agent: Fix issue with Anthropic thinking models (#33317)

cc @osyvokon 

We were seeing a bunch of errors in our backend when people were using
Claude models with thinking enabled.

In the logs we would see
> an error occurred while interacting with the Anthropic API:
invalid_request_error: messages.x.content.0.type: Expected `thinking` or
`redacted_thinking`, but found `text`. When `thinking` is enabled, a
final `assistant` message must start with a thinking block (preceeding
the lastmost set of `tool_use` and `tool_result` blocks). We recommend
you include thinking blocks from previous turns. To avoid this
requirement, disable `thinking`. Please consult our documentation at
https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking

However, this issue did not occur frequently and was not easily
reproducible. Turns out it was triggered by us not correctly handling
[Redacted Thinking
Blocks](https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#thinking-redaction).

I could constantly reproduce this issue by including this magic string:
`ANTHROPIC_MAGIC_STRING_TRIGGER_REDACTED_THINKING_46C9A13E193C177646C7398A98432ECCCE4C1253D5E2D82641AC0E52CC2876CB
` in the request, which forces `claude-3-7-sonnet` to emit redacted
thinking blocks (confusingly the magic string does not seem to be
working for `claude-sonnet-4`). As soon as we hit a tool call Anthropic
would return an error.

Thanks to @osyvokon for pointing me in the right direction 😄!


Release Notes:

- agent: Fixed an issue where Anthropic models would sometimes return an
error when thinking was enabled

crates/agent/src/thread.rs

@@ -145,6 +145,10 @@ impl Message {
         }
     }
 
+    pub fn push_redacted_thinking(&mut self, data: String) {
+        self.segments.push(MessageSegment::RedactedThinking(data));
+    }
+
     pub fn push_text(&mut self, text: &str) {
         if let Some(MessageSegment::Text(segment)) = self.segments.last_mut() {
             segment.push_str(text);
@@ -183,7 +187,7 @@ pub enum MessageSegment {
         text: String,
         signature: Option<String>,
     },
-    RedactedThinking(Vec<u8>),
+    RedactedThinking(String),
 }
 
 impl MessageSegment {
@@ -1643,6 +1647,25 @@ impl Thread {
                                     };
                                 }
                             }
+                            LanguageModelCompletionEvent::RedactedThinking {
+                                data
+                            } => {
+                                thread.received_chunk();
+
+                                if let Some(last_message) = thread.messages.last_mut() {
+                                    if last_message.role == Role::Assistant
+                                        && !thread.tool_use.has_tool_results(last_message.id)
+                                    {
+                                        last_message.push_redacted_thinking(data);
+                                    } else {
+                                        request_assistant_message_id =
+                                            Some(thread.insert_assistant_message(
+                                                vec![MessageSegment::RedactedThinking(data)],
+                                                cx,
+                                            ));
+                                    };
+                                }
+                            }
                             LanguageModelCompletionEvent::ToolUse(tool_use) => {
                                 let last_assistant_message_id = request_assistant_message_id
                                     .unwrap_or_else(|| {

crates/agent/src/thread_store.rs

@@ -731,7 +731,7 @@ pub enum SerializedMessageSegment {
         signature: Option<String>,
     },
     RedactedThinking {
-        data: Vec<u8>,
+        data: String,
     },
 }
 

crates/assistant_context/src/assistant_context.rs

@@ -2117,6 +2117,7 @@ impl AssistantContext {
                                             );
                                         }
                                     }
+                                    LanguageModelCompletionEvent::RedactedThinking { .. } => {},
                                     LanguageModelCompletionEvent::Text(mut chunk) => {
                                         if let Some(start) = thought_process_stack.pop() {
                                             let end = buffer.anchor_before(message_old_end_offset);

crates/eval/src/instance.rs

@@ -1030,6 +1030,7 @@ pub fn response_events_to_markdown(
             Ok(LanguageModelCompletionEvent::Thinking { text, .. }) => {
                 thinking_buffer.push_str(text);
             }
+            Ok(LanguageModelCompletionEvent::RedactedThinking { .. }) => {}
             Ok(LanguageModelCompletionEvent::Stop(reason)) => {
                 flush_buffers(&mut response, &mut text_buffer, &mut thinking_buffer);
                 response.push_str(&format!("**Stop**: {:?}\n\n", reason));
@@ -1126,6 +1127,7 @@ impl ThreadDialog {
 
                 // Skip these
                 Ok(LanguageModelCompletionEvent::UsageUpdate(_))
+                | Ok(LanguageModelCompletionEvent::RedactedThinking { .. })
                 | Ok(LanguageModelCompletionEvent::StatusUpdate { .. })
                 | Ok(LanguageModelCompletionEvent::StartMessage { .. })
                 | Ok(LanguageModelCompletionEvent::Stop(_)) => {}

crates/language_model/src/language_model.rs

@@ -67,6 +67,9 @@ pub enum LanguageModelCompletionEvent {
         text: String,
         signature: Option<String>,
     },
+    RedactedThinking {
+        data: String,
+    },
     ToolUse(LanguageModelToolUse),
     StartMessage {
         message_id: String,
@@ -359,6 +362,7 @@ pub trait LanguageModel: Send + Sync {
                                 Ok(LanguageModelCompletionEvent::StartMessage { .. }) => None,
                                 Ok(LanguageModelCompletionEvent::Text(text)) => Some(Ok(text)),
                                 Ok(LanguageModelCompletionEvent::Thinking { .. }) => None,
+                                Ok(LanguageModelCompletionEvent::RedactedThinking { .. }) => None,
                                 Ok(LanguageModelCompletionEvent::Stop(_)) => None,
                                 Ok(LanguageModelCompletionEvent::ToolUse(_)) => None,
                                 Ok(LanguageModelCompletionEvent::UsageUpdate(token_usage)) => {

crates/language_model/src/request.rs

@@ -303,7 +303,7 @@ pub enum MessageContent {
         text: String,
         signature: Option<String>,
     },
-    RedactedThinking(Vec<u8>),
+    RedactedThinking(String),
     Image(LanguageModelImage),
     ToolUse(LanguageModelToolUse),
     ToolResult(LanguageModelToolResult),

crates/language_models/src/provider/anthropic.rs

@@ -554,9 +554,7 @@ pub fn into_anthropic(
                         }
                         MessageContent::RedactedThinking(data) => {
                             if !data.is_empty() {
-                                Some(anthropic::RequestContent::RedactedThinking {
+                                Some(anthropic::RequestContent::RedactedThinking { data })
-                                    data: String::from_utf8(data).ok()?,
-                                })
                             } else {
                                 None
                             }
@@ -730,10 +728,8 @@ impl AnthropicEventMapper {
                         signature: None,
                     })]
                 }
-                ResponseContent::RedactedThinking { .. } => {
+                ResponseContent::RedactedThinking { data } => {
-                    // Redacted thinking is encrypted and not accessible to the user, see:
+                    vec![Ok(LanguageModelCompletionEvent::RedactedThinking { data })]
-                    // https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking#suggestions-for-handling-redacted-thinking-in-production
-                    Vec::new()
                 }
                 ResponseContent::ToolUse { id, name, .. } => {
                     self.tool_uses_by_index.insert(