Configuration

UniConfig is the second argument of both streaming methods. Pass an empty object to take the defaults.

{
  "max_tokens": 1024,
  "temperature": 1.0,
  "system_prompt": "You are a helpful assistant.",
  "tools": [
    {
      "name": "get_weather",
      "description": "Gets the current weather for a given location.",
      "parameters": {
        "type": "object",
        "properties": {"location": {"type": "string", "description": "The city name"}},
        "required": ["location"]
      }
    }
  ],
  "tool_choice": "auto",
  "thinking_level": "high",
  "thinking_summary": true,
  "prompt_caching": "enable",
  "fast_mode": false,
  "image_config": {"aspect_ratio": "4:3", "image_size": "1K"},
  "tts_config": [{"voice": "Kore"}],
  "embedding_config": {"dimensions": 768},
  "trace_id": "agent1/conversation_001"
}

Fields

Field	Type	Meaning
max_tokens	integer	The limit on output tokens.
temperature	number	The sampling temperature.
system_prompt	string	The system instruction.
tools	list	The tools the model may call: name, description, and parameters as a JSON Schema.
tool_choice	"auto", "required", "none", or a list of tool names	Which tools the model may or must call.
thinking_level	"none", "low", "medium", "high", "xhigh", "max"	How much the model thinks.
thinking_summary	boolean	Ask for a summary of the thinking. Whether one comes back depends on the model.
prompt_caching	"enable", "disable", "enhance"	How the prompt is cached.
fast_mode	boolean	Ask for the provider's faster tier.
image_config	object	For Gemini image output: aspect_ratio (1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, 21:9) and image_size (1K, 2K).
tts_config	list	For Gemini speech: one entry per speaker, with voice and an optional speaker. With two speakers, the text is a script with one Name: line per turn.
embedding_config	object	For embeddings: dimensions, the size of the vector.
trace_id	string	Saves the turn under this id. See Tracer and playground.

Values a model rejects

A value a client cannot send raises UnsupportedParameterError before the request goes out, naming the client and the parameter. A key a client has no use for is ignored. The tables below give both, per client.

thinking_level never raises. Where a model cannot turn thinking off, "none" asks for the least of it: the lowest effort on Claude Opus 5.5, Haiku 5.5, Fable 5 and Fable 5.1, GPT-6 Astra and GPT-6.1 Sol, MiniMax-M3.1-Flash and GLM-5.3; no up-front thinking on Claude Sonnet 5.5; minimal or low on Gemini, whichever the model takes. GPT-6 Luna, Kimi K3 and Kimi K3.1 Pro take "none" as it is.

What each client sends

Clients are named by their client_type (Models and endpoints); openai-official serves the text-embedding-* models through openai-embedding. On both Google clients, a speech model is one whose id contains tts and an embedding model one whose id contains embedding. The mmsp client sends the configuration unchanged, and the server applies it through the model's own client, so that client's rows hold.

max_tokens

Client	Sent as
anthropic-official, ant-messages	max_tokens; 64000 when not set, because the Messages API requires it
zai-official	max_tokens
openai-official, openai-responses, deepseek-official, minimax-official, google-official, google-genai	max_output_tokens, speech models included
openai-chat, openai-chat-vllm-adapter, moonshot-official	max_completion_tokens
openai-embedding, Gemini embedding models	ignored

temperature

Client	Sent as
ant-messages, openai-responses, openai-chat, openai-chat-vllm-adapter, zai-official	temperature, as given
minimax-official	temperature from 0 to 1; outside that range it raises
anthropic-official, openai-official, deepseek-official, moonshot-official	nothing for 1.0; any other value raises
google-official, google-genai	any value raises, speech models included
openai-embedding, Gemini embedding models	ignored

system_prompt

Client	Sent as
anthropic-official, ant-messages	system
openai-official, openai-responses, deepseek-official, minimax-official	instructions
google-official, google-genai	system_instruction; ignored by speech models
openai-chat, openai-chat-vllm-adapter, moonshot-official, zai-official	a first message with the role system; nothing for an empty string
openai-embedding, Gemini embedding models	ignored

tools

Client	Sent as
anthropic-official, ant-messages	tools, with parameters as input_schema
openai-official, openai-responses, deepseek-official, minimax-official, google-official	tools, each with type: "function"
openai-chat, openai-chat-vllm-adapter, moonshot-official, zai-official	tools, each as the function of a type: "function" tool
google-genai	one tool of function_declarations
openai-embedding, Gemini speech and embedding models	ignored

tool_choice

A cell is the value sent: the type on the Anthropic clients, the function calling mode on google-genai.

Client	"auto"	"none"	"required"	A list of names
anthropic-official	auto	none	any	any with the name; one name only
ant-messages	auto	none	any	tool with the name; one name only
openai-official, openai-responses	auto	none	required	mode: "required" with the tools
openai-chat, openai-chat-vllm-adapter	auto	none	required	allowed_tools with mode: "auto"
google-official	auto	none	any	allowed_tools with mode: "any"
google-genai	AUTO	NONE	ANY	ANY with allowed_function_names
moonshot-official	auto	none	required on Kimi K3 and K3.1 Pro; raises on Kimi K2	raises
deepseek-official	auto	none	raises	raises
minimax-official	auto	none; raises on MiniMax-M3.1-Flash	raises	raises
zai-official	auto	raises	raises	raises

Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 cannot be forced to call a tool, so "required" and a list raise on them. The Google clients send tool_choice only with tools, and speech models ignore it. openai-embedding and Gemini embedding models ignore it.

thinking_level

Each client sends the level as its protocol's effort: output_config.effort with adaptive thinking on the Anthropic clients, reasoning.effort on the Responses clients, reasoning_effort on Kimi K3 and K3.1 Pro and on GLM (with thinking enabled), thinking_level on Gemini. Kimi K2 gets a thinking type instead; disabled and between_tools are thinking types too. openai-chat sends no level, so the endpoint runs its default. The first row that matches applies.

Client and model	none	low	medium	high	xhigh	max
anthropic-official: Claude Opus 5.5, Haiku 5.5, Fable 5, Fable 5.1	low	low	medium	high	xhigh	max
anthropic-official: Claude Sonnet 5.5	between_tools	low	medium	high	xhigh	max
anthropic-official: Claude 4.6	disabled	low	medium	high	high	max
anthropic-official: other Claude models; ant-messages	disabled	low	medium	high	xhigh	max
openai-official, openai-responses: GPT-6 models except GPT-6 Luna	low	low	medium	high	xhigh	max
openai-official, openai-responses: other models	none	low	medium	high	xhigh	max
openai-chat	not sent	not sent	not sent	not sent	not sent	not sent
deepseek-official	none	low	high	high	high	max
minimax-official: MiniMax-M3.1-Flash	low	low	medium	high	xhigh	max
minimax-official: MiniMax-M3	none	low	medium	high	high	high
moonshot-official: Kimi K2	disabled	enabled	enabled	enabled	enabled	enabled
moonshot-official: Kimi K3, Kimi K3.1 Pro	none	low	high	high	max	max
zai-official: GLM-5.3, GLM-5.3-Flash	low	low	high	high	max	max
zai-official: other GLM models	disabled	low	medium	high	xhigh	max
google-official, google-genai: image models	minimal	minimal	high	high	high	high
google-official, google-genai: Gemini 3 Pro	low	low	high	high	high	high
google-official, google-genai: other Pro models, Gemini 3.7, 3.8, 4 Argon	low	low	medium	high	high	high
google-official, google-genai: other Gemini models	minimal	low	medium	high	high	high

openai-chat-vllm-adapter sends no reasoning_effort either. It sets the switch the served model's chat template reads, as chat_template_kwargs:

Served model	chat_template_kwargs
Qwen3.8-27B, Qwen3.8-Flash-Next	enable_thinking: false for none; reasoning_effort low and medium as given, xhigh for high and above
DeepSeek-V4-Flash-Vision-Exp	nothing for none, so no thinking; thinking: true with reasoning_effort low for low, high for medium to xhigh, max for max
DeepSeek-V4-Pro, DeepSeek-V4-Flash	the same, but reasoning_effort high for low too
Qwen3.6-35B-A3B, Qwen3.5-0.8B, Qwen3.5-9B and any other model	enable_thinking: false for none, true otherwise

openai-embedding and Gemini speech and embedding models ignore thinking_level.

thinking_summary

Client	true	false
anthropic-official, ant-messages	display: "summarized" on the thinking block	display: "omitted"
openai-official, deepseek-official	reasoning.summary: "concise"	nothing
openai-responses	reasoning.summary: "concise" with a thinking_level; nothing without one	nothing
google-official	thinking_summaries: "auto"	thinking_summaries: "none"
google-genai	include_thoughts: true	include_thoughts: false
openai-chat, openai-chat-vllm-adapter, minimax-official, moonshot-official, zai-official	ignored	ignored

On the Anthropic clients, a summary asked for without a thinking_level turns adaptive thinking on; with thinking disabled or between_tools, nothing is sent. openai-embedding and Gemini speech and embedding models ignore it.

prompt_caching

Client	"enable"	"enhance"	"disable"
anthropic-official	cache_control: {"type": "ephemeral"}, also when not set	the same with ttl: "1h"	nothing
anthropic-official on Bedrock	the same marker with ttl: "5m" on the last block of the last user message, also when not set	the same with ttl: "1h"	nothing
moonshot-official on Kimi K3 and K3.1 Pro	nothing	prompt_cache_options: {"mode": "implicit", "ttl": "1h"}	raises
Every other client; moonshot-official on Kimi K2	nothing	raises	raises
openai-embedding, Gemini embedding models	ignored	ignored	ignored

fast_mode

false or not set sends nothing.

Client	true
openai-official, openai-responses, openai-chat, openai-chat-vllm-adapter, google-official	service_tier: "priority"
minimax-official	service_tier: "priority" on MiniMax-M3; raises on MiniMax-M3.1-Flash
ant-messages	speed: "fast" with the beta fast-mode-2026-02-01
anthropic-official	the same; raises on Bedrock and on Claude 4.6, Sonnet 5.5, Haiku 5.5 and Fable 5.1
google-genai, deepseek-official, moonshot-official, zai-official, openai-embedding	raises
Gemini embedding models	ignored

image_config

Client	Sent as
google-genai	image_config
google-official	response_format: a text entry, then an image entry carrying aspect_ratio and image_size
Every other client; Gemini speech and embedding models	ignored

tts_config

Gemini speech models, on both Google clients:

tts_config	Sent as
Not set	the voice Kore
One entry	its voice
Two entries	one voice per speaker; an entry without a speaker raises. Gemini 3.8 speech models get the script as one part per turn, with the speaker as metadata.
Three or more	raises

Every other model and client ignores it.

embedding_config

Client	Sent as
openai-embedding	dimensions
google-official, google-genai: embedding models	output_dimensionality
Every other client and model	ignored

trace_id

Client	What it does
Every client	Saves the turn under this id.
mmsp	The server saves the turn; the client saves nothing.
moonshot-official on Kimi K2	Also sends it as prompt_cache_key.

In Python and TypeScript

The field names are the same in both languages. Both packages export ThinkingLevel and PromptCaching. Python takes the plain strings as well; TypeScript takes the enum members, such as ThinkingLevel.HIGH and PromptCaching.ENABLE.