Instruction file imported from ncosentino/needlr (.github/instructions/genesis/agentic/agent-observability.instructions.md). Copyright stays with the author.

Agent Observability Rules

Apply these rules to telemetry for AI/LLM agents, model calls, model-driven workflows, and model-invoked tools. A non-agent class that merely matches a generic name such as *Tool follows ordinary telemetry guidance instead.

• Record the model identifier, latency, input/output token counts, cached and reasoning tokens when reported, iterations, tool calls, retries, and termination reason.
• Trace model calls, tool invocations, and other costly or failure-prone boundaries. Do not create a span for every pure prompt-construction helper.
• Keep operation names and metric dimensions bounded. Put run-specific identifiers only in allowlisted trace or log fields, never metric labels.
• Do not record full prompts, model responses, retrieved documents, tool payloads, secrets, credentials, or PII by default.
• Correlate model calls, tools, workflow stages, logs, and traces through supported context propagation rather than copying payloads between signals.
• Calculate monetary cost from versioned configuration or provider data; never hardcode pricing into instrumentation logic.