Minimize audit data

Send only policy fields, choose audit_mode, and configure retention on your evaluation server.

Evaluate uses the full intent for policy matching. Audit storage keeps a privacy-safe record. Default mode is minimal — decision fields only, not open metadata bags or chat dumps.

Send only policy fields

Send:

action_type, resource, optional cost

Do not send API keys, bearer tokens, passwords, or full transcripts in metadata or record details.

Call client helpers before you log or ship telemetry yourself:

import {
  redactActionIntent,
  redactDetails,
  projectIntentForAudit,
} from "@limetry/sdk"

const safe = redactActionIntent(intent)
const forAudit = projectIntentForAudit(intent, "minimal")
const safeDetails = redactDetails(details)

Server-side scrubbing still runs even if a client forgets.

Choose an audit mode

Mode	Stored after scrubbing
`minimal` (default)	`action_type`, scrubbed `resource` (no query/userinfo), ids, cost, decision/outcome, reasons
`forensics`	Same plus redacted `metadata` and (on record) redacted `details`

import { createSlimActionPolicy } from "@limetry/sdk"

const policy = createSlimActionPolicy({
  agentId: "ops",
  allowedActionTypes: ["http_get", "http_post"],
  auditMode: "forensics",
})

Environment default when a policy omits audit_mode:

LIMETRY_DEFAULT_AUDIT_MODE=minimal

Apply scrubbing rules

Sensitive metadata/details keys matching token/password/secret patterns → [REDACTED]
resource URLs → strip userinfo, query string, and hash
Embedded Bearer …, ?token=, live secret-key substrings (sk_live_…) → redacted

Policy matching still uses the original intent at evaluate time. Digests for decision receipts are computed on the original intent so enforcement stays bound to what was approved.

Set retention

With USE_POSTGRES_STORE=true:

LIMETRY_AUDIT_RETENTION_DAYS=90
LIMETRY_AUDIT_PURGE_INTERVAL_MS=3600000

Set LIMETRY_AUDIT_RETENTION_DAYS=0 to disable automatic purge on a self-run node.