codex

What it does

The skill wraps the OpenAI Codex CLI to provide three operational modes: a code review mode that gates outputs with a pass/fail verdict, a challenge/adversarial mode that attempts to find failures or edge cases, and a consult mode to ask Codex questions with session continuity. It supports invoking either the Codex CLI tools directly (e.g., codex review, codex exec) or prompting Codex for focused analysis.

How it works

• It starts by ensuring the Codex binary is available and reports if it is not.
• It parses user input to determine the mode:
  • /codex review or /codex review <instructions> triggers Review mode, which runs:
    • Create a temp file for error capture: TMPERR=$(mktemp /tmp/codex-err-XXXXXX.txt)
    • Execute codex review --base <base> -c 'model_reasoning_effort="high"' --enable web_search_cached 2>"$TMPERR" (with a 5-minute timeout on the Bash call)
    • If user provided extra instructions, pass them as the prompt argument to codex review.
    • Parse the output cost from TMPERR by grepping for "tokens used"; if not found, tokens: unknown.
    • Determine gate verdict: if output contains [P1], gate is FAIL; else gate is PASS.
    • Present the full codex output verbatim within a CODEX SAYS block and append a GATE line with PASS/FAIL and token/cost info.
    • If a prior /review from Claude exists in the session, perform a cross-model comparison showing overlapping findings.
    • Persist the review result to ~/.claude/skills/gstack/bin/gstack-review-log with fields: skill "codex-review", timestamp, status (clean or issues_found), gate (pass or fail), and findings count.
    • Clean up temporary files.
  • /codex challenge or /codex challenge <focus> triggers Challenge mode, which constructs an adversarial prompt to identify failures, then runs codex exec with JSONL output and a 5-minute timeout. It captures reasoning traces and tool calls, extracting potential failure points and presenting them verbatim.
  • /codex with no arguments triggers Auto-detect: checks for a diff against the base branch. If a diff exists, asks the user to choose among reviewing, challenging, or prompting for a different task. If no diff, searches plan files for project-scoped matches and may propose reviewing a plan or asking Codex a question.
  • /codex <anything else> triggers Consult mode, where the remaining text is treated as the prompt to Codex.

When to use it

• Use when you want an independent, brute-force second opinion on code changes via Codex.
• Use code review mode to gate quality with a formal pass/fail, especially when you need deterministic acceptance criteria.
• Use challenge mode to aggressively test for edge cases, security issues, or robustness concerns.
• Use consult mode for a general Codex-based Q&A with session continuity.

What it can touch

• Commands: "codex review" and "codex exec" with the needed flags and prompts.
• Files: creates a temporary error file TMPERR and writes to ~/.claude/skills/gstack/bin/gstack-review-log depending on results.
• Output: prints the full Codex output verbatim within a structured CODEX SAYS block, then a GATE result.

Caveats

• Relies on the presence of the Codex CLI binary; if not found, the workflow stops with a user-facing message directing installation.
• Gate verdicts depend on presence of "[P1]" in Codex output for FAIL; otherwise, PASS.
• Cross-model comparison depends on prior Claude review output in the session.
• Logging to gstack-review-log uses a specific JSON structure and may require environment setup to persist files.