Observability

Rensei ships three complementary observability layers: Sentry for error and exception capture, OpenTelemetry (OTel) for distributed tracing of agent verb execution, and SIEM export for forwarding audit events to your security information platform. All three are opt-in via environment variables and operate independently.

Sentry error capture

Sentry is instrumented at all three Next.js execution boundaries: server, client, and edge. Configuration lives in sentry.server.config.ts, sentry.client.config.ts, and sentry.edge.config.ts.

Setup

Sentry integration is configured by the platform operator at the deployment level. Deployment-level Sentry setup (DSN, project credentials) is covered in the operator docs. Errors are enriched with user context (org ID and user ID) to enable per-tenant issue filtering.

What Sentry captures

Unhandled exceptions in API route handlers
React component render errors (via Error Boundary integration)
Edge runtime errors (middleware, edge API routes)
Manual Sentry.captureException() calls for known error paths (e.g., webhook signature failures, policy evaluation errors)

OpenTelemetry tracing

The OTel exporter subscribes to the Layer-6 verb bus and emits one span per agent tool invocation. This means every pre-verb / post-verb pair produces a span - no manual tracer.startSpan() calls are scattered through the codebase.

Enabling OTel traces

OTel tracing is activated by the platform operator via deployment-level environment variables. The exporter sends OTLP HTTP spans to any OTLP-compatible endpoint (LangSmith, Langfuse, Arize, Honeycomb, Grafana Cloud, etc.). Deployment-level OTel configuration is covered in the operator docs.

Span attributes

Every verb span includes:

Attribute	Description
`org_id`	Organization ID
`agent_card_id`	The agent card that owned the session
`session_id`	Platform session identifier
`work_type`	SDLC work type (e.g. `development`, `qa`) when present
`tool`	The tool/verb name
`latency_ms`	Duration from `pre-verb` to `post-verb`
`agent_url`	URL of the A2A agent
`skill_id`	Skill being invoked

Session events

In addition to verb spans, the exporter can emit one span per SessionEvent (session start, activity, completion). The projectSessionEvent projector maps session events to OtelTraceSpan format.

Exporter failure isolation

The OTel subscriber runs inside the verb bus's subscriber crash-isolation wrapper. A thrown or rejected span export does not block other subscribers (audit, cost, eval) and is logged at warn level.

SIEM export

After a current-path audit append succeeds, fanOutToSIEM starts a non-blocking delivery attempt for each configured destination that matches the event filter. A destination failure is logged and does not roll back the audit entry. This is not a guarantee that every destination received every entry, and the audit trail itself does not prove that every security-relevant action was captured.

Supported formats

Format	Protocol	Compatible platforms
`json`	HTTP POST, JSON body	Generic JSON webhook collectors
`cef`	HTTP POST, ArcSight CEF string	ArcSight, HP ArcSight Logger
`leef`	HTTP POST, IBM LEEF string	IBM QRadar

Configuring a SIEM destination

SIEM destinations are managed per workspace via the API:

# Add a generic JSON webhook destination
curl -X POST https://app.rensei.ai/api/audit/siem/destinations \
  -H "Authorization: Bearer rsk_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "type": "http",
    "url": "https://siem.example.com/events",
    "format": "json",
    "apiKey": "<bearer-token>",
    "name": "Production SIEM",
    "filter": ["agent.spawned", "agent.completed", "approval.gate_created"]
  }'

The filter array is optional. When omitted, all audit event types are forwarded. When present, only matching eventType values are delivered.

When apiKey is present, the sender uses Authorization: Bearer <apiKey>. Vendor-specific authorization schemes, including direct Splunk HEC Authorization: Splunk ..., are not implemented by this generic sender.

Event types forwarded to SIEM

The current audit writer supports the following representative event types; this list is not a capture-completeness claim:

agent.spawned          agent.completed         agent.failed
user.login             user.logout             user.invited
user.deprovisioned     org.member_role_changed
approval.gate_created  approval.decision_submitted  approval.gate_resolved
policy.created         policy.evaluated
audit.chain_break_detected  audit.key_rotated  audit.anchored
sso.connection_created      scim.user_provisioned
m2m.client_created          m2m.token_issued
credential.rotated          workflow.deployed

See Audit Trail for the event taxonomy, explicit verifier, and proof limits.

SIEM destination health

Test delivery parameters before saving a destination:

curl -X POST https://app.rensei.ai/api/audit/siem/test \
  -H "Authorization: Bearer rsk_live_..." \
  -H "Content-Type: application/json" \
  -d '{
    "url": "https://siem.example.com/events",
    "format": "json",
    "apiKey": "<bearer-token>"
  }'
# Returns: { "ok": true } or { "ok": false, "error": "..." }

Destinations are stored in Redis under workspace:{id}:siem:destinations. Removing the destination stops further delivery immediately.

CLI observability commands

# Tail live audit events
rensei observability events tail --org acme-corp

# Query historical events by type
rensei observability events query \
  --org acme-corp \
  --type agent.spawned \
  --since 2026-06-01 \
  --until 2026-06-02 \
  --json

Audit Trail - hash-chain integrity, proof limits, retention, and optional anchoring
Cedar Policies - policy evaluation events that appear in the audit trail
Sessions Inspector - per-session 7-tab debug view including OTel spans

On this page