Observability in Talon

Talon uses OpenTelemetry for traces and metrics, and zerolog for structured logs. Evidence records are stored in SQLite for compliance; traces and metrics are for operational observability.

Enabling OpenTelemetry

OTel is off by default. Enable it in one of these ways:

• Flag: talon --otel run "query" or talon --otel serve
• Environment: TALON_OTEL_ENABLED=true
• Verbose: talon -v ... also enables OTel (for development)

Example (production-style, without turning on verbose logs):

export TALON_OTEL_ENABLED=true
talon serve

Or:

talon --otel serve

What is exported

• Traces: Span hierarchy for agent runs, policy evaluation, LLM calls, evidence store, secrets, memory, and HTTP requests (when using talon serve). Spans include correlation_id, tenant_id, agent_id, and GenAI attributes where applicable.
• Metrics: 25+ OTel instruments across all subsystems (see Metrics reference below). Exported via OTLP or stdout depending on configuration.
• Logs: Structured JSON or console via zerolog. Key log lines include trace_id and span_id when OTel is enabled so logs can be correlated with traces in a backend.

Export destination is stdout by default. Use the OTLP exporter to send traces and metrics to a collector (e.g. Jaeger, Prometheus, Grafana). See examples/observability for a ready-made local stack.

---

Metrics reference

All metrics are registered via the OpenTelemetry Go SDK. The tables below group them by subsystem. Every metric follows the naming convention talon.<subsystem>.<metric> for custom metrics or GenAI Semantic Conventions for standard LLM telemetry.

LLM / GenAI

Registered by internal/llm. Emitted on every LLM call.

Metric	Type	Unit	Attributes	Description
talon.cost.request	Float64Histogram	eur	agent, model, degraded	Cost in EUR per LLM request.
gen_ai.client.token.usage	Int64Histogram	{token}	gen_ai.system, gen_ai.request.model, gen_ai.token.type	Token usage per LLM request (input/output). GenAI SemConv.
gen_ai.client.operation.duration	Float64Histogram	s	gen_ai.system, gen_ai.request.model	End-to-end LLM operation duration. GenAI SemConv.
gen_ai.server.time_to_first_token	Float64Histogram	s	gen_ai.system, gen_ai.request.model	Time from request sent to first content token (streaming). GenAI SemConv.
gen_ai.server.time_per_output_token	Float64Histogram	s	gen_ai.system, gen_ai.request.model	Time per output token after first token (streaming decode phase). GenAI SemConv.
talon.provider.availability	Float64Gauge	1	provider	Provider availability (1 = up, 0 = down).
talon.provider.failover.total	Int64Counter	{failover}	original_model, fallback_model, reason	Provider failover events (cost degradation or unavailability).

Gateway

Registered by internal/gateway. Emitted for every request through the LLM API gateway.

Metric	Type	Unit	Attributes	Description
talon.gateway.requests.total	Int64Counter	{request}	caller, model, gen_ai.system, status	Total gateway proxy requests.
talon.gateway.errors.total	Int64Counter	{error}	error_type	Gateway errors by type (auth, policy, provider, timeout).
talon.data_tier.requests	Int64Counter	{request}	tier, caller	Requests by data classification tier (0/1/2).
talon.tools.governance.total	Int64Counter	{decision}	tool, action	Tool governance decisions (allow, block, filter).
talon.cache.hits	Int64Counter	{hit}	tenant_id	Semantic cache hits (request served from cache).
talon.cache.misses	Int64Counter	{miss}	tenant_id	Semantic cache misses (forwarded to LLM).
talon.shadow.violations.total	Int64Counter	{violation}	violation_type	Shadow mode violations (would-have-blocked in enforce mode).
talon.gateway.egress.decisions	Int64Counter	{decision}	tenant_id, tier, gen_ai.system, region, decision	Egress policy decisions (destination × data tier), decision is allow or deny.
talon.budget.utilization	Float64Gauge	%	tenant_id, period	Current budget utilization as a percentage.
talon.budget.alerts.total	Int64Counter	{alert}	tenant_id, threshold	Budget threshold breach alerts.

When an egress policy is configured, the gateway request span also carries talon.egress.* attributes: caller, correlation_id, data_tier, destination_provider, destination_region, decision, and reason (machine code, empty on allow).

Routing spans

The llm.route / llm.graceful_route spans (agent runs) carry talon.data.tier and, when compliance-aware routing is active, talon.routing.sovereignty_mode, talon.provider.jurisdiction, talon.provider.region, talon.routing.rejected_count, and talon.routing.selection_reason.

Policy engine

Registered by internal/policy. Emitted on every policy evaluation (OPA).

Metric	Type	Unit	Attributes	Description
talon.policy.evaluations.total	Int64Counter	{evaluation}	decision, tenant_id, agent_id	Policy evaluation count by decision (allow/deny).
talon.policy.evaluation.duration	Float64Histogram	ms	tenant_id	Policy evaluation latency in milliseconds.

PII classifier

Registered by internal/classifier. Emitted when PII is detected or redacted.

Metric	Type	Unit	Attributes	Description
talon.pii.detections.total	Int64Counter	{detection}	pii_type, direction, action	PII entities detected (email, IBAN, phone, etc.).
talon.pii.redactions.total	Int64Counter	{redaction}	pii_type, direction	PII entities redacted before forwarding.

Evidence store

Registered by internal/evidence. Emitted on every evidence write or verification.

Metric	Type	Unit	Attributes	Description
talon.evidence.records.total	Int64Counter	{record}	type	Evidence records stored (LLM, tool, secret, memory).
talon.evidence.signature_verifications	Int64Counter	{verification}	result	HMAC signature verification attempts (success/failure).

Secrets vault

Registered by internal/secrets. Emitted on every secret access attempt.

Metric	Type	Unit	Attributes	Description
talon.secrets.access.total	Int64Counter	{access}	secret_name, agent_id, outcome	Secret access attempts (granted/denied).

Attachment scanner

Registered by internal/attachment. Emitted when prompt injection patterns are detected.

Metric	Type	Unit	Attributes	Description
talon.injection.attempts.total	Int64Counter	{attempt}	detection_type, action	Prompt injection attempts detected in attachments.

Agent memory

Registered by internal/memory. Emitted on memory operations.

Metric	Type	Unit	Attributes	Description
memory.writes.total	Int64Counter	{write}	category, tenant_id	Memory entries written.
memory.writes.denied	Int64Counter	{write}	reason	Memory writes denied (PII, forbidden category).
memory.conflicts.detected	Int64Counter	{conflict}	resolution	Memory conflicts detected during consolidation.
memory.reads.total	Int64Counter	{read}	source, tenant_id	Memory entries read (for prompt injection).
memory.entries.count	Int64Gauge	{entry}	tenant_id, agent_id	Current memory entry count per agent.
memory.dedup.skips	Int64Counter	{skip}	tenant_id	Duplicate memory writes suppressed by dedup window.
memory.consolidation.noops	Int64Counter	{noop}	—	Consolidation runs that found nothing to merge.
memory.consolidation.invalidations	Int64Counter	{invalidation}	—	Entries invalidated during consolidation.
memory.consolidation.updates	Int64Counter	{update}	—	Entries updated during consolidation.
talon.memory.poisoning.blocked	Int64Counter	{block}	reason, agent_id	Memory poisoning attempts blocked by governance.

---

Gateway dashboard metrics

In addition to OTel metrics, Talon provides a real-time runtime dashboard with an in-memory metrics collector. The collector is fed from successful evidence.Store.Store() commits (all invocation types), then periodically reconciled from evidence as an authoritative repair path.

The dashboard snapshot is available at:

• HTML dashboard: GET /gateway/dashboard — single-page HTML with auto-refreshing charts.
• JSON API: GET /api/v1/metrics — full snapshot for programmatic access.
• SSE stream: GET /api/v1/metrics/stream — Server-Sent Events, one snapshot every 5 seconds.

The dashboard snapshot includes:

Field	Type	Description
summary.total_requests	int	Total requests processed.
summary.blocked_requests	int	Requests denied by policy.
summary.pii_detections	int	PII entities found across all requests.
summary.pii_redactions	int	PII entities redacted.
summary.tools_filtered	int	Tool calls filtered by governance.
summary.total_cost_eur	float	Cumulative cost in EUR.
summary.avg_latency_ms	int	Average request latency.
summary.p99_latency_ms	int	P99 request latency.
summary.error_rate	float	Error rate (0.0–1.0).
summary.active_runs	int	Currently executing agent runs.
summary.pending_plans	int	Plans awaiting human review.
summary.approved_plans	int	Plans approved by reviewers.
summary.rejected_plans	int	Plans rejected by reviewers.
summary.modified_plans	int	Plans approved with modifications.
summary.dispatched_plans	int	Approved plans already dispatched/executed.
summary.plan_dispatch_errors	int	Dispatched plans that recorded execution/dispatch errors.
requests_timeline	array	5-minute bucketed request counts.
pii_timeline	array	5-minute bucketed PII detection counts.
cost_timeline	array	5-minute bucketed cost in EUR.
caller_stats	array	Per-caller aggregates (requests, PII, blocked, cost, latency).
pii_breakdown	array	Detections broken down by PII type (email, IBAN, phone, etc.).
model_breakdown	array	Requests and cost broken down by LLM model.
tool_governance	object	Tool filtering stats (total, filtered, by risk level, anomalous agents).
shadow_summary	object	Shadow mode violation summary (only present in shadow mode).
budget_status	object	Budget utilization (daily/monthly used, limit, percentage).
cache_stats	object	Semantic cache performance (hits, hit rate, cost saved).
plan_stats	object	Plan lifecycle counters (pending/approved/rejected/modified/dispatched/failures).
dropped_events	int	Collector events dropped due to in-process backpressure.

See Gateway dashboard reference for full configuration, authentication, and API details.

Operational-event projection

Projection: Evidence → OperationalEvent → Metrics / UI / CLI. Metrics emit only after evidence persists, and live increments are driven by store post-commit observer notifications. Collector overflow is exposed as dropped_events in the snapshot, metrics_events_dropped in /v1/status, and OTel counter talon.metrics.events_dropped.total. Periodic reconciliation from evidence is bounded and idempotent, used to repair drift rather than define an alternate source of truth.

Surface	Endpoint / source	Scope	Parity expectation
Evidence list	/v1/evidence	all evidence rows in tenant window	authoritative ordering: timestamp DESC, id DESC
Events API	/api/v1/events/recent, /api/v1/events/stream	terminal_plus_lifecycle_subset (evidence-backed only)	same ordering and evidence-linked fields
Dashboard	/dashboard recent-events table	mirrors events API rows, with compact signal chips	reflects events API without manual refresh
Metrics snapshot	/api/v1/metrics	all_activity collector projection	consistent summary/breakdown invariants and reconciliation health
Status	/v1/status	reliability contract fields	exposes metrics_events_dropped, events_stream_gaps, events_replay_misses, events_backlog_drops, reconcile status

---

Log–trace correlation

When OTel is enabled, critical log events (e.g. agent_run_started, agent_run_completed, failed_to_generate_evidence) include trace_id and span_id. In an observability backend that ingests both logs and traces, you can jump from a log line to the corresponding trace.

Structured log fields

Logs consistently include:

• correlation_id – unique per agent run
• tenant_id – tenant scope
• agent_id – agent name
• trace_id / span_id – when OTel is enabled

Use these for filtering and correlation in your log aggregation (e.g. Elasticsearch, Loki).

HTTP server tracing

When running talon serve, the chi router uses OTel middleware. Each HTTP request (health, status, webhooks) gets a root span; agent runs triggered by webhooks appear as child spans, so you can see the full request → run → LLM/tool chain in one trace.

Local observability stack

A ready-made Docker Compose stack is provided in examples/observability/. It includes:

• OpenTelemetry Collector — receives OTLP from Talon, exports to Prometheus
• Prometheus — scrapes metrics from the Collector
• Grafana — pre-built dashboard for all talon.* and gen_ai.* metrics

Start the stack:

cd examples/observability
docker compose up -d

Then configure Talon to export to the Collector:

export TALON_OTEL_ENABLED=true
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317
talon serve --gateway

Open Grafana at http://localhost:3000 (admin/admin) to see the pre-built Talon Gateway dashboard.

See examples/observability/README.md for details.