feat: add openai service emulator

[mvanhorn]

Summary

Adds an OpenAI API emulator with chat completions (non-streaming + SSE streaming), deterministic embeddings, model listing, and an interactive playground UI. Responses are seeded via config patterns - same input always produces the same output.

Why this matters

The Vercel AI SDK (22,900 stars, 8.8M weekly npm downloads) is Vercel's flagship AI product. Every tutorial starts with OPENAI_API_KEY. Block Engineering wrote: "We don't run live LLM tests in CI because it's too expensive, too slow, and too flaky."

The mocking space is fragmented (8+ tools, none dominant). OpenAI's SDK supports OPENAI_BASE_URL for drop-in emulator redirect, and openai-python #398 (19 thumbs-up) requests mock support directly.

Changes

New package: @internal/openai (~400 lines across 13 files)

Endpoints:

• POST /v1/chat/completions - non-streaming JSON response AND SSE streaming with proper wire format (data: {json}\n\n chunks, data: [DONE]\n\n terminator)
• POST /v1/embeddings - deterministic vectors via sha256 hash (same input = same 1536-dim output)
• GET /v1/models / GET /v1/models/:id - list and retrieve seeded models

Streaming wire format matches the real OpenAI API exactly:

data: {"choices":[{"delta":{"role":"assistant"}}]}\n\n
data: {"choices":[{"delta":{"content":"Hello!"}}]}\n\n
data: {"choices":[{"delta":{"content":" World"}}]}\n\n
data: {"choices":[{"delta":{},"finish_reason":"stop"}]}\n\n
data: [DONE]\n\n

Playground UI at /playground:

Shows seeded completion patterns with a form to test prompts interactively. Uses only core CSS classes (zero inline styles).

Seed config maps regex patterns to canned responses:

openai:
  completions:
    - pattern: "hello|hi|hey"
      content: "Hello! I'm the emulated assistant."
    - pattern: ".*"
      content: "This is a mock response from the emulated OpenAI API."

Integration: All 7 points wired in start.ts + list.ts SERVICE_DESCRIPTIONS updated.

Testing

9 tests covering non-streaming completion, SSE streaming wire format verification, deterministic embeddings (same input = same output), tool calls with finish_reason: "tool_calls", model 404 error format, and playground rendering.

Dogfooding

Ran the emulator, tested non-streaming + streaming completions via curl, verified SSE chunks arrive word-by-word with proper data: prefix and [DONE] terminator. Verified 5 default models listed. Playground screenshot above is from the running emulator.

This contribution was developed with AI assistance (Claude Code).

[vercel]

@mvanhorn is attempting to deploy a commit to the Vercel Labs Team on Vercel.

A member of the Team first needs to authorize it.