Skip to content

feat: improve review recovery and migrate Codex roles - #91

Merged
bioinformatist merged 1 commit into
mainfrom
codex/issue89-review-recovery
Nov 20, 2026
Merged

bioinformatist merged 1 commit into
mainfrom
codex/issue89-review-recovery

Conversation

@bioinformatist

Copy link
Copy Markdown
Owner

Summary

Improve now constrains review reports to the frozen plan and retains reports and diagnostics when a turn fails or is cancelled. A confirmed interrupted implementation can reach independent review once the existing connection, stopped writers, exact candidate and required evidence have been verified; a failed or unknown review still cannot grant approval.

This addresses the requirements in #89:

Issue coverage Change
1–4: review inputs and recovery Plan-derived schema/report constraints with specific coverage errors; bounded report preservation, including disconnects before the start response; interrupted-implementation eligibility; one previous-review snapshot and an explicit continuation dossier.
5–6 and 7a: planning acceptance Trace the hardest acceptance case before READY, distinguish installed bootstrap capabilities from delivered capabilities, and verify native planning tools without spending model quota on dedicated experiments.
7–8: Codex and role configuration Pin Codex 0.170.0, use 6.2 Sol for implementation/scouting and 6.2 Astra for writing/review/visual roles, while retaining Luna economy. Remove the ineffective managed personality setting and validate actual generated/merged configuration against the pinned schema. Update bilingual guidance and routing diagrams.
9: source audit and proof boundaries Retain the seven third-party pins after checking imported paths. Protect proof specifications and necessary dependencies, distinguish development-wrapper cache results from direct proof verdicts, and preserve the historical Bend/Lean evidence.

The combined protocol fixtures preserve the reviewed report-recovery cases alongside the new model and generated-role assertions. Overlapping planning and workflow text retains both sets of requirements. Generated skills come from the canonical sources.

Verification

  • nix run .#sync-vendored-skills -- --check
  • nix flake check --allow-import-from-derivation on the captured combined export
  • tests/plugin-portability.bash on the packaged combined runtime in a temporary Home
  • Generated diffs, bilingual docs, static scenarios and SVG renders independently reviewed
  • No credentials, private identifiers, absolute home paths or Nix store paths added to distributable files

The four newly bound English/Chinese previews at 1200 and 720 pixels are byte-identical to the previously approved images. Unchanged source audits and input facts were reused; the changed runtime/configuration/fixture combination received fresh checks.

Independent combined review returned APPROVE with no findings and complete scope, check and review coverage. Commit b11737e6ba69053993166651742b4e4ff5ce0e5a contains exactly the reviewed tree 47dd3484d70854cf308075385463470e45b3f585. GitHub Actions check and plugin-portability remain the final CI acceptance. Dedicated model-quality, quota-limit and proof-runtime experiments were not run; live model capabilities and newer Bend/Lean behavior are not claimed as tested.

Closes #89.

@bioinformatist
bioinformatist merged commit 243067a into main Nov 20, 2026
2 checks passed
@bioinformatist
bioinformatist deleted the codex/issue89-review-recovery branch November 20, 2026 06:35
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.

改进 Improve 的独立审查、中断恢复与计划验收(Plan 042 复盘)

1 participant