agent-governance-impl:voice-governance@proposed
Voice Tool-Loop Governance (proposed) overview
PROPOSED out-of-band governance for a realtime voice agent's mid-conversation tool/decision loop. babysitter is a deterministic, fsync-per-step, replay-by-re-execution orchestrator and CANNOT sit in the sub-second audio turn; instead each governed tool is an async MCP tool whose handler drives an in-process babysitter run off the hot path. The run decomposes the workflow (ctx.task/ctx.parallel), enforces policy (PolicyEngine deny-on-dispatch), gates sensitive actions with ctx.breakpoint (auth./destroy. postures force owner-level human approval), and emits an append-only, checksummed, replayable journal as the audit trail. Governance latency is hidden behind filler speech + async/background tools so the audio path never blocks. See docs/research/voice-governance-bridge-spec.md.
Attributes
- voice-agent-tool-call-loop
- mid-conversation-workflow-decomposition
- sensitive-action-execution
- consequential-visual-tools (canvas-content, screen-share-vnc, external-video-metadata)
- policy-engine-deny-on-dispatch
- breakpoint-approval-gates (auth./destroy. -> owner)
- human-in-the-loop-mid-call
- irreversible-action-guarding (disallow_interruptions/wait_for_playout pairing)
- append-only-checksummed-journal
- per-decision-replayable-audit-trail
- voicecall-crd-governanceRuns-status
Outgoing edges
- agent:babysitter·AgentProductbabysitter
- source-ref:babysitter-sdk·SourceRef
- agent-runtime-impl:voice-agent@proposed·AgentRuntimeImplVoice Agent Runtime (proposed)
- extension-interface:governance·ExtensionInterfaceGovernance
- layer:14-governance·LayerGovernance