fix: removed temperature modifier in librarian agent to mitigate invisible errors

assets/agents/librarian/graph.yaml

@@ -88,7 +88,6 @@ nodes:
     prompt: |
       Research prompt: {{initial_prompt}}
     tools: []
-    temperature: 0.1
     output_schema:
       type: object
       properties:
@@ -180,7 +179,6 @@ nodes:
     tools:
       - mcp:ddg-search
     max_iterations: 15
-    temperature: 0.1
     state_updates:
       search_output: "{{output}}"
     fallback: synthesize
@@ -251,7 +249,6 @@ nodes:
     tools:
       - mcp:personal-github
     max_iterations: 15
-    temperature: 0.1
     state_updates:
       oss_output: "{{output}}"
     fallback: synthesize
@@ -338,7 +335,6 @@ nodes:
     tools:
       - fetch_url_via_curl
     max_iterations: 20
-    temperature: 0.1
     state_updates:
       findings: "{{output}}"
     fallback: final_format

src/client/bedrock.rs

@@ -1,3 +1,5 @@
+use std::mem;
+
 use super::*;
 
 use crate::utils::{base64_decode, encode_uri, hex_encode, hmac_sha256, sha256, strip_think_tag};
@@ -275,10 +277,11 @@ async fn chat_completions_streaming(
                                     format!("Tool call '{function_name}' has non-JSON arguments '{function_arguments}'")
                                 })?;
                                 handler.tool_call(ToolCall::new(
-                                    function_name.clone(),
+                                    mem::take(&mut function_name),
                                     arguments,
-                                    Some(function_id.clone()),
+                                    Some(mem::take(&mut function_id)),
                                 ))?;
+                                function_arguments.clear();
                             }
                         }
                         _ => {}
@@ -529,7 +532,11 @@ fn extract_chat_completions(data: &Value) -> Result<ChatCompletionsOutput> {
         bail!("Invalid response data: {data}");
     }
 
-    let output = ChatCompletionsOutput { text, tool_calls, ..Default::default() };
+    let output = ChatCompletionsOutput {
+        text,
+        tool_calls,
+        ..Default::default()
+    };
     Ok(output)
 }
 

src/client/claude.rs

@@ -1,3 +1,5 @@
+use std::mem;
+
 use super::access_token::get_access_token;
 use super::claude_oauth::ClaudeOAuthProvider;
 use super::oauth::{self, OAuthProvider};
@@ -232,8 +234,8 @@ pub async fn claude_chat_completions_streaming(
                         handler.text("\n</think>\n\n")?;
                         reasoning_state = 0;
                         handler.thinking_block(ThinkingBlock::Thinking {
-                            thinking: std::mem::take(&mut thinking_text),
+                            thinking: mem::take(&mut thinking_text),
-                            signature: std::mem::take(&mut thinking_signature),
+                            signature: mem::take(&mut thinking_signature),
                         });
                     }
                     if !function_name.is_empty() {
@@ -245,10 +247,11 @@ pub async fn claude_chat_completions_streaming(
                             })?
                         };
                         handler.tool_call(ToolCall::new(
-                            function_name.clone(),
+                            mem::take(&mut function_name),
                             arguments,
-                            Some(function_id.clone()),
+                            Some(mem::take(&mut function_id)),
                         ))?;
+                        function_arguments.clear();
                     }
                 }
                 _ => {}

src/graph/structured.rs

@@ -88,11 +88,21 @@ async fn run_one_shot(prompt: &str, ctx: &mut RequestContext) -> Result<String>
 }
 
 fn try_parse_json(raw: &str) -> Option<Value> {
-    let cleaned = strip_code_fences(raw.trim());
+    let cleaned = strip_code_fences(strip_thinking_blocks(raw.trim()));
-
     serde_json::from_str(cleaned).ok()
 }
 
+fn strip_thinking_blocks(s: &str) -> &str {
+    let mut s = s.trim_start();
+    while s.starts_with("<think>") {
+        match s.find("</think>") {
+            Some(end) => s = s[end + "</think>".len()..].trim_start(),
+            None => break,
+        }
+    }
+    s
+}
+
 fn strip_code_fences(s: &str) -> &str {
     let after_open = s
         .strip_prefix("```json")
@@ -148,6 +158,38 @@ mod tests {
         assert_eq!(v, json!({"x": true}));
     }
 
+    #[test]
+    fn try_parse_json_strips_thinking_blocks() {
+        let raw = "<think>\nsome reasoning\n</think>\n{\"a\": 1}";
+
+        let v = try_parse_json(raw).unwrap();
+
+        assert_eq!(v, json!({"a": 1}));
+    }
+
+    #[test]
+    fn try_parse_json_strips_empty_thinking_block() {
+        let raw = "<think>\n\n</think>\n{\"a\": 1}";
+
+        let v = try_parse_json(raw).unwrap();
+
+        assert_eq!(v, json!({"a": 1}));
+    }
+
+    #[test]
+    fn try_parse_json_strips_multiple_thinking_blocks() {
+        let raw = "<think>first</think>\n<think>second</think>\n{\"a\": 1}";
+
+        let v = try_parse_json(raw).unwrap();
+
+        assert_eq!(v, json!({"a": 1}));
+    }
+
+    #[test]
+    fn try_parse_json_unclosed_think_tag_returns_none() {
+        assert!(try_parse_json("<think>unclosed {\"a\": 1}").is_none());
+    }
+
     #[test]
     fn try_parse_json_returns_none_on_prose() {
         assert!(try_parse_json("Here is the result: it's good").is_none());