Providers registered by pi extensions are missing from Settings → Providers and the model list until a chat loads the extension #1351

Description

A model provider registered by a pi extension (pi.registerProvider, e.g. @pgilad/pi-vertex-anthropic or pi-provider-litellm) is invisible to ThinkRail until some chat in the current process happens to load the extension, and it disappears again after every restart:

• The shared catalog never loads user extensions. It is built with packages: "none" and noExtensions: true (packages/server/src/agent/piRuntime.ts:139-151), and model.list, model.default, restore and provider.status all read it (agentSessionManager.ts:2993-3015).
• Extension providers reach that catalog only as a side effect of a chat starting. Chats reuse the shared runtime; pi adds the queued provider registrations in the AgentSession constructor, after the chat's model has been chosen (pi sdk.js:91-117, agent-session.js:2944). pi's own CLI registers them first via createAgentSessionServices, which ThinkRail skips. So a new chat gets "Select model", an explicit extension model is rejected as Unknown or unavailable model (agentSessionManager.ts:1471), and restore resolves the saved model too early (agentSessionManager.ts:2140-2155) — the remaining part of Restoring a chat fails with "saved model is unavailable" for extension-provided models (e.g. commandcode) #365.
• Settings → Providers never lists them. provider.status shows only provider ids recorded when the runtime was created, plus custom-* (packages/server/src/auth/providerStatus.ts:82-99), so an extension provider is never a row. For a user whose only model source is such an extension, the onboarding wizard keeps saying no provider is connected (apps/web/src/panels/useProviderStatus.ts:6-8, WelcomePanel.tsx:46-56).

This explains the reports in #996 (vertex-anthropic: "no model providers connected") and the LiteLLM-extension comment on #1039 ("no models"), and shares its root cause with #365.

Steps to reproduce

1. Install a provider extension, e.g. pi install npm:@pgilad/pi-vertex-anthropic and sign in with pi /login (or install pi-provider-litellm with its key).
2. Start ThinkRail. Settings → Providers has no row for that provider; the model picker and Settings → Models list none of its models. If it is the only model source, onboarding says no provider is connected.
3. Open any chat — it starts with "Select model". Open the model picker again: the extension's models now appear.
4. Restart ThinkRail: the models are gone again, and a restored chat that used one of them opens as "Select model".

Expected behavior

Providers and models registered by installed pi extensions are available from startup — in the model picker, Settings → Models, new chats and restored chats — and the provider is shown in Settings → Providers (and counts as a model source for onboarding), as in pi's own CLI.

Actual behavior

They exist only after a chat has loaded the extension in the current process, are never listed in Settings → Providers, and vanish on restart or a JetBrains AI runtime rebuild.

Environment

• OS: any
• Bun version (bun --version): 1.4.4
• Node version (node --version): n/a
• ThinkRail version / commit: main 52ae54782, pi 1.5.2 (reproduced with a test using the production runtime factory and boot order and a throwaway global extension calling pi.registerProvider)