codex

name	codex
description	Translates task requirements into Codex CLI commands. Used by codex-driver agent to execute coding tasks via OpenAI Codex.

Codex Skill Guide

PATH Setup (CRITICAL)

Subagents run with a minimal PATH. Always include homebrew path:

export PATH="/opt/homebrew/bin:$HOME/.local/bin:$PATH" && codex <args>

Baseline Rules

Always apply these to every codex exec command:

• --skip-git-repo-check — required for all commands
• 2>/dev/null — suppresses thinking tokens (omit only if debugging)
• --full-auto — required when using workspace-write or danger-full-access sandbox

Command Templates

New Task

export PATH="/opt/homebrew/bin:$HOME/.local/bin:$PATH" && codex exec --skip-git-repo-check --sandbox <MODE> [options] "<prompt>" 2>/dev/null

Resume Session

export PATH="/opt/homebrew/bin:$HOME/.local/bin:$PATH" && echo "<prompt>" | codex exec --skip-git-repo-check resume --last 2>/dev/null

Note: When resuming, do not include model/sandbox flags — the session inherits its original settings. Flags must go between exec and resume.

Sandbox Modes

Task Type	Mode	Notes
Analysis, review, Q&A	--sandbox read-only	No file modifications allowed
Create or edit files	--sandbox workspace-write --full-auto	Standard for most coding tasks
Network access needed	--sandbox danger-full-access --full-auto	Confirm with user first

Model Selection

When the calling agent specifies requirements, translate to flags:

Requirement	Flag	Notes
Default	-m gpt-5.2-codex	Optimized for code tasks
General-purpose	-m gpt-5.2	Non-code or mixed tasks

Complexity	Flag	Notes
Complex / multi-step	--config model_reasoning_effort="xhigh"	Architectural decisions, large refactors
Standard (default)	--config model_reasoning_effort="high"	Most coding tasks
Simple	--config model_reasoning_effort="medium"	Straightforward edits
Trivial	--config model_reasoning_effort="low"	Quick fixes, formatting

If not specified, use defaults: gpt-5.2-codex with high reasoning.

Interpreting Results

Success indicators

• Codex reports files created/modified
• Output describes completed actions matching the request
• No error messages or stack traces

Failure indicators

• Non-zero exit code
• Error messages in output
• "I cannot" or "I'm unable to" language
• Partial completion ("I've started but...")

Scope creep indicators

• Mentions of "I also..." or "While I was at it..."
• Changes to files not mentioned in the original request
• Added features, tests, or documentation not requested

Redirection indicators

• Output describes different work than requested
• "Instead of X, I did Y..."
• Solving a different problem than specified

After Completion

Report to user: "You can resume this Codex session by saying 'codex resume'."

Error Handling

• If command exits non-zero: stop and report the error
• If output contains warnings: summarize and ask how to proceed
• Before danger-full-access: confirm with user unless pre-authorized