Codex Delegate

# Codex Delegate

Use this skill to decide **whether** Codex should be used, **what** part should be delegated, and **what must remain with the main agent**.

## Core rule

Codex is a **specialist executor**, not the long-term owner of the human relationship.
Use Codex for heavy technical work, isolated analysis, and context offloading.
Keep final judgment, risk trade-offs, environment acceptance, and user-facing conclusion in the main agent.

## When to delegate to Codex

Delegate when the task is:
- code-heavy
- repo-heavy
- multi-file
- CLI-heavy
- deep review / audit work
- likely to consume too much token/context in the main thread
- better handled in an isolated, clean context

Typical examples:
- repository analysis
- skill/package review
- deep code review
- bug investigation
- patch drafting
- dependency and config analysis
- log/error analysis

## When NOT to delegate fully

Do not fully hand off tasks that depend on:
- long-term memory
- user preference and tone alignment
- final risk judgment
- environment acceptance across shell / gateway / cron / LaunchAgent
- upgrade timing and change management
- whether a temporary workaround becomes the official solution

Codex may assist, but the main agent must still decide.

## Delegation split

### Codex should do
- static inspection
- file and repo structure analysis
- code/package consistency checks
- focused error analysis
- technical option comparison
- isolated patch or remediation proposals
- heavy token-consuming analysis that would otherwise bloat the main session

### Main agent should do
- define the real goal
- set scope and constraints
- choose whether Codex is the right tool
- evaluate user risk tolerance and workflow fit
- validate dynamic/runtime reality
- decide whether something is truly "done"
- deliver the final recommendation in the human's context

## Validation rule

Never confuse these layers:
- reviewed by Codex
- technically plausible
- works in current shell
- works in gateway
- works...