Agent Cli Delegation Operations

When to Use

Use when launching or supervising autonomous CLI agents, recovering stalled Codex/Codex runs, handling delegate_task limitations, routing failed Codex lanes to Codex, or maintaining plugin identity scope.

Class-Level Workflow

1. Prefer explicit background process tracking and absolute logs over shell-disowned runs.
2. Close stdin for Codex modes that wait for EOF; salvage Codex print stalls by checking process state and logs.
   • If a Codex background log contains only Reading additional input from stdin..., treat it as likely waiting for EOF or prompt delivery rather than useful progress. Re-check current codex exec --help, relaunch with a known-good stdin-close/input pattern, and verify output artifacts advance.
3. Poll by captured session IDs after launch; generic process lists can be empty even while known background sessions are still running. Treat process list as advisory, not authoritative, when session IDs are available.
4. For Gemini CLI lanes launched inside a repository, remember Gemini may auto-load .gemini/agents/*.md. Startup validation errors such as unsupported permissionMode keys can prevent useful work even while the wrapper remains alive; capture the error, run from a clean context or fix the agent definitions through a validated repo change.
5. Treat delegate_task lane limits as an orchestration constraint.
6. When a worker patches canonical skills, verify file paths and resulting content before reporting success.
7. Keep provider/agent fallback decisions evidence-based.

Consolidated Session Learnings

The references/ directory contains archived narrow skills absorbed during the 2026-04-29 umbrella consolidation pass. Use the subsections below as the class-level index, then open the named reference when a case-specific recipe is needed.

Absorbed Narrow Skills (2026-04-29)