agent-observabilitylisted

# Agent Observability Make agent behavior explainable from request entry through model, retrieval, tool, handoff, approval, retry, and final-response spans. This Skill designs the evidence contract; it does not enable production capture, export data, or widen telemetry access without authorization. ## Define the boundary and questions Record the workflow graph, runtime boundaries, incident questions, traffic and failure expectations, telemetry stack, data classification, retention policy, sampling limits, owners, and what cannot be observed. Start with decisions such as “Which tool causes timeouts?” or “Why did cost per completed task rise?” Do not collect fields without a documented use. ## Trace and event contract Produce one trace per user-visible attempt. Define spans for model calls, retrieval, tools, handoffs, approvals, retries, and final validation. Preserve parent-child relationships and propagate a correlation ID across queues. For each field specify its meaning, unit, owner, cardinality, retention, and redaction rule. Useful stable metadata includes versions, status, duration, token and cost measures, retry counts, tool names, policy outcomes, and evaluation tags. Separate content from metadata. Default to content-free telemetry. Prompt, response, or tool-payload capture requires explicit authorization, reliable redaction, least-privilege access, a retention limit, and an audit trail. Never record secrets, tokens, raw credentials, payment data, or unapproved