Audit Trail — Immutable Action Log#

Purpose#

Provide a tamper-evident, complete record of every agent action for forensic investigation, compliance auditing, and accountability.

Integration#

Always-on hook on ALL agent actions:

• Tool use (exec, file read/write, network)
• Skill invocations
• Channel messages (in/out)
• Memory reads/writes
• Configuration changes
• Error events

Log Format (JSONL, append-only)#

jsonl

{"id":"ACT-20260331-000001","ts":"2026-03-31T14:30:00.123Z","agent":"openclaw-main","session":"sess_abc123","type":"tool_use","tool":"exec","args":{"cmd":"npm test"},"skill":"contract-tester","channel":"telegram","user_hash":"sha256:a1b2c3...","outcome":"success","duration_ms":4200,"prev_hash":"sha256:000000","hash":"sha256:d4e5f6..."}

Fields#

FieldTypeDescriptionidstringUnique sequential IDtsISO 8601Microsecond timestampagentstringAgent identifiersessionstringSession IDtypeenumtool_use, skill_invoke, channel_in, channel_out, memory_write, config_change, error``toolstringTool name (if applicable)argsobjectSanitized arguments (secrets redacted)skillstringInvoking skillchannelstringSource channeluser_hashstringSHA-256 of user identifier (never raw)outcomeenumsuccess, failure, timeout, blocked``duration_msnumberExecution timeprev_hashstringSHA-256 of previous log entry (chain)hashstringSHA-256 of this entry (including prev_hash)

Storage#

text

.security/audit-trail/
├── 2026-03-31.jsonl        (today, active)
├── 2026-03-30.jsonl        (yesterday)
├── 2026-03-29.jsonl.gz     (compressed, >7 days)
└── integrity-check.log     (chain verification results)

Integrity Verification#

• Each entry's hash = SHA-256( id + ts + type + tool + outcome + prev_hash)
• Chain validation: entry[n].prev_hash == entry[n-1].hash
• Run verification: jq -r '.hash' | sha256sum --check
• Tampering detection: broken chain → CRITICAL alert

Retention Policy#

AgeStorageAccess0-7 daysRaw JSONLDirect read7-90 daysCompressed JSONL.gzDecompress on query90-365 daysArchive (if configured)Restore on request>365 daysPurge (manual only, human approval)—

Query Examples#

bash

# All actions by a specific skill
jq 'select(.skill=="contract-tester")' .security/audit-trail/2026-03-31.jsonl

# All failures
jq 'select(.outcome=="failure")' .security/audit-trail/*.jsonl

# Actions in a time window
jq 'select(.ts >= "2026-03-31T14:00" and .ts < "2026-03-31T15:00")' .security/audit-trail/2026-03-31.jsonl

# Channel activity summary
jq -s 'group_by(.channel) | map({channel: .[0].channel, count: length})' .security/audit-trail/2026-03-31.jsonl

Guardrails#

• Log file is APPEND-ONLY — agent cannot delete or modify entries
• Secrets in arguments are redacted BEFORE logging (using agent-firewall patterns)
• User identifiers are hashed, never stored in plaintext
• Log integrity verified on every read
• Manual purge requires human approval + logged as audit action itself