feat(console): operator UI mounted on the gateway; retire visualizer.py

[jeffreysijuntan]

Summary

Stacked on #557. The largest PR in the stack — replaces the standalone rllm/eval/visualizer.py (~1k LOC) with an SPA + panel-registry backend mounted onto the model gateway. rllm view boots a view-only gateway (TraceStore + console; no proxy workers) and serves the console at /console.

What changes

Frontend (rllm-console/) — bun + Vite 6 + React 19 + RR7 + Tailwind v4 + TanStack Query 5

• Sessions panel: cross-run trace feed (gateway TraceStore) with filter bar (run / model / harness / has_error / latency) and chat-message rendering. The schema-v2 denormalised columns from feat(model-gateway): trace store schema v2 #551 are what makes this filter-heavy view fast.
• Runs panel: per-eval-run grid → episode viewer → live Step view, reading from ~/.rllm/eval_results/<run>/ and the gateway's shared db.
• Datasets panel: registry browser, entry inspector, pull-from-UI with PTY-streamed progress.
• Settings panel: config viewer + env-var manager (writes to ~/.rllm/.env; presets in known.py).
• Sandboxes / Eval Launcher / Training panels: placeholders for follow-up iterations.

Backend (rllm/console/)

• mount_console(app, eval_results_root, url_prefix=\"/console\") attaches the panel registry to any FastAPI app — typically rllm-model-gateway, but works on any host.
• Panels are Panel(id, title, icon, nav_order, router, placeholder) dataclasses registered via register_panel. Each panel ships its own api.py (FastAPI APIRouter) and a frontend panels/<id>/ module mounted by id.

CLI

• rllm/cli/view.py rewritten to launch the gateway+console pair.
• rllm/eval/trace_loader.py (new) loads run/episode/step data for the Runs panel.
• rllm/eval/visualizer.py (deleted, ~1k LOC).

Packaging

• pyproject.toml: console/static/** added to package-data so the built SPA ships in the wheel; visualizer.py ruff exclusion dropped.
• .gitignore: console/static/ (built SPA) + rllm-console/ build artefacts.

Stack

• #551 → #552 → #553 — gateway
• #556 — eval pipeline flip
• #557 — harnesses + Harbor 0.5
• [PR-4] this — console operator UI

Test plan

• python -m pytest tests/console/ tests/eval/test_trace_loader.py — 77 passed
• Full suite across all stacked PRs: python -m pytest tests/eval/ tests/sandbox/ tests/cli/test_eval_command.py tests/harnesses/ tests/console/ — 577 passed
• rllm view boots cleanly; SPA shell + panel API endpoints reachable

Follow-ups (separate PRs)

• CI workflow to cd rllm-console && bun install && bun run build before packaging the wheel — until that lands, pip install -e users have an empty console/static/ until rllm view auto-builds.
• Eval Launcher panel (placeholder today) — kicks off rllm eval runs from the UI.
• Sandboxes panel (placeholder today) — live sandbox orchestration view.
• Training panel (placeholder today) — surfaces training-run state from the gateway's worker-pool side.

🤖 Generated with Claude Code