What problem does it solve? When an AI agent misbehaves, the evidence is scattered across four live systems — Postgres checkpoints hold the conversation, Loki holds the event stream, Tempo holds the span tree, and Grafana holds the browsable view. This Skill provides the correlation know-how to go from a symptom like "agent 3048 behaved badly around 14:00" to the exact conversation, events, spans, and a shareable trace link. ## Core Features & Use Cases - Run discovery: Narrow to one agent's run by agent_id and time window in Loki, or query Postgres checkpoints for conversation content and history across compaction segments. - Trace toolchain: Three chained scripts (fetch_trace.py, read_trace.py, render_trace.py) find traces via TraceQL, fetch full span sets from the local mirror or Tempo, and render a self-contained HTML waterfall report. - Failure diagnosis recipes: Ready-made LogQL queries for exec failures, LLM spend, turn outcomes, halt reasons, and crash loops. - Use Case: A user reports an agent looped and burned tokens yesterday. You query Loki for the agent's event river, extract the trace_id of the bad turn, fetch its spans from the mirror, join the checkpoint messages via the gateway, and hand the user a Grafana Explore link. ## Quick Start Ask the assistant to inspect what agent 3048 did around 14:00 yesterday and produce a trace report with a Grafana link.
