feat: add local run dashboard

[mikasenghaas]

Summary

• Add dashboard/, a local web dashboard for run output directories: ./dashboard/dashboard.py <output_dir> (default outputs/), with a run selector for every run in the directory and four tabs: Metrics · Config · Traces · Logs.
• Run overview card on every tab: a step N headline (naturally fits unbounded runs; of max shown when set) with STATUS (running/completed/stopped, inferred from metrics.jsonl freshness), DURATION (live-ticking), MODEL, RESUMES (attempt count), TRAIN/EVAL ENVS (long lists fold into +N with a tooltip), and CREATED.
• Design: Prime Intellect language from prime-context (colors_and_type.css / brand.md) — #111 canvas, lime #B6FF3C accent used categorically, mono-dominant type with mono-caps labels, hairline dividers, zero border radii, butterfly mark top-left. Chart lines default to the brand palette purple (#B7A6FA); extra series pull from the rest of the palette. ABC Favorit is licensed and not vendored — the stack falls back to JetBrains Mono / system mono.
• Metrics (default screen): replicates the W&B overview sections (monitors/wandb/overview.py, RL and SFT flavors) from metrics.jsonl. Rows merge by step; the trainer/orchestrator time/step and time/save_ckpt collisions render as separate labeled series. Charts show a hover popover with x/y values, panel titles show the matched keys, sections are collapsible, panes reorder by dragging their header (persisted), a plain click resets drag-zoom, and a rolling-average smoothing slider (window 1–50) overlays the smoothed line on a faint raw ghost; and wandb-style resize handles on every pane (bottom edge = height, right edge = width, corner = both) resize all panes together — both persisted in localStorage. The all-metrics mode charts every logged key in nested collapsible sections along key paths (train → env → all/effective) behind a prominent regex filter; off-screen charts mount lazily. Reward/advantage columns in the trace table are green/red coded, episodes scroll without pagination, every empty case (no runs, no configs, no rollouts, no matching lines) shows a shared fixed-width empty-state box, and the live indicator is a pulsing dot. Inference metrics now fan out to every registered monitor as time-keyed rows (step: null) — W&B, the file monitor, and the Prime monitor alike (monitors.log(..., step=None); Monitor.log_metrics accepts step: int | None, and the inference/* define_metric moved into the W&B monitor init); the dashboard plots the overview inference panels (fleet aggregate + cross-engine tail) against relative wall time with s/m/h/d ticks. Step charts use integer x ticks and axes anchored at zero.
• Config: the run's resolved config JSONs (configs/**/*.json, rl.json first), pretty-printed with syntax highlighting and a regex search that prunes the JSON to matching subtrees (a matching key keeps its whole dict) with highlighted hits and a count.
• Traces: per-step rollout viewer over rollouts/step_N/{train,eval}/{all,effective}/traces.jsonl with a full-width block step selector (one block per step, filled through the selection, eval steps marked) + prev/next, fixed-width reward/advantage columns, env and error filters, env/sort/errors-only filters in a popup, and a platform-style trace viewer: prev/next step navigation with a step header, rollout selector with search and prev/next on the left, numbered collapsible message entries (per-message copy, collapse/expand all, reasoning and tool-call sub-blocks) in the center, and an overview pane (reward with per-component rewards and env metrics, copyable ids incl. runtime, turns and tool calls, token usage, tool definitions, state, timing tree) on the right — plus a per-token heatmap (advantage / logprob / mask / is_content) via server-side tokenizer decode.
• Logs: one merged pane by default, interleaving the selected components by timestamp, with a components dropdown (trainer / orchestrator / inference / evals / envs) and a merge/split toggle that switches to stacked per-component panes (per-pane file select, maximize). Masters are the default files, the inference view is engine-only (single-node router lines hidden, viewable via a virtual "router" entry), env logs merge, follow is always on, shared regex + level filters, attempt selector (latest by default), abbreviated line counts, byte-range reads only (an 8 MB inference.log is never read whole). ANSI colors render as HTML and torchrun tee prefixes are stripped.
• Works for SFT runs (SFT overview flavor; traces are eval-only) and tolerates the older episode schema (run.metadata, no token arrays).
• Standalone PEP 723 uv script (FastAPI + uvicorn + tokenizers); project pyproject.toml untouched. uPlot is vendored, so the browser needs no internet. Live updates poll incrementally, so it works over a plain SSH tunnel (ssh -L 7788:localhost:7788 <host>).
• Document the dashboard in the monitor-run skill.

Verification

The orchestrator-side inference-metrics mirroring is import-checked and exercised in the dashboard via synthetic time-keyed rows appended to a copy of a real metrics.jsonl; it has not yet run against a live training run (GPUs were occupied).

Against real runs on disk (no GPU involved): outputs/info-logs and outputs/bcast2-nccl (RL, current schema, 2 attempts) and an older run with eval rollouts. Exercised all API endpoints with curl and drove the UI with headless Chromium (all four tabs, overview card, smoothing slider, corner drag-resize, all-metrics regex filtering, merged multi-file logs, episode drawer, token heatmap) with zero console errors. Overview values match the trainer's own step logs (grad norm 0.0304, MFU 24.0%, entropy 0.0118 at step 10 of info-logs).

🤖 Generated with Claude Code

[github-advanced-security]

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