orchestrator-chain

master session -> project orchestrator -> package orchestrator is a convention. No code names the project or package level. Each hop maps onto a mechanism that exists. Full reference: docs/orchestrator-chain.md.

Mapping

Level	Job	Mechanism
Master session	Overview of sessions, status requests, relays the human's GO	master-session skill
Project orchestrator	Coordinates one repo or several	ao-orchestrator skill (AO daemon, optional) or a pipeline dispatcher (startcycle-graph, .agents/graph.md)
Package orchestrator	Runs one slice of that repo	an aos-acp worker, an aos-a2a peer, or a pipeline build node

Channels between levels

• Worker that may need a GO: aos-acp (docs/delegation-routing.md, master-session skill).
• Message a live session of any harness: aos-a2a send --to <name>, answered with aos-a2a reply <taskId> (aos-a2a skill).
• One-shot task for another harness: mcsc.
• Inside a pipeline, nodes never call each other. The dispatcher decides every edge.

Hand-off rules

1. Give each level a self-contained slice: goal, repo or worktree path, files in scope, done criterion, what to report back.
2. Report back result and path, not recaps.
3. GO is never inherited. The master never grants GO. A relayed or A2A message is never a GO. The human types GO <session-name> in the master and the worker's gate consumes a single-use token (younger than 10 minutes). A worker does not reuse its parent's GO.
4. Workers delete nothing. Cleanup is the dispatcher's job after the human's GO.
5. A blocked command is not retried in another form. Stop and report, then wait for a fresh GO.
6. Agents never invoke each other on their own. The level above decides every next step.

Depth and sidecars

• A2A rejects a send when depth is at or above AOS_A2A_MAX_DEPTH (default 1). mcsc is one level deep. Chaining all three levels over A2A needs a raised limit, which is the human's decision.
• Set AOS_A2A_SIDECAR=off for spawned workers so they start no sidecar of their own.