LLM Requests¶
This reference covers request fields and the built-in provider mappings. For the integration flow and a construction example, read LLM-Related Protocols. For the reply, continue to Responses and Streams.
Request envelope¶
An algorithm receives Request. Its llm_request contains the conversation;
the remaining envelope data belongs to the application.
| Field and type | Meaning |
|---|---|
Request.llm_request: LlmRequest |
The conversation and settings used by the algorithm and model client. |
Request.metadata: Option<Metadata> |
Optional session, correlation, and routing metadata. |
Request.raw_request: Option<Value> |
An optional original body retained by the host. Translation codecs do not read this field. |
Codec replay uses LlmRequest.preservation, separately from the envelope's raw
body. See preservation and edits
before changing a decoded request.
Conversation fields¶
These mappings describe the built-in codecs. Supported features also depend on the destination model. Select a Rust field or type for its full crate definition.
| Protocol field and type | Meaning | Anthropic Messages | OpenAI Responses | Chat Completions |
|---|---|---|---|---|
LlmRequest.model: Option<String> |
Addressed model; the host can replace it with the selected target. | model |
model |
model |
LlmRequest.instructions: Vec<InstructionBlock> |
Top-level system and developer instructions. | Top-level system; no distinct developer authority |
instructions |
Leading system/developer messages in messages |
LlmRequest.messages: Vec<Message> |
Ordered turns, each with a role and content blocks. | messages[]; inline system/developer messages are added to top-level system on encode |
Messages, including system/developer messages, and tool items in input |
messages[], including later system/developer messages |
LlmRequest.tools: Vec<Tool> |
Available functions, MCP toolsets, or retained provider definitions. | tools[] |
tools[] |
tools[]; function definitions are nested in function |
LlmRequest.tool_choice: Option<ToolChoice> |
Let the model choose, require a call, forbid calls, or select a named tool. | auto, any, none, or tool |
auto, required, none, or function |
auto, required, none, or function |
LlmRequest.stream: bool |
Request incremental output. | stream |
stream |
stream |
LlmRequest.extensions: ProviderExtensions |
Provider fields without a first-class neutral field. | Source fields retained by the codec | Source fields retained by the codec | Source fields retained by the codec |
LlmRequest.preservation: PreservationMetadata |
Original JSON used for same-format replay. | Original parsed body | Original parsed body | Original parsed body |
Messages and tool content¶
Message.role is a Role; Message.content is an ordered Vec<ContentBlock>.
The content variants cover text, reasoning, media, tools, MCP activity, refusals,
and unknown provider blocks. An unknown block retains its source format and raw
JSON. Its retention does not make it usable in a different API.
| Protocol field and type | Meaning | Anthropic Messages | OpenAI Responses | Chat Completions |
|---|---|---|---|---|
ToolDefinition.parameters: Value |
JSON Schema for function arguments. | input_schema |
Function parameters |
tools[].function.parameters |
ToolCall.id: String |
Identifier that pairs a call with its result. | tool_use.id |
function_call.call_id |
tool_calls[].id |
ToolCall.name: String |
Function name. | tool_use.name |
function_call.name |
tool_calls[].function.name |
ToolCall.arguments: Value |
Parsed argument value. | JSON-valued input |
arguments, commonly a JSON-encoded string |
function.arguments, commonly a JSON-encoded string |
ToolResult.tool_call_id: String |
Identifier of the call being answered. | tool_result.tool_use_id |
function_call_output.call_id |
Tool message tool_call_id |
ToolResult.content: Vec<ContentBlock> |
Ordered tool output. | tool_result.content |
function_call_output.output |
Tool message content |
ToolResult.is_error: Option<bool> |
Whether tool execution failed, if reported. | tool_result.is_error |
No dedicated error flag | No standard error flag |
Tool definitions describe available work. A ToolCall records a model's request
to execute that work; it does not execute the tool. Keep call IDs and content
order intact when writing an adapter.
Tool::Provider retains an unsupported tool definition for its source format.
Built-in provider tools and reasoning signatures may have no valid counterpart
in another API. Consult the translation limits
instead of assuming a successful parse implies cross-provider support.
Generation settings¶
The request groups these fields in SamplingParams, OutputParams, and
ReasoningParams. Optional values remain unspecified when absent; individual
providers apply their own defaults and accepted ranges.
| Protocol field and type | Meaning | Anthropic Messages | OpenAI Responses | Chat Completions |
|---|---|---|---|---|
SamplingParams.temperature: Option<f64> |
Sampling temperature. | (omitted; refer to the note below) | temperature |
temperature |
SamplingParams.top_p: Option<f64> |
Nucleus-sampling probability mass. | top_p |
top_p |
top_p |
SamplingParams.top_k: Option<i64> |
Maximum candidate-token count. | top_k |
No standard field | No standard field |
OutputParams.max_output_tokens: Option<u64> |
Output token budget. | max_tokens |
max_output_tokens |
max_completion_tokens, or legacy max_tokens |
OutputParams.response_format: Option<Value> |
Requested structured-output shape. | output_config.format |
text.format |
response_format |
OutputParams.is_schema_enforced: Option<bool> |
Override whether the provider must enforce the schema; absence retains native behavior. | Schema-backed output configuration | Schema strictness in text.format |
Schema strictness in response_format |
ReasoningParams.effort: Option<String> |
Requested reasoning effort. | output_config.effort |
reasoning.effort |
reasoning_effort |
ReasoningParams.raw: Option<Value> |
Provider-specific reasoning configuration. | thinking controls |
reasoning object |
No standard equivalent for the whole object |
Anthropic temperature: Switchyard omits temperature when rebuilding Anthropic
requests. Older models still support this parameter under API version 2023-06-01,
but models released after Claude Opus 4.6 reject values other than the default
1.0. For model restrictions, refer to the Anthropic parameter documentation.
API version 2026-10-21 removes the parameter entirely and rejects it even at its
default value. For removal details, refer to Anthropic's migration guide.
Switchyard scales OpenAI temperatures from the 0.0–2.0 range to Anthropic's
0.0–1.0 range before comparing with the default. An OpenAI temperature of 2.0
scales to Anthropic's default of 1.0 and is omitted without diagnostics or rejection.
Omitting a non-default value after scaling to Anthropic's range follows the
lossy-conversion policy. Exact preserved Anthropic
requests can still retain temperature.
For older Anthropic models that support non-default temperatures under API version
2023-06-01, add these settings to the target's TOML configuration:
After translation, Switchyard removes any existing temperature and sets it to
0.5 in Anthropic's units, without further scaling. This also overrides temperature
in preserved requests. Refer to the target configuration reference
for these settings.
Preserved same-format bodies take precedence over edits to these settings, including schema enforcement. See encoding an edited request.
Provider contracts¶
For complete provider schemas, see Anthropic Messages, OpenAI Responses, and Chat Completions. For response fields and event ownership, continue to Responses and Streams.