Codex Assistant

Execute tasks using OpenAI Codex CLI and analyze results.

Important Limitations

Codex runs in read-only sandbox mode. It cannot:

• Create or modify files
• Execute scripts or commands that write to disk
• Run generated code

Always ask Codex to output results directly to stdout. Do not request file creation or modification - instead, ask Codex to print/output the code or results, which Claude can then save if needed.

Command

codex exec --model gpt-5.2-codex "PROMPT"

Prompt Format

Structure prompts with three sections:

# Role
[Describe the role/persona for the task]

# Context
[Include relevant information:]
- File paths and links to documentation
- Memory files or conversation context
- Additional task-relevant details

# Task
[Clear task description]
[Expected deliverables]

Workflow

1. Construct prompt - Build structured prompt with Role, Context, Task
2. Execute codex - Run command with the formatted prompt
3. Analyze output - Review results for:
   • Completeness of task execution
   • Quality and accuracy of output
   • Issues or recommendations identified
   • Actionable items to address

Example

codex exec --model gpt-5.2-codex "# Role
Senior code reviewer with expertise in security and best practices.

# Context
Repository: /path/to/project
Files changed: src/auth.py, src/api.py
Branch: feature/user-auth

# Task
Review changes against develop branch.
Search for security issues, bugs, and suggest improvements.
Output your findings and recommendations to stdout."

Note: Always phrase tasks to output results rather than create/modify files. For code generation, ask Codex to "print the code" or "output the script" instead of "create a file".

Output Analysis

After codex completes, evaluate:

• Findings - Key issues or insights discovered
• Recommendations - Suggested improvements or fixes
• Action items - Concrete next steps to implement
• Quality - Overall assessment of the analysis