codec/anthropicapi targets POST /v1/messages. Anthropic has no in-thread system role, so request system text and system messages fold into top-level system.

Messages

The encoder requires max_tokens, defaulting to 4096 when effective sampling does not provide a positive value. User and assistant blocks map to the native tagged union. Tool results are a user message containing tool_result, and IsError is preserved. Required tool choice becomes tool_choice: {type:"any"}, and a named tool choice becomes tool_choice: {type:"tool", name:"..."}. Thinking is requested only for a model with Caps.Thinking and a non-empty effort, and its shape comes from Caps.ThinkingDialect: adaptive emits thinking: {type:"adaptive"} plus output_config.effort, and budget emits thinking: {type:"enabled", budget_tokens:N}. An undeclared dialect is an UndeclaredThinkingDialectError, never a guess. Whichever variant is emitted, temperature and top_p are omitted.

body, err := anthropicapi.EncodeRequest(req, false)
if err != nil {
	return err
}
fmt.Println(len(body))

Caching

The request encoder emits cache markers only when req.Model.Caps.PromptCaching is true and the request carries no non-empty transient system message, because such a message folds into the top-level system prefix and caching either breakpoint would capture transient context. It then emits at most two cache_control: {type:"ephemeral"} breakpoints: one on the system block, and one on the last cacheable block of the committed message history, so multi-turn requests accrue incremental cache hits. Transient messages are never marked, and thinking and redacted-thinking blocks are skipped. With the capability off, system remains a plain string and no marker is emitted. The response decoder normalizes separate cache_read_input_tokens and cache_creation_input_tokens fields.

Source and proof

Run go test ./codec/anthropicapi.