Dashboard
Tracer's dashboard is the progress and trajectory monitor for rollout jobs. Run
/tracer:dashboard to inspect jobs, task outcomes, trajectory previews, and the
optional SFT conversion output.
A demo dashboard is available here. The checked-in demo is a static UI preview; full-trace loading requires a generated dashboard with trajectory shards.
Deployment mode
Both modes read the same files, so the numbers never depend on which one you pick.
- Local (default): run
/tracer:dashboard. The board is built and served on the machine that ran the job, so prompts, tool output, model responses and failure traces stay here. - Cloud: run
/tracer:dashboard publish to Cloudflare. SetCLOUDFLARE_API_TOKEN,CLOUDFLARE_ACCOUNT_IDand the target Pages project, for examplelegoflow-tracer. Full online trajectory loading also needs the matching R2 bucket binding; without it, publish metrics and bounded previews only.
Where it reads from
The sources are fixed rather than configured. config.yaml does not name them,
and editing it will not change what the board shows:
| Path | What it holds |
|---|---|
artifacts/tasks/ | Staged task batches, one per immediate subdirectory. |
artifacts/jobs/ | Harbor jobs, one per immediate subdirectory. |
artifacts/sft_data/ | Converted datasets, one per immediate subdirectory. Optional. |
artifacts/index.yaml | Archived run history retained in the dashboard's source report. |
artifacts/processed_tasks.yaml | The consumption ledger, which records what has already been run. |
To see exactly what the board resolved, without rendering anything:
./dashboard/progress_monitor.py --report-onlyThat prints the absolute path of every source and the batches found under each.
/tracer:dashboard runs it and waits for your confirmation before rendering.
By default the trajectories come from rollouts run on this machine. To bring in trajectories generated elsewhere:
- Arrange them in the Harbor job layout: one directory per job, one directory
per task inside it, and the agent output under
<task>/agent/, matching an existing job underartifacts/jobs/. - Put that job directory under
artifacts/jobs/, or symlink it there. The directory name becomes the job name on the board. - Convert it if you want the SFT panels, following
Trajectory Conversion.
Conversion writes
artifacts/sft_data/<job>/lf.json, and it is also what computes the quality scores.
If you already hold LF-format SFT data and no raw trajectories, put it at
artifacts/sft_data/<name>/lf.json and the SFT panels will pick it up on their
own.
Import External Data
The board only reads artifacts/. Symlink an outside job into artifacts/jobs/,
or an LF dataset into artifacts/sft_data/, and the link name becomes the name
on the board.
SFT data is optional
Trajectory scores are computed during conversion, so a block that has not converted anything has no scores yet. The board then leaves out the score surfaces, meaning the score distribution, the score matrix and the scoring rubric, rather than drawing them empty. Jobs, trials, pass rate and task coverage still render. An absent score panel means nothing was scored, not that the trajectories scored zero.
What it monitors
| Monitor | What it tells you |
|---|---|
| Harbor jobs | Job-level task counts, reward summaries, retries, and errors. |
| Task coverage | Which tasks were attempted, completed, skipped, failed, or already processed. |
| Agent trajectories | Per-task messages, tool calls, observations, bounded previews, and failure traces. |
| SFT conversion | Whether successful rollouts were converted into LF data for Trainer. (only with SFT data) |
| Quality signals | Score distributions, low-score examples, tool-error rates, and segment comparisons. (only with SFT data) |
When to Use It
Use /tracer:dashboard when you want to answer questions like:
- How many tasks resolved successfully?
- Which tasks failed repeatedly and should be excluded?
- Which scaffold or model is producing better trajectories?
- Did the run produce Trainer-ready SFT data?
- Are errors coming from the agent, environment, verifier, or model endpoint?
For rollout mechanics, see Design. For artifact contracts, see Output Format.