[HouMinXi]

Summary

• A model name that is in neither MODEL_MAP nor THINKING_MAP was forwarded upstream as a raw preference. This rejects it locally with a 404 and model_not_found, so a typo never leaves the process.
• Split out of feat(mcp): add a deep research tool that is off by default #17317 at the reviewer's request, because it changes existing chat behaviour and deserves its own review.

Test plan

• An unmapped model returns 404 and the fetch counter stays at 0.
• A mapped model (pplx-sonar) still reaches the counter.
• The web cookie sweep passes, since the 404 result carries url, headers and transformedBody.

[diegosouzapw]

Thanks @HouMinXi, and thanks for splitting this out of #17317 the way we asked. This is a clean, well-scoped change. A typo'd model name no longer leaves the process as a raw model_preference, and the 404 model_not_found return keeps the same {response, url, headers, transformedBody} shape as the neighbouring early returns, so the web-cookie sweep stays happy. I checked a few things on your head (8e77008): all 11 models in the perplexity-web registry are in MODEL_MAP, so valid aliases are unaffected. Your new test is red when I revert only the source hunk (the 404 assertion fails) and green with it. The "mapped model still reaches the counter" guard is a nice touch, because it stops a counter stuck at zero from passing the test. perplexity-web, model-mappings, cookie-rotation and upstream-status all pass on your branch, and ESLint is clean.

Three small things we will fix in place on your branch so your commit and credit stay as they are:

1. Prettier formatting on tests/unit/perplexity-web-unmapped-model.test.ts (the dynamic import destructure on lines 6-8).
2. A changelog fragment, changelog.d/fixes/17822-perplexity-web-reject-unmapped-model.md.
3. Building the 404 body with buildErrorBody() from open-sse/utils/error.ts, as claude-web does, instead of hand-written JSON, so the message goes through our error sanitization.

One thing for you, when you get to #17317: please make sure it doesn't carry a copy of this hunk. If it registers a new Perplexity model id, for example a deep-research one, that id needs a MODEL_MAP entry too, because it would now get a 404 here. Your test uses pplx-deep-research as the unmapped example, so this one is easy to trip over.

Last note: removing the raw-preference passthrough is a behaviour change, so the maintainer has the final call on it. I found nothing in the tests or docs that treated it as a supported feature.