT MEASURED, with reasons:

• node scripts/check-test-completeness.mjs — exit 3, its own text: "PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named … the local reading for this gate is NOT MEASURED. ⛔ It is not a red." CI tees the log and passes it; unreachable there.
• The changeset-derived families (9 of them) do not apply: this PR is skip-changeset — .claude/**, skills/** and scripts/** publish nothing from any package.
• The CI-measured-only family reads the workflow event payload; no local run can produce a verdict for it.
• The always-runs tail (workflows with no path filter) is CI's, not measured here.

6. Out-of-scope findings — NOT fixed here, for the seat to triage

Each is a claim that this change makes false, in a file this flight does not hold.

• F1scripts/pm/dispatch-gates.mjs:8090 — why: '… the dev-agent definition is protocol semantics — every dispatched dev runs under it, and it carries an enforced copy of the decision frame'. The last clause is now false; the mandate itself still holds on the first clause, and the self-test still passes (it asserts the glob exists, not the reason).
• F2scripts/pm/dispatch-gates.mjs:8095 — why: '… the published PM skill carries two enforced copies of the decision frame (check:skill-frame-sync COPIES) …'. One copy now.
• F3scripts/pm/check-skill-line-ratchet.mjs:455 — // This file is one of the four check:skill-frame-sync COPIES, so a rule that …, in the os-dev.md ceiling ledger. That file is no longer a COPIES entry, which also weakens the stated reason for the +3 raise recorded there. Item 5 holds this file.
• F4scripts/check-skills-token-ratchet.mjs:334, 345, 359 — three ledger comments stating the published SKILL.md "carries TWO enforced frame copies" and that "the two gate-pinned copies of the decision frame stay in this file". Historically true; now stale as a present-tense claim.
• F5 the internal lane has no {decision_frame} slot. os-dev.md now expects the frame to arrive in the dispatch prompt, but .claude/skills/pm-dispatch/references/dispatch-runbook.md (派发词构造细则) does not tell the internal PM to paste it, and .claude/skills/pm-dispatch/SKILL.md is explicitly outside this flight. Until that lands, the internal dev's only protection is the stop-and-ask clause added in section 2. This is the one finding with live consequences.
• F6 the removed os-dev.md block carried a verbatim maintainer quotation — the 2026-09-01 axis-weighting ruling — so this PR deletes a quotation site that item 2's quotation-policy flight may have surveyed. The ruling text survives at .claude/skills/pm-dispatch/SKILL.md:904 and .claude/skills/pm-dispatch/references/decision-analysis.md:50, and the RULE itself survives in both published and internal PM copies of the frame.

7. Posture

Draft, governed surface. No reviewers requested, no ready flip, no enqueue, no auto-merge. Labels skip-changeset and needs:contract-review.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Part of #14685
Part of #13597

Item 4 of decision batch 3, ruled B. From the ruling comment (5520452691), verbatim:

B — the four-axis decision frame keeps two copies (internal-pm and published-pm); the dev-side copies are replaced by a {decision_frame} placeholder the dispatching PM fills at dispatch time; check:skill-frame-sync watches two sites. Confidence gap carried: whoever executes verifies that the published-dev block has no reader other than the placeholder path before removing it.

---

1. The confidence gap, discharged BEFORE any edit

The ruling made removal conditional on a reader census. Run at the merge base 5ff5f9576, over .claude/ skills/ scripts/ .github/ packages/ docs/, for four things: the published-dev start marker (Analyze every option on ... fixed axes:), the literal published-dev, {decision_frame}, and the internal-dev start marker. Every hit, and what it is:

site	what it is	depends on the published-dev BLOCK?
skills/objectstack-pm-dispatch/SKILL.md:633-662	the block itself	it IS the block
skills/objectstack-pm-dispatch/SKILL.md:623	"with the block below in place of its {decision_frame} line"	YES — the placeholder path, and the only one
skills/objectstack-pm-dispatch/rules/dev-template.md:99	the {decision_frame} line the instruction above fills	via that instruction only
scripts/check-skill-frame-sync.mjs:136-143	the published-dev COPIES entry	the gate that watches it
scripts/check-skill-frame-sync.mjs:714,719,723,724,727	self-test fixtures addressing the published-dev id	the gate's own fixtures
scripts/check-skill-frame-freshness.mjs	imports COPIES / AXIS_MAP wholesale; five fixtures spelled .claude/agents/os-dev.md	reads the TABLE, never the block text
scripts/pm/dispatch-gates.mjs:15382-15397	spawned import reading the FILE list out of COPIES for the tier mandate	reads FILES, not copies
scripts/pm/check-dispatch-gates.mjs:202	FRAME_MODULE watch-hint constant	names the module path only
scripts/check-skills-token-ratchet.mjs:334,345,359	ceiling-ledger prose	comments, no behaviour
scripts/pm/check-skill-line-ratchet.mjs:455	ceiling-ledger prose	comments, no behaviour

No test, generator, workflow, doc or fixture quotes the block by fingerprint. The gap resolves in favour of removal: the placeholder path and the two frame gates are the whole reader set, and both gates read the COPIES table rather than the prose.

The same census over the internal-dev start marker found exactly one non-gate hit — the block itself at .claude/agents/os-dev.md:400.

2. What changed, site by site

skills/objectstack-pm-dispatch/SKILL.md — the embedded dev template loses the frame copy: the declaring sentence through the binding sentence (old :633-662), which is precisely the span the gate's start and binding anchors delimit.

The stop trigger above it (old :627-632, "When to STOP instead of coding …") stays, and that is a judgement call worth naming. It is not part of the frame — the gate's anchors do not reach it — and it is the escalation TRIGGER rather than the analysis. rules/dev-template.md carries it nowhere else (grepped: the template mentions needs_decision only at :34 inside the contract-first rule and at :112 in the report schema), so deleting it with the frame would have removed the dev's instruction to stop at all, which the ruling did not ask for.

The paste instruction now names the surviving copy by its heading:

Paste rules/dev-template.md verbatim into every dispatch prompt; the PM fills
its {…} placeholders. {decision_frame} takes the stop trigger below, followed
by the whole `#### The four-axis decision frame (binding)` section above: the
frame has ONE copy in this skill and the PM pastes it at dispatch time, so no
dev-side copy can drift from it.

.claude/agents/os-dev.md — the frame block (old :400-421, declaring sentence through the binding sentence, which ran into the same paragraph as the unrelated blocked rule) collapses to one sentence. The blocked rule is preserved byte-for-byte. Neither :232 nor :303 is touched (PR #14808 holds them); the diff is one hunk at :400.

Three judgement calls in that one sentence:

1. It says the frame comes from the PM's own copy, not specifically from rules/dev-template.md, because that template is the PUBLISHED skill's artifact. The internal PM composes dispatch prompts per .claude/skills/pm-dispatch/references/dispatch-runbook.md and has no {decision_frame} slot (grep: the token exists only under skills/objectstack-pm-dispatch/). The published slot is named parenthetically, as the brief asked, without claiming the internal lane uses it. See finding F5.
2. It adds "if the dispatch prompt did not carry it, stop and ask; do not invent your own axes." os-dev.md's own doctrine is that standard clauses live in that file and prompt silence is normal — so moving the frame OUT of it opens a hole that a silent prompt would otherwise walk through.
3. It is written in Chinese and deliberately avoids the phrases fixed axes, 固定评估轴, 条评估轴 and any WORD-axis decision frame form. Those are FINGERPRINTS in the sync gate's anti-dormancy scan, whose population is per file: os-dev.md stops being a declared file with this change, so any of those phrasings would have turned the gate red. Verified after the edit — grep -nP for all four patterns over os-dev.md returns nothing.

scripts/check-skill-frame-sync.mjs — COPIES 4 → 2.

- internal-pm .claude/skills/pm-dispatch/SKILL.md KEPT
- internal-dev .claude/agents/os-dev.md REMOVED
- published-pm skills/objectstack-pm-dispatch/SKILL.md KEPT
- published-dev skills/objectstack-pm-dispatch/SKILL.md REMOVED

Every spelled count is now derived: the header ("FOUR hand-written copies across three files"), the COPIES docblock, the axisEntryStarts docblock, the freshness-gate cross-reference, and the count-disagreement message (the four copies do not agree becomes the ${results.length} copies do not agree). report() and the success line already derived theirs from COPIES.length.

The same-file overlap guard is kept even though it is vacuous today — no two copies share a file any more. Re-declaring such a pair is one edit away, and that guard is what would make such a pair prove it compares two different sections rather than the same one twice.

scripts/check-skill-frame-freshness.mjs — it imported COPIES but its self-test SPELLED .claude/agents/os-dev.md in five fixtures (cases 1, 2, 6, 7 and the two withUnreadableCopy calls in 8 and 9). Those broke on the reduction, exactly as that file's own #8024 rule predicts. They now derive SAMPLE_COPY_ID / SAMPLE_FRAME_FILE from COPIES[0], so the next change to the table re-points the fixtures instead of stranding them.

3. The anti-dormancy scan still refuses a returning copy — measured, both directions

Committed first, then mutated, with an EXIT INT TERM trap restoring absolute paths; each mutation proved on disk by counting the injected marker before reading any verdict; each restore proved by git diff HEAD empty AND by comparing git hash-object against the path's HEAD blob.

Direction A — the removed text returns to the now-undeclared .claude/agents/os-dev.md. REFUSED. Marker occurrences after mutation: 1 (was 0). node scripts/check-skill-frame-sync.mjs exit 1:

✗ check-skill-frame-sync: 1 problem(s).
• .claude/agents/os-dev.md
looks like ANOTHER copy of the decision frame (matched /fixed axes/) but is
not declared in COPIES, so nothing keeps it in sync.

Restore: git diff HEAD empty; worktree blob 968d4fd10a653f013635a8a48bdee25a25efc648 equals the HEAD blob.

Direction B — the same text returns to the still-declared skills/objectstack-pm-dispatch/SKILL.md. NOT REFUSED (exit 0). This is a real, measured limit and it is not new: the scan skips any file present in declaredFiles, so a second undeclared copy inside an already-declared file has always been invisible to it. Restore proved the same way (blob 3cbbed0f501736a0db3a5ecdd9e6cda43d44b434).

What that limit means for this PR, stated rather than implied: what keeps a second published copy from being written is the paste instruction naming the surviving section by heading, plus review — not the scan. Closing it mechanically needs a per-occurrence criterion the gate does not have (the published file legitimately carries two more fixed axes fingerprints, at :534 and in the :539 heading). A comment next to the new self-test case says exactly this, so the next reader does not mistake direction B for a bug introduced here.

A self-test case pins direction A against the REAL path this change undeclared, rather than a synthetic one.

4. Counts (merge base 5ff5f9576 → head)

file	bytes	lines	tokens (ceil bytes/4)
skills/objectstack-pm-dispatch/SKILL.md	38807 → 36898 (−1909)	753 → 726 (−27)	9702 → 9225 (−477), ceiling 9708
.claude/agents/os-dev.md	42743 → 41196 (−1547)	469 → 450 (−19)	— (line-ratcheted), ceiling 469
scripts/check-skill-frame-sync.mjs	40961 → 43180 (+2219)	960 → 990 (+30)	not ratcheted
scripts/check-skill-frame-freshness.mjs	46344 → 47480 (+1136)	1065 → 1085 (+20)	not ratcheted

Whole published package skills/objectstack-pm-dispatch/** (all markdown): 11540 → 11063 tokens, −477.

Two ratcheted files, both shrink; no ceiling row is edited, in either ratchet. The measured token saving is −477, not the −630 the card estimated — the difference is the stop trigger, kept for the reason in section 2. The script growth (+50 lines across two gates) is entirely the derivation constants, the new regression case and the comments recording the measured limit; neither script is under a ratchet.

5. Gates

Union derived at the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands:

dispatch-gates: gate list derived from the tree of 'objectstack-ai/objectstack' at commit ae616fbc91 (/home/user/objectstack-14685-item4).
dispatch-gates: change set derived from git — 4 path(s) vs merge base 5ff5f9576 of 'origin/main' and HEAD

Every exit code captured by redirect before any pipe.

All 32 re-run at the final head ae616fbc91, after the last commit. 31 green, 1 NOT MEASURED, 0 red. The load-bearing ones, quoted from their own verdict lines:

command	exit	its own verdict line
pnpm check:skill-frame-sync	0	✓ check-skill-frame-sync self-test: 13 cases pass, plus 5 dispatch-gates declaration cases. / ✓ check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
node scripts/check-skill-frame-freshness.mjs	0	✓ check-skill-frame-freshness: the decision frame in this tree is current with origin/main (fetched just now). (12 self-test cases pass)
node scripts/check-skills-token-ratchet.mjs	0	✓ check-skills-token-ratchet: skills/objectstack-pm-dispatch/SKILL.md is 9225 tokens (ceiling 9708; headroom 483).
pnpm check:pm-skill-ratchet	0	✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 450 lines (ceiling 469; headroom 19).
node scripts/check-self-test-wired.mjs	0	✓ check-self-test-wired: every one of the 164 script(s) CI runs that ship a --self-test has that self-test run by CI.
pnpm check:pm-dispatch-gates	0	✓ dispatch-gates self-test: 1241 cases pass.
pnpm --filter @objectstack/spec run check:skill-docs	0	✓ skills/README.md / ✓ content/docs/ai/skills-reference.mdx — nothing to regenerate
pnpm check:ratchet-remedy-authority	0	green
pnpm check:pm-skill-id-lint	0	green
pnpm check:nul-bytes	0	green (plus a direct grep -naP for control bytes over both edited markdown files: no match)
pnpm check:doc-authoring, check:role-word, check:agent-model-declared, check:agent-test-spelling, check:corpus-claim-drift, check:skill-compatibility, check:skill-identifier-liveness, check:watch-hint-literal, check:entry-guard, check:parse-guard, check:cli-command-ids, check:bash32-floor, check:pnpm-filter-targets, check:pm-governed-merges, check:cross-package-test-inputs	0	green
node scripts/check-ci-filter-parity.mjs, check-cross-package-test-inputs.mjs, check-shard-attestation.mjs, check-whole-set-label-write.mjs, scripts/pm/bare-root-worklist.mjs --self-test	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	first run answered PREREQUISITE NOT MET (exit 3, @objectstack/formula then @objectstack/lint unbuilt). Prerequisite supplied via the shared verify lock (turbo run build --filter=@objectstack/formula, then --filter=@objectstack/lint) and re-run to a real verdict: ✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 426 files / 1365 TS blocks judged clean
node scripts/check-test-completeness.mjs	3	NOT MEASURED — see below

One red was found and fixed rather than reported: on the first pass check:pm-skill-ratchet said ✗ check-skill-line-ratchet: .claude/agents/os-dev.md has 1 line(s) over the 120-byte budget: L400 (121B). The ratchet prices a per-line byte budget as well as a line ceiling. The replacement sentence was rewrapped at legal CJK break points, content byte-identical, and the whole union was re-derived and re-run at the resulting head — which is the head every number above comes from.

NOT MEASURED, with reasons:

• node scripts/check-test-completeness.mjs — exit 3, its own text: "PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named … the local reading for this gate is NOT MEASURED. ⛔ It is not a red." CI tees the log and passes it; unreachable there.
• The changeset-derived families (9 of them) do not apply: this PR is skip-changeset — .claude/**, skills/** and scripts/** publish nothing from any package.
• The CI-measured-only family reads the workflow event payload; no local run can produce a verdict for it.
• The always-runs tail (workflows with no path filter) is CI's, not measured here.

6. Out-of-scope findings — NOT fixed here, for the seat to triage

Each is a claim that this change makes false, in a file this flight does not hold.

• F1scripts/pm/dispatch-gates.mjs:8090 — why: '… the dev-agent definition is protocol semantics — every dispatched dev runs under it, and it carries an enforced copy of the decision frame'. The last clause is now false; the mandate itself still holds on the first clause, and the self-test still passes (it asserts the glob exists, not the reason).
• F2scripts/pm/dispatch-gates.mjs:8095 — why: '… the published PM skill carries two enforced copies of the decision frame (check:skill-frame-sync COPIES) …'. One copy now.
• F3scripts/pm/check-skill-line-ratchet.mjs:455 — // This file is one of the four check:skill-frame-sync COPIES, so a rule that …, in the os-dev.md ceiling ledger. That file is no longer a COPIES entry, which also weakens the stated reason for the +3 raise recorded there. Item 5 holds this file.
• F4scripts/check-skills-token-ratchet.mjs:334, 345, 359 — three ledger comments stating the published SKILL.md "carries TWO enforced frame copies" and that "the two gate-pinned copies of the decision frame stay in this file". Historically true; now stale as a present-tense claim.
• F5 the internal lane has no {decision_frame} slot. os-dev.md now expects the frame to arrive in the dispatch prompt, but .claude/skills/pm-dispatch/references/dispatch-runbook.md (派发词构造细则) does not tell the internal PM to paste it, and .claude/skills/pm-dispatch/SKILL.md is explicitly outside this flight. Until that lands, the internal dev's only protection is the stop-and-ask clause added in section 2. This is the one finding with live consequences.
• F6 the removed os-dev.md block carried a verbatim maintainer quotation — the 2026-09-01 axis-weighting ruling — so this PR deletes a quotation site that item 2's quotation-policy flight may have surveyed. The ruling text survives at .claude/skills/pm-dispatch/SKILL.md:904 and .claude/skills/pm-dispatch/references/decision-analysis.md:50, and the RULE itself survives in both published and internal PM copies of the frame.

7. Posture

Draft, governed surface. No reviewers requested, no ready flip, no enqueue, no auto-merge. Labels skip-changeset and needs:contract-review.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Part of #14685
Part of #13597

Item 4 of decision batch 3, ruled B. From the ruling comment (5520452691), verbatim:

B — the four-axis decision frame keeps two copies (internal-pm and published-pm); the dev-side copies are replaced by a {decision_frame} placeholder the dispatching PM fills at dispatch time; check:skill-frame-sync watches two sites. Confidence gap carried: whoever executes verifies that the published-dev block has no reader other than the placeholder path before removing it.

---

1. The confidence gap, discharged BEFORE any edit

The ruling made removal conditional on a reader census. Run at the merge base 5ff5f9576, over .claude/ skills/ scripts/ .github/ packages/ docs/, for four things: the published-dev start marker (Analyze every option on ... fixed axes:), the literal published-dev, {decision_frame}, and the internal-dev start marker. Every hit, and what it is:

site	what it is	depends on the published-dev BLOCK?
skills/objectstack-pm-dispatch/SKILL.md:633-662	the block itself	it IS the block
skills/objectstack-pm-dispatch/SKILL.md:623	"with the block below in place of its {decision_frame} line"	YES — the placeholder path, and the only one
skills/objectstack-pm-dispatch/rules/dev-template.md:99	the {decision_frame} line the instruction above fills	via that instruction only
scripts/check-skill-frame-sync.mjs:136-143	the published-dev COPIES entry	the gate that watches it
scripts/check-skill-frame-sync.mjs:714,719,723,724,727	self-test fixtures addressing the published-dev id	the gate's own fixtures
scripts/check-skill-frame-freshness.mjs	imports COPIES / AXIS_MAP wholesale; five fixtures spelled .claude/agents/os-dev.md	reads the TABLE, never the block text
scripts/pm/dispatch-gates.mjs:15382-15397	spawned import reading the FILE list out of COPIES for the tier mandate	reads FILES, not copies
scripts/pm/check-dispatch-gates.mjs:202	FRAME_MODULE watch-hint constant	names the module path only
scripts/check-skills-token-ratchet.mjs:334,345,359	ceiling-ledger prose	comments, no behaviour
scripts/pm/check-skill-line-ratchet.mjs:455	ceiling-ledger prose	comments, no behaviour

No test, generator, workflow, doc or fixture quotes the block by fingerprint. The gap resolves in favour of removal: the placeholder path and the two frame gates are the whole reader set, and both gates read the COPIES table rather than the prose.

The same census over the internal-dev start marker found exactly one non-gate hit — the block itself at .claude/agents/os-dev.md:400.

2. What changed, site by site

skills/objectstack-pm-dispatch/SKILL.md — the embedded dev template loses the frame copy: the declaring sentence through the binding sentence (old :633-662), which is precisely the span the gate's start and binding anchors delimit.

The stop trigger above it (old :627-632, "When to STOP instead of coding …") stays, and that is a judgement call worth naming. It is not part of the frame — the gate's anchors do not reach it — and it is the escalation TRIGGER rather than the analysis. rules/dev-template.md carries it nowhere else (grepped: the template mentions needs_decision only at :34 inside the contract-first rule and at :112 in the report schema), so deleting it with the frame would have removed the dev's instruction to stop at all, which the ruling did not ask for.

The paste instruction now names the surviving copy by its heading:

Paste rules/dev-template.md verbatim into every dispatch prompt; the PM fills
its {…} placeholders. {decision_frame} takes the stop trigger below, followed
by the whole `#### The four-axis decision frame (binding)` section above: the
frame has ONE copy in this skill and the PM pastes it at dispatch time, so no
dev-side copy can drift from it.

.claude/agents/os-dev.md — the frame block (old :400-421, declaring sentence through the binding sentence, which ran into the same paragraph as the unrelated blocked rule) collapses to one sentence. The blocked rule is preserved byte-for-byte. Neither :232 nor :303 is touched (PR #14808 holds them); the diff is one hunk at :400.

Three judgement calls in that one sentence:

1. It says the frame comes from the PM's own copy, not specifically from rules/dev-template.md, because that template is the PUBLISHED skill's artifact. The internal PM composes dispatch prompts per .claude/skills/pm-dispatch/references/dispatch-runbook.md and has no {decision_frame} slot (grep: the token exists only under skills/objectstack-pm-dispatch/). The published slot is named parenthetically, as the brief asked, without claiming the internal lane uses it. See finding F5.
2. It adds "if the dispatch prompt did not carry it, stop and ask; do not invent your own axes." os-dev.md's own doctrine is that standard clauses live in that file and prompt silence is normal — so moving the frame OUT of it opens a hole that a silent prompt would otherwise walk through.
3. It is written in Chinese and deliberately avoids the phrases fixed axes, 固定评估轴, 条评估轴 and any WORD-axis decision frame form. Those are FINGERPRINTS in the sync gate's anti-dormancy scan, whose population is per file: os-dev.md stops being a declared file with this change, so any of those phrasings would have turned the gate red. Verified after the edit — grep -nP for all four patterns over os-dev.md returns nothing.

scripts/check-skill-frame-sync.mjs — COPIES 4 → 2.

- internal-pm .claude/skills/pm-dispatch/SKILL.md KEPT
- internal-dev .claude/agents/os-dev.md REMOVED
- published-pm skills/objectstack-pm-dispatch/SKILL.md KEPT
- published-dev skills/objectstack-pm-dispatch/SKILL.md REMOVED

Every spelled count is now derived: the header ("FOUR hand-written copies across three files"), the COPIES docblock, the axisEntryStarts docblock, the freshness-gate cross-reference, and the count-disagreement message (the four copies do not agree becomes the ${results.length} copies do not agree). report() and the success line already derived theirs from COPIES.length.

The same-file overlap guard is kept even though it is vacuous today — no two copies share a file any more. Re-declaring such a pair is one edit away, and that guard is what would make such a pair prove it compares two different sections rather than the same one twice.

scripts/check-skill-frame-freshness.mjs — it imported COPIES but its self-test SPELLED .claude/agents/os-dev.md in five fixtures (cases 1, 2, 6, 7 and the two withUnreadableCopy calls in 8 and 9). Those broke on the reduction, exactly as that file's own #8024 rule predicts. They now derive SAMPLE_COPY_ID / SAMPLE_FRAME_FILE from COPIES[0], so the next change to the table re-points the fixtures instead of stranding them.

3. The anti-dormancy scan still refuses a returning copy — measured, both directions

Committed first, then mutated, with an EXIT INT TERM trap restoring absolute paths; each mutation proved on disk by counting the injected marker before reading any verdict; each restore proved by git diff HEAD empty AND by comparing git hash-object against the path's HEAD blob.

Direction A — the removed text returns to the now-undeclared .claude/agents/os-dev.md. REFUSED. Marker occurrences after mutation: 1 (was 0). node scripts/check-skill-frame-sync.mjs exit 1:

✗ check-skill-frame-sync: 1 problem(s).
• .claude/agents/os-dev.md
looks like ANOTHER copy of the decision frame (matched /fixed axes/) but is
not declared in COPIES, so nothing keeps it in sync.

Restore: git diff HEAD empty; worktree blob 968d4fd10a653f013635a8a48bdee25a25efc648 equals the HEAD blob.

Direction B — the same text returns to the still-declared skills/objectstack-pm-dispatch/SKILL.md. NOT REFUSED (exit 0). This is a real, measured limit and it is not new: the scan skips any file present in declaredFiles, so a second undeclared copy inside an already-declared file has always been invisible to it. Restore proved the same way (blob 3cbbed0f501736a0db3a5ecdd9e6cda43d44b434).

What that limit means for this PR, stated rather than implied: what keeps a second published copy from being written is the paste instruction naming the surviving section by heading, plus review — not the scan. Closing it mechanically needs a per-occurrence criterion the gate does not have (the published file legitimately carries two more fixed axes fingerprints, at :534 and in the :539 heading). A comment next to the new self-test case says exactly this, so the next reader does not mistake direction B for a bug introduced here.

A self-test case pins direction A against the REAL path this change undeclared, rather than a synthetic one.

4. Counts (merge base 5ff5f9576 → head)

file	bytes	lines	tokens (ceil bytes/4)
skills/objectstack-pm-dispatch/SKILL.md	38807 → 36898 (−1909)	753 → 726 (−27)	9702 → 9225 (−477), ceiling 9708
.claude/agents/os-dev.md	42743 → 41196 (−1547)	469 → 450 (−19)	— (line-ratcheted), ceiling 469
scripts/check-skill-frame-sync.mjs	40961 → 43180 (+2219)	960 → 990 (+30)	not ratcheted
scripts/check-skill-frame-freshness.mjs	46344 → 47480 (+1136)	1065 → 1085 (+20)	not ratcheted

Whole published package skills/objectstack-pm-dispatch/** (all markdown): 11540 → 11063 tokens, −477.

Two ratcheted files, both shrink; no ceiling row is edited, in either ratchet. The measured token saving is −477, not the −630 the card estimated — the difference is the stop trigger, kept for the reason in section 2. The script growth (+50 lines across two gates) is entirely the derivation constants, the new regression case and the comments recording the measured limit; neither script is under a ratchet.

5. Gates

Union derived at the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands:

dispatch-gates: gate list derived from the tree of 'objectstack-ai/objectstack' at commit ae616fbc91 (/home/user/objectstack-14685-item4).
dispatch-gates: change set derived from git — 4 path(s) vs merge base 5ff5f9576 of 'origin/main' and HEAD

Every exit code captured by redirect before any pipe.

All 32 re-run at the final head ae616fbc91, after the last commit. 31 green, 1 NOT MEASURED, 0 red. The load-bearing ones, quoted from their own verdict lines:

command	exit	its own verdict line
pnpm check:skill-frame-sync	0	✓ check-skill-frame-sync self-test: 13 cases pass, plus 5 dispatch-gates declaration cases. / ✓ check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
node scripts/check-skill-frame-freshness.mjs	0	✓ check-skill-frame-freshness: the decision frame in this tree is current with origin/main (fetched just now). (12 self-test cases pass)
node scripts/check-skills-token-ratchet.mjs	0	✓ check-skills-token-ratchet: skills/objectstack-pm-dispatch/SKILL.md is 9225 tokens (ceiling 9708; headroom 483).
pnpm check:pm-skill-ratchet	0	✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 450 lines (ceiling 469; headroom 19).
node scripts/check-self-test-wired.mjs	0	✓ check-self-test-wired: every one of the 164 script(s) CI runs that ship a --self-test has that self-test run by CI.
pnpm check:pm-dispatch-gates	0	✓ dispatch-gates self-test: 1241 cases pass.
pnpm --filter @objectstack/spec run check:skill-docs	0	✓ skills/README.md / ✓ content/docs/ai/skills-reference.mdx — nothing to regenerate
pnpm check:ratchet-remedy-authority	0	green
pnpm check:pm-skill-id-lint	0	green
pnpm check:nul-bytes	0	green (plus a direct grep -naP for control bytes over both edited markdown files: no match)
pnpm check:doc-authoring, check:role-word, check:agent-model-declared, check:agent-test-spelling, check:corpus-claim-drift, check:skill-compatibility, check:skill-identifier-liveness, check:watch-hint-literal, check:entry-guard, check:parse-guard, check:cli-command-ids, check:bash32-floor, check:pnpm-filter-targets, check:pm-governed-merges, check:cross-package-test-inputs	0	green
node scripts/check-ci-filter-parity.mjs, check-cross-package-test-inputs.mjs, check-shard-attestation.mjs, check-whole-set-label-write.mjs, scripts/pm/bare-root-worklist.mjs --self-test	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	first run answered PREREQUISITE NOT MET (exit 3, @objectstack/formula then @objectstack/lint unbuilt). Prerequisite supplied via the shared verify lock (turbo run build --filter=@objectstack/formula, then --filter=@objectstack/lint) and re-run to a real verdict: ✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 426 files / 1365 TS blocks judged clean
node scripts/check-test-completeness.mjs	3	NOT MEASURED — see below

One red was found and fixed rather than reported: on the first pass check:pm-skill-ratchet said ✗ check-skill-line-ratchet: .claude/agents/os-dev.md has 1 line(s) over the 120-byte budget: L400 (121B). The ratchet prices a per-line byte budget as well as a line ceiling. The replacement sentence was rewrapped at legal CJK break points, content byte-identical, and the whole union was re-derived and re-run at the resulting head — which is the head every number above comes from.

NOT MEASURED, with reasons:

• node scripts/check-test-completeness.mjs — exit 3, its own text: "PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named … the local reading for this gate is NOT MEASURED. ⛔ It is not a red." CI tees the log and passes it; unreachable there.
• The changeset-derived families (9 of them) do not apply: this PR is skip-changeset — .claude/**, skills/** and scripts/** publish nothing from any package.
• The CI-measured-only family reads the workflow event payload; no local run can produce a verdict for it.
• The always-runs tail (workflows with no path filter) is CI's, not measured here.

6. Out-of-scope findings — NOT fixed here, for the seat to triage

Each is a claim that this change makes false, in a file this flight does not hold.

• F1scripts/pm/dispatch-gates.mjs:8090 — why: '… the dev-agent definition is protocol semantics — every dispatched dev runs under it, and it carries an enforced copy of the decision frame'. The last clause is now false; the mandate itself still holds on the first clause, and the self-test still passes (it asserts the glob exists, not the reason).
• F2scripts/pm/dispatch-gates.mjs:8095 — why: '… the published PM skill carries two enforced copies of the decision frame (check:skill-frame-sync COPIES) …'. One copy now.
• F3scripts/pm/check-skill-line-ratchet.mjs:455 — // This file is one of the four check:skill-frame-sync COPIES, so a rule that …, in the os-dev.md ceiling ledger. That file is no longer a COPIES entry, which also weakens the stated reason for the +3 raise recorded there. Item 5 holds this file.
• F4scripts/check-skills-token-ratchet.mjs:334, 345, 359 — three ledger comments stating the published SKILL.md "carries TWO enforced frame copies" and that "the two gate-pinned copies of the decision frame stay in this file". Historically true; now stale as a present-tense claim.
• F5 the internal lane has no {decision_frame} slot. os-dev.md now expects the frame to arrive in the dispatch prompt, but .claude/skills/pm-dispatch/references/dispatch-runbook.md (派发词构造细则) does not tell the internal PM to paste it, and .claude/skills/pm-dispatch/SKILL.md is explicitly outside this flight. Until that lands, the internal dev's only protection is the stop-and-ask clause added in section 2. This is the one finding with live consequences.
• F6 the removed os-dev.md block carried a verbatim maintainer quotation — the 2026-09-01 axis-weighting ruling — so this PR deletes a quotation site that item 2's quotation-policy flight may have surveyed. The ruling text survives at .claude/skills/pm-dispatch/SKILL.md:904 and .claude/skills/pm-dispatch/references/decision-analysis.md:50, and the RULE itself survives in both published and internal PM copies of the frame.

7. Posture

Draft, governed surface. No reviewers requested, no ready flip, no enqueue, no auto-merge. Labels skip-changeset and needs:contract-review.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Part of #14685
Part of #13597

Item 4 of decision batch 3, ruled B. From the ruling comment (5520452691), verbatim:

B — the four-axis decision frame keeps two copies (internal-pm and published-pm); the dev-side copies are replaced by a {decision_frame} placeholder the dispatching PM fills at dispatch time; check:skill-frame-sync watches two sites. Confidence gap carried: whoever executes verifies that the published-dev block has no reader other than the placeholder path before removing it.

---

1. The confidence gap, discharged BEFORE any edit

The ruling made removal conditional on a reader census. Run at the merge base 5ff5f9576, over .claude/ skills/ scripts/ .github/ packages/ docs/, for four things: the published-dev start marker (Analyze every option on ... fixed axes:), the literal published-dev, {decision_frame}, and the internal-dev start marker. Every hit, and what it is:

site	what it is	depends on the published-dev BLOCK?
skills/objectstack-pm-dispatch/SKILL.md:633-662	the block itself	it IS the block
skills/objectstack-pm-dispatch/SKILL.md:623	"with the block below in place of its {decision_frame} line"	YES — the placeholder path, and the only one
skills/objectstack-pm-dispatch/rules/dev-template.md:99	the {decision_frame} line the instruction above fills	via that instruction only
scripts/check-skill-frame-sync.mjs:136-143	the published-dev COPIES entry	the gate that watches it
scripts/check-skill-frame-sync.mjs:714,719,723,724,727	self-test fixtures addressing the published-dev id	the gate's own fixtures
scripts/check-skill-frame-freshness.mjs	imports COPIES / AXIS_MAP wholesale; five fixtures spelled .claude/agents/os-dev.md	reads the TABLE, never the block text
scripts/pm/dispatch-gates.mjs:15382-15397	spawned import reading the FILE list out of COPIES for the tier mandate	reads FILES, not copies
scripts/pm/check-dispatch-gates.mjs:202	FRAME_MODULE watch-hint constant	names the module path only
scripts/check-skills-token-ratchet.mjs:334,345,359	ceiling-ledger prose	comments, no behaviour
scripts/pm/check-skill-line-ratchet.mjs:455	ceiling-ledger prose	comments, no behaviour

No test, generator, workflow, doc or fixture quotes the block by fingerprint. The gap resolves in favour of removal: the placeholder path and the two frame gates are the whole reader set, and both gates read the COPIES table rather than the prose.

The same census over the internal-dev start marker found exactly one non-gate hit — the block itself at .claude/agents/os-dev.md:400.

2. What changed, site by site

skills/objectstack-pm-dispatch/SKILL.md — the embedded dev template loses the frame copy: the declaring sentence through the binding sentence (old :633-662), which is precisely the span the gate's start and binding anchors delimit.

The stop trigger above it (old :627-632, "When to STOP instead of coding …") stays, and that is a judgement call worth naming. It is not part of the frame — the gate's anchors do not reach it — and it is the escalation TRIGGER rather than the analysis. rules/dev-template.md carries it nowhere else (grepped: the template mentions needs_decision only at :34 inside the contract-first rule and at :112 in the report schema), so deleting it with the frame would have removed the dev's instruction to stop at all, which the ruling did not ask for.

The paste instruction now names the surviving copy by its heading:

Paste rules/dev-template.md verbatim into every dispatch prompt; the PM fills
its {…} placeholders. {decision_frame} takes the stop trigger below, followed
by the whole `#### The four-axis decision frame (binding)` section above: the
frame has ONE copy in this skill and the PM pastes it at dispatch time, so no
dev-side copy can drift from it.

.claude/agents/os-dev.md — the frame block (old :400-421, declaring sentence through the binding sentence, which ran into the same paragraph as the unrelated blocked rule) collapses to one sentence. The blocked rule is preserved byte-for-byte. Neither :232 nor :303 is touched (PR #14808 holds them); the diff is one hunk at :400.

Three judgement calls in that one sentence:

1. It says the frame comes from the PM's own copy, not specifically from rules/dev-template.md, because that template is the PUBLISHED skill's artifact. The internal PM composes dispatch prompts per .claude/skills/pm-dispatch/references/dispatch-runbook.md and has no {decision_frame} slot (grep: the token exists only under skills/objectstack-pm-dispatch/). The published slot is named parenthetically, as the brief asked, without claiming the internal lane uses it. See finding F5.
2. It adds "if the dispatch prompt did not carry it, stop and ask; do not invent your own axes." os-dev.md's own doctrine is that standard clauses live in that file and prompt silence is normal — so moving the frame OUT of it opens a hole that a silent prompt would otherwise walk through.
3. It is written in Chinese and deliberately avoids the phrases fixed axes, 固定评估轴, 条评估轴 and any WORD-axis decision frame form. Those are FINGERPRINTS in the sync gate's anti-dormancy scan, whose population is per file: os-dev.md stops being a declared file with this change, so any of those phrasings would have turned the gate red. Verified after the edit — grep -nP for all four patterns over os-dev.md returns nothing.

scripts/check-skill-frame-sync.mjs — COPIES 4 → 2.

- internal-pm .claude/skills/pm-dispatch/SKILL.md KEPT
- internal-dev .claude/agents/os-dev.md REMOVED
- published-pm skills/objectstack-pm-dispatch/SKILL.md KEPT
- published-dev skills/objectstack-pm-dispatch/SKILL.md REMOVED

Every spelled count is now derived: the header ("FOUR hand-written copies across three files"), the COPIES docblock, the axisEntryStarts docblock, the freshness-gate cross-reference, and the count-disagreement message (the four copies do not agree becomes the ${results.length} copies do not agree). report() and the success line already derived theirs from COPIES.length.

The same-file overlap guard is kept even though it is vacuous today — no two copies share a file any more. Re-declaring such a pair is one edit away, and that guard is what would make such a pair prove it compares two different sections rather than the same one twice.

scripts/check-skill-frame-freshness.mjs — it imported COPIES but its self-test SPELLED .claude/agents/os-dev.md in five fixtures (cases 1, 2, 6, 7 and the two withUnreadableCopy calls in 8 and 9). Those broke on the reduction, exactly as that file's own #8024 rule predicts. They now derive SAMPLE_COPY_ID / SAMPLE_FRAME_FILE from COPIES[0], so the next change to the table re-points the fixtures instead of stranding them.

3. The anti-dormancy scan still refuses a returning copy — measured, both directions

Committed first, then mutated, with an EXIT INT TERM trap restoring absolute paths; each mutation proved on disk by counting the injected marker before reading any verdict; each restore proved by git diff HEAD empty AND by comparing git hash-object against the path's HEAD blob.

Direction A — the removed text returns to the now-undeclared .claude/agents/os-dev.md. REFUSED. Marker occurrences after mutation: 1 (was 0). node scripts/check-skill-frame-sync.mjs exit 1:

✗ check-skill-frame-sync: 1 problem(s).
• .claude/agents/os-dev.md
looks like ANOTHER copy of the decision frame (matched /fixed axes/) but is
not declared in COPIES, so nothing keeps it in sync.

Restore: git diff HEAD empty; worktree blob 968d4fd10a653f013635a8a48bdee25a25efc648 equals the HEAD blob.

Direction B — the same text returns to the still-declared skills/objectstack-pm-dispatch/SKILL.md. NOT REFUSED (exit 0). This is a real, measured limit and it is not new: the scan skips any file present in declaredFiles, so a second undeclared copy inside an already-declared file has always been invisible to it. Restore proved the same way (blob 3cbbed0f501736a0db3a5ecdd9e6cda43d44b434).

What that limit means for this PR, stated rather than implied: what keeps a second published copy from being written is the paste instruction naming the surviving section by heading, plus review — not the scan. Closing it mechanically needs a per-occurrence criterion the gate does not have (the published file legitimately carries two more fixed axes fingerprints, at :534 and in the :539 heading). A comment next to the new self-test case says exactly this, so the next reader does not mistake direction B for a bug introduced here.

A self-test case pins direction A against the REAL path this change undeclared, rather than a synthetic one.

4. Counts (merge base 5ff5f9576 → head)

file	bytes	lines	tokens (ceil bytes/4)
skills/objectstack-pm-dispatch/SKILL.md	38807 → 36898 (−1909)	753 → 726 (−27)	9702 → 9225 (−477), ceiling 9708
.claude/agents/os-dev.md	42743 → 41196 (−1547)	469 → 450 (−19)	— (line-ratcheted), ceiling 469
scripts/check-skill-frame-sync.mjs	40961 → 43180 (+2219)	960 → 990 (+30)	not ratcheted
scripts/check-skill-frame-freshness.mjs	46344 → 47480 (+1136)	1065 → 1085 (+20)	not ratcheted

Whole published package skills/objectstack-pm-dispatch/** (all markdown): 11540 → 11063 tokens, −477.

Two ratcheted files, both shrink; no ceiling row is edited, in either ratchet. The measured token saving is −477, not the −630 the card estimated — the difference is the stop trigger, kept for the reason in section 2. The script growth (+50 lines across two gates) is entirely the derivation constants, the new regression case and the comments recording the measured limit; neither script is under a ratchet.

5. Gates

Union derived at the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands:

dispatch-gates: gate list derived from the tree of 'objectstack-ai/objectstack' at commit ae616fbc91 (/home/user/objectstack-14685-item4).
dispatch-gates: change set derived from git — 4 path(s) vs merge base 5ff5f9576 of 'origin/main' and HEAD

Every exit code captured by redirect before any pipe.

All 32 re-run at the final head ae616fbc91, after the last commit. 31 green, 1 NOT MEASURED, 0 red. The load-bearing ones, quoted from their own verdict lines:

command	exit	its own verdict line
pnpm check:skill-frame-sync	0	✓ check-skill-frame-sync self-test: 13 cases pass, plus 5 dispatch-gates declaration cases. / ✓ check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
node scripts/check-skill-frame-freshness.mjs	0	✓ check-skill-frame-freshness: the decision frame in this tree is current with origin/main (fetched just now). (12 self-test cases pass)
node scripts/check-skills-token-ratchet.mjs	0	✓ check-skills-token-ratchet: skills/objectstack-pm-dispatch/SKILL.md is 9225 tokens (ceiling 9708; headroom 483).
pnpm check:pm-skill-ratchet	0	✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 450 lines (ceiling 469; headroom 19).
node scripts/check-self-test-wired.mjs	0	✓ check-self-test-wired: every one of the 164 script(s) CI runs that ship a --self-test has that self-test run by CI.
pnpm check:pm-dispatch-gates	0	✓ dispatch-gates self-test: 1241 cases pass.
pnpm --filter @objectstack/spec run check:skill-docs	0	✓ skills/README.md / ✓ content/docs/ai/skills-reference.mdx — nothing to regenerate
pnpm check:ratchet-remedy-authority	0	green
pnpm check:pm-skill-id-lint	0	green
pnpm check:nul-bytes	0	green (plus a direct grep -naP for control bytes over both edited markdown files: no match)
pnpm check:doc-authoring, check:role-word, check:agent-model-declared, check:agent-test-spelling, check:corpus-claim-drift, check:skill-compatibility, check:skill-identifier-liveness, check:watch-hint-literal, check:entry-guard, check:parse-guard, check:cli-command-ids, check:bash32-floor, check:pnpm-filter-targets, check:pm-governed-merges, check:cross-package-test-inputs	0	green
node scripts/check-ci-filter-parity.mjs, check-cross-package-test-inputs.mjs, check-shard-attestation.mjs, check-whole-set-label-write.mjs, scripts/pm/bare-root-worklist.mjs --self-test	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	first run answered PREREQUISITE NOT MET (exit 3, @objectstack/formula then @objectstack/lint unbuilt). Prerequisite supplied via the shared verify lock (turbo run build --filter=@objectstack/formula, then --filter=@objectstack/lint) and re-run to a real verdict: ✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 426 files / 1365 TS blocks judged clean
node scripts/check-test-completeness.mjs	3	NOT MEASURED — see below

One red was found and fixed rather than reported: on the first pass check:pm-skill-ratchet said ✗ check-skill-line-ratchet: .claude/agents/os-dev.md has 1 line(s) over the 120-byte budget: L400 (121B). The ratchet prices a per-line byte budget as well as a line ceiling. The replacement sentence was rewrapped at legal CJK break points, content byte-identical, and the whole union was re-derived and re-run at the resulting head — which is the head every number above comes from.

NOT MEASURED, with reasons:

• node scripts/check-test-completeness.mjs — exit 3, its own text: "PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named … the local reading for this gate is NOT MEASURED. ⛔ It is not a red." CI tees the log and passes it; unreachable there.
• The changeset-derived families (9 of them) do not apply: this PR is skip-changeset — .claude/**, skills/** and scripts/** publish nothing from any package.
• The CI-measured-only family reads the workflow event payload; no local run can produce a verdict for it.
• The always-runs tail (workflows with no path filter) is CI's, not measured here.

6. Out-of-scope findings — NOT fixed here, for the seat to triage

Each is a claim that this change makes false, in a file this flight does not hold.

• F1scripts/pm/dispatch-gates.mjs:8090 — why: '… the dev-agent definition is protocol semantics — every dispatched dev runs under it, and it carries an enforced copy of the decision frame'. The last clause is now false; the mandate itself still holds on the first clause, and the self-test still passes (it asserts the glob exists, not the reason).
• F2scripts/pm/dispatch-gates.mjs:8095 — why: '… the published PM skill carries two enforced copies of the decision frame (check:skill-frame-sync COPIES) …'. One copy now.
• F3scripts/pm/check-skill-line-ratchet.mjs:455 — // This file is one of the four check:skill-frame-sync COPIES, so a rule that …, in the os-dev.md ceiling ledger. That file is no longer a COPIES entry, which also weakens the stated reason for the +3 raise recorded there. Item 5 holds this file.
• F4scripts/check-skills-token-ratchet.mjs:334, 345, 359 — three ledger comments stating the published SKILL.md "carries TWO enforced frame copies" and that "the two gate-pinned copies of the decision frame stay in this file". Historically true; now stale as a present-tense claim.
• F5 the internal lane has no {decision_frame} slot. os-dev.md now expects the frame to arrive in the dispatch prompt, but .claude/skills/pm-dispatch/references/dispatch-runbook.md (派发词构造细则) does not tell the internal PM to paste it, and .claude/skills/pm-dispatch/SKILL.md is explicitly outside this flight. Until that lands, the internal dev's only protection is the stop-and-ask clause added in section 2. This is the one finding with live consequences.
• F6 the removed os-dev.md block carried a verbatim maintainer quotation — the 2026-09-01 axis-weighting ruling — so this PR deletes a quotation site that item 2's quotation-policy flight may have surveyed. The ruling text survives at .claude/skills/pm-dispatch/SKILL.md:904 and .claude/skills/pm-dispatch/references/decision-analysis.md:50, and the RULE itself survives in both published and internal PM copies of the frame.

7. Posture

Draft, governed surface. No reviewers requested, no ready flip, no enqueue, no auto-merge. Labels skip-changeset and needs:contract-review.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Part of #14685
Part of #13597

Item 4 of decision batch 3, ruled B. From the ruling comment (5520452691), verbatim:

B — the four-axis decision frame keeps two copies (internal-pm and published-pm); the dev-side copies are replaced by a {decision_frame} placeholder the dispatching PM fills at dispatch time; check:skill-frame-sync watches two sites. Confidence gap carried: whoever executes verifies that the published-dev block has no reader other than the placeholder path before removing it.

---

1. The confidence gap, discharged BEFORE any edit

The ruling made removal conditional on a reader census. Run at the merge base 5ff5f9576, over .claude/ skills/ scripts/ .github/ packages/ docs/, for four things: the published-dev start marker (Analyze every option on ... fixed axes:), the literal published-dev, {decision_frame}, and the internal-dev start marker. Every hit, and what it is:

site	what it is	depends on the published-dev BLOCK?
skills/objectstack-pm-dispatch/SKILL.md:633-662	the block itself	it IS the block
skills/objectstack-pm-dispatch/SKILL.md:623	"with the block below in place of its {decision_frame} line"	YES — the placeholder path, and the only one
skills/objectstack-pm-dispatch/rules/dev-template.md:99	the {decision_frame} line the instruction above fills	via that instruction only
scripts/check-skill-frame-sync.mjs:136-143	the published-dev COPIES entry	the gate that watches it
scripts/check-skill-frame-sync.mjs:714,719,723,724,727	self-test fixtures addressing the published-dev id	the gate's own fixtures
scripts/check-skill-frame-freshness.mjs	imports COPIES / AXIS_MAP wholesale; five fixtures spelled .claude/agents/os-dev.md	reads the TABLE, never the block text
scripts/pm/dispatch-gates.mjs:15382-15397	spawned import reading the FILE list out of COPIES for the tier mandate	reads FILES, not copies
scripts/pm/check-dispatch-gates.mjs:202	FRAME_MODULE watch-hint constant	names the module path only
scripts/check-skills-token-ratchet.mjs:334,345,359	ceiling-ledger prose	comments, no behaviour
scripts/pm/check-skill-line-ratchet.mjs:455	ceiling-ledger prose	comments, no behaviour

No test, generator, workflow, doc or fixture quotes the block by fingerprint. The gap resolves in favour of removal: the placeholder path and the two frame gates are the whole reader set, and both gates read the COPIES table rather than the prose.

The same census over the internal-dev start marker found exactly one non-gate hit — the block itself at .claude/agents/os-dev.md:400.

2. What changed, site by site

skills/objectstack-pm-dispatch/SKILL.md — the embedded dev template loses the frame copy: the declaring sentence through the binding sentence (old :633-662), which is precisely the span the gate's start and binding anchors delimit.

The stop trigger above it (old :627-632, "When to STOP instead of coding …") stays, and that is a judgement call worth naming. It is not part of the frame — the gate's anchors do not reach it — and it is the escalation TRIGGER rather than the analysis. rules/dev-template.md carries it nowhere else (grepped: the template mentions needs_decision only at :34 inside the contract-first rule and at :112 in the report schema), so deleting it with the frame would have removed the dev's instruction to stop at all, which the ruling did not ask for.

The paste instruction now names the surviving copy by its heading:

Paste rules/dev-template.md verbatim into every dispatch prompt; the PM fills
its {…} placeholders. {decision_frame} takes the stop trigger below, followed
by the whole `#### The four-axis decision frame (binding)` section above: the
frame has ONE copy in this skill and the PM pastes it at dispatch time, so no
dev-side copy can drift from it.

.claude/agents/os-dev.md — the frame block (old :400-421, declaring sentence through the binding sentence, which ran into the same paragraph as the unrelated blocked rule) collapses to one sentence. The blocked rule is preserved byte-for-byte. Neither :232 nor :303 is touched (PR #14808 holds them); the diff is one hunk at :400.

Three judgement calls in that one sentence:

1. It says the frame comes from the PM's own copy, not specifically from rules/dev-template.md, because that template is the PUBLISHED skill's artifact. The internal PM composes dispatch prompts per .claude/skills/pm-dispatch/references/dispatch-runbook.md and has no {decision_frame} slot (grep: the token exists only under skills/objectstack-pm-dispatch/). The published slot is named parenthetically, as the brief asked, without claiming the internal lane uses it. See finding F5.
2. It adds "if the dispatch prompt did not carry it, stop and ask; do not invent your own axes." os-dev.md's own doctrine is that standard clauses live in that file and prompt silence is normal — so moving the frame OUT of it opens a hole that a silent prompt would otherwise walk through.
3. It is written in Chinese and deliberately avoids the phrases fixed axes, 固定评估轴, 条评估轴 and any WORD-axis decision frame form. Those are FINGERPRINTS in the sync gate's anti-dormancy scan, whose population is per file: os-dev.md stops being a declared file with this change, so any of those phrasings would have turned the gate red. Verified after the edit — grep -nP for all four patterns over os-dev.md returns nothing.

scripts/check-skill-frame-sync.mjs — COPIES 4 → 2.

- internal-pm .claude/skills/pm-dispatch/SKILL.md KEPT
- internal-dev .claude/agents/os-dev.md REMOVED
- published-pm skills/objectstack-pm-dispatch/SKILL.md KEPT
- published-dev skills/objectstack-pm-dispatch/SKILL.md REMOVED

Every spelled count is now derived: the header ("FOUR hand-written copies across three files"), the COPIES docblock, the axisEntryStarts docblock, the freshness-gate cross-reference, and the count-disagreement message (the four copies do not agree becomes the ${results.length} copies do not agree). report() and the success line already derived theirs from COPIES.length.

The same-file overlap guard is kept even though it is vacuous today — no two copies share a file any more. Re-declaring such a pair is one edit away, and that guard is what would make such a pair prove it compares two different sections rather than the same one twice.

scripts/check-skill-frame-freshness.mjs — it imported COPIES but its self-test SPELLED .claude/agents/os-dev.md in five fixtures (cases 1, 2, 6, 7 and the two withUnreadableCopy calls in 8 and 9). Those broke on the reduction, exactly as that file's own #8024 rule predicts. They now derive SAMPLE_COPY_ID / SAMPLE_FRAME_FILE from COPIES[0], so the next change to the table re-points the fixtures instead of stranding them.

3. The anti-dormancy scan still refuses a returning copy — measured, both directions

Committed first, then mutated, with an EXIT INT TERM trap restoring absolute paths; each mutation proved on disk by counting the injected marker before reading any verdict; each restore proved by git diff HEAD empty AND by comparing git hash-object against the path's HEAD blob.

Direction A — the removed text returns to the now-undeclared .claude/agents/os-dev.md. REFUSED. Marker occurrences after mutation: 1 (was 0). node scripts/check-skill-frame-sync.mjs exit 1:

✗ check-skill-frame-sync: 1 problem(s).
• .claude/agents/os-dev.md
looks like ANOTHER copy of the decision frame (matched /fixed axes/) but is
not declared in COPIES, so nothing keeps it in sync.

Restore: git diff HEAD empty; worktree blob 968d4fd10a653f013635a8a48bdee25a25efc648 equals the HEAD blob.

Direction B — the same text returns to the still-declared skills/objectstack-pm-dispatch/SKILL.md. NOT REFUSED (exit 0). This is a real, measured limit and it is not new: the scan skips any file present in declaredFiles, so a second undeclared copy inside an already-declared file has always been invisible to it. Restore proved the same way (blob 3cbbed0f501736a0db3a5ecdd9e6cda43d44b434).

What that limit means for this PR, stated rather than implied: what keeps a second published copy from being written is the paste instruction naming the surviving section by heading, plus review — not the scan. Closing it mechanically needs a per-occurrence criterion the gate does not have (the published file legitimately carries two more fixed axes fingerprints, at :534 and in the :539 heading). A comment next to the new self-test case says exactly this, so the next reader does not mistake direction B for a bug introduced here.

A self-test case pins direction A against the REAL path this change undeclared, rather than a synthetic one.

4. Counts (merge base 5ff5f9576 → head)

file	bytes	lines	tokens (ceil bytes/4)
skills/objectstack-pm-dispatch/SKILL.md	38807 → 36898 (−1909)	753 → 726 (−27)	9702 → 9225 (−477), ceiling 9708
.claude/agents/os-dev.md	42743 → 41196 (−1547)	469 → 450 (−19)	— (line-ratcheted), ceiling 469
scripts/check-skill-frame-sync.mjs	40961 → 43180 (+2219)	960 → 990 (+30)	not ratcheted
scripts/check-skill-frame-freshness.mjs	46344 → 47480 (+1136)	1065 → 1085 (+20)	not ratcheted

Whole published package skills/objectstack-pm-dispatch/** (all markdown): 11540 → 11063 tokens, −477.

Two ratcheted files, both shrink; no ceiling row is edited, in either ratchet. The measured token saving is −477, not the −630 the card estimated — the difference is the stop trigger, kept for the reason in section 2. The script growth (+50 lines across two gates) is entirely the derivation constants, the new regression case and the comments recording the measured limit; neither script is under a ratchet.

5. Gates

Union derived at the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands:

dispatch-gates: gate list derived from the tree of 'objectstack-ai/objectstack' at commit ae616fbc91 (/home/user/objectstack-14685-item4).
dispatch-gates: change set derived from git — 4 path(s) vs merge base 5ff5f9576 of 'origin/main' and HEAD

Every exit code captured by redirect before any pipe.

All 32 re-run at the final head ae616fbc91, after the last commit. 31 green, 1 NOT MEASURED, 0 red. The load-bearing ones, quoted from their own verdict lines:

command	exit	its own verdict line
pnpm check:skill-frame-sync	0	✓ check-skill-frame-sync self-test: 13 cases pass, plus 5 dispatch-gates declaration cases. / ✓ check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
node scripts/check-skill-frame-freshness.mjs	0	✓ check-skill-frame-freshness: the decision frame in this tree is current with origin/main (fetched just now). (12 self-test cases pass)
node scripts/check-skills-token-ratchet.mjs	0	✓ check-skills-token-ratchet: skills/objectstack-pm-dispatch/SKILL.md is 9225 tokens (ceiling 9708; headroom 483).
pnpm check:pm-skill-ratchet	0	✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 450 lines (ceiling 469; headroom 19).
node scripts/check-self-test-wired.mjs	0	✓ check-self-test-wired: every one of the 164 script(s) CI runs that ship a --self-test has that self-test run by CI.
pnpm check:pm-dispatch-gates	0	✓ dispatch-gates self-test: 1241 cases pass.
pnpm --filter @objectstack/spec run check:skill-docs	0	✓ skills/README.md / ✓ content/docs/ai/skills-reference.mdx — nothing to regenerate
pnpm check:ratchet-remedy-authority	0	green
pnpm check:pm-skill-id-lint	0	green
pnpm check:nul-bytes	0	green (plus a direct grep -naP for control bytes over both edited markdown files: no match)
pnpm check:doc-authoring, check:role-word, check:agent-model-declared, check:agent-test-spelling, check:corpus-claim-drift, check:skill-compatibility, check:skill-identifier-liveness, check:watch-hint-literal, check:entry-guard, check:parse-guard, check:cli-command-ids, check:bash32-floor, check:pnpm-filter-targets, check:pm-governed-merges, check:cross-package-test-inputs	0	green
node scripts/check-ci-filter-parity.mjs, check-cross-package-test-inputs.mjs, check-shard-attestation.mjs, check-whole-set-label-write.mjs, scripts/pm/bare-root-worklist.mjs --self-test	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	first run answered PREREQUISITE NOT MET (exit 3, @objectstack/formula then @objectstack/lint unbuilt). Prerequisite supplied via the shared verify lock (turbo run build --filter=@objectstack/formula, then --filter=@objectstack/lint) and re-run to a real verdict: ✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 426 files / 1365 TS blocks judged clean
node scripts/check-test-completeness.mjs	3	NOT MEASURED — see below

One red was found and fixed rather than reported: on the first pass check:pm-skill-ratchet said ✗ check-skill-line-ratchet: .claude/agents/os-dev.md has 1 line(s) over the 120-byte budget: L400 (121B). The ratchet prices a per-line byte budget as well as a line ceiling. The replacement sentence was rewrapped at legal CJK break points, content byte-identical, and the whole union was re-derived and re-run at the resulting head — which is the head every number above comes from.

NOT MEASURED, with reasons:

• node scripts/check-test-completeness.mjs — exit 3, its own text: "PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named … the local reading for this gate is NOT MEASURED. ⛔ It is not a red." CI tees the log and passes it; unreachable there.
• The changeset-derived families (9 of them) do not apply: this PR is skip-changeset — .claude/**, skills/** and scripts/** publish nothing from any package.
• The CI-measured-only family reads the workflow event payload; no local run can produce a verdict for it.
• The always-runs tail (workflows with no path filter) is CI's, not measured here.

6. Out-of-scope findings — NOT fixed here, for the seat to triage

Each is a claim that this change makes false, in a file this flight does not hold.

• F1scripts/pm/dispatch-gates.mjs:8090 — why: '… the dev-agent definition is protocol semantics — every dispatched dev runs under it, and it carries an enforced copy of the decision frame'. The last clause is now false; the mandate itself still holds on the first clause, and the self-test still passes (it asserts the glob exists, not the reason).
• F2scripts/pm/dispatch-gates.mjs:8095 — why: '… the published PM skill carries two enforced copies of the decision frame (check:skill-frame-sync COPIES) …'. One copy now.
• F3scripts/pm/check-skill-line-ratchet.mjs:455 — // This file is one of the four check:skill-frame-sync COPIES, so a rule that …, in the os-dev.md ceiling ledger. That file is no longer a COPIES entry, which also weakens the stated reason for the +3 raise recorded there. Item 5 holds this file.
• F4scripts/check-skills-token-ratchet.mjs:334, 345, 359 — three ledger comments stating the published SKILL.md "carries TWO enforced frame copies" and that "the two gate-pinned copies of the decision frame stay in this file". Historically true; now stale as a present-tense claim.
• F5 the internal lane has no {decision_frame} slot. os-dev.md now expects the frame to arrive in the dispatch prompt, but .claude/skills/pm-dispatch/references/dispatch-runbook.md (派发词构造细则) does not tell the internal PM to paste it, and .claude/skills/pm-dispatch/SKILL.md is explicitly outside this flight. Until that lands, the internal dev's only protection is the stop-and-ask clause added in section 2. This is the one finding with live consequences.
• F6 the removed os-dev.md block carried a verbatim maintainer quotation — the 2026-09-01 axis-weighting ruling — so this PR deletes a quotation site that item 2's quotation-policy flight may have surveyed. The ruling text survives at .claude/skills/pm-dispatch/SKILL.md:904 and .claude/skills/pm-dispatch/references/decision-analysis.md:50, and the RULE itself survives in both published and internal PM copies of the frame.

7. Posture

Draft, governed surface. No reviewers requested, no ready flip, no enqueue, no auto-merge. Labels skip-changeset and needs:contract-review.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Part of #14685
Part of #13597

Item 4 of decision batch 3, ruled B. From the ruling comment (5520452691), verbatim:

B — the four-axis decision frame keeps two copies (internal-pm and published-pm); the dev-side copies are replaced by a {decision_frame} placeholder the dispatching PM fills at dispatch time; check:skill-frame-sync watches two sites. Confidence gap carried: whoever executes verifies that the published-dev block has no reader other than the placeholder path before removing it.

---

1. The confidence gap, discharged BEFORE any edit

The ruling made removal conditional on a reader census. Run at the merge base 5ff5f9576, over .claude/ skills/ scripts/ .github/ packages/ docs/, for four things: the published-dev start marker (Analyze every option on ... fixed axes:), the literal published-dev, {decision_frame}, and the internal-dev start marker. Every hit, and what it is:

site	what it is	depends on the published-dev BLOCK?
skills/objectstack-pm-dispatch/SKILL.md:633-662	the block itself	it IS the block
skills/objectstack-pm-dispatch/SKILL.md:623	"with the block below in place of its {decision_frame} line"	YES — the placeholder path, and the only one
skills/objectstack-pm-dispatch/rules/dev-template.md:99	the {decision_frame} line the instruction above fills	via that instruction only
scripts/check-skill-frame-sync.mjs:136-143	the published-dev COPIES entry	the gate that watches it
scripts/check-skill-frame-sync.mjs:714,719,723,724,727	self-test fixtures addressing the published-dev id	the gate's own fixtures
scripts/check-skill-frame-freshness.mjs	imports COPIES / AXIS_MAP wholesale; five fixtures spelled .claude/agents/os-dev.md	reads the TABLE, never the block text
scripts/pm/dispatch-gates.mjs:15382-15397	spawned import reading the FILE list out of COPIES for the tier mandate	reads FILES, not copies
scripts/pm/check-dispatch-gates.mjs:202	FRAME_MODULE watch-hint constant	names the module path only
scripts/check-skills-token-ratchet.mjs:334,345,359	ceiling-ledger prose	comments, no behaviour
scripts/pm/check-skill-line-ratchet.mjs:455	ceiling-ledger prose	comments, no behaviour

No test, generator, workflow, doc or fixture quotes the block by fingerprint. The gap resolves in favour of removal: the placeholder path and the two frame gates are the whole reader set, and both gates read the COPIES table rather than the prose.

The same census over the internal-dev start marker found exactly one non-gate hit — the block itself at .claude/agents/os-dev.md:400.

2. What changed, site by site

skills/objectstack-pm-dispatch/SKILL.md — the embedded dev template loses the frame copy: the declaring sentence through the binding sentence (old :633-662), which is precisely the span the gate's start and binding anchors delimit.

The stop trigger above it (old :627-632, "When to STOP instead of coding …") stays, and that is a judgement call worth naming. It is not part of the frame — the gate's anchors do not reach it — and it is the escalation TRIGGER rather than the analysis. rules/dev-template.md carries it nowhere else (grepped: the template mentions needs_decision only at :34 inside the contract-first rule and at :112 in the report schema), so deleting it with the frame would have removed the dev's instruction to stop at all, which the ruling did not ask for.

The paste instruction now names the surviving copy by its heading:

Paste rules/dev-template.md verbatim into every dispatch prompt; the PM fills
its {…} placeholders. {decision_frame} takes the stop trigger below, followed
by the whole `#### The four-axis decision frame (binding)` section above: the
frame has ONE copy in this skill and the PM pastes it at dispatch time, so no
dev-side copy can drift from it.

.claude/agents/os-dev.md — the frame block (old :400-421, declaring sentence through the binding sentence, which ran into the same paragraph as the unrelated blocked rule) collapses to one sentence. The blocked rule is preserved byte-for-byte. Neither :232 nor :303 is touched (PR #14808 holds them); the diff is one hunk at :400.

Three judgement calls in that one sentence:

1. It says the frame comes from the PM's own copy, not specifically from rules/dev-template.md, because that template is the PUBLISHED skill's artifact. The internal PM composes dispatch prompts per .claude/skills/pm-dispatch/references/dispatch-runbook.md and has no {decision_frame} slot (grep: the token exists only under skills/objectstack-pm-dispatch/). The published slot is named parenthetically, as the brief asked, without claiming the internal lane uses it. See finding F5.
2. It adds "if the dispatch prompt did not carry it, stop and ask; do not invent your own axes." os-dev.md's own doctrine is that standard clauses live in that file and prompt silence is normal — so moving the frame OUT of it opens a hole that a silent prompt would otherwise walk through.
3. It is written in Chinese and deliberately avoids the phrases fixed axes, 固定评估轴, 条评估轴 and any WORD-axis decision frame form. Those are FINGERPRINTS in the sync gate's anti-dormancy scan, whose population is per file: os-dev.md stops being a declared file with this change, so any of those phrasings would have turned the gate red. Verified after the edit — grep -nP for all four patterns over os-dev.md returns nothing.

scripts/check-skill-frame-sync.mjs — COPIES 4 → 2.

- internal-pm .claude/skills/pm-dispatch/SKILL.md KEPT
- internal-dev .claude/agents/os-dev.md REMOVED
- published-pm skills/objectstack-pm-dispatch/SKILL.md KEPT
- published-dev skills/objectstack-pm-dispatch/SKILL.md REMOVED

Every spelled count is now derived: the header ("FOUR hand-written copies across three files"), the COPIES docblock, the axisEntryStarts docblock, the freshness-gate cross-reference, and the count-disagreement message (the four copies do not agree becomes the ${results.length} copies do not agree). report() and the success line already derived theirs from COPIES.length.

The same-file overlap guard is kept even though it is vacuous today — no two copies share a file any more. Re-declaring such a pair is one edit away, and that guard is what would make such a pair prove it compares two different sections rather than the same one twice.

scripts/check-skill-frame-freshness.mjs — it imported COPIES but its self-test SPELLED .claude/agents/os-dev.md in five fixtures (cases 1, 2, 6, 7 and the two withUnreadableCopy calls in 8 and 9). Those broke on the reduction, exactly as that file's own #8024 rule predicts. They now derive SAMPLE_COPY_ID / SAMPLE_FRAME_FILE from COPIES[0], so the next change to the table re-points the fixtures instead of stranding them.

3. The anti-dormancy scan still refuses a returning copy — measured, both directions

Committed first, then mutated, with an EXIT INT TERM trap restoring absolute paths; each mutation proved on disk by counting the injected marker before reading any verdict; each restore proved by git diff HEAD empty AND by comparing git hash-object against the path's HEAD blob.

Direction A — the removed text returns to the now-undeclared .claude/agents/os-dev.md. REFUSED. Marker occurrences after mutation: 1 (was 0). node scripts/check-skill-frame-sync.mjs exit 1:

✗ check-skill-frame-sync: 1 problem(s).
• .claude/agents/os-dev.md
looks like ANOTHER copy of the decision frame (matched /fixed axes/) but is
not declared in COPIES, so nothing keeps it in sync.

Restore: git diff HEAD empty; worktree blob 968d4fd10a653f013635a8a48bdee25a25efc648 equals the HEAD blob.

Direction B — the same text returns to the still-declared skills/objectstack-pm-dispatch/SKILL.md. NOT REFUSED (exit 0). This is a real, measured limit and it is not new: the scan skips any file present in declaredFiles, so a second undeclared copy inside an already-declared file has always been invisible to it. Restore proved the same way (blob 3cbbed0f501736a0db3a5ecdd9e6cda43d44b434).

What that limit means for this PR, stated rather than implied: what keeps a second published copy from being written is the paste instruction naming the surviving section by heading, plus review — not the scan. Closing it mechanically needs a per-occurrence criterion the gate does not have (the published file legitimately carries two more fixed axes fingerprints, at :534 and in the :539 heading). A comment next to the new self-test case says exactly this, so the next reader does not mistake direction B for a bug introduced here.

A self-test case pins direction A against the REAL path this change undeclared, rather than a synthetic one.

4. Counts (merge base 5ff5f9576 → head)

file	bytes	lines	tokens (ceil bytes/4)
skills/objectstack-pm-dispatch/SKILL.md	38807 → 36898 (−1909)	753 → 726 (−27)	9702 → 9225 (−477), ceiling 9708
.claude/agents/os-dev.md	42743 → 41196 (−1547)	469 → 450 (−19)	— (line-ratcheted), ceiling 469
scripts/check-skill-frame-sync.mjs	40961 → 43180 (+2219)	960 → 990 (+30)	not ratcheted
scripts/check-skill-frame-freshness.mjs	46344 → 47480 (+1136)	1065 → 1085 (+20)	not ratcheted

Whole published package skills/objectstack-pm-dispatch/** (all markdown): 11540 → 11063 tokens, −477.

Two ratcheted files, both shrink; no ceiling row is edited, in either ratchet. The measured token saving is −477, not the −630 the card estimated — the difference is the stop trigger, kept for the reason in section 2. The script growth (+50 lines across two gates) is entirely the derivation constants, the new regression case and the comments recording the measured limit; neither script is under a ratchet.

5. Gates

Union derived at the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands:

dispatch-gates: gate list derived from the tree of 'objectstack-ai/objectstack' at commit ae616fbc91 (/home/user/objectstack-14685-item4).
dispatch-gates: change set derived from git — 4 path(s) vs merge base 5ff5f9576 of 'origin/main' and HEAD

Every exit code captured by redirect before any pipe.

All 32 re-run at the final head ae616fbc91, after the last commit. 31 green, 1 NOT MEASURED, 0 red. The load-bearing ones, quoted from their own verdict lines:

command	exit	its own verdict line
pnpm check:skill-frame-sync	0	✓ check-skill-frame-sync self-test: 13 cases pass, plus 5 dispatch-gates declaration cases. / ✓ check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
node scripts/check-skill-frame-freshness.mjs	0	✓ check-skill-frame-freshness: the decision frame in this tree is current with origin/main (fetched just now). (12 self-test cases pass)
node scripts/check-skills-token-ratchet.mjs	0	✓ check-skills-token-ratchet: skills/objectstack-pm-dispatch/SKILL.md is 9225 tokens (ceiling 9708; headroom 483).
pnpm check:pm-skill-ratchet	0	✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 450 lines (ceiling 469; headroom 19).
node scripts/check-self-test-wired.mjs	0	✓ check-self-test-wired: every one of the 164 script(s) CI runs that ship a --self-test has that self-test run by CI.
pnpm check:pm-dispatch-gates	0	✓ dispatch-gates self-test: 1241 cases pass.
pnpm --filter @objectstack/spec run check:skill-docs	0	✓ skills/README.md / ✓ content/docs/ai/skills-reference.mdx — nothing to regenerate
pnpm check:ratchet-remedy-authority	0	green
pnpm check:pm-skill-id-lint	0	green
pnpm check:nul-bytes	0	green (plus a direct grep -naP for control bytes over both edited markdown files: no match)
pnpm check:doc-authoring, check:role-word, check:agent-model-declared, check:agent-test-spelling, check:corpus-claim-drift, check:skill-compatibility, check:skill-identifier-liveness, check:watch-hint-literal, check:entry-guard, check:parse-guard, check:cli-command-ids, check:bash32-floor, check:pnpm-filter-targets, check:pm-governed-merges, check:cross-package-test-inputs	0	green
node scripts/check-ci-filter-parity.mjs, check-cross-package-test-inputs.mjs, check-shard-attestation.mjs, check-whole-set-label-write.mjs, scripts/pm/bare-root-worklist.mjs --self-test	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	first run answered PREREQUISITE NOT MET (exit 3, @objectstack/formula then @objectstack/lint unbuilt). Prerequisite supplied via the shared verify lock (turbo run build --filter=@objectstack/formula, then --filter=@objectstack/lint) and re-run to a real verdict: ✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 426 files / 1365 TS blocks judged clean
node scripts/check-test-completeness.mjs	3	NOT MEASURED — see below

One red was found and fixed rather than reported: on the first pass check:pm-skill-ratchet said ✗ check-skill-line-ratchet: .claude/agents/os-dev.md has 1 line(s) over the 120-byte budget: L400 (121B). The ratchet prices a per-line byte budget as well as a line ceiling. The replacement sentence was rewrapped at legal CJK break points, content byte-identical, and the whole union was re-derived and re-run at the resulting head — which is the head every number above comes from.

NOT MEASURED, with reasons:

• node scripts/check-test-completeness.mjs — exit 3, its own text: "PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named … the local reading for this gate is NOT MEASURED. ⛔ It is not a red." CI tees the log and passes it; unreachable there.
• The changeset-derived families (9 of them) do not apply: this PR is skip-changeset — .claude/**, skills/** and scripts/** publish nothing from any package.
• The CI-measured-only family reads the workflow event payload; no local run can produce a verdict for it.
• The always-runs tail (workflows with no path filter) is CI's, not measured here.

6. Out-of-scope findings — NOT fixed here, for the seat to triage

Each is a claim that this change makes false, in a file this flight does not hold.

• F1scripts/pm/dispatch-gates.mjs:8090 — why: '… the dev-agent definition is protocol semantics — every dispatched dev runs under it, and it carries an enforced copy of the decision frame'. The last clause is now false; the mandate itself still holds on the first clause, and the self-test still passes (it asserts the glob exists, not the reason).
• F2scripts/pm/dispatch-gates.mjs:8095 — why: '… the published PM skill carries two enforced copies of the decision frame (check:skill-frame-sync COPIES) …'. One copy now.
• F3scripts/pm/check-skill-line-ratchet.mjs:455 — // This file is one of the four check:skill-frame-sync COPIES, so a rule that …, in the os-dev.md ceiling ledger. That file is no longer a COPIES entry, which also weakens the stated reason for the +3 raise recorded there. Item 5 holds this file.
• F4scripts/check-skills-token-ratchet.mjs:334, 345, 359 — three ledger comments stating the published SKILL.md "carries TWO enforced frame copies" and that "the two gate-pinned copies of the decision frame stay in this file". Historically true; now stale as a present-tense claim.
• F5 the internal lane has no {decision_frame} slot. os-dev.md now expects the frame to arrive in the dispatch prompt, but .claude/skills/pm-dispatch/references/dispatch-runbook.md (派发词构造细则) does not tell the internal PM to paste it, and .claude/skills/pm-dispatch/SKILL.md is explicitly outside this flight. Until that lands, the internal dev's only protection is the stop-and-ask clause added in section 2. This is the one finding with live consequences.
• F6 the removed os-dev.md block carried a verbatim maintainer quotation — the 2026-09-01 axis-weighting ruling — so this PR deletes a quotation site that item 2's quotation-policy flight may have surveyed. The ruling text survives at .claude/skills/pm-dispatch/SKILL.md:904 and .claude/skills/pm-dispatch/references/decision-analysis.md:50, and the RULE itself survives in both published and internal PM copies of the frame.

7. Posture

Draft, governed surface. No reviewers requested, no ready flip, no enqueue, no auto-merge. Labels skip-changeset and needs:contract-review.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Part of #14685
Part of #13597

Item 4 of decision batch 3, ruled B. From the ruling comment (5520452691), verbatim:

B — the four-axis decision frame keeps two copies (internal-pm and published-pm); the dev-side copies are replaced by a {decision_frame} placeholder the dispatching PM fills at dispatch time; check:skill-frame-sync watches two sites. Confidence gap carried: whoever executes verifies that the published-dev block has no reader other than the placeholder path before removing it.

---

1. The confidence gap, discharged BEFORE any edit

The ruling made removal conditional on a reader census. Run at the merge base 5ff5f9576, over .claude/ skills/ scripts/ .github/ packages/ docs/, for four things: the published-dev start marker (Analyze every option on ... fixed axes:), the literal published-dev, {decision_frame}, and the internal-dev start marker. Every hit, and what it is:

site	what it is	depends on the published-dev BLOCK?
skills/objectstack-pm-dispatch/SKILL.md:633-662	the block itself	it IS the block
skills/objectstack-pm-dispatch/SKILL.md:623	"with the block below in place of its {decision_frame} line"	YES — the placeholder path, and the only one
skills/objectstack-pm-dispatch/rules/dev-template.md:99	the {decision_frame} line the instruction above fills	via that instruction only
scripts/check-skill-frame-sync.mjs:136-143	the published-dev COPIES entry	the gate that watches it
scripts/check-skill-frame-sync.mjs:714,719,723,724,727	self-test fixtures addressing the published-dev id	the gate's own fixtures
scripts/check-skill-frame-freshness.mjs	imports COPIES / AXIS_MAP wholesale; five fixtures spelled .claude/agents/os-dev.md	reads the TABLE, never the block text
scripts/pm/dispatch-gates.mjs:15382-15397	spawned import reading the FILE list out of COPIES for the tier mandate	reads FILES, not copies
scripts/pm/check-dispatch-gates.mjs:202	FRAME_MODULE watch-hint constant	names the module path only
scripts/check-skills-token-ratchet.mjs:334,345,359	ceiling-ledger prose	comments, no behaviour
scripts/pm/check-skill-line-ratchet.mjs:455	ceiling-ledger prose	comments, no behaviour

No test, generator, workflow, doc or fixture quotes the block by fingerprint. The gap resolves in favour of removal: the placeholder path and the two frame gates are the whole reader set, and both gates read the COPIES table rather than the prose.

The same census over the internal-dev start marker found exactly one non-gate hit — the block itself at .claude/agents/os-dev.md:400.

2. What changed, site by site

skills/objectstack-pm-dispatch/SKILL.md — the embedded dev template loses the frame copy: the declaring sentence through the binding sentence (old :633-662), which is precisely the span the gate's start and binding anchors delimit.

The stop trigger above it (old :627-632, "When to STOP instead of coding …") stays, and that is a judgement call worth naming. It is not part of the frame — the gate's anchors do not reach it — and it is the escalation TRIGGER rather than the analysis. rules/dev-template.md carries it nowhere else (grepped: the template mentions needs_decision only at :34 inside the contract-first rule and at :112 in the report schema), so deleting it with the frame would have removed the dev's instruction to stop at all, which the ruling did not ask for.

The paste instruction now names the surviving copy by its heading:

Paste rules/dev-template.md verbatim into every dispatch prompt; the PM fills
its {…} placeholders. {decision_frame} takes the stop trigger below, followed
by the whole `#### The four-axis decision frame (binding)` section above: the
frame has ONE copy in this skill and the PM pastes it at dispatch time, so no
dev-side copy can drift from it.

.claude/agents/os-dev.md — the frame block (old :400-421, declaring sentence through the binding sentence, which ran into the same paragraph as the unrelated blocked rule) collapses to one sentence. The blocked rule is preserved byte-for-byte. Neither :232 nor :303 is touched (PR #14808 holds them); the diff is one hunk at :400.

Three judgement calls in that one sentence:

1. It says the frame comes from the PM's own copy, not specifically from rules/dev-template.md, because that template is the PUBLISHED skill's artifact. The internal PM composes dispatch prompts per .claude/skills/pm-dispatch/references/dispatch-runbook.md and has no {decision_frame} slot (grep: the token exists only under skills/objectstack-pm-dispatch/). The published slot is named parenthetically, as the brief asked, without claiming the internal lane uses it. See finding F5.
2. It adds "if the dispatch prompt did not carry it, stop and ask; do not invent your own axes." os-dev.md's own doctrine is that standard clauses live in that file and prompt silence is normal — so moving the frame OUT of it opens a hole that a silent prompt would otherwise walk through.
3. It is written in Chinese and deliberately avoids the phrases fixed axes, 固定评估轴, 条评估轴 and any WORD-axis decision frame form. Those are FINGERPRINTS in the sync gate's anti-dormancy scan, whose population is per file: os-dev.md stops being a declared file with this change, so any of those phrasings would have turned the gate red. Verified after the edit — grep -nP for all four patterns over os-dev.md returns nothing.

scripts/check-skill-frame-sync.mjs — COPIES 4 → 2.

- internal-pm .claude/skills/pm-dispatch/SKILL.md KEPT
- internal-dev .claude/agents/os-dev.md REMOVED
- published-pm skills/objectstack-pm-dispatch/SKILL.md KEPT
- published-dev skills/objectstack-pm-dispatch/SKILL.md REMOVED

Every spelled count is now derived: the header ("FOUR hand-written copies across three files"), the COPIES docblock, the axisEntryStarts docblock, the freshness-gate cross-reference, and the count-disagreement message (the four copies do not agree becomes the ${results.length} copies do not agree). report() and the success line already derived theirs from COPIES.length.

The same-file overlap guard is kept even though it is vacuous today — no two copies share a file any more. Re-declaring such a pair is one edit away, and that guard is what would make such a pair prove it compares two different sections rather than the same one twice.

scripts/check-skill-frame-freshness.mjs — it imported COPIES but its self-test SPELLED .claude/agents/os-dev.md in five fixtures (cases 1, 2, 6, 7 and the two withUnreadableCopy calls in 8 and 9). Those broke on the reduction, exactly as that file's own #8024 rule predicts. They now derive SAMPLE_COPY_ID / SAMPLE_FRAME_FILE from COPIES[0], so the next change to the table re-points the fixtures instead of stranding them.

3. The anti-dormancy scan still refuses a returning copy — measured, both directions

Committed first, then mutated, with an EXIT INT TERM trap restoring absolute paths; each mutation proved on disk by counting the injected marker before reading any verdict; each restore proved by git diff HEAD empty AND by comparing git hash-object against the path's HEAD blob.

Direction A — the removed text returns to the now-undeclared .claude/agents/os-dev.md. REFUSED. Marker occurrences after mutation: 1 (was 0). node scripts/check-skill-frame-sync.mjs exit 1:

✗ check-skill-frame-sync: 1 problem(s).
• .claude/agents/os-dev.md
looks like ANOTHER copy of the decision frame (matched /fixed axes/) but is
not declared in COPIES, so nothing keeps it in sync.

Restore: git diff HEAD empty; worktree blob 968d4fd10a653f013635a8a48bdee25a25efc648 equals the HEAD blob.

Direction B — the same text returns to the still-declared skills/objectstack-pm-dispatch/SKILL.md. NOT REFUSED (exit 0). This is a real, measured limit and it is not new: the scan skips any file present in declaredFiles, so a second undeclared copy inside an already-declared file has always been invisible to it. Restore proved the same way (blob 3cbbed0f501736a0db3a5ecdd9e6cda43d44b434).

What that limit means for this PR, stated rather than implied: what keeps a second published copy from being written is the paste instruction naming the surviving section by heading, plus review — not the scan. Closing it mechanically needs a per-occurrence criterion the gate does not have (the published file legitimately carries two more fixed axes fingerprints, at :534 and in the :539 heading). A comment next to the new self-test case says exactly this, so the next reader does not mistake direction B for a bug introduced here.

A self-test case pins direction A against the REAL path this change undeclared, rather than a synthetic one.

4. Counts (merge base 5ff5f9576 → head)

file	bytes	lines	tokens (ceil bytes/4)
skills/objectstack-pm-dispatch/SKILL.md	38807 → 36898 (−1909)	753 → 726 (−27)	9702 → 9225 (−477), ceiling 9708
.claude/agents/os-dev.md	42743 → 41196 (−1547)	469 → 450 (−19)	— (line-ratcheted), ceiling 469
scripts/check-skill-frame-sync.mjs	40961 → 43180 (+2219)	960 → 990 (+30)	not ratcheted
scripts/check-skill-frame-freshness.mjs	46344 → 47480 (+1136)	1065 → 1085 (+20)	not ratcheted

Whole published package skills/objectstack-pm-dispatch/** (all markdown): 11540 → 11063 tokens, −477.

Two ratcheted files, both shrink; no ceiling row is edited, in either ratchet. The measured token saving is −477, not the −630 the card estimated — the difference is the stop trigger, kept for the reason in section 2. The script growth (+50 lines across two gates) is entirely the derivation constants, the new regression case and the comments recording the measured limit; neither script is under a ratchet.

5. Gates

Union derived at the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands:

dispatch-gates: gate list derived from the tree of 'objectstack-ai/objectstack' at commit ae616fbc91 (/home/user/objectstack-14685-item4).
dispatch-gates: change set derived from git — 4 path(s) vs merge base 5ff5f9576 of 'origin/main' and HEAD

Every exit code captured by redirect before any pipe.

All 32 re-run at the final head ae616fbc91, after the last commit. 31 green, 1 NOT MEASURED, 0 red. The load-bearing ones, quoted from their own verdict lines:

command	exit	its own verdict line
pnpm check:skill-frame-sync	0	✓ check-skill-frame-sync self-test: 13 cases pass, plus 5 dispatch-gates declaration cases. / ✓ check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
node scripts/check-skill-frame-freshness.mjs	0	✓ check-skill-frame-freshness: the decision frame in this tree is current with origin/main (fetched just now). (12 self-test cases pass)
node scripts/check-skills-token-ratchet.mjs	0	✓ check-skills-token-ratchet: skills/objectstack-pm-dispatch/SKILL.md is 9225 tokens (ceiling 9708; headroom 483).
pnpm check:pm-skill-ratchet	0	✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 450 lines (ceiling 469; headroom 19).
node scripts/check-self-test-wired.mjs	0	✓ check-self-test-wired: every one of the 164 script(s) CI runs that ship a --self-test has that self-test run by CI.
pnpm check:pm-dispatch-gates	0	✓ dispatch-gates self-test: 1241 cases pass.
pnpm --filter @objectstack/spec run check:skill-docs	0	✓ skills/README.md / ✓ content/docs/ai/skills-reference.mdx — nothing to regenerate
pnpm check:ratchet-remedy-authority	0	green
pnpm check:pm-skill-id-lint	0	green
pnpm check:nul-bytes	0	green (plus a direct grep -naP for control bytes over both edited markdown files: no match)
pnpm check:doc-authoring, check:role-word, check:agent-model-declared, check:agent-test-spelling, check:corpus-claim-drift, check:skill-compatibility, check:skill-identifier-liveness, check:watch-hint-literal, check:entry-guard, check:parse-guard, check:cli-command-ids, check:bash32-floor, check:pnpm-filter-targets, check:pm-governed-merges, check:cross-package-test-inputs	0	green
node scripts/check-ci-filter-parity.mjs, check-cross-package-test-inputs.mjs, check-shard-attestation.mjs, check-whole-set-label-write.mjs, scripts/pm/bare-root-worklist.mjs --self-test	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	first run answered PREREQUISITE NOT MET (exit 3, @objectstack/formula then @objectstack/lint unbuilt). Prerequisite supplied via the shared verify lock (turbo run build --filter=@objectstack/formula, then --filter=@objectstack/lint) and re-run to a real verdict: ✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 426 files / 1365 TS blocks judged clean
node scripts/check-test-completeness.mjs	3	NOT MEASURED — see below

One red was found and fixed rather than reported: on the first pass check:pm-skill-ratchet said ✗ check-skill-line-ratchet: .claude/agents/os-dev.md has 1 line(s) over the 120-byte budget: L400 (121B). The ratchet prices a per-line byte budget as well as a line ceiling. The replacement sentence was rewrapped at legal CJK break points, content byte-identical, and the whole union was re-derived and re-run at the resulting head — which is the head every number above comes from.

NOT MEASURED, with reasons:

• node scripts/check-test-completeness.mjs — exit 3, its own text: "PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named … the local reading for this gate is NOT MEASURED. ⛔ It is not a red." CI tees the log and passes it; unreachable there.
• The changeset-derived families (9 of them) do not apply: this PR is skip-changeset — .claude/**, skills/** and scripts/** publish nothing from any package.
• The CI-measured-only family reads the workflow event payload; no local run can produce a verdict for it.
• The always-runs tail (workflows with no path filter) is CI's, not measured here.

6. Out-of-scope findings — NOT fixed here, for the seat to triage

Each is a claim that this change makes false, in a file this flight does not hold.

• F1scripts/pm/dispatch-gates.mjs:8090 — why: '… the dev-agent definition is protocol semantics — every dispatched dev runs under it, and it carries an enforced copy of the decision frame'. The last clause is now false; the mandate itself still holds on the first clause, and the self-test still passes (it asserts the glob exists, not the reason).
• F2scripts/pm/dispatch-gates.mjs:8095 — why: '… the published PM skill carries two enforced copies of the decision frame (check:skill-frame-sync COPIES) …'. One copy now.
• F3scripts/pm/check-skill-line-ratchet.mjs:455 — // This file is one of the four check:skill-frame-sync COPIES, so a rule that …, in the os-dev.md ceiling ledger. That file is no longer a COPIES entry, which also weakens the stated reason for the +3 raise recorded there. Item 5 holds this file.
• F4scripts/check-skills-token-ratchet.mjs:334, 345, 359 — three ledger comments stating the published SKILL.md "carries TWO enforced frame copies" and that "the two gate-pinned copies of the decision frame stay in this file". Historically true; now stale as a present-tense claim.
• F5 the internal lane has no {decision_frame} slot. os-dev.md now expects the frame to arrive in the dispatch prompt, but .claude/skills/pm-dispatch/references/dispatch-runbook.md (派发词构造细则) does not tell the internal PM to paste it, and .claude/skills/pm-dispatch/SKILL.md is explicitly outside this flight. Until that lands, the internal dev's only protection is the stop-and-ask clause added in section 2. This is the one finding with live consequences.
• F6 the removed os-dev.md block carried a verbatim maintainer quotation — the 2026-09-01 axis-weighting ruling — so this PR deletes a quotation site that item 2's quotation-policy flight may have surveyed. The ruling text survives at .claude/skills/pm-dispatch/SKILL.md:904 and .claude/skills/pm-dispatch/references/decision-analysis.md:50, and the RULE itself survives in both published and internal PM copies of the frame.

7. Posture

Draft, governed surface. No reviewers requested, no ready flip, no enqueue, no auto-merge. Labels skip-changeset and needs:contract-review.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code

[os-litant]

Part of #14685
Part of #13597

Item 4 of decision batch 3, ruled B. From the ruling comment (5520452691), verbatim:

B — the four-axis decision frame keeps two copies (internal-pm and published-pm); the dev-side copies are replaced by a {decision_frame} placeholder the dispatching PM fills at dispatch time; check:skill-frame-sync watches two sites. Confidence gap carried: whoever executes verifies that the published-dev block has no reader other than the placeholder path before removing it.

---

1. The confidence gap, discharged BEFORE any edit

The ruling made removal conditional on a reader census. Run at the merge base 5ff5f9576, over .claude/ skills/ scripts/ .github/ packages/ docs/, for four things: the published-dev start marker (Analyze every option on ... fixed axes:), the literal published-dev, {decision_frame}, and the internal-dev start marker. Every hit, and what it is:

site	what it is	depends on the published-dev BLOCK?
skills/objectstack-pm-dispatch/SKILL.md:633-662	the block itself	it IS the block
skills/objectstack-pm-dispatch/SKILL.md:623	"with the block below in place of its {decision_frame} line"	YES — the placeholder path, and the only one
skills/objectstack-pm-dispatch/rules/dev-template.md:99	the {decision_frame} line the instruction above fills	via that instruction only
scripts/check-skill-frame-sync.mjs:136-143	the published-dev COPIES entry	the gate that watches it
scripts/check-skill-frame-sync.mjs:714,719,723,724,727	self-test fixtures addressing the published-dev id	the gate's own fixtures
scripts/check-skill-frame-freshness.mjs	imports COPIES / AXIS_MAP wholesale; five fixtures spelled .claude/agents/os-dev.md	reads the TABLE, never the block text
scripts/pm/dispatch-gates.mjs:15382-15397	spawned import reading the FILE list out of COPIES for the tier mandate	reads FILES, not copies
scripts/pm/check-dispatch-gates.mjs:202	FRAME_MODULE watch-hint constant	names the module path only
scripts/check-skills-token-ratchet.mjs:334,345,359	ceiling-ledger prose	comments, no behaviour
scripts/pm/check-skill-line-ratchet.mjs:455	ceiling-ledger prose	comments, no behaviour

No test, generator, workflow, doc or fixture quotes the block by fingerprint. The gap resolves in favour of removal: the placeholder path and the two frame gates are the whole reader set, and both gates read the COPIES table rather than the prose.

The same census over the internal-dev start marker found exactly one non-gate hit — the block itself at .claude/agents/os-dev.md:400.

2. What changed, site by site

skills/objectstack-pm-dispatch/SKILL.md — the embedded dev template loses the frame copy: the declaring sentence through the binding sentence (old :633-662), which is precisely the span the gate's start and binding anchors delimit.

The stop trigger above it (old :627-632, "When to STOP instead of coding …") stays, and that is a judgement call worth naming. It is not part of the frame — the gate's anchors do not reach it — and it is the escalation TRIGGER rather than the analysis. rules/dev-template.md carries it nowhere else (grepped: the template mentions needs_decision only at :34 inside the contract-first rule and at :112 in the report schema), so deleting it with the frame would have removed the dev's instruction to stop at all, which the ruling did not ask for.

The paste instruction now names the surviving copy by its heading:

Paste rules/dev-template.md verbatim into every dispatch prompt; the PM fills
its {…} placeholders. {decision_frame} takes the stop trigger below, followed
by the whole `#### The four-axis decision frame (binding)` section above: the
frame has ONE copy in this skill and the PM pastes it at dispatch time, so no
dev-side copy can drift from it.

.claude/agents/os-dev.md — the frame block (old :400-421, declaring sentence through the binding sentence, which ran into the same paragraph as the unrelated blocked rule) collapses to one sentence. The blocked rule is preserved byte-for-byte. Neither :232 nor :303 is touched (PR #14808 holds them); the diff is one hunk at :400.

Three judgement calls in that one sentence:

1. It says the frame comes from the PM's own copy, not specifically from rules/dev-template.md, because that template is the PUBLISHED skill's artifact. The internal PM composes dispatch prompts per .claude/skills/pm-dispatch/references/dispatch-runbook.md and has no {decision_frame} slot (grep: the token exists only under skills/objectstack-pm-dispatch/). The published slot is named parenthetically, as the brief asked, without claiming the internal lane uses it. See finding F5.
2. It adds "if the dispatch prompt did not carry it, stop and ask; do not invent your own axes." os-dev.md's own doctrine is that standard clauses live in that file and prompt silence is normal — so moving the frame OUT of it opens a hole that a silent prompt would otherwise walk through.
3. It is written in Chinese and deliberately avoids the phrases fixed axes, 固定评估轴, 条评估轴 and any WORD-axis decision frame form. Those are FINGERPRINTS in the sync gate's anti-dormancy scan, whose population is per file: os-dev.md stops being a declared file with this change, so any of those phrasings would have turned the gate red. Verified after the edit — grep -nP for all four patterns over os-dev.md returns nothing.

scripts/check-skill-frame-sync.mjs — COPIES 4 → 2.

- internal-pm .claude/skills/pm-dispatch/SKILL.md KEPT
- internal-dev .claude/agents/os-dev.md REMOVED
- published-pm skills/objectstack-pm-dispatch/SKILL.md KEPT
- published-dev skills/objectstack-pm-dispatch/SKILL.md REMOVED

Every spelled count is now derived: the header ("FOUR hand-written copies across three files"), the COPIES docblock, the axisEntryStarts docblock, the freshness-gate cross-reference, and the count-disagreement message (the four copies do not agree becomes the ${results.length} copies do not agree). report() and the success line already derived theirs from COPIES.length.

The same-file overlap guard is kept even though it is vacuous today — no two copies share a file any more. Re-declaring such a pair is one edit away, and that guard is what would make such a pair prove it compares two different sections rather than the same one twice.

scripts/check-skill-frame-freshness.mjs — it imported COPIES but its self-test SPELLED .claude/agents/os-dev.md in five fixtures (cases 1, 2, 6, 7 and the two withUnreadableCopy calls in 8 and 9). Those broke on the reduction, exactly as that file's own #8024 rule predicts. They now derive SAMPLE_COPY_ID / SAMPLE_FRAME_FILE from COPIES[0], so the next change to the table re-points the fixtures instead of stranding them.

3. The anti-dormancy scan still refuses a returning copy — measured, both directions

Committed first, then mutated, with an EXIT INT TERM trap restoring absolute paths; each mutation proved on disk by counting the injected marker before reading any verdict; each restore proved by git diff HEAD empty AND by comparing git hash-object against the path's HEAD blob.

Direction A — the removed text returns to the now-undeclared .claude/agents/os-dev.md. REFUSED. Marker occurrences after mutation: 1 (was 0). node scripts/check-skill-frame-sync.mjs exit 1:

✗ check-skill-frame-sync: 1 problem(s).
• .claude/agents/os-dev.md
looks like ANOTHER copy of the decision frame (matched /fixed axes/) but is
not declared in COPIES, so nothing keeps it in sync.

Restore: git diff HEAD empty; worktree blob 968d4fd10a653f013635a8a48bdee25a25efc648 equals the HEAD blob.

Direction B — the same text returns to the still-declared skills/objectstack-pm-dispatch/SKILL.md. NOT REFUSED (exit 0). This is a real, measured limit and it is not new: the scan skips any file present in declaredFiles, so a second undeclared copy inside an already-declared file has always been invisible to it. Restore proved the same way (blob 3cbbed0f501736a0db3a5ecdd9e6cda43d44b434).

What that limit means for this PR, stated rather than implied: what keeps a second published copy from being written is the paste instruction naming the surviving section by heading, plus review — not the scan. Closing it mechanically needs a per-occurrence criterion the gate does not have (the published file legitimately carries two more fixed axes fingerprints, at :534 and in the :539 heading). A comment next to the new self-test case says exactly this, so the next reader does not mistake direction B for a bug introduced here.

A self-test case pins direction A against the REAL path this change undeclared, rather than a synthetic one.

4. Counts (merge base 5ff5f9576 → head)

file	bytes	lines	tokens (ceil bytes/4)
skills/objectstack-pm-dispatch/SKILL.md	38807 → 36898 (−1909)	753 → 726 (−27)	9702 → 9225 (−477), ceiling 9708
.claude/agents/os-dev.md	42743 → 41196 (−1547)	469 → 450 (−19)	— (line-ratcheted), ceiling 469
scripts/check-skill-frame-sync.mjs	40961 → 43180 (+2219)	960 → 990 (+30)	not ratcheted
scripts/check-skill-frame-freshness.mjs	46344 → 47480 (+1136)	1065 → 1085 (+20)	not ratcheted

Whole published package skills/objectstack-pm-dispatch/** (all markdown): 11540 → 11063 tokens, −477.

Two ratcheted files, both shrink; no ceiling row is edited, in either ratchet. The measured token saving is −477, not the −630 the card estimated — the difference is the stop trigger, kept for the reason in section 2. The script growth (+50 lines across two gates) is entirely the derivation constants, the new regression case and the comments recording the measured limit; neither script is under a ratchet.

5. Gates

Union derived at the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack --commands:

dispatch-gates: gate list derived from the tree of 'objectstack-ai/objectstack' at commit ae616fbc91 (/home/user/objectstack-14685-item4).
dispatch-gates: change set derived from git — 4 path(s) vs merge base 5ff5f9576 of 'origin/main' and HEAD

Every exit code captured by redirect before any pipe.

All 32 re-run at the final head ae616fbc91, after the last commit. 31 green, 1 NOT MEASURED, 0 red. The load-bearing ones, quoted from their own verdict lines:

command	exit	its own verdict line
pnpm check:skill-frame-sync	0	✓ check-skill-frame-sync self-test: 13 cases pass, plus 5 dispatch-gates declaration cases. / ✓ check-skill-frame-sync: 2 copies of the decision frame are structurally isomorphic across 2 files
node scripts/check-skill-frame-freshness.mjs	0	✓ check-skill-frame-freshness: the decision frame in this tree is current with origin/main (fetched just now). (12 self-test cases pass)
node scripts/check-skills-token-ratchet.mjs	0	✓ check-skills-token-ratchet: skills/objectstack-pm-dispatch/SKILL.md is 9225 tokens (ceiling 9708; headroom 483).
pnpm check:pm-skill-ratchet	0	✓ check-skill-line-ratchet: .claude/agents/os-dev.md is 450 lines (ceiling 469; headroom 19).
node scripts/check-self-test-wired.mjs	0	✓ check-self-test-wired: every one of the 164 script(s) CI runs that ship a --self-test has that self-test run by CI.
pnpm check:pm-dispatch-gates	0	✓ dispatch-gates self-test: 1241 cases pass.
pnpm --filter @objectstack/spec run check:skill-docs	0	✓ skills/README.md / ✓ content/docs/ai/skills-reference.mdx — nothing to regenerate
pnpm check:ratchet-remedy-authority	0	green
pnpm check:pm-skill-id-lint	0	green
pnpm check:nul-bytes	0	green (plus a direct grep -naP for control bytes over both edited markdown files: no match)
pnpm check:doc-authoring, check:role-word, check:agent-model-declared, check:agent-test-spelling, check:corpus-claim-drift, check:skill-compatibility, check:skill-identifier-liveness, check:watch-hint-literal, check:entry-guard, check:parse-guard, check:cli-command-ids, check:bash32-floor, check:pnpm-filter-targets, check:pm-governed-merges, check:cross-package-test-inputs	0	green
node scripts/check-ci-filter-parity.mjs, check-cross-package-test-inputs.mjs, check-shard-attestation.mjs, check-whole-set-label-write.mjs, scripts/pm/bare-root-worklist.mjs --self-test	0	green
pnpm --filter @objectstack/lint run check:doc-formula-expressions	0	first run answered PREREQUISITE NOT MET (exit 3, @objectstack/formula then @objectstack/lint unbuilt). Prerequisite supplied via the shared verify lock (turbo run build --filter=@objectstack/formula, then --filter=@objectstack/lint) and re-run to a real verdict: ✓ check:doc-formula-expressions: 22 record-scoped formula example(s) across 426 files / 1365 TS blocks judged clean
node scripts/check-test-completeness.mjs	3	NOT MEASURED — see below

One red was found and fixed rather than reported: on the first pass check:pm-skill-ratchet said ✗ check-skill-line-ratchet: .claude/agents/os-dev.md has 1 line(s) over the 120-byte budget: L400 (121B). The ratchet prices a per-line byte budget as well as a line ceiling. The replacement sentence was rewrapped at legal CJK break points, content byte-identical, and the whole union was re-derived and re-run at the resulting head — which is the head every number above comes from.

NOT MEASURED, with reasons:

• node scripts/check-test-completeness.mjs — exit 3, its own text: "PREREQUISITE NOT MET — this gate grades a saved turbo run test log, and no log was named … the local reading for this gate is NOT MEASURED. ⛔ It is not a red." CI tees the log and passes it; unreachable there.
• The changeset-derived families (9 of them) do not apply: this PR is skip-changeset — .claude/**, skills/** and scripts/** publish nothing from any package.
• The CI-measured-only family reads the workflow event payload; no local run can produce a verdict for it.
• The always-runs tail (workflows with no path filter) is CI's, not measured here.

6. Out-of-scope findings — NOT fixed here, for the seat to triage

Each is a claim that this change makes false, in a file this flight does not hold.

• F1scripts/pm/dispatch-gates.mjs:8090 — why: '… the dev-agent definition is protocol semantics — every dispatched dev runs under it, and it carries an enforced copy of the decision frame'. The last clause is now false; the mandate itself still holds on the first clause, and the self-test still passes (it asserts the glob exists, not the reason).
• F2scripts/pm/dispatch-gates.mjs:8095 — why: '… the published PM skill carries two enforced copies of the decision frame (check:skill-frame-sync COPIES) …'. One copy now.
• F3scripts/pm/check-skill-line-ratchet.mjs:455 — // This file is one of the four check:skill-frame-sync COPIES, so a rule that …, in the os-dev.md ceiling ledger. That file is no longer a COPIES entry, which also weakens the stated reason for the +3 raise recorded there. Item 5 holds this file.
• F4scripts/check-skills-token-ratchet.mjs:334, 345, 359 — three ledger comments stating the published SKILL.md "carries TWO enforced frame copies" and that "the two gate-pinned copies of the decision frame stay in this file". Historically true; now stale as a present-tense claim.
• F5 the internal lane has no {decision_frame} slot. os-dev.md now expects the frame to arrive in the dispatch prompt, but .claude/skills/pm-dispatch/references/dispatch-runbook.md (派发词构造细则) does not tell the internal PM to paste it, and .claude/skills/pm-dispatch/SKILL.md is explicitly outside this flight. Until that lands, the internal dev's only protection is the stop-and-ask clause added in section 2. This is the one finding with live consequences.
• F6 the removed os-dev.md block carried a verbatim maintainer quotation — the 2026-09-01 axis-weighting ruling — so this PR deletes a quotation site that item 2's quotation-policy flight may have surveyed. The ruling text survives at .claude/skills/pm-dispatch/SKILL.md:904 and .claude/skills/pm-dispatch/references/decision-analysis.md:50, and the RULE itself survives in both published and internal PM copies of the frame.

7. Posture

Draft, governed surface. No reviewers requested, no ready flip, no enqueue, no auto-merge. Labels skip-changeset and needs:contract-review.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LraLgQVGq8egUwfYZpbYt1

---

Generated by Claude Code