src/provider/anthropic.rs

This file holds AnthropicClient, the Messages API client behind the anthropic feature, and it is shared by the Z.ai profile and Bedrock’s native path.

It runs to about 3,640 lines with tests.

Key items

from_env() and the builder give you a client. with_max_tokens defaults to 8192 and rejects 0 at build time, replay_reasoning defaults to true since 0.4.0, and prompt_caching(bool) arrived in 0.4.0 and is off by default.

With prompt caching you get up to four cache_control breakpoints per request: the last tool definition, the system prompt, and the last two eligible tail blocks. Thinking blocks are skipped as breakpoints.

Your headers are x-api-key plus anthropic-version: 2023-06-01.

RequestBodySpec and build_request_body keep system always a top-level field, folded from inline system messages, with "" when empty, and a marked text block when you enable prompt_caching.

Tool use and tool result blocks ride along, and is_error: true travels on the wire.

The shared StreamEmitter, which Bedrock also reuses, gives you typed SSE events. Usage is latched at message_start and max-merged at message_delta.

Since 0.4.0 the input total is normalized to raw plus cache reads plus creation, and cache_read_input_tokens fills Usage::cached_input_tokens.

Thinking and redacted-thinking lanes exist on their own indexes, and signatures latch since 0.4.0, meaning signature and the redacted payload ride MessagePart::Thinking for your continuation replay. message_stop is the terminal event.

Forced-tool response_format synthesis lives here, along with the two up-front rejections for strict formats and grammar constraints.

pub(super) DEFAULT_MAX_TOKENS is shared with Bedrock.

Behavior notes

Tool-input deltas carry no index, so the emitter remembers the open block from content_block_start.

Late events after message_stop are swallowed as a desync guard, and a late error after a clean stop is dropped.

Since 0.4.0 your prior-turn thinking rides the request verbatim through replay_reasoning, and non-streaming replies carry their thinking blocks too, so the old strip-on-boundary behavior is gone.

Images are dropped by the converter.

For the deep dive, see Anthropic.