Workflow Orchestrator

Read the full workflow from .claude/commands/workflow.md and execute it.

Quick Summary

1. Decompose — break the user's request into discrete subtasks with dependencies
2. Dispatch — send subtasks to other agents via dispatch_task (with /do instructions) and verify each task started with the intended title, profile, and constraints
3. Wait — persist state and a stable workflow-step waitKey, register wait_for_subtasks, and end the turn
4. React — dispatch dependent tasks as predecessors complete, retry failures
5. Complete — summarize results when all subtasks finish

Why This Exists

SAM owns durable child-task status and can wake a sleeping parent exactly once when a registered wait resolves. The orchestrator therefore releases its prompt turn instead of maintaining an expensive foreground poller. Bounded foreground polling is a compatibility fallback only when the connected server does not advertise wait_for_subtasks or reports that durable wake delivery is disabled.

Staging Debugging Access

All agents have access to $CF_TOKEN for direct Cloudflare API queries against staging. When monitoring subtasks that deploy to staging, use the CF API to verify their work landed correctly — query D1 for data state, read KV for feature flags, check DNS for routing. See .claude/rules/32-cf-api-debugging.md for the full cheat sheet.

State Persistence

Maintain .workflow-state.md (gitignored) as external memory. Write it before registering a wait and re-read it whenever SAM wakes the session. This survives sleep, recovery, and context compaction. See .claude/commands/workflow.md for the full state file format.