AgentSkillsCN

subcodex

通过命名会话与 Codex（GPT-5.2）展开协作。触发方式：用户提出希望与 Codex 合作、使用 Codex、向 Codex 提问、让 Codex 参与等。 重要提示：请通过 Bash 使用 Subcodex CLI，而非 Task 子代理工具。

SKILL.md
--- frontmatter
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

bash
# 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:

  1. •Tell the user you're waiting for Codex
  2. •Stop immediately — don't poll or check
  3. •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:

bash
./skills/subcodex/subcodex resume my-task-x7k2 "continue with..."

subcodex list shows: status (running/stopped), tool_calls, cwd, duration.

Code Reviews

bash
# 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:

  1. •Claude proposes approach
  2. •Codex reviews/critiques
  3. •Claude decides
  4. •Codex implements
  5. •Claude reviews result

Push for simple code: less abstraction, fail-fast, no "just in case" code.