Codex Oracle

Use OpenAI Codex CLI as a planning oracle and code reviewer. Codex provides analysis and recommendations; Claude synthesizes and presents to the user.

Critical: This skill is for planning and review ONLY. Never use Codex to implement changes.

Agent Compatibility

Claude Code only. This skill depends on Claude Code features (ExitPlanMode, AskUserQuestion, plan mode). If you are not Claude Code — e.g., you are Codex CLI, GPT, or another agent — stop immediately and inform the user that this skill requires Claude Code.

Prerequisites

Before invoking Codex, validate availability:

scripts/check-codex.sh

If the script exits non-zero, display the error message and stop. Do not proceed without Codex CLI.

For codex exec, prefer the wrapper script (handles flag compatibility and surfaces errors):

scripts/run-codex-exec.sh

Configuration Defaults

Model Allowlist

Only the following models may be passed to -m. Reject any model not in this list — do not infer or substitute models from the user’s prompt (e.g., “o3”, “gpt-4o”).

Timeout Guidelines

When invoking codex exec via the Bash tool, always set an appropriate timeout:

• Minimum: 5 minutes (300000ms) for any Codex operation
• Simple queries (single file review, focused question): 5 minutes (300000ms)
• Moderate complexity (multi-file review, feature planning): 10 minutes (600000ms)
• High complexity (high reasoning): 15 minutes minimum (900000ms)
• Maximum complexity (xhigh reasoning): 20 minutes (1200000ms)

Reasoning Effort Guidelines

Select reasoning effort based on task complexity:

Selection heuristics:

• low: Task involves <3 files, simple question, or quick validation
• medium: Task involves 3-10 files or requires moderate analysis
• high: Task spans multiple modules or requires architectural thinking
• xhigh: Task requires comprehensive codebase understanding or critical decisions

For available models and reasoning levels, consult references/codex-flags.md.

Workflow

1. Validate Prerequisites

Run the check script. On failure, report the installation instructions and abort.

2. Determine Mode

• Planning mode: User wants architecture, implementation approach, or design decisions
• Review mode: User wants code analysis, bug detection, or improvement suggestions

Plan mode detection: If Claude Code is running in plan mode (a plan file path was provided in the system prompt), the review target changes:

• Normal mode → review uncommitted changes (git diff + git diff --staged)
• Plan mode → read the plan file and review its contents instead of uncommitted changes

3. Construct Prompt

Build a focused prompt for Codex based on mode:

Planning prompt template:

Analyze this codebase and provide a detailed implementation plan for: [user request]

Focus on:
- Architecture decisions and trade-offs
- Files to create or modify
- Implementation sequence
- Potential risks or blockers

Do not implement anything. Provide analysis and recommendations only.

Review prompt template (normal mode — uncommitted changes):

Review the following code changes for:
- Bugs and logic errors
- Security vulnerabilities
- Performance issues
- Code quality and maintainability
- Adherence to best practices

[git diff output]

Provide specific, actionable feedback with file locations and line references.

Review prompt template (plan mode — plan file):

Review the following implementation plan for:
- Completeness and feasibility
- Architectural soundness
- Missing edge cases or risks
- Ordering and dependency issues
- Alignment with stated goals

[plan file contents]

Provide specific, actionable feedback on the plan. Identify gaps, risks, and improvements.

4. Assess Complexity and Execute Codex

Before executing, assess task complexity to select appropriate reasoning effort:

1. Count files involved in the query
2. Evaluate scope (single module vs cross-cutting)
3. Consider depth (surface review vs architectural analysis)

Use HEREDOC syntax for safe prompt handling. Always use the Bash tool’s timeout parameter (minimum 300000ms / 5 minutes).

Redirect output to a temp file to avoid context bloat and race conditions.

Step 1: Generate a unique temp file path using $RANDOM for randomness:

CODEX_OUTPUT="/tmp/codex-${RANDOM}${RANDOM}.txt"

Step 2: Execute Codex via the wrapper (detects supported flags and surfaces stderr):

# Select EFFORT based on complexity assessment (low/medium/high/xhigh)
# Bash tool timeout: 5-20 minutes based on complexity
MODEL="${MODEL:-gpt-5.3-codex}" \
EFFORT="${EFFORT}" \
CODEX_OUTPUT="$CODEX_OUTPUT" \
scripts/run-codex-exec.sh <<'EOF'
[constructed prompt]
EOF

Important flags:

• -m: Model selection
• -c model_reasoning_effort=: Reasoning depth (low/medium/high/xhigh) — select based on Reasoning Effort Guidelines
• -s read-only: Prevents any file modifications (non-negotiable)
• --skip-git-repo-check: Works outside git repositories (exec-specific)
• -o <file>: Write output to file (cleaner than shell redirection)
• --search: Enable web search for queries needing current information (optional)
• 2>/dev/null: Suppresses stderr (progress/thinking output)

Bash tool timeout: Estimate based on task complexity (see Timeout Guidelines above). Never use the default 2-minute timeout for Codex operations.

5. Present Codex Output

Step 3: Read the analysis from the temp file and display to the user:

cat "$CODEX_OUTPUT"

Format the output with clear attribution:

## Codex Analysis

[Codex output from the temp file]

---
Model: gpt-5.3-codex | Reasoning: [selected effort level]

For very large outputs (>5000 lines), summarize key sections rather than displaying everything.

6. Synthesize and Plan

After presenting Codex output:

1. Synthesize key insights from Codex analysis
2. Identify actionable items and critical decisions
3. If Codex’s analysis presents multiple viable approaches or significant trade-offs, consider using AskUserQuestion to clarify user preferences before finalizing the plan
4. Write a structured plan
5. Call ExitPlanMode to present the plan for user approval

When to use AskUserQuestion:

• Codex proposes multiple architectures with different trade-offs
• Technology or library choices need user input
• Scope decisions (minimal vs comprehensive) are ambiguous

Skip clarification when:

• Codex’s recommendations are clear and unambiguous
• User’s original request already specified preferences
• Only one viable approach exists

Error Handling

The check script returns specific exit codes:

The codex exec command returns:

Runtime errors:

Usage Examples

Planning Request

User: “Ask Codex to plan how to add authentication to this app”

1. Validate Codex CLI available
2. Gather relevant codebase context
3. Assess complexity → auth spans multiple modules → high reasoning
4. Construct planning prompt with auth requirements
5. Execute Codex with gpt-5.3-codex and high
6. Present Codex’s architecture recommendations
7. Synthesize into Claude plan format

Code Review Request

User: “Have Codex review the changes in src/auth/”

1. Validate Codex CLI available
2. Read files in src/auth/ directory
3. Assess complexity → single directory, focused review → medium reasoning
4. Construct review prompt with file contents and diff context
5. Execute via run-codex-exec.sh with EFFORT=medium
6. Present findings with file/line references
7. Summarize critical issues and recommendations

Additional Resources

Reference Files

• references/codex-flags.md – Complete model and flag documentation

Scripts

• scripts/check-codex.sh – Prerequisite validation (run before any Codex command)