korchasa/supervise

Supervised Workflow Run Entry

Drive one flowai-workflow run to a clear terminal state by delegating the run to supervisor.

Delegate

Launch supervisor with the host's real subagent mechanism:

• Claude Code: use the Agent / Task tool with subagent_type=supervisor;
• OpenCode: use the @supervisor <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 supervisor skill and return the SUPERVISOR_REPORT block as its final message. The worker, not the parent, reads all run artifacts.

Pass:

• workflow folder path;
• run id or --resume <run-id>, if the user provided one;
• requested polling cadence or stop limit, if provided;
• the user's exact supervision goal.

Ask the supervisor to return only a concise stop report: workflow, run id, final state, failed/stalled node, evidence paths read, fix surface, resume command, and remaining blocker.

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

Parent Boundaries

The parent session may locate the requested workflow folder, but must not:

• choose among multiple workflows without the user naming one;
• read .flowai-workflow/ORCHESTRATION.md;
• interpret orchestration policy;
• diagnose node artifacts in the parent context;
• run flowai-workflow directly unless the user explicitly rejects subagent delegation and asks for inline supervision.

Out of Scope

• Workflow policy loops -> use orchestrate.
• First-time workflow creation or broad adaptation -> use scaffold.
• Post-mortem of a completed run without intent to resume -> use an investigation workflow.
• Editing the flowai-workflow engine source unless the user explicitly asks.