feat: add local run dashboard

[mikasenghaas]

Summary

Adds a local run dashboard for SFT / RL / Eval runs. Start with

# default (all runs in outputs/, port 7788)
uv run dashboard 

For example, populate with

# rl run
uv run rl @ examples/basic/reverse-text/rl.toml --run.name reverse-text-rl

# sft run
uv run sft @ examples/basic/reverse-text/sft.toml --max-steps 10 --run.name reverse-text-sft

# eval run (multi-agent, multi-branch)
uv run eval \
  --env.id agentic-judge \
  --env.taskset.id terminal-bench-2 \
  --env.taskset.tasks '["fix-git"]' \
  --env.solver.harness.id rlm \
  --env.solver.harness.summarize-at-tokens 4000 \
  --env.judge.harness.id bash \
  --run.name tb2-agentic-judge

Combines the best things from W&B + PI platform. Fully isolated agent-maintained logic in prime_rl.dashboard. Fully supersedes the previously advertised tmux viewer. Should basically be always on when working with prl locally.

Gallery

Configs

Metrics

Trace Overview

Trace Viewer

Logs

Breaking

• scripts/tmux.sh is removed — the dashboard replaces the tmux log-pane workflow (docs, example walkthroughs, and the monitor-run skill updated).

• Resolved per-component configs move from {run_dir}/configs/*.json to {run_dir}/configs/resolved/*.json; configs/ now holds the launch TOML (rl.toml/sft.toml). Anything reading the old paths must switch (the dashboard reads both layouts).

🤖 Generated with Claude Code

---

Note

Low Risk
Changes are additive for training (new CLI and docs) plus removal of the tmux helper; the config path split may break external tooling that still reads {run_dir}/configs/*.json directly.

Overview
This PR adds uv run dashboard — a local web UI (default http://localhost:7788) for watching SFT/RL/eval runs from output directories only: metrics (W&B-style overview from metrics.jsonl), configs (launch TOML vs merged resolved JSON), rollout traces with per-token advantage/logprob views, and merged component logs. It ships as a dashboard optional extra (FastAPI/uvicorn) or as uv run --script src/prime_rl/dashboard/server.py on machines without full GPU deps.

scripts/tmux.sh is removed; docs, basic examples, and the monitor-run skill now recommend two terminals plus the dashboard instead of tmux panes. Side-by-side ablations on one node can use uv run dashboard outputs to track multiple experiments.

Resolved configs from dry-run and launches are documented at {run_dir}/configs/resolved/*.json, with the launch TOML at configs/*.toml; the dashboard server still reads the old flat configs/*.json layout for older runs.

^{Reviewed by Cursor Bugbot for commit 99d9a8f. Bugbot is set up for automated code reviews on this repo. Configure here.}

[github-advanced-security]

CodeQL found more than 20 potential problems in the proposed changes. Check the Files changed tab for more details.

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 99d9a8f. Configure here.}

[cursor]

Stale partial-line episode summaries

High Severity

episode_summaries only invalidates by line count, while line_offsets intentionally rescans a possibly partial last line when the file grows. A live read of a large in-progress traces.jsonl can summarize a torn write as unparseable; once that count is cached (and written to the sidecar), later completion of the same line never reparses it, so the episode stays broken in the table for the rest of the session and across restarts.

Additional Locations (1)

• src/prime_rl/dashboard/server.py#L469-L499

 

^{Reviewed by Cursor Bugbot for commit 99d9a8f. Configure here.}