openhuman/docs/tinyagents-full-migration-plan/07-subagents/03-steering-recursion.md

07.3 — Steering surface + one recursion authority

Steering

Current status (2026-07-02): harness/run_queue/ is still live. OpenHuman drains its own Steer/Collect lanes and forwards them as TinyAgents SteeringCommand::InjectMessage; TinyAgents drains those SteeringHandle commands at loop checkpoints. The queue is also the product adapter for web-channel followup and parallel semantics and for detached sub-agent steer/collect lookup through running_subagents. The local openhuman::tinyagents::orchestration seam now re-exports SteeringRegistry, TaskId, SteeringCommand, SteeringPolicy, and SteeringHandle, and sub-agent TinyAgents runs register their live SteeringHandle in a process-local SteeringRegistry while the run is active. Detached steer/collect controls now prefer that registry handle and fall back to RunQueue when no handle is available. The shared OpenHuman turn seam installs a local steering policy that accepts only InjectMessage and Pause until a product control surface owns additional crate command kinds. Do not delete the directory until the remaining detached control modes and the web-channel followup/parallel lanes either have TinyAgents-owned equivalents or move into local owners. Hard cancel/prune paths also deregister task handles so aborted runs do not leave stale registry entries. The unused DomainEvent::RunQueueMessageDelivered projection has been removed; queued/interrupt/followup events remain live.

1. Map product tools onto crate commands: steer_subagent → SteeringCommand::InjectMessage via the registered SteeringRegistry is live for active TinyAgents sub-agent runs, with RunQueue fallback. Next: map redirect/pause/resume/cancel modes to corresponding variants while preserving delivery-at-safe-boundaries semantics (crate drains before each model call).
2. Tighten policy by run class: the shared turn seam now installs an InjectMessage/Pause allowlist; future background-only runs can accept Cancel without also accepting transcript injection.
3. Accepted/rejected steering emits AgentEvent::Steered; the bridge now logs accepted/rejected command kinds, and UI projection remains pending. Delete bespoke acknowledgment plumbing in run_queue/ after parity.
4. harness/run_queue/ (317 total; 174 non-test): first split it by ownership. Detached sub-agent Steer/Collect should collapse to a SteeringRegistry lookup/registration path; web-channel Followup/Parallel remain product turn orchestration unless a crate-owned replacement is introduced.

Recursion

5. One cap: crate RunLimits.max_depth/RecursionPolicy becomes authoritative; crate-internal spawn_depth_context.rs (66) becomes a reader/projector of crate depth (ToolExecutionContext.depth) for product error wording, or is deleted if the wording can wrap TinyAgentsError::SubAgentDepth. Current code now threads MAX_SPAWN_DEPTH = 3 into TinyAgents RunPolicy.limits.max_depth, RunConfig.max_depth, and MCP agent.run_subagent via mcp_server/subagent_depth.rs, but OpenHuman's own spawn_depth_context still rejects beyond the same cap before the TinyAgents run.
6. One error shape: SubAgentDepth/RecursionLimit now preserve the existing SpawnDepthExceeded error surface before the subagent graph wraps provider failures.

Deletions

• harness/run_queue/ queue mechanics, only after detached steer/collect no longer use Arc<RunQueue> and web-channel followup/parallel behavior has a replacement or local owner.
• crate-internal harness/spawn_depth_context.rs (if step 5 wording-wrap suffices).

Acceptance

• Mid-flight steer lands only at loop boundaries (existing steering tests); depth-exceeded produces the same user-facing error as today.