Add OpenRouter as a setup connection with searchable model pickers

[JeremySNR]

What does this change?

Adds OpenRouter as an AI connection in the first-run wizard, replacing the "Claude subscription" card, which didn't do anything. OpenRouter is also a choice under Settings → AI connection.

• Key: OpenRouter API key field, encrypted with safeStorage the same way as the OpenAI key. Check key calls OpenRouter's /key endpoint, which makes no model request.
• Clip-finding model (LLM): a searchable dropdown built from OpenRouter's public /models catalogue. Suggested models sit at the top: GPT-5.4 Mini (the default), Gemini 3.8 Flash, Claude Sonnet 5 and GPT-5.5. Suggestions the live catalogue doesn't carry are hidden. Each row shows price, context size and a Vision tag, and a full custom model id can be typed.
• Transcription: local Whisper (pinned first, with the existing install fields) or one of OpenRouter's three Whisper models: openai/whisper-1, openai/whisper-large-v3-turbo and openai/whisper-large-v3. Nothing else is allowed, and a stored or submitted unsupported id is refused.
• Routing: with OpenRouter selected, chat and transcription go to https://openrouter.ai/api/v1 with the OpenRouter key. The OpenAI base-URL settings and OPENAI_BASE_URL don't apply on this route.
• Key guard: every request checks its key against the credential of the currently selected connection. If the provider is switched in Settings while a job runs, the job stops instead of sending one vendor's key to the other's endpoint.
• Transcription request: uses OpenRouter's documented JSON body: base64 input_audio, verbose_json, and timestamp_granularities: ["word","segment"]. Chunks are 5 minutes, down from 20, to stay inside OpenRouter's 60-second upstream timeout. A reply with speech but no word timestamps fails once with a clear message and isn't retried.
• Offline: if the catalogue can't be fetched, the suggested models stand in. A sample catalogue fixture (tests/fixtures/openrouter-catalog.json, prices illustrative) drives the offline smoke walk, which now also captures the OpenRouter wizard screens.

Why?

Users asked to pick any LLM and a transcription model through one OpenRouter key, while keeping local transcription as an option.

Transcription is limited to the Whisper models because captions, cut tightening and clip timing need per-word timestamps. Through OpenRouter, only the Whisper models served by OpenAI-compatible upstreams (OpenAI, Groq, Together) reliably return them. Others either reject verbose_json with HTTP 400 (openai/gpt-4o-transcribe, microsoft/mai-transcribe-1.5, Google Chirp), return no word timings, or accept only short or WAV-only clips (AssemblyAI sync, Meta Muse Voice). The reasoning is recorded next to OPENROUTER_TRANSCRIPTION_MODELS in src/shared/openrouter.ts.

Release notes

• OpenRouter is now a connection in first-run setup and Settings. Add an OpenRouter key (encrypted with your system keychain), then pick the clip-finding model from OpenRouter's full model list. The list is searchable, suggests models at the top, and shows prices, context size and image support. Transcription can use local Whisper or one of OpenRouter's Whisper models, which are the only OpenRouter transcription models that return the per-word timestamps captions need.
• OpenRouter transcription sends five-minute audio chunks so each request finishes inside OpenRouter's 60-second limit.
• The setup wizard no longer shows the unavailable "Claude subscription" card. Claude models can be used through OpenRouter instead.

How did you test it?

• npm test (723 passing, including the new tests/openrouter.test.ts: catalogue parsing, allowlist filtering, JSON request body, WAV labelling, chunk length, missing-word errors without retries, and the key guard on provider switches)
• npm run typecheck
• npm run lint
• Checked the change in the running app: wizard and Settings screenshots under Xvfb, using the sample catalogue
• scripts/test-pipeline.ts and scripts/test-resilience.ts (offline), plus a manual check that an 11-minute file splits into three 5-minute chunks with correct keep-windows

Anything to watch out for?

• Not live-tested. The dev sandbox couldn't reach openrouter.ai, so no real OpenRouter request was made. The request shapes follow OpenRouter's documented schema, but a real-key check of one clip-finding run and one Whisper transcription is still needed.
• Suggested model ids came from recent web sources. Any that OpenRouter doesn't list are hidden automatically.
• No context prompt: OpenRouter's transcription schema has no top-level prompt, so chunks are transcribed without the previous chunk's text. The 8-second overlap and seam repair still cover the joins.
• Preview and export are untouched.

🤖 Generated with Claude Code

https://claude.ai/code/session_012fRPuZ4E7P2EcTR4c9BzhS

[cursor]

Stale Bugbot comment from a previous run.

[JeremySNR]

Independent review summary. A separate read-only reviewer went through this PR and found nothing blocking in normal use. a4f4438 fixes the following:

• Mixed credentials after a mid-job switch (security): jobs read their key once at the start, but endpoints are read on every request. Switching provider in Settings while a job ran could send an OpenAI key to OpenRouter, or the reverse. Requests now check the key against the current route and stop with "The AI connection changed… Start it again."
• Retried word-timing failure: a missing-word-timestamps reply was retried 4× (same as the Bugbot thread). It's now checked after the retry loop.
• Wrong credential named in errors: the caption and clip-finding errors and the top-bar prompt now name the credential for the selected provider.
• Stale highlight in the model picker: it now tracks the highlighted row by id, so a catalogue that arrives while the list is open can't move Enter onto a different model.
• Catalogue load failure: if the IPC call rejects, the dropdowns fall back to the suggested models.

Deliberately left as they are (these match the existing API route's behaviour, or are follow-ups):

• Shared localTranscription flag: one flag serves both the API and OpenRouter routes, so changing it on one changes the other. The UI shows the current state.
• No local-Whisper check at finish: the wizard doesn't require a local Whisper check before finishing OpenRouter + local. The API route behaves the same.
• No way to clear a stored key: the OpenAI key field works the same way.
• Report label: source-discovery and editorial reports label the provider api for OpenRouter. Cache keys already include the OpenRouter base URL.
• Missing tests: there are no unit tests for the settings.ts provider branches, because they would need an Electron safeStorage mock.

---

Generated by Claude Code

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

Bugbot Autofix prepared a fix for the issue found in the latest run.

• ✅ Resolved by another fix: Key guard skips empty credentials
  • Already fixed on this branch by fbb4cd3, which reads the route credential live and keeps the guard armed when the new route has no key.

_{You can send follow-ups to the cloud agent here.}

^{Reviewed by Cursor Bugbot for commit a4f4438. Configure here.}