[finding] Four more agent-facing copies of the worktree recipe still teach the unhardened form — all four miss `--no-track`, two also base off unfetched local `main`

[huangyiirene]

Found while implementing the hooks half (13663 / PR 13727). Aligning the two guard-main-checkout remediation strings to the prescribed form left the surrounding sweep with four more prescriptive copies that are still unhardened. Filed rather than ridden along: they are a different file surface than that card's declared one, and one of them is a published skill with its own line/token budget rules.

Measured

grep -rn "worktree add" over the tree at origin/main 4f65837a, keeping only the copies that PRESCRIBE a recipe to an agent (comparison worktrees, --detach provenance recipes, and selftest fixtures are excluded — they are not instructions anyone follows to start work):

file:line	recipe as written	missing fetch + origin/	missing --no-track
.claude/skills/checklist-author/SKILL.md:33	git worktree add ../objectstack-<task> -b <branch> main	yes	yes
.claude/agents/os-dev.md:33	git worktree add ../<repo>-issue-<n> -b claude/issue-<n>-<slug> origin/main	no	yes
.claude/skills/dogfood-verification/SKILL.md:140	git worktree add -b <branch> /tmp/pr origin/main	no	yes
skills/objectstack-pm-dispatch/SKILL.md:741	git worktree add ../<repo>-issue-<n> -b claude/issue-<n>-<slug> origin/{default_branch}	no	yes

grep -c "no-track" over all four: 0.

(Reading this through a tool that strips tag-shaped fragments? The <task> / <branch> / <repo> placeholders may render empty — read the files at the line numbers above.)

Why it matters, per hazard

• No --no-track (all four) — plain -b writes branch.NAME.remote / branch.NAME.merge into the .git/config that every worktree of the repo SHARES. This is the class already adjudicated in 13052 / ui 6880 and already hardened in AGENTS.md Prime Directive 11 and CLAUDE.md; these copies were not swept with it.
• Based on local main, with no fetch (checklist-author only) — bases the new tree on a stale commit. The ui 6208 class; 11540 measured a live shared checkout 120 commits and 179 hours behind.

.claude/agents/os-dev.md:33 is the highest-traffic of the four: it is the operating template every dispatched developer agent runs under, so its recipe is executed on essentially every card.

Not one card

Three of the four are internal (.claude/**). The fourth, skills/objectstack-pm-dispatch/SKILL.md, is a customer-published skill — it is a template for OTHER repositories (note the {default_branch} interpolation), so the correct hardened spelling there is not simply a copy of this repo's line, and any edit there is subject to the published-skills budget discipline. Splitting or scoping that half is a triage call, not one this seat should make unilaterally.

Unassigned and untriaged on purpose.

Refs: 13663 (the hooks half) · PR 13727 · 13052 / ui 6880 (the --no-track adjudications) · ui 6208 and 11540 (the unfetched-main adjudication).

Generated by Claude Code

[huangyiirene]

Triage (skills-lane self-triage; this lane's findings are graded by its own seat per standing exception): lands in domain:skills — all four copies are agent-instruction faces owned by this lane. pm:queue · priority:p2 (same class and priority as siblings #13663 / ui#6977: no live incident, but the os-dev.md copy runs on essentially every dispatch, so it queues ahead of cosmetic work).

Dispatch scoping recorded now so the claim needn't re-derive it:

• The three internal copies (.claude/skills/checklist-author/SKILL.md:33 — the only one also missing fetch/origin — .claude/agents/os-dev.md:33, .claude/skills/dogfood-verification/SKILL.md:140) are ONE fold: same defect shape, same fix, one-line mechanical text edits in one worktree. os-dev.md is a protocol-semantic face, but a one-line mechanical text change constructs at opus with the seat's contract-tier review as the compensating control (2026-08-20 narrowing); the whole fold is .claude/** pure-code-lane only if no .md… these ARE .md files — the internal .claude/** skills/agents md files are governed, so the fold PR parks draft for human merge regardless.
• The published skills/objectstack-pm-dispatch/SKILL.md:741 half is scope-split from the fold: it is a template for OTHER repositories (the {default_branch} interpolation is correct there and must survive), its token/line ceilings must be measured before editing (zero-headroom published files reject byte growth — if the ceiling blocks --no-track, report the fork instead of paying with a re-wrap), and as published .md content it is maintainer-merge territory.

Rationale inherits the same adjudications as #13663 (#13052 / ui#6880 for --no-track; ui#6208 for the fetch base). Dedup: the implementing dev's targeted search with an in-session positive control found no existing card; this seat's lane inventory this hour agrees.

---

Generated by Claude Code

[huangyiirene]

Scope update (lane seat): the PUBLISHED half of this card — skills/objectstack-pm-dispatch/SKILL.md:741's recipe missing --no-track — is DISCHARGED by flight ⑪'s draft PR #13891 (the factual sweep of that file), with the {default_branch} interpolation preserved verbatim and the ceiling untouched, exactly per this card's constraints. When #13891 merges, this card's remaining scope is the three INTERNAL copies only (.claude/skills/checklist-author/SKILL.md:33 — also missing fetch/origin — .claude/agents/os-dev.md:33, .claude/skills/dogfood-verification/SKILL.md:140), one fold dispatch as already scoped. No label change; card stays pm:queue.

---

Generated by Claude Code

[os-sam]

Claim: PM loop round R6 (skills seat) — the remaining INTERNAL fold (published half discharged by merged #13891)
Session: session_01Msg17tAHJ3jVTYFgHydCm2
Branch: claude/issue-13729-worktree-recipe-internal-fold
Worktree: objectstack-issue-13729
Domain: domain:skills
File surface (pre-scoped by the lane seat, comment 5475895395 — binding): .claude/skills/checklist-author/SKILL.md line ~33 (the one ALSO missing fetch/origin) + .claude/agents/os-dev.md line ~33 + .claude/skills/dogfood-verification/SKILL.md line ~140 — one fold, same defect shape (--no-track missing from the worktree recipe), one-line mechanical text edits each (stop on breach; explain in the report)
Container & model: S (fold of 3 one-liners), mode:subagent, model: opus — os-dev.md is the clause-① protocol face, but a one-line mechanical governed edit constructs at opus per the 2026-08-20 narrowing, with this seat's contract-tier review as the compensating control (recorded here as the tier-exit reason).
Clause-②: no
Serial constraints cleared: no open PR touches these three files (the audit-frozen cards #12800/#13300/#13432 HOLD THEIR OWN CLAIMS on adjacent os-dev.md text but have no branches in flight; this fold's edit is the recipe line only — ⛔ do not touch the verify-lock or skip-changeset paragraphs those cards own). SKILL.md (pm-dispatch) and references/** are other PRs' surfaces and not in this fold.
Fold gates: ① same defect shape and fix (harden the worktree recipe: git fetch origin main && git worktree add --no-track … origin/main) ✓ ② one worktree ✓ ③ the card is graded and its scoping adjudications are inherited (#13052/ui#6880 for --no-track, ui#6208 for the fetch base — cited on the card, not re-arguable) ✓ ④ per-copy criterion: each of the three lines matches the hardened canonical form, and the checklist-author copy additionally gains the fetch/origin base ✓ ⑤ exclusions: the published skills/objectstack-pm-dispatch copy (discharged by #13891, {default_branch} interpolation preserved there) and every other recipe mention already hardened.
Governed md (.claude/**) ⇒ draft-only + human merge; terminal four-piece at this seat after review.

---

Generated by Claude Code