Coding-Agent CLI Automation

Method — step-by-step

Run these nine steps in order for every new automation. The sections below are the toolbox; this list is the procedure.

1. Discover runtime. Which CLIs, skills, MCPs, and models are available on the host? Fail fast when a dependency is missing. See references/introspection.md.
2. Pick category — local execution or cloud-delegated PR (§Two categories, five CLIs).
3. Pick CLI, then load its reference BEFORE writing any invocation. Default to Claude Code; use another only if the user named it. For non-Claude targets, loading references/{codex,gemini,copilot,github-agent-task}.md is mandatory — examples in this file are Claude-syntax. See §Do not transliterate.
4. Define the I/O contract — inputs (positional args, env vars, stdin) and outputs (stdout JSON, exit codes, artifact files). Everything a caller needs to re-run the script lives here (§Reusability & observability).
5. Pick shape — single-turn extract, agentic loop, multi-turn, parallel fan-out, or cloud delegation (§The five canonical shapes).
6. Apply non-negotiable flags — auto mode + --allowedTools + --output-format json + schema + success verification (§Non-negotiable flags).
7. Wire reusability — parameterize every path, model, timeout; no hardcoded constants (§Reusability & observability).
8. Wire observability — capture session ID, cost, live progress, structured exit codes (§Reusability & observability).
9. Deliver — header block documenting usage, env, permission posture, tunables.

Two categories, five CLIs