Module protocol

Expand description

The Azure Responses API wire types: what a request carries and what a streamed response deserializes into.

Structs§

AISearchOutput
IncompleteReason: Why a response is incomplete, e.g. max_output_tokens or content_filter.
InputTokensDetails: How much of the input was served from Azure’s prompt cache. The only way to tell whether the prompt prefix is actually stable, since a perturbed prefix fails silently by costing more.
LLMRequest
LLMRequestResponseFormatParam
MistralParams: Mistral’s request parameters are not decided yet; this field exists only so LLMRequestParams::Mistral has a shape of its own to serialize and deserialize as, distinct from the other two variants.
NonThinkingParams
OutputTokensDetails: How much of the output was thinking, which is what moves when the reasoning context widens or narrows.
Reasoning
ReasoningOutput
RequestTextOptions
Response: Response received from LLM API
ResponseError: Error object returned by the LLM API on a failed response. Fields are optional so any error shape deserializes rather than crashing the stream parser.
ResponseOutput: One line of the Azure event stream. Which fields are set depends on response_type: a text delta, a completed item, a whole response, or an error.
ResponseReasoning: The reasoning settings a response reports back, as opposed to the ones the request asked for.
ThinkingParams
Usage: What the request was billed for. Optional throughout: Azure reports the cache fields only on some deployment types, and PTU-M never reports cache_write_tokens at all.

InputItem
LLMRequestParams: Each variant’s own struct rejects a field belonging to another, which is what tells them apart on the way back in: untagged tries them in order and they otherwise all match any object.
LLMToolChoice
OutputItem
ReasoningContext: How far back the model reuses its own reasoning.
ReceivedOutputItem: An output item as it arrived, which is not necessarily one this code knows.
SummaryType: Untagged would serialize these unit variants as null, which asks Azure for no summary at all.

azure_stream_error 🔒: Builds the error the stream parsers raise when Azure reports a failure on the line they are reading, attaching err as the error’s Azure source.
check_response_complete 🔒: Errors once a response has finished streaming if it stopped short of a complete one, whether output names the reason or the caller only knows a response.incomplete event fired with none.
check_response_output 🔒: Logs output’s usage under context, then errors if it reports a failure — the two checks both stream parsers make on every line, before either looks at its own content.
log_response_usage 🔒: Logs the usage a response line carries, if any, under context.
reported_azure_error 🔒: The error Azure reported on this line, if any: the one nested under response when a response object came with it, otherwise the top-level one. fallback_response_id is used only for the latter, since the former carries its own response id.
reported_incomplete_reason 🔒: Why the response this line carries is incomplete, if it says so. A response that stopped short still streams as if it had finished, so this is the only thing that tells the two apart.
urls_that_parse 🔒: The search results whose url this code can actually fetch. One unparseable entry must not drop every other citation from the same search.