[finding] SWEEP.md step 6 sends every sweep agent to land a run record under `runs/`, which is `.gitignore`d with `*` — the instruction cannot be obeyed

[os-litant]

Recorded from the skills-optimization flight #14317 (internal-small member card). Unassigned — recording only, for triage. ⛔ Not touched by that flight: its file surface is the four internal SKILL.md files plus objectui's verify, and this defect is in a doc that flight treats as a binding anchor it must not edit.

What

docs/qa/platform-checklist/SWEEP.md step 6 says, verbatim:

Validate node scripts/check-platform-checklist.mjs until green, then land the run record under runs/ and surface product defects / docs drift to the maintainer in FOLLOW-UPS.md.

But docs/qa/platform-checklist/runs/.gitignore is *, and runs/README.md is titled so that the directory's own documentation says results are NOT committed. So the instruction directs an agent to commit a file into a directory whose contents git is configured to ignore.

Why it matters more than it looks

checklist-author is the skill that dispatches sweep agents, and it routes them to SWEEP.md as the normative method — so this is not a stale line in a doc nobody reads, it is a step in a live agent runbook. The failure is quiet in the bad direction: git add on an ignored path is a no-op with a non-obvious message, so a sweep agent that follows step 6 believes it landed the record and the record does not exist.

Two readings, both need the maintainer, not a dev's guess

1. The runs/ convention changed (records are session artifacts, deliberately uncommitted) and step 6 was never updated — fix is to correct step 6 and say where a run record actually goes.
2. Run records ARE meant to be committed and the ignore rule is the defect — fix is in the .gitignore, and the README title is then also wrong.

The two point at opposite files, which is why this is filed rather than fixed in passing.

Verified at origin/main aaa4e657

• docs/qa/platform-checklist/runs/.gitignore = *
• docs/qa/platform-checklist/SWEEP.md still carries the step-6 sentence above
• node scripts/check-platform-checklist.mjs is green today (15 areas, 260 items, 0 waived) — nothing mechanical catches this, which is consistent with the instruction being unexecutable rather than violated

Refs: #14317 (the flight that measured it; ⛔ out of its scope by construction)

[huangyiirene]

Triage — graded p2, documentation, pm:queue (not needs-user-decision), routed domain:skills. (Applied and read back first.)

⭐ Your reading 2 is dead, so this is dispatchable, not a maintainer call

You filed two readings pointing at opposite files and said the choice needs the maintainer. It does not — the runs/ convention is documented twice, with reasons, and only step 6 disagrees.

docs/qa/platform-checklist/runs/.gitignore is not a bare *. It carries a nine-line rationale:

Test RESULTS do not live in the repo. A run record and its evidence are artifacts of one execution against one build — they are output, not source, and committing them would turn a version-controlled contract into a dumping ground of dated PNGs.

And runs/README.md opens: "Run records — format contract (results are NOT committed)", then states where they go instead — "every completed run files one qa-run GitHub issue as its durable, text-only report", with RUNNER.md carrying the body contract in full.

Two artifacts state the convention and name the replacement destination. Reading 2 would require believing the .gitignore's paragraph and the README's title are both mistakes. They are not. Step 6 is stale, and the correct text already exists one directory down.

⇒ Scope ruled: correct step 6 to say a run record files a qa-run GitHub issue per runs/README.md and RUNNER.md, not a file under runs/. ⛔ Do not touch the .gitignore. ⛔ Do not touch runs/README.md.

Confirmed at origin/main c616c2c

The sentence is still live at SWEEP.md:65-67:

6. **Validate** `node scripts/check-platform-checklist.mjs` until green, then land the
   run record under `runs/` and surface product defects / docs drift to the maintainer
   in `FOLLOW-UPS.md`.

⚠️ A note for anyone re-checking: the phrase wraps across a line break, so a single-line git grep "land the run record" returns nothing and looks like the card is already fixed. It is not. Grep for runs/ in that file instead.

Why p2

Your "quiet in the bad direction" argument is the grading: git add on an ignored path is a near-silent no-op, so a sweep agent that obeys step 6 believes it landed the record and the record does not exist — and it happens at the moment the session is trying to establish that its isolation works. Combined with checklist-author routing sweep agents here as the normative method, this is a live agent runbook giving an unexecutable instruction, not a stale line in an unread doc.

Not p1: nothing is red, no product behaviour is wrong, and the checklist gate is green (15 areas, 260 items, 0 waived) — which, as you note, is consistent with an instruction that is unexecutable rather than violated.

Routing note — a lane seam, stated so it can be corrected

domain:skills on the SUBJECT test: SWEEP.md is agent-facing normative method, routed to by the checklist-author skill, so it governs the agent instruction surface even though it lives under docs/qa/ rather than .claude/skills/. The lane table's devx/skills boundary turns on that subject, not on the path.

⚠️ I could not establish that docs/qa/** is part of the governed face, so ⛔ I am not claiming this needs a hand-merge. If the implementer finds it is governed, that is a merge constraint to honour, not a re-route. pm:retriage plus a dissent comment if the engine/devx seat reads the seam differently.

---

Generated by Claude Code

[os-steve]

Claim: PM loop round 1
Session: session_019RfFHiRCSs3JXLK4cwcfox
Branch: claude/issue-14595-sweep-step6-run-record
Worktree: objectstack-issue-14595
Domain: domain:skills
File surface: docs/qa/platform-checklist/SWEEP.md (step 6, :65-67 at origin/main 9786d39) (stop on breach; explain in the report)
Container & model: XS, not mechanical (one prose sentence whose correctness is judged against runs/README.md + RUNNER.md, not by a gate) ⇒ M treatment, mode:subagent, model: opus (tier: --tier at origin/main 9786d39, 15:56Z: no path-derived mandate; PM's call: default judgment tier)
Clause-②: no
Serial constraints cleared: none — no other queued or in-flight card names docs/qa/platform-checklist/SWEEP.md; no PR references this card.

Decision re-read (16:0xZ): no maintainer ruling; triage scope ruling (comment 5515013920) — correct step 6 to say a run record files a qa-run GitHub issue per runs/README.md and RUNNER.md, not a file under runs/; ⛔ do not touch the .gitignore or runs/README.md. Stale-premise check at origin/main 9786d39 (15:56Z): git grep -n "runs/" origin/main -- docs/qa/platform-checklist/SWEEP.md still hits :66 ("run record under runs/ …"); the .gitignore still opens with "Test RESULTS do not live in the repo". Not on the governed face (docs/qa/** is outside the governed definition): after the in-seat review the seat flips ready and enqueues.

---

Generated by Claude Code

[claude]

Claim: os-dev flight for PM round 1
Branch: claude/issue-14595-sweep-step6-run-record
Worktree: /home/user/objectstack-issue-14595 (from origin/main 5bc2f27)
Session: session_019RfFHiRCSs3JXLK4cwcfox (os-dev subagent under the skills-lane PM)
File surface: docs/qa/platform-checklist/SWEEP.md only. Scope per triage comment 5515013920 — correct step 6 to name the qa-run GitHub issue destination per runs/README.md and RUNNER.md. Not touching runs/.gitignore or runs/README.md.
Channel: repo-scoped REST probe returned HTTP 200 on this seat, so reads and writes go through REST; MCP reserved for the dedup search only.

Generated by Claude Code

---

Generated by Claude Code