Runtime Wiring

In normal application code, telemetry is usually driven by runtime events rather than manual openRoot() calls.

The pattern is:

1. create one shared event dispatcher
2. create one shared Telemetry instance
3. create the projectors for the runtimes you use
4. attach them through RuntimeEventBridge
5. build your runtime objects with the same event dispatcher

Minimal Wiring Example

use Cognesy\Agents\Telemetry\AgentsTelemetryProjector;
use Cognesy\Events\Dispatchers\EventDispatcher;
use Cognesy\Http\Telemetry\HttpClientTelemetryProjector;
use Cognesy\Polyglot\Telemetry\PolyglotTelemetryProjector;
use Cognesy\Telemetry\Application\Projector\CompositeTelemetryProjector;
use Cognesy\Telemetry\Application\Projector\RuntimeEventBridge;

$events = new EventDispatcher('app');

(new RuntimeEventBridge(new CompositeTelemetryProjector([
    new AgentsTelemetryProjector($telemetry),
    new PolyglotTelemetryProjector($telemetry),
    new HttpClientTelemetryProjector($telemetry),
])))->attachTo($events);
// @doctest id="71c0"

Pass $events into the runtime objects that should emit telemetry.

Which Projectors To Add

Add only the projectors for the packages you actually use:

• agents: AgentsTelemetryProjector
• agent control: AgentCtrlTelemetryProjector
• instructor: InstructorTelemetryProjector
• polyglot: PolyglotTelemetryProjector
• http client: HttpClientTelemetryProjector

If a package has no matching projector attached, its events will not be turned into telemetry.

Metric Catalog

Each projector also emits metrics through Telemetry::metric(), using the canonical types from Cognesy\Metrics\Data\* (Counter, Gauge, Histogram, Timer). This is what the two runtime projectors emit today.

PolyglotTelemetryProjector

Metric	Type	Emitted on	Tags
inference.client.token.usage.input	Histogram	InferenceUsageReported (skipped when the token count is null)	inference.execution.id, inference.response.model, inference.usage.final
inference.client.token.usage.output	Histogram	same	same
inference.client.token.usage.total	Histogram	same, and EmbeddingsResponseReceived	same / embeddings attribute bag
inference.client.operation.count	Counter	InferenceCompleted -> inference.outcome=success; InferenceFailed -> inference.outcome=failure	inference.outcome, inference.finish_reason (success), http.response.status_code (failure)
inference.client.operation.duration	Timer	InferenceCompleted, when a non-negative duration is available	same as the success counter
inference.client.attempt.count	Counter	InferenceAttemptSucceeded -> success; InferenceAttemptFailed -> failure	inference.outcome, inference.finish_reason (success), error.type + http.response.status_code + inference.retry (failure)
inference.client.attempt.duration	Timer	either attempt event, when a non-negative duration is available	same as its counter
inference.embeddings.operation.count	Counter	EmbeddingsResponseReceived -> success; EmbeddingsFailed -> failure	inference.outcome, inference.response.model (success), http.response.status_code (failure)

PolyglotTelemetryProjector emits no Gauge. There is no point-in-time resource level on the inference path — per-call sizes are distributions, not levels — and forcing one would misrepresent the data. Gauges belong where a loop actually carries state, which is the agents projector below.

AgentsTelemetryProjector

Metric	Type	Emitted on	Tags
inference.client.token.usage.total	Histogram	TokenUsageReported	agent.id, inference.execution.id, agent.operation, agent.parent_id (+ envelope attributes)
agent.context.message_count	Gauge	AgentStepStarted	agent.is_subagent
agent.step.count	Counter	AgentStepCompleted	agent.has_tool_calls, inference.finish_reason, agent.is_subagent
agent.step.duration	Timer	AgentStepCompleted, when a non-negative duration is available	same
agent.execution.count	Counter	AgentExecutionCompleted -> agent.outcome=success; AgentExecutionFailed -> agent.outcome=failure	agent.status, agent.is_subagent, agent.outcome, error.type (failure only)
agent.execution.steps	Histogram	same two events	same
agent.tool_call.count	Counter	ToolCallCompleted -> agent.outcome=success\|error; ToolCallBlocked -> agent.outcome=blocked	agent.tool, agent.outcome
agent.tool_call.duration	Timer	ToolCallCompleted, when a non-negative duration is available	agent.tool, agent.outcome
agent.subagent.depth	Gauge	SubagentSpawning	agent.subagent
agent.subagent.count	Counter	SubagentCompleted	agent.subagent, agent.subagent.status

Type Policy

• event and failure totals -> Counter
• durations -> Timer
• point-in-time state (current context size, current subagent nesting depth) -> Gauge
• distributions (token counts, steps per run) -> Histogram

Timer::create() throws on a negative duration, so every duration Timer above is guarded: a projector must not let an observability path take down the call it is observing just because a clock went backwards. Each Timer is skipped, not clamped, when the duration is missing or negative.

Tag Discipline

Metric tags are aggregation dimensions, not span attributes, so every tag value must be low-cardinality: tool names, subagent names, statuses, outcomes, booleans. Per-run identifiers (agent.id, execution ids) belong on spans, not tags — one tag value per run means one time series per run in the metrics backend.

inference.client.token.usage.* is the single deliberate exception, and it is exempt from the rule as a whole rather than for one tag: it is a correlation-carrying metric, not an aggregation-friendly one. It carries inference.execution.id because LangfusePayloadMapper::matchesMetric() correlates a metric back to its span by that id, which already makes it one series per run — so the agent.id the agents projector also puts on it costs nothing further and stays useful for grouping token spend by agent. Every other metric above carries no id tag at all, so it never matches an observation and never reaches attributesForMetric() — which returns [] for unknown names anyway.

Cognesy\Telemetry\Domain\Metric\MetricNames holds the token-usage metric names and the inference.execution.id / inference.usage.final tag keys shared by PolyglotTelemetryProjector, AgentsTelemetryProjector and LangfusePayloadMapper, so the three cannot drift apart. Metric names used by a single producer (agent.*, inference.embeddings.*) stay private to that producer and are not in MetricNames.

Practical Examples

The examples directory has working end-to-end setups:

• examples/D05_AgentTroubleshooting/TelemetryLangfuse/run.php
• examples/D05_AgentTroubleshooting/TelemetryLogfire/run.php
• examples/D05_AgentTroubleshooting/SubagentTelemetryLangfuse/run.php
• examples/D05_AgentTroubleshooting/SubagentTelemetryLogfire/run.php

These examples keep normal console output and add telemetry export at the same time.