overengineering:audit: make the neutral-name-then-rename write the default for unattended runs and apply it to every per-layer write

[kyle-sexton]

The audit skill requires the findings artifact to be written after each layer, so an interrupted run keeps its completed layers (skills/audit/SKILL.md:155; skills/audit/context/surface-walk.md:101). It also allows a delegated executor to write to a neutral filename and then rename, for harnesses that refuse a report-shaped filename (SKILL.md:76-81). But it presents that route as a conditional fallback. It does not make it the default for unattended runs, even though line 77 names the unattended caller as exactly that case. It also does not say the route applies to every per-layer write, including the re-read-and-merge in SKILL.md:158-163. And nothing in "Before the walk" step 1 (SKILL.md:121-130) tests the write path, so a refusal first shows up at the first write.

Observed on 2026-09-27: an unattended run, dispatched as a general-purpose subagent against melodic-software/.github, was blocked by the harness from writing findings.md directly ("Subagents should return findings as text"). The subagent used the rename route but wrote the artifact once, at the end of the walk. The checkpoint guarantee was lost without anyone noticing.

Proposed fix:

1. When unattended is set, use the neutral-name-then-rename route from the first write.
2. State that every per-layer write, including the re-read-and-merge, goes through that route.
3. Add a write-path probe to "Before the walk" step 1, so a refusal shows up before any layer is walked.

Acceptance: a delegated unattended run leaves a partial artifact on disk after each completed layer, with scope naming only the layers walked so far.

Found in: overengineering 0.4.13.

🤖 Generated with Claude Code

https://claude.ai/code/session_01S5b57FoPJkJMe1BDwhExiG

[kyle-sexton]

This was generated by AI during triage.

Triage: verified, briefed

Verified against main (overengineering 0.4.13). Every cited passage matches:

• The audit skill's "The walk" section requires a per-layer artifact write, re-reading the artifact from disk before each write and merging against that copy. The surface-walk context's "Incremental artifact writes" section says the same and ties scope to completed layers.
• The "Writing the artifact from a delegated run" paragraph names the unattended caller as the case where a harness refuses a report-shaped filename. It still presents the neutral-name-then-rename route as conditional, and says nothing about applying it to every per-layer write.
• "Before the walk" step 1 resolves the branch identity and the artifact home, then emits the read-only opening line. Nothing in it tests whether the executor can write to that home.

The observed failure, a delegated unattended run that wrote once at the end and so lost its checkpoints, is what these gaps allow. Classified as a Bug: the skill promises per-layer checkpoints, and a delegated run does not keep that promise.

Brief

Goal. An unattended audit run whose harness refuses report-shaped filenames still checkpoints after every completed layer.

Required behavior

1. When unattended is set, every findings-artifact write uses the file-write tool: write to a neutral filename in the artifact's directory, then rename it to the contract filename. This starts with the first write and does not wait for a refusal. Attended runs keep the current behavior.
2. The skill states explicitly that this route covers every per-layer write, including the re-read-and-merge write, not only the first or last one.
3. "Before the walk" step 1 gains a write-path probe that runs after the home is resolved and before any layer is walked. It writes a neutral file in the artifact's directory and renames it, then cleans up after itself. On failure, it reports the refusal and stops before walking, the same as the existing "neither route available" stop. The probe must leave no stray file behind, and it counts as part of the one sanctioned artifact write, not as a new auxiliary write.
4. The existing ban on a shell content-write stays.

Acceptance criteria

• A delegated unattended run leaves a partial artifact on disk after each completed layer, and its scope names only the layers walked so far.
• A run whose write path is refused stops before walking the first layer and names the refusal.
• An attended run behaves as it does today.
• The plugin version is bumped per repo convention. Any volatile specific the edited skill body restates follows the skill-body verification-record rule.

Out of scope. The justify and delta lanes and the realign skill, unless they share the same write paragraph (check, do not assume).

Type: Bug. Priority: medium. Nothing is blocked, there is no deadline, and the loss is recoverable by re-running.

[kyle-sexton]

This was generated by AI during work-loop.

Ratify C3 dispatch? The work-loop admission gate (instance melo-lap-001-wsl-2) reads this item (work-class: scoped, type Bug) as C3, bug-fix-shaped. The fix makes the neutral-name-then-rename write the default for unattended overengineering:audit runs and applies it to every per-layer write, following the defaulted decision in the triage brief.

Intended dispatch: one /work-items:work worker implements the brief and opens a draft PR.

The lane has not finished its first drain and this item carries no recorded ratification, so it is queued here instead of dispatched. Ratify through /work-items:attend-queue.

[kyle-sexton]

[kyle-sexton]

This was generated by AI during attend-queue.

Ratified by the operator on 2026-09-27: C3 (work-class: scoped) dispatch approved for autonomous work. Flipping to agent-ready; a worker lane picks it up on its next cycle.