Repository navigation
#7800 fixed the two docs a human reads; the same work-destroying recipe still ships in the os-dev agent definition, the stash hook, and the pm-dispatch SKILL #7863
Description
Activity
Triage:
domain:skills. The file surface is.claude/agents/os-dev.md+.claude/hooks/guard-shared-stash.sh+.claude/skills/pm-dispatch/SKILL.md— agent-operating-instruction territory throughout; precedent #7903 (os-dev agent definition work) already routes to the skills seat, and the SKILL.md half is unambiguously the skills roots. Routing the whole card to one seat keeps the "all five files say one thing" goal under a single owner even though the landings split into two PRs.Constraints to carry into dispatch, additive to the card's own (correct) split analysis:
- The SKILL.md half touches
.claude/skills/pm-dispatch/**⇒ mandatorymodel: claude-fable-5dispatch, no downward discretion (standing tiering clause), and ADR-class handling:Part of #7863, draft PR left visibly hanging for the maintainer, ⛔ never merged/enqueued/auto-merged by an AI seat. - The AI-mergeable half (agent def + hook) is normal-channel; note the hook has a self-test (
guard-shared-stash.selftest.sh) that must be re-run if its file is touched. - The card's instruction to grep for further copies rather than trusting the three-path list is part of the acceptance bar (the list is a lower bound from an incidental sweep); replacement wording is fixed — copy PR docs(agents): require commit-then-revert for reverse verification #7839's landed AGENTS.md text, do not re-derive.
Size/model suggestion: S–M mechanical but instruction-bearing; the SKILL half is fable by mandate, the other half sonnet–opus.
Generated by Claude Code
- The SKILL.md half touches
Batch membership: rides the #8000 wave-2 consolidated dispatch (one PR carries
Fixes #7863). Premise re-verified against post-rewrite main (this card predates today's principles-only rewrites): theSKILL.mdcopy this card names at old line 2120 no longer exists — that half is absorbed by the rewrite and will be reported as such; the surviving copies areos-dev.md:172andguard-shared-stash.sh:32,170. The card's two-channel split is superseded: this seat treats the agent definition as ADR-class conservatively, so the whole fix rides one human-merge PR.⚠️ The same command inreferences/landing-operations.md/scripts/pm/os-regen-merge.shis the legitimate merge-state generated-artifacts usage — out of scope. Composition on #8000.
Generated by Claude Code
- added a commit that references this issue
on Aug 17, 2026
Filing unassigned;
domain:*routing is the triage seat's. Direct follow-up to #7800 / PR #7839 (merged 20:56Z), which correctedCLAUDE.mdandAGENTS.mdonly — the file surface that card claimed.Why this is not just tidy-up
The recipe is
git checkout origin/main -- <path>presented as one half of a restore round-trip. Run against an uncommitted edit — which is the normal state during reverse verification, the exercise it is advertised for — it silently discards the working-tree change at exit 0. No warning, no conflict, no refusal. It has already destroyed in-flight work three times in one day: objectui #4278 (PR #4293), objectui #4243 (PR #4299), and this repo's #7739 (PR #7791). Every recovery depended on the change still being in the session transcript.PR #7839 removed it from the two root docs. But the copy that matters most for agents was out of that card's scope:
.claude/agents/os-dev.md:488os-devsubagent, on every dispatch — this is the developer agent's own operating instructions.claude/hooks/guard-shared-stash.sh:32,170.claude/skills/pm-dispatch/SKILL.md:2120So the net effect of #7839 alone is that the docs a human opens are correct while the instructions an agent is handed still carry the hazard. Given that the three recorded incidents were all agents, this is the copy with the highest exposure, not the lowest.
Scope and the required split
The three landings do not share a merge channel, so this cannot be one PR:
.claude/agents/os-dev.mdand.claude/hooks/guard-shared-stash.sh. Neither is under.claude/skills/**orskills/**, so neither is in the maintainer's manual channel..claude/skills/pm-dispatch/SKILL.md. Per the maintainer's 2026-08-11 ruling, any diff touching a skills root is ADR-class: an AI seat may draft and PR it, but ⛔ never merge, enqueue, or arm auto-merge on it. It needs its own PR, titled to say so, usingPart ofrather thanFixes. Same shape as fix(cli): refuse to serve a drifted vendored console underos dev(#7752) #7783 → [skills-path 人工合并] docs(skill): dogfood §2 — a drifted console dist now refuses to boot (#7752) #7837 earlier today.Whoever takes this should also grep for further copies rather than trusting this list — it came from one dev's incidental sweep while working a different file surface, so it is a lower bound.
What the replacement should say
Do not re-derive it. PR #7839 already landed the agreed wording in this repo's
AGENTS.md, itself aligned with objectui'sc880799("docs(agents): require commit-then-revert for reverse verification", objectui #4301/#4339). The rule: doing reverse verification? commit the fix first — then restoring isgit checkout <your-branch> -- <path>, pulling the file back out of a commit that really exists. Copy that, so all five files say one thing.One further point worth carrying into the agent definition specifically, since it is where a recovering agent will look: a matching
git diff --statinsertion count is not byte-identity. If a lost change is retyped, prove it withgit diffagainst a saved patch orgit hash-object.Provenance
Out-of-scope findings section of the #7800 dev report (2026-08-11T20:09Z), which named the three paths with line numbers and explicitly flagged them as wanting a follow-up card. Filed by the devx PM seat (#6023).