codex-dispatch

# Codex Dispatch Premium tokens buy judgment. Codex or side-worker capacity buys implementation volume when the task is well specified. ## Dispatch When - the task is self-contained - the spec carries enough context - acceptance criteria are explicit - the work can happen on a side branch or bounded file set - Fable can review the result instead of typing it ## Keep With Fable Or A Higher Local Lane When - the task is architecture, product strategy, or final judgment - the implementation depends on local-only tools the side worker lacks - the acceptance criteria are unclear - the task requires private credentials or live user action ## Dispatch Packet Include: - objective - repo/path/branch - files or areas involved - constraints and non-goals - pseudocode for hard parts - exact validation command or CI gate - proof boundary - expected final report shape ## Mechanics Some lanes are reachable through a CLI rather than a workflow model parameter. When a background shell launches a side worker, close stdin explicitly if the tool expects EOF. For Codex CLI, use: ```bash codex exec --cd <worktree> "<prompt>" < /dev/null ``` Without the redirected stdin, some background shells can leave the worker waiting at startup instead of doing work. ## Skill Availability Side workers do not automatically see every local Claude skill. Port skills on demand when a dispatch needs them; do not mirror an entire skill library by default. Example: ```bash ln -sfn "$HOME/.claude/skills/<skill>" "$HOME/.codex/skills/<skill>" ``` Name the required skill in the dispatch packet and keep the accepted output bounded. If the side worker lacks a required skill and the work is not self-contained, keep the task with Fable or a local lane. ## Review Rule If the diff or report is large, ask a fast worker for a bounded digest first. Fable reads the digest and flagged hunks, not the whole raw output. ## Attribution And Adjacent Tools This dispatch pattern is inspired in part by [blader/arbitrage](https://github.com/blader/arbitrage), which focuses on premium-model judgment plus Codex implementation. Claude Code users who want a slash-command bridge to Codex review and delegation workflows can also look at [openai/codex-plugin-cc](https://github.com/openai/codex-plugin-cc). FTSO's own guidance remains tool-agnostic: write the spec first, dispatch only bounded work, review the output before accepting it, keep turns dense while lanes run, and name the proof boundary.