Module llm 

One provider-neutral API layer for the assistant (design: docs/src/research/agent-interface.md).

This module knows nothing about TeamCraft (no Actor, no mcp): it holds the neutral conversation types, the pure Provider trait (encode a request, decode a reply, map an error – no I/O, so recorded provider bodies test it directly) and the Transport trait with its one HTTP implementation in http. The wire formats live in one file each (anthropic).

The tool loop in crate::assistant only sees these types.

Modules

anthropic

The Anthropic Messages API: request encoding (prompt caching, the server-side fallback beta), reply decoding and the error table.

http

The one HTTP client of the LLM layer: timeouts, no redirects (a 3xx is an error, so the API key never goes to a second host), the key only in its request header.

openai

Chat Completions (POST {base}/chat/completions), the format OpenAI, Azure OpenAI (v1), Mistral, OpenRouter, Groq, Ollama, vLLM and LM Studio speak. Stateless like the Messages API, so the loop stays the same.

Structs

HttpCall

A request ready to send: the path below the base URL, headers (without the API key) and the JSON body.

HttpResponse

An HTTP answer.

Reply

A decoded model reply.

Request

One model call.

ToolCall

One tool call of the model.

ToolResult

The result of one tool call, handed back to the model.

ToolSpec

A tool the model may call. schema is the JSON Schema of the input.

Usage

Token counts of one call. input means all prompt tokens (cached ones included).

Enums

AuthStyle

How the API key is sent.

LlmError

Why a model call failed.

Stop

Why the model stopped.

Turn

One entry of the conversation.

Traits

Provider

One wire format. Pure: no network, no clock, no database.

Transport

Sends one HTTP call and returns the response as it is (any status).

Functions

complete

One model call, one attempt: encode, send, map errors, decode.

cut

Cuts s to at most max bytes at a character boundary.

error_message

The error.message of a JSON error body, cut short.

retry_after

retry-after in seconds (the HTTP-date form is not used by the APIs).

status_error

The one error table of all providers (design page): 401/403 are an authentication error, 402 and a billing-quota 429 a quota error (no retry), other 429s are rate limits with their retry-after, 5xx (also 529) an overloaded service, a context-length rejection its own error, everything else a rejected request.