docs(agents): harden the three internal worktree recipes to the canonical form

[claude]

Fixes#13729

The three INTERNAL copies that PRESCRIBE a worktree recipe to an agent still taught the unhardened spelling. Each now reads as the form AGENTS.md Prime Directive 11 and CLAUDE.md already teach:

git fetch origin main && git worktree add --no-track ../REPO-TASK -b BRANCH origin/main

Exactly one recipe sentence per file; nothing else in these files is touched.

What changed, per copy

file	before	after
.claude/skills/checklist-author/SKILL.md L33	git worktree add ../objectstack-TASK -b BRANCH main	gains --no-trackand the git fetch origin main + origin/main base
.claude/agents/os-dev.md L32-33	git worktree add ../REPO-issue-N -b claude/issue-N-SLUG origin/main	gains --no-track and the fetch
.claude/skills/dogfood-verification/SKILL.md L140	git worktree add -b BRANCH /tmp/pr origin/main	gains --no-track and the fetch

Placeholder metavariables are spelled in caps above because this body is sanitized; the files themselves keep each copy's own angle-bracket placeholder naming verbatim (../objectstack-TASK stays the checklist-author spelling, claude/issue-N-SLUG stays the os-dev spelling, /tmp/pr stays the dogfood one).

Why

Plain -b writes branch.NAME.remote / branch.NAME.merge into the .git/config that every worktree of the repo SHARES — the class already adjudicated in #13052 / objectstack-ai/objectui#6880 and already hardened in AGENTS.md PD#11 and CLAUDE.md; these copies were never swept with it. The checklist-author copy was additionally the only one still basing a new tree on local main with no fetch — the objectstack-ai/objectui#6208 class.

The published fourth copy (skills/objectstack-pm-dispatch/SKILL.md) is NOT in this diff: it was discharged separately in #13891, which has landed, with its {default_branch} interpolation preserved. Verified on today's origin/main before editing.

Two ratchets bind here, so the edits are reflows rather than rewraps

check:pm-skill-ratchet constrains these files twice, and the obvious spellings break one each:

• Zero line-count headroom. All three files sit exactly at their ceiling (os-dev.md 466/466, checklist-author 62/62, dogfood-verification 157/157). Wrapping the longer recipe onto a new line failed the gate on all three.
• 120-byte per-line budget. Keeping the longer recipe on its single existing line failed the other half — 139B / 133B / 134B respectively.

So each recipe is reflowed within the lines its sentence already occupied: the break moves, the line count does not, and every touched line lands under budget (97/81, 105/106, 92/106 bytes). Neither ratchet moves and no ceiling is raised, so no maintainer ruling is needed for either.

Verification

Gate union derived on the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (no hand-built path list) at f977b0b54 — 10 families, all run at that same head:

• ✅ check:pm-skill-ratchet — .claude/agents/os-dev.md is 466 lines (ceiling 466; headroom 0), checklist-author/SKILL.md is 62 lines (ceiling 62; headroom 0), dogfood-verification/SKILL.md is 157 lines (ceiling 157; headroom 0); widest-table-row pins all unchanged at 0
• ✅ check:agent-model-declared · check:agent-test-spelling · check:doc-authoring · check:nul-bytes · check:pm-governed-merges · check:pm-skill-id-lint · check:skill-frame-sync
• ✅ pnpm --filter @objectstack/lint run check:doc-formula-expressions — self-test: 58 cases passed, 22 record-scoped formula example(s) across 427 files / 1451 TS blocks judged clean (needed @objectstack/formula and @objectstack/lint built first; its own PREREQUISITE NOT MET refusal is not a finding)
• ⚪ node scripts/pm/check-governed-queue-guard.mjs — NOT MEASURED locally, by construction: it reads GITHUB_EVENT_PATH and nothing else, so outside a workflow run it exits 1 with could not read GITHUB_EVENT_PATH. That is its refusal path, not a finding. CI measures it.

pnpm lint (repo-wide ESLint) is a declared narrowing, measured on all three axes: (1) the population comes from ESLint's own resolved config, not a guess — all three files report File ignored because no matching configuration was supplied., i.e. markdown is outside the linted population entirely; (2) the count comes from --format json — 3 files reported, 0 errors; (3) invariance for untouched files holds because the diff contains no eslint.config.mjs, no tsconfig and no TypeScript at all, so no untouched file's verdict can move.

No changeset: the diff is .claude/** only and publishes nothing from any package, so skip-changeset applies.

Scope guard

.claude/agents/os-dev.md carries adjacent text owned by other in-flight cards — the verify-lock and skip-changeset paragraphs of #12800 / #13300 / #13432. Those are deliberately untouched here; this diff is the recipe sentence only. Those three cards are not addressed by this PR and remain open.

Draft on purpose: governed .claude/** markdown, human merge.

---

Generated by Claude Code

[claude]

Fixes#13729

The three INTERNAL copies that PRESCRIBE a worktree recipe to an agent still taught the unhardened spelling. Each now reads as the form AGENTS.md Prime Directive 11 and CLAUDE.md already teach:

git fetch origin main && git worktree add --no-track ../REPO-TASK -b BRANCH origin/main

Exactly one recipe sentence per file; nothing else in these files is touched.

What changed, per copy

file	before	after
.claude/skills/checklist-author/SKILL.md L33	git worktree add ../objectstack-TASK -b BRANCH main	gains --no-trackand the git fetch origin main + origin/main base
.claude/agents/os-dev.md L32-33	git worktree add ../REPO-issue-N -b claude/issue-N-SLUG origin/main	gains --no-track and the fetch
.claude/skills/dogfood-verification/SKILL.md L140	git worktree add -b BRANCH /tmp/pr origin/main	gains --no-track and the fetch

Placeholder metavariables are spelled in caps above because this body is sanitized; the files themselves keep each copy's own angle-bracket placeholder naming verbatim (../objectstack-TASK stays the checklist-author spelling, claude/issue-N-SLUG stays the os-dev spelling, /tmp/pr stays the dogfood one).

Why

Plain -b writes branch.NAME.remote / branch.NAME.merge into the .git/config that every worktree of the repo SHARES — the class already adjudicated in #13052 / objectstack-ai/objectui#6880 and already hardened in AGENTS.md PD#11 and CLAUDE.md; these copies were never swept with it. The checklist-author copy was additionally the only one still basing a new tree on local main with no fetch — the objectstack-ai/objectui#6208 class.

The published fourth copy (skills/objectstack-pm-dispatch/SKILL.md) is NOT in this diff: it was discharged separately in #13891, which has landed, with its {default_branch} interpolation preserved. Verified on today's origin/main before editing.

Two ratchets bind here, so the edits are reflows rather than rewraps

check:pm-skill-ratchet constrains these files twice, and the obvious spellings break one each:

• Zero line-count headroom. All three files sit exactly at their ceiling (os-dev.md 466/466, checklist-author 62/62, dogfood-verification 157/157). Wrapping the longer recipe onto a new line failed the gate on all three.
• 120-byte per-line budget. Keeping the longer recipe on its single existing line failed the other half — 139B / 133B / 134B respectively.

So each recipe is reflowed within the lines its sentence already occupied: the break moves, the line count does not, and every touched line lands under budget (97/81, 105/106, 92/106 bytes). Neither ratchet moves and no ceiling is raised, so no maintainer ruling is needed for either.

Verification

Gate union derived on the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (no hand-built path list) at f977b0b54 — 10 families, all run at that same head:

• ✅ check:pm-skill-ratchet — .claude/agents/os-dev.md is 466 lines (ceiling 466; headroom 0), checklist-author/SKILL.md is 62 lines (ceiling 62; headroom 0), dogfood-verification/SKILL.md is 157 lines (ceiling 157; headroom 0); widest-table-row pins all unchanged at 0
• ✅ check:agent-model-declared · check:agent-test-spelling · check:doc-authoring · check:nul-bytes · check:pm-governed-merges · check:pm-skill-id-lint · check:skill-frame-sync
• ✅ pnpm --filter @objectstack/lint run check:doc-formula-expressions — self-test: 58 cases passed, 22 record-scoped formula example(s) across 427 files / 1451 TS blocks judged clean (needed @objectstack/formula and @objectstack/lint built first; its own PREREQUISITE NOT MET refusal is not a finding)
• ⚪ node scripts/pm/check-governed-queue-guard.mjs — NOT MEASURED locally, by construction: it reads GITHUB_EVENT_PATH and nothing else, so outside a workflow run it exits 1 with could not read GITHUB_EVENT_PATH. That is its refusal path, not a finding. CI measures it.

pnpm lint (repo-wide ESLint) is a declared narrowing, measured on all three axes: (1) the population comes from ESLint's own resolved config, not a guess — all three files report File ignored because no matching configuration was supplied., i.e. markdown is outside the linted population entirely; (2) the count comes from --format json — 3 files reported, 0 errors; (3) invariance for untouched files holds because the diff contains no eslint.config.mjs, no tsconfig and no TypeScript at all, so no untouched file's verdict can move.

No changeset: the diff is .claude/** only and publishes nothing from any package, so skip-changeset applies.

Scope guard

.claude/agents/os-dev.md carries adjacent text owned by other in-flight cards — the verify-lock and skip-changeset paragraphs of #12800 / #13300 / #13432. Those are deliberately untouched here; this diff is the recipe sentence only. Those three cards are not addressed by this PR and remain open.

Draft on purpose: governed .claude/** markdown, human merge.

---

Generated by Claude Code

[claude]

Fixes#13729

The three INTERNAL copies that PRESCRIBE a worktree recipe to an agent still taught the unhardened spelling. Each now reads as the form AGENTS.md Prime Directive 11 and CLAUDE.md already teach:

git fetch origin main && git worktree add --no-track ../REPO-TASK -b BRANCH origin/main

Exactly one recipe sentence per file; nothing else in these files is touched.

What changed, per copy

file	before	after
.claude/skills/checklist-author/SKILL.md L33	git worktree add ../objectstack-TASK -b BRANCH main	gains --no-trackand the git fetch origin main + origin/main base
.claude/agents/os-dev.md L32-33	git worktree add ../REPO-issue-N -b claude/issue-N-SLUG origin/main	gains --no-track and the fetch
.claude/skills/dogfood-verification/SKILL.md L140	git worktree add -b BRANCH /tmp/pr origin/main	gains --no-track and the fetch

Placeholder metavariables are spelled in caps above because this body is sanitized; the files themselves keep each copy's own angle-bracket placeholder naming verbatim (../objectstack-TASK stays the checklist-author spelling, claude/issue-N-SLUG stays the os-dev spelling, /tmp/pr stays the dogfood one).

Why

Plain -b writes branch.NAME.remote / branch.NAME.merge into the .git/config that every worktree of the repo SHARES — the class already adjudicated in #13052 / objectstack-ai/objectui#6880 and already hardened in AGENTS.md PD#11 and CLAUDE.md; these copies were never swept with it. The checklist-author copy was additionally the only one still basing a new tree on local main with no fetch — the objectstack-ai/objectui#6208 class.

The published fourth copy (skills/objectstack-pm-dispatch/SKILL.md) is NOT in this diff: it was discharged separately in #13891, which has landed, with its {default_branch} interpolation preserved. Verified on today's origin/main before editing.

Two ratchets bind here, so the edits are reflows rather than rewraps

check:pm-skill-ratchet constrains these files twice, and the obvious spellings break one each:

• Zero line-count headroom. All three files sit exactly at their ceiling (os-dev.md 466/466, checklist-author 62/62, dogfood-verification 157/157). Wrapping the longer recipe onto a new line failed the gate on all three.
• 120-byte per-line budget. Keeping the longer recipe on its single existing line failed the other half — 139B / 133B / 134B respectively.

So each recipe is reflowed within the lines its sentence already occupied: the break moves, the line count does not, and every touched line lands under budget (97/81, 105/106, 92/106 bytes). Neither ratchet moves and no ceiling is raised, so no maintainer ruling is needed for either.

Verification

Gate union derived on the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (no hand-built path list) at f977b0b54 — 10 families, all run at that same head:

• ✅ check:pm-skill-ratchet — .claude/agents/os-dev.md is 466 lines (ceiling 466; headroom 0), checklist-author/SKILL.md is 62 lines (ceiling 62; headroom 0), dogfood-verification/SKILL.md is 157 lines (ceiling 157; headroom 0); widest-table-row pins all unchanged at 0
• ✅ check:agent-model-declared · check:agent-test-spelling · check:doc-authoring · check:nul-bytes · check:pm-governed-merges · check:pm-skill-id-lint · check:skill-frame-sync
• ✅ pnpm --filter @objectstack/lint run check:doc-formula-expressions — self-test: 58 cases passed, 22 record-scoped formula example(s) across 427 files / 1451 TS blocks judged clean (needed @objectstack/formula and @objectstack/lint built first; its own PREREQUISITE NOT MET refusal is not a finding)
• ⚪ node scripts/pm/check-governed-queue-guard.mjs — NOT MEASURED locally, by construction: it reads GITHUB_EVENT_PATH and nothing else, so outside a workflow run it exits 1 with could not read GITHUB_EVENT_PATH. That is its refusal path, not a finding. CI measures it.

pnpm lint (repo-wide ESLint) is a declared narrowing, measured on all three axes: (1) the population comes from ESLint's own resolved config, not a guess — all three files report File ignored because no matching configuration was supplied., i.e. markdown is outside the linted population entirely; (2) the count comes from --format json — 3 files reported, 0 errors; (3) invariance for untouched files holds because the diff contains no eslint.config.mjs, no tsconfig and no TypeScript at all, so no untouched file's verdict can move.

No changeset: the diff is .claude/** only and publishes nothing from any package, so skip-changeset applies.

Scope guard

.claude/agents/os-dev.md carries adjacent text owned by other in-flight cards — the verify-lock and skip-changeset paragraphs of #12800 / #13300 / #13432. Those are deliberately untouched here; this diff is the recipe sentence only. Those three cards are not addressed by this PR and remain open.

Draft on purpose: governed .claude/** markdown, human merge.

---

Generated by Claude Code

[claude]

Fixes#13729

The three INTERNAL copies that PRESCRIBE a worktree recipe to an agent still taught the unhardened spelling. Each now reads as the form AGENTS.md Prime Directive 11 and CLAUDE.md already teach:

git fetch origin main && git worktree add --no-track ../REPO-TASK -b BRANCH origin/main

Exactly one recipe sentence per file; nothing else in these files is touched.

What changed, per copy

file	before	after
.claude/skills/checklist-author/SKILL.md L33	git worktree add ../objectstack-TASK -b BRANCH main	gains --no-trackand the git fetch origin main + origin/main base
.claude/agents/os-dev.md L32-33	git worktree add ../REPO-issue-N -b claude/issue-N-SLUG origin/main	gains --no-track and the fetch
.claude/skills/dogfood-verification/SKILL.md L140	git worktree add -b BRANCH /tmp/pr origin/main	gains --no-track and the fetch

Placeholder metavariables are spelled in caps above because this body is sanitized; the files themselves keep each copy's own angle-bracket placeholder naming verbatim (../objectstack-TASK stays the checklist-author spelling, claude/issue-N-SLUG stays the os-dev spelling, /tmp/pr stays the dogfood one).

Why

Plain -b writes branch.NAME.remote / branch.NAME.merge into the .git/config that every worktree of the repo SHARES — the class already adjudicated in #13052 / objectstack-ai/objectui#6880 and already hardened in AGENTS.md PD#11 and CLAUDE.md; these copies were never swept with it. The checklist-author copy was additionally the only one still basing a new tree on local main with no fetch — the objectstack-ai/objectui#6208 class.

The published fourth copy (skills/objectstack-pm-dispatch/SKILL.md) is NOT in this diff: it was discharged separately in #13891, which has landed, with its {default_branch} interpolation preserved. Verified on today's origin/main before editing.

Two ratchets bind here, so the edits are reflows rather than rewraps

check:pm-skill-ratchet constrains these files twice, and the obvious spellings break one each:

• Zero line-count headroom. All three files sit exactly at their ceiling (os-dev.md 466/466, checklist-author 62/62, dogfood-verification 157/157). Wrapping the longer recipe onto a new line failed the gate on all three.
• 120-byte per-line budget. Keeping the longer recipe on its single existing line failed the other half — 139B / 133B / 134B respectively.

So each recipe is reflowed within the lines its sentence already occupied: the break moves, the line count does not, and every touched line lands under budget (97/81, 105/106, 92/106 bytes). Neither ratchet moves and no ceiling is raised, so no maintainer ruling is needed for either.

Verification

Gate union derived on the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (no hand-built path list) at f977b0b54 — 10 families, all run at that same head:

• ✅ check:pm-skill-ratchet — .claude/agents/os-dev.md is 466 lines (ceiling 466; headroom 0), checklist-author/SKILL.md is 62 lines (ceiling 62; headroom 0), dogfood-verification/SKILL.md is 157 lines (ceiling 157; headroom 0); widest-table-row pins all unchanged at 0
• ✅ check:agent-model-declared · check:agent-test-spelling · check:doc-authoring · check:nul-bytes · check:pm-governed-merges · check:pm-skill-id-lint · check:skill-frame-sync
• ✅ pnpm --filter @objectstack/lint run check:doc-formula-expressions — self-test: 58 cases passed, 22 record-scoped formula example(s) across 427 files / 1451 TS blocks judged clean (needed @objectstack/formula and @objectstack/lint built first; its own PREREQUISITE NOT MET refusal is not a finding)
• ⚪ node scripts/pm/check-governed-queue-guard.mjs — NOT MEASURED locally, by construction: it reads GITHUB_EVENT_PATH and nothing else, so outside a workflow run it exits 1 with could not read GITHUB_EVENT_PATH. That is its refusal path, not a finding. CI measures it.

pnpm lint (repo-wide ESLint) is a declared narrowing, measured on all three axes: (1) the population comes from ESLint's own resolved config, not a guess — all three files report File ignored because no matching configuration was supplied., i.e. markdown is outside the linted population entirely; (2) the count comes from --format json — 3 files reported, 0 errors; (3) invariance for untouched files holds because the diff contains no eslint.config.mjs, no tsconfig and no TypeScript at all, so no untouched file's verdict can move.

No changeset: the diff is .claude/** only and publishes nothing from any package, so skip-changeset applies.

Scope guard

.claude/agents/os-dev.md carries adjacent text owned by other in-flight cards — the verify-lock and skip-changeset paragraphs of #12800 / #13300 / #13432. Those are deliberately untouched here; this diff is the recipe sentence only. Those three cards are not addressed by this PR and remain open.

Draft on purpose: governed .claude/** markdown, human merge.

---

Generated by Claude Code

[claude]

Fixes#13729

The three INTERNAL copies that PRESCRIBE a worktree recipe to an agent still taught the unhardened spelling. Each now reads as the form AGENTS.md Prime Directive 11 and CLAUDE.md already teach:

git fetch origin main && git worktree add --no-track ../REPO-TASK -b BRANCH origin/main

Exactly one recipe sentence per file; nothing else in these files is touched.

What changed, per copy

file	before	after
.claude/skills/checklist-author/SKILL.md L33	git worktree add ../objectstack-TASK -b BRANCH main	gains --no-trackand the git fetch origin main + origin/main base
.claude/agents/os-dev.md L32-33	git worktree add ../REPO-issue-N -b claude/issue-N-SLUG origin/main	gains --no-track and the fetch
.claude/skills/dogfood-verification/SKILL.md L140	git worktree add -b BRANCH /tmp/pr origin/main	gains --no-track and the fetch

Placeholder metavariables are spelled in caps above because this body is sanitized; the files themselves keep each copy's own angle-bracket placeholder naming verbatim (../objectstack-TASK stays the checklist-author spelling, claude/issue-N-SLUG stays the os-dev spelling, /tmp/pr stays the dogfood one).

Why

Plain -b writes branch.NAME.remote / branch.NAME.merge into the .git/config that every worktree of the repo SHARES — the class already adjudicated in #13052 / objectstack-ai/objectui#6880 and already hardened in AGENTS.md PD#11 and CLAUDE.md; these copies were never swept with it. The checklist-author copy was additionally the only one still basing a new tree on local main with no fetch — the objectstack-ai/objectui#6208 class.

The published fourth copy (skills/objectstack-pm-dispatch/SKILL.md) is NOT in this diff: it was discharged separately in #13891, which has landed, with its {default_branch} interpolation preserved. Verified on today's origin/main before editing.

Two ratchets bind here, so the edits are reflows rather than rewraps

check:pm-skill-ratchet constrains these files twice, and the obvious spellings break one each:

• Zero line-count headroom. All three files sit exactly at their ceiling (os-dev.md 466/466, checklist-author 62/62, dogfood-verification 157/157). Wrapping the longer recipe onto a new line failed the gate on all three.
• 120-byte per-line budget. Keeping the longer recipe on its single existing line failed the other half — 139B / 133B / 134B respectively.

So each recipe is reflowed within the lines its sentence already occupied: the break moves, the line count does not, and every touched line lands under budget (97/81, 105/106, 92/106 bytes). Neither ratchet moves and no ceiling is raised, so no maintainer ruling is needed for either.

Verification

Gate union derived on the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (no hand-built path list) at f977b0b54 — 10 families, all run at that same head:

• ✅ check:pm-skill-ratchet — .claude/agents/os-dev.md is 466 lines (ceiling 466; headroom 0), checklist-author/SKILL.md is 62 lines (ceiling 62; headroom 0), dogfood-verification/SKILL.md is 157 lines (ceiling 157; headroom 0); widest-table-row pins all unchanged at 0
• ✅ check:agent-model-declared · check:agent-test-spelling · check:doc-authoring · check:nul-bytes · check:pm-governed-merges · check:pm-skill-id-lint · check:skill-frame-sync
• ✅ pnpm --filter @objectstack/lint run check:doc-formula-expressions — self-test: 58 cases passed, 22 record-scoped formula example(s) across 427 files / 1451 TS blocks judged clean (needed @objectstack/formula and @objectstack/lint built first; its own PREREQUISITE NOT MET refusal is not a finding)
• ⚪ node scripts/pm/check-governed-queue-guard.mjs — NOT MEASURED locally, by construction: it reads GITHUB_EVENT_PATH and nothing else, so outside a workflow run it exits 1 with could not read GITHUB_EVENT_PATH. That is its refusal path, not a finding. CI measures it.

pnpm lint (repo-wide ESLint) is a declared narrowing, measured on all three axes: (1) the population comes from ESLint's own resolved config, not a guess — all three files report File ignored because no matching configuration was supplied., i.e. markdown is outside the linted population entirely; (2) the count comes from --format json — 3 files reported, 0 errors; (3) invariance for untouched files holds because the diff contains no eslint.config.mjs, no tsconfig and no TypeScript at all, so no untouched file's verdict can move.

No changeset: the diff is .claude/** only and publishes nothing from any package, so skip-changeset applies.

Scope guard

.claude/agents/os-dev.md carries adjacent text owned by other in-flight cards — the verify-lock and skip-changeset paragraphs of #12800 / #13300 / #13432. Those are deliberately untouched here; this diff is the recipe sentence only. Those three cards are not addressed by this PR and remain open.

Draft on purpose: governed .claude/** markdown, human merge.

---

Generated by Claude Code

[claude]

Fixes#13729

The three INTERNAL copies that PRESCRIBE a worktree recipe to an agent still taught the unhardened spelling. Each now reads as the form AGENTS.md Prime Directive 11 and CLAUDE.md already teach:

git fetch origin main && git worktree add --no-track ../REPO-TASK -b BRANCH origin/main

Exactly one recipe sentence per file; nothing else in these files is touched.

What changed, per copy

file	before	after
.claude/skills/checklist-author/SKILL.md L33	git worktree add ../objectstack-TASK -b BRANCH main	gains --no-trackand the git fetch origin main + origin/main base
.claude/agents/os-dev.md L32-33	git worktree add ../REPO-issue-N -b claude/issue-N-SLUG origin/main	gains --no-track and the fetch
.claude/skills/dogfood-verification/SKILL.md L140	git worktree add -b BRANCH /tmp/pr origin/main	gains --no-track and the fetch

Placeholder metavariables are spelled in caps above because this body is sanitized; the files themselves keep each copy's own angle-bracket placeholder naming verbatim (../objectstack-TASK stays the checklist-author spelling, claude/issue-N-SLUG stays the os-dev spelling, /tmp/pr stays the dogfood one).

Why

Plain -b writes branch.NAME.remote / branch.NAME.merge into the .git/config that every worktree of the repo SHARES — the class already adjudicated in #13052 / objectstack-ai/objectui#6880 and already hardened in AGENTS.md PD#11 and CLAUDE.md; these copies were never swept with it. The checklist-author copy was additionally the only one still basing a new tree on local main with no fetch — the objectstack-ai/objectui#6208 class.

The published fourth copy (skills/objectstack-pm-dispatch/SKILL.md) is NOT in this diff: it was discharged separately in #13891, which has landed, with its {default_branch} interpolation preserved. Verified on today's origin/main before editing.

Two ratchets bind here, so the edits are reflows rather than rewraps

check:pm-skill-ratchet constrains these files twice, and the obvious spellings break one each:

• Zero line-count headroom. All three files sit exactly at their ceiling (os-dev.md 466/466, checklist-author 62/62, dogfood-verification 157/157). Wrapping the longer recipe onto a new line failed the gate on all three.
• 120-byte per-line budget. Keeping the longer recipe on its single existing line failed the other half — 139B / 133B / 134B respectively.

So each recipe is reflowed within the lines its sentence already occupied: the break moves, the line count does not, and every touched line lands under budget (97/81, 105/106, 92/106 bytes). Neither ratchet moves and no ceiling is raised, so no maintainer ruling is needed for either.

Verification

Gate union derived on the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (no hand-built path list) at f977b0b54 — 10 families, all run at that same head:

• ✅ check:pm-skill-ratchet — .claude/agents/os-dev.md is 466 lines (ceiling 466; headroom 0), checklist-author/SKILL.md is 62 lines (ceiling 62; headroom 0), dogfood-verification/SKILL.md is 157 lines (ceiling 157; headroom 0); widest-table-row pins all unchanged at 0
• ✅ check:agent-model-declared · check:agent-test-spelling · check:doc-authoring · check:nul-bytes · check:pm-governed-merges · check:pm-skill-id-lint · check:skill-frame-sync
• ✅ pnpm --filter @objectstack/lint run check:doc-formula-expressions — self-test: 58 cases passed, 22 record-scoped formula example(s) across 427 files / 1451 TS blocks judged clean (needed @objectstack/formula and @objectstack/lint built first; its own PREREQUISITE NOT MET refusal is not a finding)
• ⚪ node scripts/pm/check-governed-queue-guard.mjs — NOT MEASURED locally, by construction: it reads GITHUB_EVENT_PATH and nothing else, so outside a workflow run it exits 1 with could not read GITHUB_EVENT_PATH. That is its refusal path, not a finding. CI measures it.

pnpm lint (repo-wide ESLint) is a declared narrowing, measured on all three axes: (1) the population comes from ESLint's own resolved config, not a guess — all three files report File ignored because no matching configuration was supplied., i.e. markdown is outside the linted population entirely; (2) the count comes from --format json — 3 files reported, 0 errors; (3) invariance for untouched files holds because the diff contains no eslint.config.mjs, no tsconfig and no TypeScript at all, so no untouched file's verdict can move.

No changeset: the diff is .claude/** only and publishes nothing from any package, so skip-changeset applies.

Scope guard

.claude/agents/os-dev.md carries adjacent text owned by other in-flight cards — the verify-lock and skip-changeset paragraphs of #12800 / #13300 / #13432. Those are deliberately untouched here; this diff is the recipe sentence only. Those three cards are not addressed by this PR and remain open.

Draft on purpose: governed .claude/** markdown, human merge.

---

Generated by Claude Code

[claude]

Fixes#13729

The three INTERNAL copies that PRESCRIBE a worktree recipe to an agent still taught the unhardened spelling. Each now reads as the form AGENTS.md Prime Directive 11 and CLAUDE.md already teach:

git fetch origin main && git worktree add --no-track ../REPO-TASK -b BRANCH origin/main

Exactly one recipe sentence per file; nothing else in these files is touched.

What changed, per copy

file	before	after
.claude/skills/checklist-author/SKILL.md L33	git worktree add ../objectstack-TASK -b BRANCH main	gains --no-trackand the git fetch origin main + origin/main base
.claude/agents/os-dev.md L32-33	git worktree add ../REPO-issue-N -b claude/issue-N-SLUG origin/main	gains --no-track and the fetch
.claude/skills/dogfood-verification/SKILL.md L140	git worktree add -b BRANCH /tmp/pr origin/main	gains --no-track and the fetch

Placeholder metavariables are spelled in caps above because this body is sanitized; the files themselves keep each copy's own angle-bracket placeholder naming verbatim (../objectstack-TASK stays the checklist-author spelling, claude/issue-N-SLUG stays the os-dev spelling, /tmp/pr stays the dogfood one).

Why

Plain -b writes branch.NAME.remote / branch.NAME.merge into the .git/config that every worktree of the repo SHARES — the class already adjudicated in #13052 / objectstack-ai/objectui#6880 and already hardened in AGENTS.md PD#11 and CLAUDE.md; these copies were never swept with it. The checklist-author copy was additionally the only one still basing a new tree on local main with no fetch — the objectstack-ai/objectui#6208 class.

The published fourth copy (skills/objectstack-pm-dispatch/SKILL.md) is NOT in this diff: it was discharged separately in #13891, which has landed, with its {default_branch} interpolation preserved. Verified on today's origin/main before editing.

Two ratchets bind here, so the edits are reflows rather than rewraps

check:pm-skill-ratchet constrains these files twice, and the obvious spellings break one each:

• Zero line-count headroom. All three files sit exactly at their ceiling (os-dev.md 466/466, checklist-author 62/62, dogfood-verification 157/157). Wrapping the longer recipe onto a new line failed the gate on all three.
• 120-byte per-line budget. Keeping the longer recipe on its single existing line failed the other half — 139B / 133B / 134B respectively.

So each recipe is reflowed within the lines its sentence already occupied: the break moves, the line count does not, and every touched line lands under budget (97/81, 105/106, 92/106 bytes). Neither ratchet moves and no ceiling is raised, so no maintainer ruling is needed for either.

Verification

Gate union derived on the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (no hand-built path list) at f977b0b54 — 10 families, all run at that same head:

• ✅ check:pm-skill-ratchet — .claude/agents/os-dev.md is 466 lines (ceiling 466; headroom 0), checklist-author/SKILL.md is 62 lines (ceiling 62; headroom 0), dogfood-verification/SKILL.md is 157 lines (ceiling 157; headroom 0); widest-table-row pins all unchanged at 0
• ✅ check:agent-model-declared · check:agent-test-spelling · check:doc-authoring · check:nul-bytes · check:pm-governed-merges · check:pm-skill-id-lint · check:skill-frame-sync
• ✅ pnpm --filter @objectstack/lint run check:doc-formula-expressions — self-test: 58 cases passed, 22 record-scoped formula example(s) across 427 files / 1451 TS blocks judged clean (needed @objectstack/formula and @objectstack/lint built first; its own PREREQUISITE NOT MET refusal is not a finding)
• ⚪ node scripts/pm/check-governed-queue-guard.mjs — NOT MEASURED locally, by construction: it reads GITHUB_EVENT_PATH and nothing else, so outside a workflow run it exits 1 with could not read GITHUB_EVENT_PATH. That is its refusal path, not a finding. CI measures it.

pnpm lint (repo-wide ESLint) is a declared narrowing, measured on all three axes: (1) the population comes from ESLint's own resolved config, not a guess — all three files report File ignored because no matching configuration was supplied., i.e. markdown is outside the linted population entirely; (2) the count comes from --format json — 3 files reported, 0 errors; (3) invariance for untouched files holds because the diff contains no eslint.config.mjs, no tsconfig and no TypeScript at all, so no untouched file's verdict can move.

No changeset: the diff is .claude/** only and publishes nothing from any package, so skip-changeset applies.

Scope guard

.claude/agents/os-dev.md carries adjacent text owned by other in-flight cards — the verify-lock and skip-changeset paragraphs of #12800 / #13300 / #13432. Those are deliberately untouched here; this diff is the recipe sentence only. Those three cards are not addressed by this PR and remain open.

Draft on purpose: governed .claude/** markdown, human merge.

---

Generated by Claude Code

[claude]

Fixes#13729

The three INTERNAL copies that PRESCRIBE a worktree recipe to an agent still taught the unhardened spelling. Each now reads as the form AGENTS.md Prime Directive 11 and CLAUDE.md already teach:

git fetch origin main && git worktree add --no-track ../REPO-TASK -b BRANCH origin/main

Exactly one recipe sentence per file; nothing else in these files is touched.

What changed, per copy

file	before	after
.claude/skills/checklist-author/SKILL.md L33	git worktree add ../objectstack-TASK -b BRANCH main	gains --no-trackand the git fetch origin main + origin/main base
.claude/agents/os-dev.md L32-33	git worktree add ../REPO-issue-N -b claude/issue-N-SLUG origin/main	gains --no-track and the fetch
.claude/skills/dogfood-verification/SKILL.md L140	git worktree add -b BRANCH /tmp/pr origin/main	gains --no-track and the fetch

Placeholder metavariables are spelled in caps above because this body is sanitized; the files themselves keep each copy's own angle-bracket placeholder naming verbatim (../objectstack-TASK stays the checklist-author spelling, claude/issue-N-SLUG stays the os-dev spelling, /tmp/pr stays the dogfood one).

Why

Plain -b writes branch.NAME.remote / branch.NAME.merge into the .git/config that every worktree of the repo SHARES — the class already adjudicated in #13052 / objectstack-ai/objectui#6880 and already hardened in AGENTS.md PD#11 and CLAUDE.md; these copies were never swept with it. The checklist-author copy was additionally the only one still basing a new tree on local main with no fetch — the objectstack-ai/objectui#6208 class.

The published fourth copy (skills/objectstack-pm-dispatch/SKILL.md) is NOT in this diff: it was discharged separately in #13891, which has landed, with its {default_branch} interpolation preserved. Verified on today's origin/main before editing.

Two ratchets bind here, so the edits are reflows rather than rewraps

check:pm-skill-ratchet constrains these files twice, and the obvious spellings break one each:

• Zero line-count headroom. All three files sit exactly at their ceiling (os-dev.md 466/466, checklist-author 62/62, dogfood-verification 157/157). Wrapping the longer recipe onto a new line failed the gate on all three.
• 120-byte per-line budget. Keeping the longer recipe on its single existing line failed the other half — 139B / 133B / 134B respectively.

So each recipe is reflowed within the lines its sentence already occupied: the break moves, the line count does not, and every touched line lands under budget (97/81, 105/106, 92/106 bytes). Neither ratchet moves and no ceiling is raised, so no maintainer ruling is needed for either.

Verification

Gate union derived on the final head with node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack (no hand-built path list) at f977b0b54 — 10 families, all run at that same head:

• ✅ check:pm-skill-ratchet — .claude/agents/os-dev.md is 466 lines (ceiling 466; headroom 0), checklist-author/SKILL.md is 62 lines (ceiling 62; headroom 0), dogfood-verification/SKILL.md is 157 lines (ceiling 157; headroom 0); widest-table-row pins all unchanged at 0
• ✅ check:agent-model-declared · check:agent-test-spelling · check:doc-authoring · check:nul-bytes · check:pm-governed-merges · check:pm-skill-id-lint · check:skill-frame-sync
• ✅ pnpm --filter @objectstack/lint run check:doc-formula-expressions — self-test: 58 cases passed, 22 record-scoped formula example(s) across 427 files / 1451 TS blocks judged clean (needed @objectstack/formula and @objectstack/lint built first; its own PREREQUISITE NOT MET refusal is not a finding)
• ⚪ node scripts/pm/check-governed-queue-guard.mjs — NOT MEASURED locally, by construction: it reads GITHUB_EVENT_PATH and nothing else, so outside a workflow run it exits 1 with could not read GITHUB_EVENT_PATH. That is its refusal path, not a finding. CI measures it.

pnpm lint (repo-wide ESLint) is a declared narrowing, measured on all three axes: (1) the population comes from ESLint's own resolved config, not a guess — all three files report File ignored because no matching configuration was supplied., i.e. markdown is outside the linted population entirely; (2) the count comes from --format json — 3 files reported, 0 errors; (3) invariance for untouched files holds because the diff contains no eslint.config.mjs, no tsconfig and no TypeScript at all, so no untouched file's verdict can move.

No changeset: the diff is .claude/** only and publishes nothing from any package, so skip-changeset applies.

Scope guard

.claude/agents/os-dev.md carries adjacent text owned by other in-flight cards — the verify-lock and skip-changeset paragraphs of #12800 / #13300 / #13432. Those are deliberately untouched here; this diff is the recipe sentence only. Those three cards are not addressed by this PR and remain open.

Draft on purpose: governed .claude/** markdown, human merge.

---

Generated by Claude Code