Every run is recorded to a local trace database (database/db.sqlite). /trace expands one into the full replay the live rail abbreviates: the query, every step of the loop with its timing, the checklist if the agent wrote one, the agent's reasoning and tool choices at each pass, each tool call with its output, every gate decision, and, last, the recorded answer. It's the execution log, not a reprint of the answer.
| command | shows |
|---|---|
/trace | the last run's drill-down; /trace #id any run; /trace -l [n] lists runs |
/trace why [#id] | why it did what it did: the checklist, each pass's thought and tool choice, the evidence the answer was built from, and when each pass thought |
/trace source [n] | the full material behind source n of the last answer |
/trace invoke [#id] | every model call's input and output, with timing and tokens; --full shows whole messages, exactly what your machine sent (/trace context is the same view) |
/trace export [#id] | write the run's complete record as JSON to logging/exports/ (-o <path> to choose) |
/trace replay <file> | re-render an exported record, offline |
/trace on|off|full | live rail verbosity; recording is always on |
::the export record
/trace export (or headless --export FILE, or -q, which exports automatically) writes one self-contained JSON file: every event, every tool call and its result, every model call, and every human gate decision. saturn --replay <file> renders it anywhere, offline, with no database and no models. You can hand someone a record of exactly what the agent did, and they can replay it.
saturn --replay logging/exports/run_1.json::live, while it runs
- –The rail shows each tool call with a one-line result preview, the agent's words before a call as a leaf under its row, and every gate decision. An auto-approved call doesn't print a gate row.
- –A pass that thought shows
thought 1.8son its row and a leaf with why it thought and the opening of the thought. A thought that was cut, stopped, or came back empty says so, and the pass answered without it. - –The status bar carries
thinking 3swhile a thought is in flight (Esc stops it), the context-fill meter, tok/s, the egress count once it's above zero, and the machine:gpu 31% · mem 21.4/36 GB, GPU utilisation and unified memory in use, sampled every two seconds. Memory turns yellow at 85%. - –The receipt under each answer keeps what the bar showed once it's gone: time, passes, tools, tok/s, and the turn's thinking time, followed by the turn's trust summary.
- –The rail fails soft: a display bug prints as one line, the run stays recorded, and the answer still arrives.