Skip to main content
/trace opens a viewer over the session logs on your machine — Helix sessions and, where present, plain Pi and Claude Code sessions too. Pick a session and walk it: step by step, or as a waterfall of real durations. It is read-only — nothing in it re-runs, steers, deletes, or writes to a session file — which is what makes it safe to open on a live run. It works in every mode.
Looking for where Evaluate reads its evidence? That’s Traces — a different thing that shares the name.
The trace viewer's session list: a header counting ten indexed sessions across one harness, then ten rows spanning three projects — checkout-api, pricing-engine and docs-site — each row carrying the session id, harness, model, start time, a duration ranging from 2.6 seconds to 6 minutes 30, a turn count, and a sub-agent count on the two sessions that dispatched any.

Where /trace opens: every session on the machine, one row each — project, harness, model, duration, turns and sub-agents.

The trace viewer's waterfall screen for a demo session: a header with spans, turns, model wait and idle time, a time axis from 0s to 2m31s, four Agent dispatch spans, four long parallel sub-agent lanes of different lengths, and the assistant message spans between them.

The waterfall: four sub-agents dispatched in parallel, each lane a real duration — what ran long, what waited on what.

​
Screens

The trace viewer's trajectory screen: numbered rows for user, assistant, tool and sub-agent entries with token counts, cost and relative timestamps, including four Agent dispatch rows and their four sub-agent children.

The trajectory: the same session step by step — dispatches, sub-agent rows, and per-step tokens, cost and timing.

​
Keys