2 months ago
b8859ff`agent()` maps onto the existing AgentControl facade: spawn -> wait -> validate -> repair -> close, transcribed from the production loop in taskReview.ts with three corrections a fan-out needs and a single review does not. - StatusBus counts `errored` as terminal, so `wait()` returns timedOut:false for a child that crashed. Reading its text there would surface a failure as a plausible-looking result, so `executionState` is checked first. - `AgentWaitInspection` carries no usage fields, so wait() alone cannot fund `budget.spent()`. Each agent is inspected once before close. - A single 600s wait leaves cancellation latency up to ten minutes on the last agent, so waits are sliced into ~120s windows that re-check the abort signal. Schema-validated returns use a `<workflow_result>` envelope with exactly one repair turn, rather than widening the `<agent_report>` contract that every child already shares. The schema stays a JSON Schema literal: it has to survive a structured-clone boundary and be journal-serializable, which a live zod object is not. `onError` defaults to "fail" so failures reach the script, and task-lock or cancellation errors abort the whole run regardless — otherwise a 300-way fan-out degrades into 300 silent nulls. Concurrency is capped at 12, below MAX_ACTIVE_CHILDREN_PER_PARENT, leaving headroom for the parent turn's own spawnAgent calls. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Kk98XjGKJrKp5ynbiNFMFu
Parenta7d68f2
