# Coding-Agent Runbook (PTY + Background) This orchestrator MUST delegate implementation tasks to a coding agent. Do not hand-code feature work directly when the skill is active. Supported agents: - `codex` - `claude` - `opencode` - `pi` ## First Rule At skill start, ask: 1) Which coding agent should run tasks? 2) Which fallback agent should be used if primary fails? ## Launch Patterns ### Codex ```bash codex exec --full-auto "" ``` ### Claude ```bash claude "" ``` ### OpenCode ```bash opencode run "" ``` ### Pi ```bash pi -p "" ``` OpenClaw execution recommendation: - `pty:true` for interactive CLIs - `background:true` for long-running work - `workdir:` ## Required Orchestration Loop 1. Generate gate prompt (`generate_gate_prompt.py`). 2. Execute selected coding agent with that prompt (`agent_exec.py` or equivalent). 3. Require coding agent to update docs immediately after task completion: - docs/tasks.md - docs/progress.md - docs/change-log.md - docs/traceability.md - docs/test-results.md - docs/agent-handoff.md 4. OpenClaw agent runs verification itself: - CLI checks in terminal - Browser/manual checks for web journeys 5. If validation fails: - summarize issue clearly with command/flow + output - re-spawn coding agent with fix prompt (same task/spec) - require docs updates again - re-test 6. Only then update gate status. ## Manual Review Responsibility Even in autonomous mode, the OpenClaw agent performs manual verification itself: - Web/UI flows: run in browser tools, test critical journeys. - CLI flows: run required commands in terminal and inspect outputs. If checks fail, send concrete failure details to coding agent, request fix, and retest. ## Completion Wake Pattern For long runs, require coding agent wake messages to include task + verification handoff. "When fully done, run: openclaw gateway wake --text 'Done: | task: | handoff: see docs/agent-handoff.md for CLI+Browser checks' --mode now"