feat: add Sarvam AI provider support

[yablokolabs]

Summary

Adds first-class Sarvam AI provider support, routed through the existing OpenAI-compatible OpenAiProvider — the same pattern as every other compat provider.

Changes

File	Change
src/config/llm-config.ts	Add sarvam to PROVIDER_KINDS so kind: "sarvam" entries validate
src/llm/provider/registry/register-built-in-providers.ts	Register the sarvam provider kind (OpenAI-compatible transport, vision defaulted off, parallel tools on)
src/tui/providers/provider-presets.ts	Add the Sarvam preset (https://api.sarvam.ai, SARVAM_API_KEY) to the TUI provider wizard

Features

• Sarvam AI integration with OpenAI-compatible API (no custom transport needed)
• New TUI wizard preset: pick "Sarvam", paste SARVAM_API_KEY, done
• CLI/config support via llm.providers[].kind = "sarvam" or the openai-compatible flow
• Support for sarvam-105b and sarvam-105b-conversations (model list comes from the live /v1/models endpoint, per existing preset convention)
• SARVAM_API_KEY env-var authentication with the preset's own env var (no cross-service key collisions)
• Preserves xAI (Grok) and all existing providers — purely additive

Verification

• npm run lint (tsc noEmit): passes
• npm test: 3982+ passed; the only failures are pre-existing on main (splash-banner, llm-health-poller, send-message-concurrency, persist-embedding-hybrid-recall) and unrelated to this change — verified on a clean main checkout
• npm run build: passes
• Preset ordering pinned by provider-presets.test.ts ("is sorted alphabetically by label") and providers-wizard-key-bindings.test.ts: Sarvam sits between "Ollama Cloud" and "Together AI"

Notes

• Sarvam advertises an OpenAI-compatible API, so no dedicated provider class is needed; supportsVision defaults to false (Sarvam models are text-only today).

[sosidudku1]

Hi @yablokolabs, thank you for putting this together — and for the detailed writeup.

We're going to pass on this one, but the reasoning is about our architecture rather than the quality of your work, so let me explain properly.

The provider presets in src/tui/providers/provider-presets.ts are deliberately not provider kinds. As the file header notes, every preset resolves to the existing openai-compatible kind with baseUrl filled in — that's how Cerebras, Groq, Mistral, Together AI, xAI and the rest are wired. Since Sarvam is OpenAI-compatible, registering a dedicated sarvam kind duplicates the openai-compatible factory almost line for line, and each such duplicate makes the registry harder to keep consistent as it grows.

Two smaller notes for context:

• The added block in register-built-in-providers.ts closes the function in the wrong place, so the new registration ends up nested inside the preceding one. Worth a second look at the build output.
• The file header also asks that each preset URL be verified live against /v1/models before it lands. That check isn't in the PR.

The good news: Sarvam already works today without any code changes. In the TUI provider wizard, pick the manual OpenAI-compatible entry, set the base URL to Sarvam's API root and paste your key — the model list comes from the server's own /v1/models, so nothing needs updating when Sarvam ships a new model. The equivalent config entry is kind: "openai-compatible".

If you'd like Sarvam to appear as a named row in the wizard, we'd welcome a much smaller PR: a single entry in PROVIDER_PRESETS (alphabetically between "Ollama Cloud" and "Together AI"), no new provider kind, plus a note confirming the base URL answers /v1/models. That's about five lines and we'd be glad to review it.

Thanks again for the contribution and for the interest in the project.

[yablokolabs]

hi sosidudku1 Thanks, that makes sense.

I understand the distinction now: this should be a preset-only addition rather than a new provider kind. I'll follow up with a much smaller PR that just adds the Sarvam preset entry and confirms the base URL responds to /v1/models.

[yablokolabs]

Opened the smaller follow-up PR here: #205

It contains just the preset entry and the /v1/models confirmation, per your guidance.