fix(provider): fall back gracefully when the configured model has no provider

[Aayam Bansal (aayambansal)]

Symptom

On a machine with a saved config model (~/.config/openscience/openscience.json → {"model": "anthropic/claude-sonnet-4-6"}) but no API keys and no managed session, sending a prompt from the web workspace crashes the server loop with a raw unhandled stack in the terminal:

ProviderModelNotFoundError ... data: { providerID: "anthropic", modelID: "claude-sonnet-4-6", suggestions: [] }
    at getModel (src/provider/provider.ts:1294)

and the web UI shows nothing — the session silently never answers.

Root causes

1. src/provider/provider.ts — defaultModel() returned parseModel(cfg.model) without checking the provider/model is actually available; lastModel() in src/session/prompt.ts likewise reused the model of an old user message even if its provider is gone.
2. src/server/routes/session.ts — the /:sessionID/prompt_async route fire-and-forgets SessionPrompt.prompt(...) with no .catch, so any throw became an unhandled rejection: raw stack to the terminal, nothing to the UI.
3. src/session/prompt.ts — in loop(), Provider.getModel(lastUser.model...) threw without publishing a session error event (unlike the command path, which catches ModelNotFoundError and publishes Session.Event.Error before rethrowing).

Fix

• backend/cli/src/provider/provider.ts — defaultModel() only honors cfg.model when its provider and model exist in the available providers; otherwise it logs a warning and falls through to the existing priority-based selection. The zero-providers throw now carries an actionable message (NO_PROVIDER_HINT). Behavior is unchanged when the configured model is available.
• backend/cli/src/session/prompt.ts
  • lastModel() validates that the historical model's provider+model are still available before reusing them, falling back to Provider.defaultModel().
  • loop() catches Provider.ModelNotFoundError from getModel, publishes Session.Event.Error with an actionable message ("Model anthropic/claude-sonnet-4-6 is not available. Add an API key for a provider (openscience auth login) or connect a managed account (openscience connect login), then choose a model."), finalizes an assistant message carrying the error (same shape as the processor error path), and exits the loop cleanly. The fire-and-forget ensureTitle call now has a .catch — it was a second source of unhandled ProviderModelNotFoundError rejections.
  • prompt() publishes Session.Event.Error when user-message creation fails (e.g. truly zero providers), so the UI is never silent.
• backend/cli/src/server/routes/session.ts — the prompt_async fire-and-forget now has a .catch that logs and swallows; session-level error publishing happens inside prompt/loop.

Verification

• bun run typecheck — 6/6 packages pass.
• bun run --cwd backend/cli test — 820 pass, 0 fail.
• End-to-end repro in an isolated env (HOME/XDG_* pointed at a temp dir containing only the config model, ANTHROPIC_API_KEY=/OPENAI_API_KEY= empty, serve --port 412x, POST /session + POST /session/{id}/prompt_async {"parts":[{"type":"text","text":"hi"}]}):

Before (dev.log):

INFO  ... service=session.prompt step=0 sessionID=ses_0d18df634ffe... loop
ERROR ... service=default e=ProviderModelNotFoundError rejection
ERROR ... service=default e=ProviderModelNotFoundError rejection

No session.error event, GET /session/{id}/message returned only the user message — the UI had nothing to show.

After (same env, same requests): zero rejection lines, no raw stack, and the session receives the error:

WARN  ... service=provider providerID=anthropic modelID=claude-sonnet-4-6 configured model is not available, falling back to default selection
ERROR ... service=server sessionID=ses_0d1894987ffe... error=No model providers are available. Add an API key for a provider (`openscience auth login`) or connect a managed account (`openscience connect login`), then choose a model. prompt_async failed

/event stream:

data: {"type":"session.error","properties":{"sessionID":"ses_0d1894987ffe...","error":{"name":"UnknownError","data":{"message":"No model providers are available. Add an API key for a provider (`openscience auth login`) or connect a managed account (`openscience connect login`), then choose a model."}}}}

• Loop path (explicit unavailable model in the request body, which bypasses the lastModel fallback): GET /session/{id}/message now shows a finalized assistant message with error.data.message = "Model anthropic/claude-sonnet-4-6 is not available. Add an API key for a provider (openscience auth login) or connect a managed account (openscience connect login), then choose a model." plus the matching session.error event — and zero unhandled rejections.
• Stale-session path (session created before the fix with a stored anthropic/... user message): WARN ... last used model is no longer available, falling back to default, graceful error, no crash.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
openscience	Ready	Preview, Comment	Jul 4, 2026 6:51pm