L’interface est localisée ; les présentations et documents des Skills sont traduits automatiquement par IA dans votre langue. Le code, les commandes, chemins et termes techniques restent en version originale, et le document de l’auteur fait foi.
Explorer les Skills
Productivité · SKILL.md

tracing-and-instrumentation

tracing-and-instrumentation

Emit, inspect, adapt, and troubleshoot Agent Lightning spans, rewards, operation traces, OpenTelemetry/AgentOps tracers, and token-ID signals.

Independent GitHub snapshot · Automated classification · Original documentation governs

Documentation de la Skill

GitHub

name: tracing-and-instrumentation description: "Emit, inspect, adapt, and troubleshoot Agent Lightning spans, rewards, operation traces, OpenTelemetry/AgentOps tracers, and token-ID signals." disable-model-invocation: true metadata: disco-role: operating license: MIT

Tracing and instrumentation

Use this sub-skill when a task is about spans, rewards, trace collection, adapters, OpenTelemetry/AgentOps/Weave tracer choices, or missing token/logprob signals.

Route by task

RequestRead/run
Emit rewards, messages, objects, exceptions, or operation spansreferences/tracing-workflows.md
Inspect final reward or convert traces for algorithmsreferences/api-reference.md
Debug no spans, no active tracer, missing reward, disconnected spans, token IDsreferences/troubleshooting.md
Run a local no-service trace smokepython scripts/local_trace_smoke.py
Start stores/proxies/metrics servicesroute to ../cli-and-services/SKILL.md

Key rules

  • OtelTracer is the minimal local tracer for explicit OpenTelemetry and emitter tests.
  • AgentOpsTracer is the default trainer tracer and instruments many LLM/agent frameworks locally.
  • Emitters require an active tracer unless propagate=False is used for offline local span creation.
  • Agent Lightning reward spans are annotation spans; use find_final_reward for the last reward value.
  • Token IDs are optional signals. VERL/vLLM-style training may require token IDs from a compatible serving path; do not claim they exist unless the trace shows them.

Minimal local pattern

import agentlightning as agl

tracer = agl.OtelTracer()
store = agl.InMemoryLightningStore()
rollout = await store.start_rollout(input={"origin": "debug"})
with tracer.lifespan(store):
    async with tracer.trace_context(
        "debug-trace",
        store=store,
        rollout_id=rollout.rollout_id,
        attempt_id=rollout.attempt.attempt_id,
    ):
        agl.emit_message("hello")
        agl.emit_reward(1.0)
spans = await store.query_spans(rollout.rollout_id)
assert agl.find_final_reward(spans) == 1.0

The bundled scripts/local_trace_smoke.py runs this with extra operation, tag, link, and object assertions.

Boundary

This sub-skill owns spans and adapters. It does not own agent function signatures, store status transitions, or service startup. Route those to agent-authoring, runner-store-training, and cli-and-services.

Skills enregistrées