run

Command: run [--go]

Batch-execute every task in state/todo-ids.json (which only contains rows with Status == "To-Do" AND Exec == "Agent").

Claude Code: The Haiku subagent builds the plan from the sidecar; the orchestrator picks a model per task at runtime and spawns execution agents.

Non-Claude harnesses: The skill reads config.json to determine available models. If config.json doesn't exist, run /agency-os init first.

Auto-refresh. run always calls refresh as its first step. If refresh fails, run aborts.

Plan phase (Haiku subagent)

1. Execute scripts/query-tasks.py via Bash — this is mandatory and must happen before reading anything. Run python .claude/skills/agency-os/scripts/query-tasks.py and verify it exits 0. Only then read the freshly written state/todo-ids.json. Never read the sidecar without running the script first; the file on disk is always stale.
2. Dedup containers. For each row with has_todo_subtasks: true, skip the parent — its work IS its subtasks.
3. Resolve dependencies. Each sidecar row carries dependencies: [{id, status}]. For every dep:
   • status == "Done" -> satisfied, ignore.
   • dep id is in the current in-batch set -> record as an intra-batch edge.
   • otherwise -> external blocker: drop from dispatch plan, collect into blocked_deps[].
4. Topological stage assignment. Build a DAG from intra-batch edges; assign each task a stage = 1 + max(stage of its in-batch deps) (stage 1 = no in-batch deps). If a cycle is detected, abort.
5. Sort within each stage: Priority asc, then overdue-recurring first, then Effort asc.
6. Return the plan to the orchestrator as stages: [[(id, title, corpus, effort, parent_id, description_preview), ...], ...] plus blocked_deps.

Dispatch phase (orchestrator)

Before spawning any execution agent, the orchestrator prints the plan outline (see ### Output below) so the user sees which tasks are about to fire, in which stages, with which model per task. Then dispatching stage 1... and dispatch begins.