fix(starter): route /api/models to report-worker instead of orchestrator WS

[borisnieuwenhuis]

The /api/models rewrite in next.config.ts pointed to MODEL_ORCHESTRATOR_URL which is a WebSocket endpoint, causing Upgrade Required errors on REST requests from the frontend.

Routes to the report-worker's /reports/models REST API instead.

Analysis

Here's what I found after reading both the PR and the full codebase:

Architecture

The model-orchestrator runs two separate servers:

┌────────────────────┬───────────────────────┬──────────┬────────────────────────────────────────┐
│ Server │ Port │ Protocol │ Purpose │
├────────────────────┼───────────────────────┼──────────┼────────────────────────────────────────┤
│ WebSocket server │ configurable (e.g. │ WS │ Streams model state changes in │
│ │ 8000) │ │ real-time │
├────────────────────┼───────────────────────┼──────────┼────────────────────────────────────────┤
│ Local Deploy REST │ 8001 (hardcoded) │ HTTP │ CRUD for models — only started in │
│ API │ │ │ local/dev mode │
└────────────────────┴───────────────────────┴──────────┴────────────────────────────────────────┘

The next.config.ts rewrite /api/models → MODEL_ORCHESTRATOR_URL/models currently hits the WS
server, which explicitly rejects non-upgrade requests with 426 Upgrade Required.

What the WS server provides

Connect to ws://:/<crunch_id> and you get:

• init event on connect — all running models for that crunch
• update event on state change — the changed models
• Data shape per model:

    { deployment_id, model_id, infos: { model_name, cruncher_id, cruncher_name, ... }, state, ip,  port }                                                                                             
  

What the models table needs

The ModelsTable currently renders: ID, model_name, cruncher_name, status, desired_state (with a
toggle switch), plus log URIs for the logs dialog.

The WS messages are missing several fields the table depends on:

• desired_state (needed for the running/stopped toggle)
• statusMessage (shown in tooltip)
• builder_log_uri / runner_log_uri (needed for log viewer)
• Granular status (the WS only sends the runner state, not build statuses like BUILDER_BUILDING,
  BUILDER_FAILED, etc.)

The crunch_id problem

The WS requires /<crunch_id> in the connection path. The starter app currently has no concept of
which crunch it belongs to — it just calls GET /models and gets everything back.

My recommendation

A pure WS replacement won't work today without backend changes, because the WS messages don't
carry enough data for the table. There are two realistic paths:

Option A — Hybrid (best UX, more work)

1. Enrich the WS init/update messages in the orchestrator to include desired_state, status,
   statusMessage, and log URIs (same shape as the REST GET /models response)
2. Create a useModelsWebSocket hook in the frontend that connects to
   ws://<MODEL_ORCHESTRATOR_URL>/<crunch_id>, populates state from init, and merges update events
3. Keep REST mutations (add/update/delete) routed to the local deploy REST API on port 8001 via
   the Next.js rewrite
4. Add a NEXT_PUBLIC_CRUNCH_ID env var (or fetch it from a new endpoint) so the WS knows which
   crunch to subscribe to

This gives you instant real-time updates instead of the current 30-second polling, and the table
stays fully functional.

Option B — Accept PR #37 (quick fix, no real-time)

Route /api/models to the report-worker's REST endpoint. The table keeps polling every 30 seconds
like it does now. Simple, works, but no streaming.

My suggestion

Option A is the right direction, but it requires a small backend change first: enrich the WS
_create_model_state_message in websocket_server.py to include the missing fields (desired_state,
status, statusMessage, log URIs). That's ~15 lines of Python. After that, the frontend hook is
straightforward.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
coordinator-webapp	Ignored		Feb 19, 2026 2:32pm