OpenAI Chat Completions on a SpicyAPI text model

Accepts an OpenAI Chat Completions request. messages, tools, tool_choice, parallel_tool_calls, response_format, max_tokens (or max_completion_tokens), temperature, top_p, top_k, seed and stop map onto the model's public input fields of the same name when the model declares them (each model page lists its own set); stream, stream_options, user, metadata and store are handled by this layer. Pure sampling parameters the model does not declare (temperature, top_p, top_k, min_p, top_a, frequency_penalty, presence_penalty, repetition_penalty), n: 1, an empty stop, stop on a model without stop sequences, reasoning_effort on a model without reasoning levels, and an OpenRouter-style reasoning object that cannot be translated are ignored, and the response lists them in the X-Spicy-Ignored-Params header. Values the model declares but cannot take as sent are changed to the nearest accepted value and listed as name=value in the X-Spicy-Adjusted-Params header: a sampling value outside the model's range is clamped to the bound, a stop list longer than the model allows keeps its first entries, a reasoning_effort level the model does not offer becomes the nearest level it does (the lower one on a tie), and reasoning.effort (or reasoning.enabled: false, as none) becomes reasoning_effort. Any other parameter is forwarded to the model's input schema and rejected with 400 when the model does not declare it.

stream: true returns a Server-Sent Events stream of chat.completion.chunk objects terminated by data: [DONE]. Every chunk has a non-empty choices array and each choice ends with exactly one chunk carrying finish_reason. Usage follows OpenAI: a stream carries no usage unless stream_options.include_usage is true, in which case one extra chunk with choices: [] and the final usage comes right before [DONE] and every other chunk has usage: null. stream: false returns one chat.completion object that always has usage. The same hold, settlement and refund rules as /api/v1/jobs/stream apply; the X-Spicy-Task-Id response header names the task that holds the charge.

A non-streaming call that is still running after about 30 seconds is kept open: the 200 status and the X-Spicy-Keepalive: 1 header go out first, followed by a newline every 15 seconds and then the JSON body. A failure after that point is the usual {"error": {...}} object in the body of that 200 response, so check for an error key when the header is present. The same applies to /v1/responses, /v1/messages (Anthropic error object) and :generateContent (Gemini error object, whose error.code keeps the HTTP status the error would have had).

Errors a client can act on: an unknown model returns 404 with code: model_not_found, the ID shape (publisher/model/chat) and the closest callable chat model; a model that is not a chat model returns 400 with param: model; a field or message part the model does not accept (for example an image_url part on a text-only model) returns 400 with code: unsupported_parameter and param pointing at it; a content-policy refusal returns 400 with code: content_rejected.