Fetch gateway models directly from Electron instead of spawning preview sessions

[charlesvien]

Problem

To populate the model selector dropdown, we currently spawn a lightweight "preview session" (__preview__ task ID) via the agent package, wait for it to initialize, and extract the available models from its configOptions response. This means:

• Unnecessary latency — full agent session startup just to get a model list
• Unnecessary complexity — preview session lifecycle management, cleanup logic, the __preview__ task ID convention
• Fragility — if the agent fails to start, no models appear in the selector

Current flow

TaskInput mount
  → usePreviewSession() hook
  → sessionService.startPreviewSession({ adapter })
  → trpcClient.agent.start.mutate({ taskId: "__preview__", ... })
  → Agent initializes → Claude adapter → fetchGatewayModels()
  → configOptions returned with model list
  → Model selector populated

Proposed change

Call fetchGatewayModels() directly from a main process service and expose it via a tRPC query. The main process already has the gateway URL and API credentials — no agent session needed.

A tRPC endpoint (getGatewayModels) already exists in apps/code/src/main/trpc/routers/agent.ts that does exactly this. We should:

1. Have the renderer call this tRPC query directly to populate the model selector
2. Remove the preview session machinery (usePreviewSession, startPreviewSession, cancelPreviewSession, PREVIEW_TASK_ID)
3. Keep the 10-min TTL cache in packages/agent/src/gateway-models.ts (still used by real sessions)

Key files

• packages/agent/src/gateway-models.ts — fetchGatewayModels() with 10-min cache
• apps/code/src/main/trpc/routers/agent.ts — existing getGatewayModels tRPC query (~line 197)
• apps/code/src/main/services/agent/service.ts — getGatewayModels() (~line 1675)
• apps/code/src/renderer/features/task-detail/hooks/usePreviewSession.ts — preview session hook (to remove)
• apps/code/src/renderer/features/sessions/service/service.ts — startPreviewSession() (to remove)
• apps/code/src/renderer/features/sessions/components/ModelSelector.tsx — needs to use tRPC query instead

Context

From Josh: "for the model select issue thingy we've had open for a while, iirc we spawn some kind of preview session and then use the agent package to load the available models — why do we do that vs calling the gateway directly from electron?"

Answer: no good reason. The direct approach is simpler, faster and already partially wired up.