Skip to content

Solution for long term memory for agent coding CLIs and to facilitate handoff between different agent vendors

Rust

fix(llm): read Claude responses that start with thinking blocks

Vinícius DS Andrade0a1f618f629f
diff --git a/crates/ai-memory-llm/src/anthropic.rs b/crates/ai-memory-llm/src/anthropic.rs
index 150d2879..5bc04d62 100644
--- a/crates/ai-memory-llm/src/anthropic.rs
+++ b/crates/ai-memory-llm/src/anthropic.rs
@@ -212,6 +212,13 @@ struct AnthropicResponse {
enum AnthropicContent {
Text { text: String },
ToolUse { input: serde_json::Value },
+ /// A block this provider does not read, such as `thinking` or
+ /// `redacted_thinking`. Models that think by default (Opus 5 and later,
+ /// Sonnet 5 and later, Haiku 5.5, Fable, Mythos), and every model sent
+ /// adaptive thinking for a configured effort, can start a response with
+ /// such blocks. Without this variant one of them fails the whole parse.
+ #[serde(other)]
+ Other,
}
#[derive(Debug, Deserialize)]
@@ -247,7 +254,7 @@ impl LlmProvider for AnthropicProvider {
.iter()
.filter_map(|c| match c {
AnthropicContent::Text { text } => Some(text.as_str()),
- AnthropicContent::ToolUse { .. } => None,
+ AnthropicContent::ToolUse { .. } | AnthropicContent::Other => None,
})
.collect::<Vec<_>>()
.join("\n");
@@ -593,6 +600,8 @@ fn version_segment(s: &str) -> Option<u32> {
mod tests {
use secrecy::SecretString;
use serde_json::json;
+ use wiremock::matchers::{method, path};
+ use wiremock::{Mock, MockServer, ResponseTemplate};
use crate::types::{ChatMessage, ReasoningEffort};
use rstest::rstest;
@@ -737,6 +746,46 @@ mod tests {
assert!(plain.get("thinking").is_none());
}
+ /// A model that thinks opens its response with `thinking` blocks, empty
+ /// under the default `display: "omitted"`, and may return
+ /// `redacted_thinking`. Both completion paths must read past them.
+ #[tokio::test]
+ async fn thinking_blocks_do_not_fail_either_completion_path() {
+ let server = MockServer::start().await;
+ Mock::given(method("POST"))
+ .and(path("/v1/messages"))
+ .respond_with(ResponseTemplate::new(200).set_body_json(json!({
+ "model": "claude-opus-5",
+ "content": [
+ {"type": "thinking", "thinking": "", "signature": "sig"},
+ {"type": "redacted_thinking", "data": "opaque"},
+ {"type": "text", "text": "OK"},
+ {
+ "type": "tool_use",
+ "id": "toolu_1",
+ "name": "result",
+ "input": {"answer": "OK"}
+ }
+ ],
+ "usage": {"input_tokens": 3, "output_tokens": 5}
+ })))
+ .expect(2)
+ .mount(&server)
+ .await;
+ let provider = AnthropicProvider::new(SecretString::from("sk-ant-test"), "claude-opus-5")
+ .unwrap()
+ .with_base_url(server.uri());
+
+ let response = provider.complete(chat_request()).await.unwrap();
+ assert_eq!(response.text, "OK");
+
+ let value = provider
+ .complete_structured_raw(chat_request(), json!({"type": "object"}))
+ .await
+ .unwrap();
+ assert_eq!(value, json!({"answer": "OK"}));
+ }
+
fn body_for_effort(model: &str, effort: ReasoningEffort) -> serde_json::Value {
let provider = AnthropicProvider::new(SecretString::from("sk-ant-test"), model)
.unwrap()