Integrate Tier 1 AI providers (Groq, DeepSeek, Mistral, xAI, Cerebras)

[ralyodio]

Background

packages/ai/ has 59 provider adapters. Four are fully integrated (Claude, OpenAI, Gemini, Qwen); 55 are BYOK stubs that only collect an API key. This issue tracks integration of the first tier — the 5 highest-impact stubs.

All five speak the OpenAI Chat Completions protocol, so each is essentially: clone packages/ai/openai/src/index.ts, swap DEFAULT_BASE + defaultModel + the models list + the error label, and confirm an end-to-end call works.

Reference adapter

packages/ai/openai/src/index.ts — copy this shape. Keep tokenSetup() as-is (already wired in each stub).

Providers to integrate

Adapter dir	Base URL	Default model	Suggested models list	Env var
packages/ai/groq/	https://api.groq.com/openai	llama-3.3-70b-versatile	llama-3.3-70b-versatile, llama-3.1-8b-instant, mixtral-8x7b-32768, deepseek-r1-distill-llama-70b	GROQ_API_KEY
packages/ai/deepseek/	https://api.deepseek.com	deepseek-chat	deepseek-chat, deepseek-reasoner	DEEPSEEK_API_KEY
packages/ai/mistral/	https://api.mistral.ai	mistral-large-latest	mistral-large-latest, mistral-small-latest, codestral-latest, ministral-8b-latest	MISTRAL_API_KEY
packages/ai/xai/	https://api.x.ai	grok-3	grok-3, grok-3-mini, grok-2-latest	XAI_API_KEY
packages/ai/cerebras/	https://api.cerebras.ai	llama-3.3-70b	llama-3.3-70b, llama3.1-8b	CEREBRAS_API_KEY

Endpoint path is /v1/chat/completions for all five (same as OpenAI).

Acceptance criteria

• Each adapter's generate() makes a real HTTP call (no more [stub] log + placeholder return)
• ctx.dryRun short-circuits before the network call (match openai adapter)
• Errors include status code + first 200 chars of response body
• inputTokens / outputTokens populated from data.usage when present
• pnpm vitest run packages/ai/{groq,deepseek,mistral,xai,cerebras} passes
• Manual smoke test: sh1pt promote ai setup --platform groq accepts a key, then a minimal generate() call returns text
• One PR per adapter is fine (preferred), or one bundled PR

Out of scope

• Tier 2-5 adapters (Together, Fireworks, Perplexity, etc.) — separate issues
• Streaming responses
• Tool/function calling

References

• Reference impl: packages/ai/openai/src/index.ts
• AI provider interface: packages/core/src/ai.ts
• Setup helper: packages/core/src/setup.ts (tokenSetup pattern)

[lucianchen]

Delivered the five OpenAI-compatible adapter PRs requested here:

• Groq: feat: integrate Groq AI adapter #67
• DeepSeek: feat: integrate DeepSeek AI adapter #68
• Mistral: feat: integrate Mistral AI adapter #69
• xAI: feat: integrate xAI adapter #70
• Cerebras: feat: integrate Cerebras AI adapter #71

Each PR replaces the adapter stub with a real /v1/chat/completions request, keeps dry-run short-circuiting before network calls, maps usage tokens, and includes provider status + response-body excerpts on errors.

Local verification run for each package:

• pnpm exec vitest run packages/ai/<adapter>/src/index.test.ts
• pnpm --filter @profullstack/sh1pt-ai-<adapter> typecheck
• git diff --check

Note: the PR CodeQL analyze (javascript-typescript) job is currently failing with CodeQL analyses from advanced configurations cannot be processed when the default setup is enabled, which looks like a repository-level CodeQL default setup vs advanced workflow conflict. npm audit, gitleaks, and Socket checks are passing.