Files
Josh Avant 562391b9af feat(audit): explain denied operator approvals (#119815)
* Audit: add durable decision receipts

* Audit: route generic decision facts through writer

* Audit: satisfy strict decision count typing

* Test: align decision writer type coverage

* Refactor: isolate decision receipt projection

* Fix: preserve ambiguous approval correlation

* Fix decision coverage across pages

* Fix approval coverage across corrupt pages

* Fix decision summaries across retention and paging

* Remove superseded decision fact count path

* Keep session-derived approval links conservative

* Test decision paging at owner boundary

* Audit: bind approval receipts to exact execution

* Test: keep decision receipt coverage focused

* docs(agents): preserve decision receipt ownership

* Docs: refresh Plugin SDK split baselines

* Docs: require opt-in for future decision facts

* Fix: bound decision receipt inspection
2026-08-12 05:23:29 -05:00

15 KiB
Raw Permalink Blame History

summary, read_when, title
summary read_when title
CLI reference for activity records, execution identity, and decision receipts
You need to answer who ran an agent or tool, when it ran, and how it ended
You need content-free inbound or outbound message lifecycle metadata
You need a bounded, redaction-safe activity export
Audit records

openclaw audit

Query the Gateway's metadata-only activity ledger, discover executions that share a run correlation, or inspect immutable identity context for one exact agent execution.

Run and tool activity records are on by default. Execution identity is separately off by default on fresh installs and upgrades. Enable it explicitly:

openclaw config set logging.audit.executionIdentity true
openclaw gateway restart

Identity collection requires logging.audit.enabled to remain enabled. Message records are also separately disabled by default; set logging.audit.messages to direct or all and restart the Gateway to record them. Existing records stay queryable until they expire (30 days).

Direct local commands use the same bounded writer lifecycle as the Gateway. openclaw agent exec deletes its temporary state directory by default, so its audit evidence is intentionally discarded with the rest of that isolated run. Use agent exec --state-dir <dir> when the run state must remain available, and inspect it through a Gateway using that same state directory.

The ledger is separate from conversation transcripts: it records identity, ordering, provenance, action, status, and normalized outcome codes, but never stores content, and message identifiers appear only as installation-local keyed pseudonyms. Audit history owns the full data model, privacy semantics, storage/retention bounds, and coverage limits; this page covers the command surface.

openclaw audit
openclaw audit --agent main --status failed
openclaw audit --session "agent:main:main" --after 2026-07-01T00:00:00Z
openclaw audit --run 8c69f72e-8b11-4c54-98d5-1a3dd67450c3
openclaw audit --run 8c69f72e-8b11-4c54-98d5-1a3dd67450c3 --explain
openclaw audit --execution 5da4c4c3-e1c9-4c95-a17d-6e5c10fd45cf --explain
openclaw audit --execution 5da4c4c3-e1c9-4c95-a17d-6e5c10fd45cf --explain --json
openclaw audit --run 8c69f72e-8b11-4c54-98d5-1a3dd67450c3 --explain --json
openclaw audit --kind tool_action --limit 50 --json
openclaw audit --kind message --direction outbound --channel telegram --json

Filters

  • --agent <id>: exact agent id
  • --session <key>: exact session key
  • --run <id>: exact run id; filters activity unless --explain is also set
  • --execution <id>: exact execution id; requires --explain
  • --kind <kind>: agent_run, tool_action, or message
  • --status <status>: started, succeeded, failed, cancelled, timed_out, blocked, or unknown
  • --direction <direction>: message direction, inbound or outbound
  • --channel <channel>: exact message channel
  • --after <timestamp> / --before <timestamp>: inclusive ISO timestamp or Unix milliseconds
  • --limit <count>: activity page size from 1 to 500 (default 100), decision page size from 1 to 100, or ambiguous execution-candidate page size from 1 to 50 with --explain (default 50)
  • --cursor <sequence>: continue an activity, decision, or ambiguous execution-candidate page
  • --explain: inspect immutable execution identity and run-admission reasoning; requires exactly one of --run or --execution and accepts only --limit, --cursor, and --json
  • --json: print the bounded page as JSON

The CLI queries the versioned activity RPC so one command shows the complete configured ledger. Text output shows time, kind, direction, channel, status, agent, run, and action. Missing message provenance renders as -; OpenClaw does not invent agent or run ids. Tool actions also show the tool name. JSON output includes nextCursor when another page exists. Pass that value to --cursor to continue without reordering records that arrive during paging.

These exports remain sensitive operational metadata even though message bodies and raw message identity fields are absent. Agent, session, and run ids, timing, channels, outcomes, and stable HMAC references can correlate activity. Protect them with the same access controls and retention practices as other operator records.

The Gateway intentionally exposes retained execution-identity diagnostics to every client with operator.read in its operator domain. That scope is a trusted read-only boundary, not hostile multi-tenant isolation. Use separate Gateway trust domains when operators must not share audit identity data.

Discover and explain executions

Every admitted outer turn receives an opaque executionId. contextId identifies its immutable evidence record; the existing runId stays a possibly shared session, routing, or recovery correlation. Use --run <id> --explain to discover retained executions rather than query the best-effort activity list. One match resolves directly. Multiple matches return ambiguous, list at most 50 candidates, and tell you to select one explicitly:

openclaw audit --execution <execution-id> --explain

OpenClaw never silently selects the first or latest execution. The exact text view renders these sections:

  1. Identity: trust domain, invoker, ingress, agent principal, agent definition, runtime instance, represented subject, and sponsor.
  2. Authority: applicable grants and assurance evidence.
  3. Lineage: parent context or an explicit absent, unknown, or unsupported state.
  4. Decisions: bounded run-admission and authoritative action-decision receipts, including terminal operator approvals.
  5. Missing evidence and Next steps.

Every field includes present, absent, unknown, or unsupported; the CLI does not infer a user from a session key, device id, display name, or shared credential. A direct local run currently shows authoritative local-cli ingress, an absent invoker, and unattributed coverage. Its admission receipt says not-applicable because no identity-aware policy or grant evaluation was proven.

For Gateway runs, a resolved authenticated profile can make the invoker present and coverage attribution-only. Paired devices and shared credentials do not establish a person: without a durable profile the invoker stays absent, or unknown when authenticated user evidence promised a profile that could not be resolved. Session creation retains the live canonical durable profile id so profile linking does not orphan ownership, while run inspection consumes the immutable connection-time audit fact. Ordinary session provenance stores no display label. An optional bounded, secret-redacted label can be retained only in execution identity after that audit storage is explicitly enabled.

A terminal approval receipt shows allowed or denied, its stable reason code, enforcement state, authoritative source boundary, policy and grant references, context fields used, and remediation. Expired and cancelled approvals are denied non-actions with distinct reason codes. no-route is an enforced denial only when the approval owner recorded that terminal state. A corrupt approval is unknown. The text view labels operator_approvals as an authoritative owner-native SQLite record retained for 30 days; JSON preserves the same source owner and record reference without lossy reformatting. enforced requires the approval's immutable owner-local binding to match the selected context, execution, and run exactly. A missing, malformed, or mismatched binding reports operator_approval_execution_link_missing, operator_approval_execution_link_malformed, or operator_approval_execution_link_mismatch with unknown coverage and no grant references. The inspector never reconstructs that binding from runId, session metadata, timestamps, or the number of retained executions.

JSON output is the Gateway result without lossy reformatting. An exact result contains one bounded V1 context (maximum 16 KiB), up to 100 decision receipts, coverage and missing-evidence codes, and an optional nextDecisionCursor. An ambiguous run result instead contains at most 50 execution candidates and an optional nextExecutionCursor. Sensitive domain, runtime, invoker, assurance, ingress-source, and grant references are installation-local HMAC projections. Configured agent ids and exact run ids remain visible, as do context and execution ids, so redirected output is still private operator data.

An older Gateway produces an explicit unsupported result with gateway_upgrade_required and an upgrade-and-rerun next step. The CLI never reconstructs identity from legacy audit rows. A current Gateway distinguishes an unknown run, an unavailable pre-feature, disabled, or failed context write, an expired context, and a corrupt context without claiming that missing best-effort activity proves no execution. A newly admitted run can also be temporarily unavailable while its bounded identity envelope waits in the audit writer queue; retry inspection after the run or normal process shutdown. Admission never waits for writer readiness, schema or HMAC-key initialization, SQLite, or persistence.

Once a context is older than 30 days, the CLI returns no fields or linked decisions from it. While bounded cleanup is pending, the result is unsupported with an expiry-and-rerun next step. After cleanup it can become unknown if no separately retained activity remains; this absence does not prove that the run did not occur. Startup and hourly maintenance prune at most 1,024 identity contexts per tick and continue when collection is disabled. Queue saturation, worker/storage failure, cleanup failure, or abrupt process termination can lose best-effort evidence but never block or abort the agent run. Normal Gateway and direct-local CLI shutdown flushes accepted work when its writer lifecycle permits.

Recorded events

The Gateway projects trusted lifecycle streams into six actions:

  • agent.run.started
  • agent.run.finished
  • tool.action.started
  • tool.action.finished
  • message.inbound.processed
  • message.outbound.finished

Every returned record has a stable event id, a monotonically increasing ledger sequence, a lifecycle timestamp, actor, action, status, a schemaVersion: 1 marker, source sequence, and redaction: "metadata_only". Agent/session/run provenance and event-specific fields are present only when the trusted source provides them. Message records intentionally omit sessionKey and sessionId, so --session filters run and tool records only.

Terminal run and tool records distinguish success, failure, cancellation, timeout, and policy blocks with closed status and error codes. unknown is an explicit non-success result when an upstream runtime does not expose an authoritative terminal outcome. Tool call ids are exported only as stable fingerprints. Tool names must match the compact model-facing name contract; other values become unknown.

Message records add direction, channel, conversation kind, outcome, and optional delivery kind, failure stage, duration, result count, normalized reason code, and keyed account/conversation/message/target pseudonyms. The current inbound boundary covers accepted messages that reach core dispatch, including core duplicate and terminal processing outcomes. The outbound boundary writes one terminal row per original logical reply payload that reaches shared durable delivery; chunking and adapter fan-out are aggregated in resultCount. Queued retryable or ambiguous sends are recorded only after an acknowledgement, dead letter, or reconciliation makes the outcome terminal. Plugin-local and direct-send paths that bypass those shared boundaries are not yet covered; absence of a row does not prove that no message existed.

The audit ledger does not replace transcripts, task history, cron run history, or logs. It provides a small cross-run index for operator questions without copying conversation content into another store.

For inbound rows, durationMs measures core dispatch and resultCount counts finalized queued tool, block, and reply payloads. For outbound rows, durationMs includes delivery ownership through its terminal (and therefore queued wait time), while resultCount counts identified physical platform sends. deliveryKind, when present, describes the effective post-hook, post-render payload; suppressed and crash-ambiguous rows omit it.

Gateway RPC

audit.activity.list requires operator.read and accepts the same filters. It returns the named V1 activity event union, including run, tool, inbound-message, and outbound-message records.

openclaw gateway call audit.activity.list --params '{"channel":"telegram","limit":50}'

The result is { "events": AuditActivityEventV1[], "nextCursor"?: string }. Results are newest first and limited to 500 records per request.

audit.run.inspect also requires operator.read:

openclaw gateway call audit.run.inspect \
  --params '{"runId":"8c69f72e-8b11-4c54-98d5-1a3dd67450c3","decisionLimit":50}'

openclaw gateway call audit.run.inspect \
  --params '{"executionId":"5da4c4c3-e1c9-4c95-a17d-6e5c10fd45cf","decisionLimit":50}'

Its result is { "schemaVersion": 1, "run": ..., "identity": ..., "decisions": ..., "coverage": ..., "nextDecisionCursor"?: ..., "nextExecutionCursor"?: ... }. The closed request accepts exactly one of executionId or runId. decisionLimit is 1100 and decisionCursor is optional. Run discovery also accepts executionLimit from 150 and an optional executionCursor. A run with multiple retained executions returns the typed ambiguous identity state and no identity context or decisions until the caller selects an execution id. For one selected context, receipt paging starts with admission, then reads owner-native terminal approvals, then generic facts for boundaries without a native durable record. Approval inspection never writes a generic duplicate. Generic fact writes and projections also require the full context, execution, and run tuple to match the immutable execution context.

The activity ledger remains best-effort. By contrast, a returned approval receipt comes from the authoritative first-answer-wins approval row, and a returned generic receipt comes from the additive immutable decision-fact table. All three surfaces use 30-day retention, but absence from the activity ledger cannot prove that an approval or action did not occur. Generic fact delivery is also best-effort until its bounded worker write persists the row; owner-native approval persistence does not use that queue.

The shipped audit.list RPC remains unchanged for older run/tool clients. When audit.activity.list is unavailable on an older Gateway, the CLI retries audit.list only if every requested filter is supported by that legacy method. --kind message, --direction, and --channel fail with an upgrade message on an older Gateway instead of being silently discarded.