Skip to content
ignisignis

Chat completions

Accepts messages (role + content), model, stream, max_tokens, temperature (0..2), top_p (0..1), presence_penalty and frequency_penalty (-2..2), and a signed 64-bit seed.

  • top_k is an Ignis extension, not an OpenAI Chat Completions parameter: 0..20, where 0 selects the 20-candidate sampler cap during stochastic sampling.

  • class is an Ignis extension: the request’s lane tag, interactive (default) or agent. The same class can be stated as an @agent suffix on the model name.

  • thinking_budget is an Ignis extension: the reasoning tokens this request may spend before its close is forced. Absent takes --model-thinking-budget, 0 means no budget, and reasoning_effort: "max" ignores it and runs with no budget. The budget always leaves 2,048 tokens of max_tokens for the answer. A forced close is reported as thinking_budget_forced_at on the choice (on the finish chunk when streaming; top-level on /v1/responses): the reasoning tokens emitted when it began. The field is absent when the close was not forced.

  • A turn that ends inside its reasoning answers. With thinking on, an end of turn the model draws while its reasoning block is still open becomes the block’s </think>, and the model goes on to answer — usually with the call or the text it was writing in its reasoning, now on the content channel. Such a turn used to come back as reasoning only, with no content and no tool call. The same holds on /v1/responses. The request log’s ignis.request.done line marks it with reasoning_redirected_at, the output index of that </think>. With thinking off, or ignore_eos, the end of turn ends the turn as it always has.

  • A loop in the reasoning is taken back (--model-loop-guard, rewind by default, in the table above). usage.completion_tokens_details.rewound_tokens (usage.output_tokens_details.rewound_tokens on /v1/responses) counts the tokens taken back, which completion_tokens (output_tokens) still counts; it is absent when nothing was. On /v1/responses, include: ["reasoning.encrypted_content"] gives each reasoning item an encrypted_content with the reasoning as the model saw it, and a reasoning input item that carries one is rendered from it in place of its text.

  • Values outside the ranges above are rejected with invalid_sampling_parameter; they are never silently clamped.

  • Absent sampling fields take the Qwen3.8 model card’s values for the request’s mode, each field on its own (spec server/12):

    mode temperature top_p top_k presence_penalty frequency_penalty
    thinking 1.0 0.95 20 0 0
    enable_thinking: false 0.7 0.80 20 1.5 0

    An absent seed is a fresh one per request, so two identical requests are two independent samples; a sent seed is kept. Sampling runs on the GPU, and its RNG is counter-based, keyed by the seed, the sequence’s own position and the draw’s purpose (prefill or decode): what a request generates depends on its own seed alone, never on which requests shared its decode round, and no RNG state travels with a sequence that is moved off the device and back. Greedy is temperature: 0: the fields such a request does not send are neutral, and because the leaf’s greedy branch does not read stochastic filters or penalties, a non-neutral top_p, top_k, presence_penalty or frequency_penalty sent with it is rejected rather than ignored.

  • A request that sends no max_tokens (nor max_completion_tokens, nor max_output_tokens on /v1/responses) generates at most --model-default-max-tokens tokens, its reasoning included. The flag is 38,912 unless you set it (IGNIS_MODEL_DEFAULT_MAX_TOKENS, make DEFAULT_MAX_TOKENS). Such a request ends with finish_reason: "length" when it gets there (incomplete with max_output_tokens on /v1/responses). Before, it could generate to the end of the context. Send max_tokens for more, or start the server with --model-default-max-tokens 0 for the old behaviour.

  • max_completion_tokens is max_tokens under OpenAI’s current name — the one the current SDKs send. Both with different values is a 400.

  • stop (a string, or 1 to 4 non-empty strings) ends the answer before the first sequence to appear in its content, with finish_reason: "stop". The sequence is never emitted, streaming included; it is never matched in the reasoning, nor inside a tool call.

  • tool_choice takes "auto" (the default), "none", "required", and a named function ({"type": "function", "function": {"name": "..."}}, or {"type": "function", "name": "..."} on the responses API). The last two force the call’s opening — <tool_call> and the function tag, with the name when one is named — and the model writes the arguments. With thinking on (the default) the opening is forced right after the reasoning block closes, so the model still thinks first; with enable_thinking: false it is the first thing generated. After a </think> the model wrote itself, the first forced token is the opening’s second, its joiner (the \n after <tool_call>): the token before it was drawn unseen, and may itself have opened the call. The call comes back as any other call does, with finish_reason: "tool_calls". Naming a function that is not in tools, "required" with no tools, and a max_tokens shorter than the opening are a 400; so is a tokenizer that cannot force the opening (tool_choice_unsupported — with thinking on it needs </think> and <tool_call> to be single tokens, and enable_thinking: false is the way out). With thinking on the reasoning spends the same max_tokens: a block that never closes ends length with no call, which the thinking budget’s answer reserve prevents.

  • Non-streaming returns choices[].message.content plus usage; streaming emits chat.completion.chunk SSE frames (token deltas, a final finish_reason chunk, then [DONE]). usage.prompt_tokens_details.cached_tokens is the prompt the request resumed from retained state instead of prefilling (0 when none).

  • Every other OpenAI field is accepted as inert when its value asks for what the server already does (n: 1, store: false, logprobs: false, response_format: {"type": "text"}, metadata, user, …) and refused with a 400 unsupported_value naming it in param otherwise. The full table is in the API reference at /v1/docs/; a field outside it is ignored.