orchestrate

name	orchestrate
description	Run the flowai-workflow orchestration policy loop. Delegates policy execution to the workflow orchestrator agent, which selects workflows from ORCHESTRATION.md and supervisor summaries.
argument-hint	optional .flowai-workflow path or loop limit
effort	high

Workflow Orchestration Entry

Start or continue a project-local flowai-workflow orchestration loop. This is the public entry point and dispatch bridge. Policy decisions belong to orchestrator; run recovery belongs to supervisor.

Dispatch Loop

1. Launch orchestrator with the host's real subagent mechanism:
   • Claude Code: use the Agent / Task tool with subagent_type=orchestrator;
   • OpenCode: use the @orchestrator <task prompt> mention syntax.
   • Codex: spawn a native worker subagent (the parent dispatches; Codex max_depth=1 forbids nested spawns) and instruct it — by skill name — to invoke the orchestrator skill and return the SUPERVISOR_DELEGATION block as its final message. The worker, not the parent, performs all policy reads.
2. Ask it to read policy/history, append its decision, and return either:
   • STOP: <reason>; or
   • a SUPERVISOR_DELEGATION block with workflow path, optional run id, reason, and requested short-result fields.
3. If it returns SUPERVISOR_DELEGATION, launch supervisor from the parent context using the real subagent mechanism:
   • Claude Code: Agent / Task with subagent_type=supervisor;
   • OpenCode: @supervisor <task prompt>.
   • Codex: spawn a native worker subagent and instruct it — by skill name — to invoke the supervisor skill with the delegation prompt and return the SUPERVISOR_REPORT block.
4. Pass the supervisor only one workflow/run and ask for a short result: workflow, run_id, status, fixes, repeat, blocker.
5. Feed that short result into the next orchestrator call if the policy loop should continue. Stop when the policy stop condition, supervisor blocker, or user loop limit is reached.

Pass the orchestrator:

• the user's request;
• repository root;
• optional loop limit, if the user gave one;
• prior supervisor summary, when continuing after a supervised run;
• instruction to return concise progress plus a structured delegation block.

If the host has neither a native subagent mechanism (Claude subagent_type, OpenCode @mention) nor Codex-style worker dispatch, say that context-isolated orchestration is unavailable in this IDE and stop. Do not inline the loop in the parent context. Do not simulate delegation with bash, echo, heredocs, or placeholder text.

Parent Boundaries

The parent session may identify that .flowai-workflow/ exists, but must not:

• choose workflows by policy;
• read .flowai-workflow/ORCHESTRATION.md deeply;
• read workflow runs/, state.json, journal.jsonl, logs/, or node artifacts;
• patch workflow files;
• run flowai-workflow directly for a selected workflow.

Those actions belong to the orchestrator subagent and, for a single run, the supervisor subagent. The parent only dispatches between them because some hosts do not allow nested subagent calls.

Expected Subagent Result

Ask the orchestrator to report:

• decisions appended to .flowai-workflow/orchestration.jsonl;
• selected workflows, supervisor delegation requests, and supervisor run ids;
• final stop condition or next pending action;
• blockers requiring a real user decision.