displayName
Babysitter task outputSchema guard
appliesTo
tool-result
phase
post-call
detectorKinds
onTrigger
block
severity
block-call
description
Per-task structural guard. The babysitter SDK accepts an
`outputSchema` on agent task definitions (see
packages/sdk/src/tasks/types.ts:46 and
packages/sdk/src/prompts/templates/task-kinds.md). At result-posting
time the SDK validates the task result against the per-task JSON
schema and rejects shape-mismatched results, surfacing a structured
error on the journal as part of EFFECT_RESOLVED with status=error.
Distinct from `output-guard:claude-code-aup-refusal` (post-call
refusal-rewrite over assistant text) — this guard is structural over
tool/effect results, not safety/AUP over assistant content.