::docs :: reference

observability & replay

/trace: the drill-down of any recorded run, why it decided what it did, what the model was actually sent, exports, and offline replay.

Every run is recorded to a local trace database. /trace expands one into the full replay the live rail abbreviates: the query, every node with its timing and metrics, the plan as it advanced, the agent's reasoning and tool-call decisions at each step, each tool call with its output, every gate decision — and, last and de-emphasized, the recorded final answer. It is the execution log, not a reprint of the answer.

commandshows
/tracethe last run's drill-down; /trace #id any run; /trace -l [n] lists runs
/trace whythe run's decisions explained — each step's reasoning and the judge's verdicts
/trace answeranswer-level provenance: each cited source's origin and trust, what left the machine, the human gate decisions
/trace source <n>the full material behind citation [n] of the last answer
/trace invokewhat each model call saw and said
/trace context [--node <name>] [--preview]exactly what your machine sent the model — every system prompt and context block, token-for-token, per node
/trace exportwrite the run's complete record as JSON to logging/exports/
/trace replay <file>re-render an exported record
/trace on|off|fulllive trace verbosity

::the export record

/trace export (or headless --export FILE, or -q automatically) writes one JSON file with the plan, every tool call and observation, every model call, and every human gate decision. saturn --replay <file> renders it anywhere, offline, with no database — a shareable, replayable artifact of exactly what the agent did. Oversized values are clipped with an explicit truncated marker naming what was dropped, never a JSON slice, and the replay discloses it under the node row.

offline replay
saturn --replay logging/exports/run_1.json

::what is not there

/trace calls, /trace cost, /trace state, and the --md export format were removed: calls duplicated the drill-down, cost measured cloud-era spend a local agent doesn't have (tok/s and context fill are live in the status bar), and the JSON export was always the one replayable record.

note

The live rail fails soft: a display bug while rendering the trace or plan prints as one line, the run stays recorded, and the answer still arrives.