ray0907 Codex Result Handling
Use this skill when presenting Codex output to the user after a codex-rescue or review run.
General rules
- Preserve the verdict, summary, findings, and next-steps structure from Codex output
- Use file paths and line numbers exactly as Codex reports them — do not normalize or guess
- Keep evidence boundaries intact: if Codex marked something as an inference, uncertainty, or open question, preserve that distinction
- If there are no findings, say so explicitly and keep any residual-risk note brief
- If Codex made file edits, say so and list the touched files when available
- If Codex output is malformed or the run failed, include the most actionable stderr lines and stop — do not guess at what Codex would have said
Review output (codex review / adversarial-review)
Present findings ordered by severity. Then stop completely.
Do not:
- Apply any fixes
- Suggest you are about to make changes
- Silently correct issues mentioned in the review
After presenting findings, ask the user explicitly:
"Which of these issues, if any, would you like me to address?"
Wait for their answer before touching any file. Auto-applying review findings — even obvious ones — is not allowed.
Rescue output (codex-rescue)
- Return Codex stdout verbatim
- Do not turn a failed or incomplete Codex run into a Claude-side implementation attempt
- If Codex was never successfully invoked, do not generate a substitute answer
Setup / auth errors
If Codex output indicates it is not installed or not authenticated, direct the user to run:
npm install -g @openai/codex # if not installed
codex login # if not authenticated
Do not improvise alternate auth flows.