OpenAI Chat Completions

codec/openaiapi targets POST /v1/chat/completions. Its Codec satisfies both the client codec.Codec and ingress codec.ServerCodec contracts.

Request

BuildChatRequest maps the model name, effective sampling, system prompt, messages, function tools, JSON-schema output, and tool choice, required or named. Stream mode adds stream: true and stream_options.include_usage: true; invoke mode omits both. Text-only user content is a string; mixed text and images become [{type:"text"}, {type:"image_url"}]. Inline image bytes become a data URI.

body, err := openaiapi.EncodeRequest(req, true)
if err != nil {
	return err
}
var wire struct {
	Model  string `json:"model"`
	Stream bool   `json:"stream"`
}
if err := json.Unmarshal(body, &wire); err != nil {
	return err
}
fmt.Println(wire.Model, wire.Stream)

Response and cache

The decoder reads choices[0], preserves reasoning, text, and tool calls, and maps finish reasons to the neutral stream values. Prompt usage is gross prompt_tokens; the decoder subtracts prompt_tokens_details.cached_tokens and cache_write_tokens into the neutral InputTokens field. The request encoder emits no cache breakpoint or cache-control field. The cache counts are accepted only because the response DTO and normalization code implement them.

Source and proof

• openaiapi/codec.go
• openaiapi/encode.go
• openaiapi/decode.go
• openaiapi/stream_test.go

Run go test ./codec/openaiapi.