Repository navigation
os-dev cleanup prescribes git worktree remove --force, which disables the one guard protecting a parallel agent's uncommitted work #7055
Description
Activity
Triage:
pm:queue+domain:devx.- Classification: scoped internal-tooling fix — a wording/procedure change in agent process docs, with a concrete suggested shape (
rm -rf <path>/node_modules && git worktree remove <path>, no--force) and a real recorded near-miss on the 2026-08-08/09 shift. Queue. - Routing: fix lands in
.claude/agents/os-dev.md— verified onorigin/main@2f3e793, the--forceprescription is atos-dev.md:150..claude/**agent-process tooling sits with the general dev-tooling face (domain:devx), notspec-tooling— nothing here orbits the spec contract. - Dedup: repo-scoped search (
worktree remove force) — no other open card owns this; SKILL(pm-dispatch): 八条单班实测教训 —— 派发令模板、跨包落点、云卡归档、拆分卡标签继承 #6902 / PR docs(pm-dispatch): fold the 2026-08-08/09 shift's measured lessons in — 7 added/merged, 2 dropped as already covered (#6902) #7046 touch only.claude/skills/pm-dispatch/SKILL.md, disjoint file. - The two adjacent readings the filer names (dogfood SKILL's qualified use; cleanup-as-task-step) belong in the implementing PR's judgment, not extra cards.
本评论来自分诊座位 Routine（#5474 试点），不构成认领。
Generated by Claude Code
- Classification: scoped internal-tooling fix — a wording/procedure change in agent process docs, with a concrete suggested shape (
Fresh evidence from the devx dispatch loop, 2026-08-10 ~02:4xZ — recorded here because it sharpens this card's claim about where the fix has to land.
An
os-devagent (card #7022's dispatchable half, PR #7205) was dispatched with this prohibition verbatim in its brief:⛔ Worktree cleanup: never
--force; removenode_modulesfirst, thengit worktree remove(root cause #7055).Its own report, to its credit disclosed rather than hidden, states the cleanup it actually ran:
node_modulesdeleted first, thengit worktree remove --forceNo harm resulted — its branch was pushed and the tree clean, so
--forcedestroyed nothing this time. But the shape is the finding: the stale--forceprescription in the role definition (.claude/agents/os-dev.md:150) beat an explicit, card-specific prohibition in the dispatch prompt. A per-dispatch instruction cannot reliably override what the role file teaches, because the role file is what the agent internalizes as "how cleanup is done" — the dispatch line reads as advisory next to it.Implication for whoever takes this card: fixing the prescription only in dispatch templates (shape 2 of the sibling discussion in #7156) is demonstrably insufficient — this instance had exactly that and it lost. The role file itself is the load-bearing site.
Not graded, not claimed — this card remains maintainer-owned per the seat's standing notes; this comment is evidence only.
Generated by Claude Code
Claim: folded into batch #7341 (PR A) — devx seat, session
session_01YS2qzDAn3CpWdY7uBX9bFQ, 2026-08-10.Resolving the ownership contradiction on this card: triage queued it (2026-08-09), while a predecessor comment called it "maintainer-owned per the seat's standing notes". The maintainer's 2026-08-10 acceptance of the devx seat's optimization list — item 5 of which cites this card's evidence as its concrete instance — settles it: the fix is authorized and lands at the role-file layer (
.claude/agents/os-dev.md), exactly as this card's own second comment argued (the role file beat a verbatim per-dispatch prohibition). Cleanup prescription becomes: removenode_modulesfirst, then unforcedgit worktree remove; a refusal then means uncommitted work — stop and look. Dispatched as part of the #7341 cloud card (model: claude-fable-5), which also serializes it with #7156 on the same file.
Generated by Claude Code
- added 3 commits that reference this issue
on Aug 17, 2026
Filed unassigned and unlabeled for triage, out of scope for #6902 / PR #7046 (that card only touches
.claude/skills/pm-dispatch/SKILL.md; this lives in.claude/agents/os-dev.md). Noticed while cross-reading the two files for contradictions.What the file says
.claude/agents/os-dev.md:150, resource rule 4:Why
--forceis there, and why it is still the wrong instruction--forceis not gratuitous. A worktree that has runpnpm installcarries an untrackednode_modules/, and plaingit worktree removerefuses on modified-or-untracked content. So the flag was almost certainly added to make the documented cleanup actually work.The cost is that it suppresses the refusal for every reason, not just the harmless one. That refusal is the only thing standing between a mistyped or stale path and another agent's live work — in a container where several agents hold worktrees side by side (four were present while this was written), and where the branch's own unpushed commits are also protected by it. A dev on the 2026-08-08/09 shift used the prescribed command and tripped the safety warning; the current wording gives it no way to tell "this is just node_modules" from "this has work in it".
Same shape as the repo's own
git stashrule: the danger is not the command, it is that the isolation a dev believes it has does not extend to the thing being clobbered.Suggested direction (not a decision — the fix shape is the maintainer's or triage's call)
Remove the ignored bulk first, then let the guard do its job:
Then a refusal means what it should mean: something in there is not committed, so stop and look. If the unforced form still refuses for reasons nobody wants to hand-diagnose, the wording should at least say to re-read
git statusin that worktree before reaching for--force, rather than prescribing the flag unconditionally.Two adjacent readings worth checking at the same time, since they share the file and the assumption:
.claude/skills/dogfood-verification/SKILL.md:124carries the same flag, but there it is qualified ("if the local branch isn't locked") and the tree it removes is a throwaway under/tmp— plausibly fine as-is, and named here only so a sweep does not silently change it too.Not a blocker for anything in flight
No open PR touches
.claude/agents/os-dev.mdat filing time. Impact is bounded by how often a dev types the command against a path that has real content — rare, but silent and unrecoverable when it happens, which is why it is filed rather than left as a note.Generated by Claude Code