openclaw/openclaw merged commit f53346944d on July 6, adding a shared-state audit event store, closed action/status/error vocabularies, one-way-hashed tool-call IDs, bounded retention, an audit.list gateway RPC, and an openclaw audit CLI surface.
For agent operators, the morning decision is to watch whether audit rows become the normal incident receipt: the strongest public source proves metadata correlation, while the caveat is that this is not yet full tool governance or a public privacy review.
Facts
- The commit records native web-search outcomes only when completion or failure is explicit; otherwise the terminal outcome remains unknown.
- The new audit path is enabled by default, but
audit.enabled=falsestops new writes while existing rows remain readable until retention expires. - The implementation adds protocol schema, gateway method, CLI command, event writer, state database schema, and tests across agent-run termination, tool errors, subagent waits, and gateway subscriptions.
Evidence
The public receipts are OpenClaw commit f53346944d, the new docs/cli/audit.md page, the gateway protocol documentation, and the audit schema/test files added under packages/gateway-protocol and src/audit.
Context
This matters because agent runs are becoming operational systems. When a search, tool call, cancellation, or error is only visible inside a transcript, incident review depends on reading chat. OpenClaw is moving that evidence into a bounded runtime ledger.
Limits
The evidence proves a metadata ledger and query surface, not a complete compliance system. What to watch: whether future commits connect audit rows to policy review, user-facing export, or redaction controls for teams running agents at scale.