codex-review-cycle
Codex Review Cycle
Overview
A simple, user-driven review-and-fix workflow. Every cycle runs one codex review, Claude verifies each finding's validity, presents a verbatim summary, and the user picks which findings to address. Claude then applies only the chosen fixes and loops. The default cap is 2 cycles; after the final cycle's fix phase Claude renders a §Final-cycle Assessment block covering addressed findings, scope-health warnings, and a continue / new-angle / end recommendation. The user decides whether to terminate, run another cycle with the same focus, or run another cycle with an <angle_request> asking codex for a materially different angle. Each user-elected extension cycle ends with the same prompt, so the loop continues only as long as the user keeps electing to extend.
The skill is deliberately simple. It does not auto-fix, does not run parallel reviewers, does not compute stall fingerprints, does not manage autonomy bands, and does not delegate to rescue subagents. Claude is the only fix applier; the user is the only arbiter of which findings matter.
Language
All user-facing output is rendered in the user's language (the language the user has been using in the conversation, or as configured in the Claude Code system-level language setting). This section is the authoritative translation contract — any per-language sample reference (e.g. references/summary-samples.ja.md) is illustrative only and MUST NOT contradict these rules.
Translate into the user's language:
- Section headings (e.g.
Cycle N review summary,Active stop signals,Review assessment,Final-cycle assessment,Findings addressed in this run,Outstanding findings,Claude's judgment,Final-cycle decision) - Per-finding bullet labels (
File,Claude's note,Recommended action,codex recommendation (verbatim)) and stop-signal table column headers (Signal,Status,Evidence) - §Final-cycle Assessment per-fix sub-bullet labels (
Invariant,Surfaces checked,Residuals,Next-cycle attack) and judgment field labels (Trend,Residuals summary,Scope-health,Verification gap,Recommendation), plus the scope-health values (healthy,warning) - Free-text fields Claude authors:
Claude's notebody,Recommended actionvalues, fleet-rate warnings, stop-signal footer prose, termination messages, post-cycle review assessment, §Final-cycle Assessment judgment-field bodies, scope-health trigger prose, and recommendation rationale - Verdict headline free-text (see
references/termination.md§Verdict Headline): the three verdict keyword labels (Clean termination,Terminated with residuals,Cap reached), the segment wordlets (cycles,fix(es) applied,applied,residual,unresolved,trend), and the three trend keyword values (converging,stable,cascading)
More from adhi-jp/agent-skills
review-scope-guard
Use when Codex or code/plan review findings need Definition-of-Done scope triage, especially to separate must-fix issues from minimal hygiene, out-of-scope hardening, repeated noise, or self-induced refinements. Also use when codex-review-cycle invokes scope guard between validity checking and summary render. Do not use for single-shot lint or unrelated changes.
34writing-style-guide
Use when generating or editing user-facing prose, including docs, comments, READMEs, changelogs, commit messages, PR descriptions, and chat replies.
32minecraft-modding-workbench
>
22review-fix-cascade-guard
Use when applying user-selected review findings after scope triage, especially inside codex-review-cycle or after a standalone Codex adversarial review, and a line-scoped fix may create follow-on valid findings. Do not use for ordinary edits, plan drafting, single-shot lint, no-review contexts, or generic refactor checklists.
19vibe-planning
>
17vibe-plan-execution
Use when the user asks to execute, implement, continue, or apply an existing implementation plan, specification, acceptance criteria, task plan, or vibe-planning output. Do not use for plan creation or coding requests with no concrete plan to bind.
17