See what the agent actually did.

The bundled CSV contains eight timestamped events. After sorting by timestamp, retry attempts are counted whenever a nonempty tool name equals the previous event tool name. Three adjacent write file events therefore create two retries and a longest streak of three, even though the third write succeeds. Errors are counted independently from status; the sample contains two error events. The sample latencies sum to two thousand seven hundred forty milliseconds. The test event contributes eight hundred, the largest sample latency. The rail allocates width proportional to each latency but enforces a minimum rectangle width, so short rectangles are not perfectly proportional. Supplied latency is summed; timestamp gaps are not substituted for it. There is no evidence of provider tokens, hidden reasoning or actual external effects. Four sample rows have type action; two have status ok. Action success is therefore fifty percent, not six successful statuses divided by all eight events. Rows with any nonempty tool name count as tool calls, giving seven calls across five unique names. The metrics are calculated from the data, but the displayed write file finding and slowest test sentence contain fixed sample wording that can be misleading for different traces. Offline parser or chart libraries may be unavailable; this draft does not certify a live audit.