Codex Companion

Codex Companion

OpenAI's official Codex plugin for Claude Code, adapted for the Agentbox container. Delegates code review, adversarial review, and rescue tasks to GPT-6 Astra via the Codex CLI.

When to Use This Skill

• Code review: Get a Codex review of your git changes (working tree or branch)
• Adversarial review: Challenge your implementation approach and design choices
• Rescue: When Claude is stuck, hand the task to Codex for a second opinion
• GPT-6 Astra delegation: Offload substantial coding, debugging, or research tasks
• Cross-model validation: Use Codex as a check on Claude's work

When Not to Use

• For simple tasks Claude can finish quickly — don't add Codex overhead
• For non-code tasks (research, content, design) — use dedicated skills instead
• For DeepSeek reasoning — use deepseek-reasoning skill instead
• If OPENAI_API_KEY is not configured — run /codex:setup first

Container Integration

Codex runs under the current tmux profile — there is no separate pseudo-user (the openai-user/as-openai/UID-1002 model is retired). The baked Codex CLI reads its credentials from CODEX_HOME (~/.codex), where the entrypoint wires [consultants.codex] auth and the agentbox-memory MCP server on first boot. Set OPENAI_API_KEY in the environment only if you are overriding that baked auth.

# Codex commands run in-place; no user switch needed
/codex:review              # Review working tree changes
/codex:adversarial-review  # Challenge the implementation approach
/codex:rescue [task]       # Hand a stuck task to Codex
/codex:status              # Check Codex job status
/codex:result              # Get last Codex result
/codex:cancel              # Cancel running Codex job

Commands

Rescue Agent

The codex-rescue subagent activates proactively when:

• Claude Code is stuck on a debugging task
• A second implementation pass would help
• Deeper root-cause investigation is needed
• A substantial coding task should be delegated

The rescue agent uses this skill's bundled prompting reference library (references/codex-prompting/ — prompt blocks, recipes, and antipatterns) to compose tight, structured Codex prompts with XML tags, output contracts, and verification loops.

Codex Prompting Patterns

The skill includes a reference library for composing effective Codex prompts:

Review Output Schema

Reviews follow a structured JSON schema (review-output.schema.json) with:

• Severity levels (critical, major, minor, suggestion)
• Category classification (bug, security, performance, design, style)
• File + line references
• Confidence scores

Hooks

• SessionStart/SessionEnd: Lifecycle management for Codex companion state
• Stop: Review gate — optionally triggers a Codex review before session ends

Environment Variables

Setup

# Auth is wired by the entrypoint (~/.codex / [consultants.codex]); no setup
# is normally needed. To (re)configure the Codex CLI or supply your own key:
/codex:setup

# Verify:
/codex:status

Two consultation paths (addendum 2026-09-08)

The consultant model is gpt-6-astra. Path 1 (MCP consultant-codex → consult) takes a pasted excerpt; Path 2 (direct codex exec) reads a whole repository or document bundle itself. Full comparison table, the verified Path 2 recipe, and the gotchas that cost a run each (read-only sandbox, approval policy, stdin hang, "Context I still need"): references/codex-exec-recipe.md.

Runtime note

The /codex:* commands run from the baked plugin at /opt/agentbox/plugins/codex-plugin-cc; behavioural fixes go through the flake overlay, not this skill.

Integration with Other Skills

Attribution

Codex Plugin for Claude Code by OpenAI. Apache 2.0 License.