---

name: observability-and-hooks description: "Guides agents configuring CrewAI telemetry, tracing, observability integrations, event listeners, execution hooks, kickoff hooks, output logs, task outputs, and security fingerprints." disable-model-invocation: true metadata: disco-role: operating license: MIT

CrewAI Observability and Hooks

Use this sub-skill when a task involves CrewAI tracing, anonymous telemetry, observability providers, event listeners, LLM/tool execution hooks, before/after kickoff hooks, output logs, latest task outputs, or component fingerprints.

Route by Intent

• Need Crew/Flow tracing enablement, telemetry disable switches, privacy boundaries, output logs, latest task outputs, or replay-observability data? Read Tracing reference.
• Need @before_kickoff, @after_kickoff, LLM hooks, tool hooks, hook registration/cleanup, event listener setup, callback signatures, or hook ordering? Read Hooks reference.
• Need OpenTelemetry exporter patterns, OpenLIT, Datadog, Langfuse, LangDB, Phoenix, Portkey, Weave, Opik, MLflow, Patronus, or other provider integration boundaries? Read Integrations.
• Need to debug duplicate/missing traces, env override confusion, endpoint/protocol mismatches, callback signature errors, hooks overriding each other, log/task-output gaps, or fingerprint changes? Read Troubleshooting.
• Need a safe read-only snapshot of tracing-related env vars and installed observability packages? Run check_tracing_config.py with --help first.

Boundaries

• Stay here for tracing, telemetry, event listeners used for monitoring, execution hooks, kickoff hooks, output logs, task output audit data, and fingerprints.
• Use flows-and-events for Flow, @start, @listen, @router, flow graph mechanics, route labels, plotting, and flow persistence semantics.
• Use cli-and-projects for exact crewai command syntax, project scaffolding, crewai traces, crewai log-tasks-outputs, replay flags, and deployment commands.
• Use core-runtime for Agent, Task, Crew, Process, CrewOutput, TaskOutput, guardrails, callbacks outside observability, and core kickoff behavior.
• Use tools-and-mcp for tool implementation, official tool classes, MCP adapters, and tool publishing beyond hook interception.
• Return to the CrewAI root skill when a request spans multiple CrewAI capability areas and needs root routing context.

Safe Defaults

• Prefer read-only inspection first: env/package checks, config review, and hook registration review before running a crew, flow, tool, LLM, or external exporter.
• Treat hosted trace dashboards and third-party observability providers as credential-bound: document required env vars, but do not send traces or initialize SDKs without approval.
• Use tracing=False on a specific Crew/Flow when a run must not emit CrewAI built-in traces; use telemetry disable env vars for OpenTelemetry-style anonymous telemetry spans.
• Keep hooks small and deterministic; avoid network calls, destructive tools, blocking prompts, or secret-printing inside global hooks unless the user explicitly approves the side effect.
• Clear or unregister global hooks in tests, notebooks, and long-lived workers to prevent duplicate behavior across repeated imports or @CrewBase instantiations.