feat(models): ModelClient registry, mock, and OpenAI client (issue #17)

[leo-aa88]

Summary

Adds the ModelClient boundary from design doc §12.2 F with a registry backed by Project.spec.providers.models (§7.1), a mock for deterministic tests, and a minimal OpenAI-compatible HTTP client.

API

Types (types.go)

• ModelClient — Generate(ctx, GenerateRequest) (GenerateResponse, error)
• GenerateRequest — Model, Messages []ChatMessage
• GenerateResponse — Content (assistant text, often JSON for structured agents), Meta (DurationMs, CostUSD placeholders per issue scope)

Registry (registry.go)

• NewRegistry(*spec.ProjectGraph) reads graph.Spec.Providers.Models
• ClientFor("namespace/model_id") returns a client and the model id segment (e.g. openai/gpt-4.1 → id gpt-4.1)
• Supported ModelProviderConfig.Type values: openai → *OpenAIClient, mock → fixed JSON mock

OpenAI (openai.go)

• NewOpenAIClientFromConfig(cfg) resolves apiKeyFrom via ResolveAPIKeyFrom
• Default base URL https://api.openai.com/v1; POST .../chat/completions with Bearer auth
• BaseURL / HTTPClient overridable (used in tests with httptest)

Key resolution (key.go)

• ResolveAPIKeyFrom("env:VAR") — returns error if variable unset (MVP)

Mock (mock.go)

• MockClient — returns configured Content and optional Meta without network I/O

Tests

• Mock / agent step: JSON output unmarshals into a structured schema; meta honored
• Registry: unknown provider namespace (anthropic/... with only openai configured); invalid modelRef format
• Registry + OpenAI: resolves *OpenAIClient with env:OPENAI_API_KEY
• OpenAI client: httptest server asserts /v1/chat/completions and Authorization
• ResolveAPIKeyFrom: success and missing env

Verification

• make fmt (go fmt ./...)
• make test (go test ./... -race)
• make vet

Closes #17

Made with Cursor