Skip to content

fix(review-anvil): safe copy-editor launches and evidence reconciliation - #12

Closed
mrshu wants to merge 1 commit into
mainfrom
fix/review-anvil-evidence-and-audit-launch
Closed

mrshu wants to merge 1 commit into
mainfrom
fix/review-anvil-evidence-and-audit-launch

Conversation

@mrshu

@mrshu mrshu commented Oct 6, 2026

Copy link
Copy Markdown
Owner

Changes

  • Add run-reviewer.sh --codex-copy-editor for Codex clarity renderers, action-lock auditors, and bounded repairs. The wrapper owns command construction: fresh private empty workdir, --skip-git-repo-check, ephemeral read-only execution, and existing model/reasoning settings. Cleanup covers success, failure, empty output, and timeout.
  • Route the documented renderer/auditor/repair calls through that mode. Keep semantic auditing, bounded repair, and exact-source fallback intact. Host configuration and environment still apply; this is not additional isolation beyond the Codex sandbox.
  • Tighten the existing synthesis check around base/head behavior, applicable contract, reachable boundaries, and consequence-based severity. Deduplicated severity is provisional; reviewer agreement cannot supply missing evidence. Newly exposed pre-existing defects remain in scope.
  • Replace incidental source/wording assertions with filesystem-backed lifecycle coverage. Add no reviewer waves, finding-schema fields, or clarity wording rules.

Verification performed

  • Actual Codex semantic audit through the new mode returned STATUS=ok; the private workspace was removed. The supplied diagnosis and requested-work preservation checks passed.
  • Wrapper lifecycle tests passed, including success/failure/empty/timeout cleanup and invalid-usage/workspace-creation boundaries.
  • Existing reproduction-policy fixtures passed.
  • All 29 clarity-output validator tests passed.
  • Skill frontmatter and manifest checks passed. Frontmatter validation used the existing CI dependency, PyYAML 6.0.3, through an isolated uv invocation.

Synthesis evaluation and limits

Retain sanitized evidence in evaluations/2026-10-06/synthesis-replay.json: ten feedback-derived/synthetic cases, two independent paired runs, and 40 stateless model calls. Cases cover intentional contracts, unrelated unchanged defects, newly exposed old defects, unsupported consensus, public APIs, developer tooling, coverage obligations, and proportional severity.

Both baseline and proposed policies passed all 20 decisions each within the predeclared disposition/severity ranges. No decision-quality improvement was demonstrated. The synthesis change consolidates existing requirements; the launch fix has direct runtime evidence. The replay is excerpt-level classification, not end-to-end review or a population accuracy estimate. Historical author claims remain unverified, synthetic facts are labeled, and the resolved backend identity behind the default model alias was unavailable.

No Hermes deployment or installed-skill changes. The unrelated local plan is excluded.

@mrshu mrshu closed this Oct 6, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant