subcodex
Collaborate with Codex (GPT-5.2) via named sessions. Trigger: user asks to work with Codex, use Codex, ask Codex, involve Codex, etc.CRITICAL: Use subcodex CLI via Bash, NOT the Task subagent tool.
When & Why to Use This Skill
Subcodex is an advanced agentic skill that enables seamless collaboration between Claude and Codex (GPT-5.2) through persistent, named sessions. It is specifically designed to handle long-running, complex programming tasks by offloading execution to a background process, allowing Claude to manage high-level strategy while Codex handles direct file system modifications, code generation, and deep architectural reviews.
Use Cases
- Executing complex, time-intensive coding tasks (up to 30+ minutes) in the background without interrupting the user's workflow.
- Conducting high-reasoning code reviews on uncommitted changes or specific git branches to identify security vulnerabilities and over-engineering.
- Maintaining persistent development context across multiple conversation sessions using named sessions (e.g., 'feature-migration-x7k2').
- Implementing large-scale refactoring or new feature sets where Claude acts as the lead architect and Codex serves as the primary execution engine.
- Automating system-level changes and script executions within a controlled, named sandbox environment.
| name | subcodex |
|---|---|
| description | | |
| Collaborate with Codex (GPT-5.2) via named sessions. Trigger | user asks to work with Codex, use Codex, ask Codex, involve Codex, etc. |
| CRITICAL | Use subcodex CLI via Bash, NOT the Task subagent tool. |
subcodex
Named Codex sessions that persist across conversation compacts.
Quick Reference
# New session
./skills/subcodex/subcodex new <name> "prompt"
# Resume session
./skills/subcodex/subcodex resume <session-name> "prompt"
# List sessions (shows last 25, use --all for all)
./skills/subcodex/subcodex list
# Code review
./skills/subcodex/subcodex review --uncommitted <name>
./skills/subcodex/subcodex review --base main <name> "focus on security..."
Critical: Background Execution
ALWAYS use run_in_background: true — Codex takes 1-30+ minutes.
After starting a background command:
- Tell the user you're waiting for Codex
- Stop immediately — don't poll or check
- You'll be notified via
<task-notification>when Codex is done
Codex Capabilities
Codex can read, write, and execute — just like you. Tell it to make changes directly, not to provide diffs or suggestions.
Options:
--reasoning low|medium|high|xhigh(default: high)--read-only— read-only sandbox--dangerous— full system access
Sessions
Session names get a random suffix: my-task becomes my-task-x7k2.
Use the full name (with suffix) when resuming:
./skills/subcodex/subcodex resume my-task-x7k2 "continue with..."
subcodex list shows: status (running/stopped), tool_calls, cwd, duration.
Code Reviews
# Review uncommitted changes
./skills/subcodex/subcodex review --uncommitted my-review
# Review against branch
./skills/subcodex/subcodex review --base main my-review
# Review specific commit
./skills/subcodex/subcodex review --commit HEAD my-review
# Add instructions
./skills/subcodex/subcodex review --uncommitted my-review "read these relevant docs first, focus on security, suggest simplifications, .."
Reviews use --reasoning xhigh and read-only mode by default.
Collaboration
Claude leads, Codex executes. Use Codex to review designs and implement complex tasks. Push back on over-engineering.
Typical flow:
- Claude proposes approach
- Codex reviews/critiques
- Claude decides
- Codex implements
- Claude reviews result
Push for simple code: less abstraction, fail-fast, no "just in case" code.