review-adjudication
Verify generated or human review findings before fixing, defending, replying, resolving, or reusing them as learned guidance. Use within a review, review-comment fix, monitor, or readiness workflow.
- License 许可
- MIT
- 最近更新
- 2026-10-23
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- review-adjudication
- description
- Verify generated or human review findings before fixing, defending, replying, resolving, or reusing them as learned guidance. Use within a review, review-comment fix, monitor, or readiness workflow.
03 SKILL.md 正文
Review adjudication
Use one evidence gate for findings from every reviewer. This procedure is extracted from the existing preflight workflow; it does not replace that workflow's specialist checks, acceptance thresholds, or coverage requirements. Review-only requests stay read-only. Existing authorization and repository instructions govern edits, commits, replies and resolutions.
Verify the claim and each remedy
Separate generation from judgment. For every candidate, record a stable ID, exact base/head or working-diff identity, affected contract/callers, proposed remedy and concrete failure sequence. Recheck the diff, surrounding implementation, tests, repository rules and verified product intent; actively seek counterevidence. Reviewer confidence, a resolved thread or an initial corrective commit is not proof.
Split a comment that proposes multiple remedies into independently judged child IDs. A real defect does not validate every suggested change. Merge same-defect reports only when one correction covers the same failure mechanism, retaining reviewer/source IDs and all affected locations.
Assign each claim/remedy:
FIX: real, in scope and sufficiently evidenced; apply the smallest root-cause correction when authorized.DEFEND: a plausible reviewer claim or remedy is wrong, inapplicable, already handled or conflicts with verified intent; preserve sound code and prepare a concise evidence-backed reply.DISCARD: duplicate, withdrawn or rejected by the applicable review guardrails; retain the reason and any link to the surviving claim.CLARIFY: missing intent, ownership, runtime or contract evidence could change the decision; investigate or ask a focused question before editing that claim.
When challenged, reopen the judgment and show supporting evidence, counterevidence, current decision and what fact would change it. A remembered judgment is a lead, never authority. Revalidate against current code, requirements, tests and reviewer baseline before reuse.
Validate and retain the outcome
Apply only authorized FIX remedies. Run proportionate checks and inspect the final diff, including sibling callers. Mark a remedy FIXED only after its correction is present and validation passes at the recorded final revision. A failed or unavailable check remains explicit; it cannot become a successful validation record.
Track each remedy separately through proposed, applied, validated, retained, rejected, removed or reverted outcomes. After later commits/review, compare the final retained code with the proposed correction. Keep rejected/reverted outcomes and superseding evidence; do not silently overwrite history or count an initial fix, thread resolution or accepted sibling remedy as final acceptance.
Only a source-verified, validated correction retained in the final reviewed revision can support a learned positive rule. If final retention is unavailable, record it as unverified and defer that learning. Rejected or reverted suggestions remain counterexamples with their actual scope; they do not imply a universal ban. Future guidance changes still require their own approval and evaluation.
Keep one local ledger
Resolve git rev-parse --git-common-dir to an absolute path. Read existing adjudication ledgers under its codex/ directory and reuse the applicable ledger, including a workflow's legacy ledger path. If none exists, use <git-common-dir>/codex/review-adjudications.md. Keep it outside tracked files; create it only for finalized reusable judgments. If storage is unavailable, disclose session-only retention.
Record finalized DEFEND, validated FIXED and materially reusable DISCARD outcomes, including rejected/reverted remedies. Keep active FIX/CLARIFY in the run report until settled. Each ledger record needs:
- Repository identity, stable parent/child IDs, semantic claim signature, symbols/contracts, reviewer/prompt family and baseline.
- Exact reviewed revision, decisive evidence and counterevidence, checks/results, relevant confirmed intent and invalidating conditions.
- Judgment, per-remedy lifecycle/final retention, source/comment/fix/final-revision references, date and superseded-entry link when applicable.
- Prepared reply when useful. Store concise reasoning and references, never credentials, secrets, PII or raw logs.
Act and report within authorization
Draft replies from current evidence. Post or resolve only when the user has authorized those actions; human-reply restrictions from the calling workflow still apply. Claim a remote fix only after the validated correction is present at the published revision. Resolve a mixed thread only when every child remedy is settled; an unresolved child stays visible.
Report unresolved FIX as NOT READY, unresolved CLARIFY or missing required validation as BLOCKED, and verified DEFEND separately as a possible false-positive comment risk. Never relabel a defense as a fixed defect. Include the ledger path, reused/added/superseded IDs, per-remedy outcomes and remaining evidence gaps. The calling workflow's stricter readiness and review gates still apply.
04 同仓库更多 Skills
sample-skill
暂无简介
synthetic-review-candidate
Candidate for smoke testing an evaluation runner only.
synthetic-review-baseline
Baseline for smoke testing an evaluation runner only.
review
Run the existing user-invoked review bundle and independently adjudicate its findings.
monitor-mr
Monitor review findings and CI using the shared adjudication procedure.
migrate-to-codex
Migrate supported instruction files, skills, agents, and MCP configuration into Codex project and global files. USE FOR: move existing agent definitions and MCP setup into Codex global configuration, migrate supported skills into Codex project files, convert supported Claude Code configuration to Codex. DO NOT USE FOR: PostgreSQL database migrations, moving a database schema to a new server, unrelated data migrations.