fix(agent): recover malformed tool_call blocks with leading text

src/agent/loop_.rs

@@ -409,10 +409,11 @@ fn extract_json_values(input: &str) -> Vec<serde_json::Value> {
 /// compatibility.
 ///
 /// Also supports JSON with `tool_calls` array from OpenAI-format responses.
-fn parse_tool_calls(response: &str) -> (String, Vec<ParsedToolCall>) {
+fn parse_tool_calls(response: &str) -> (String, Vec<ParsedToolCall>, bool) {
     let mut text_parts = Vec::new();
     let mut calls = Vec::new();
     let mut remaining = response;
+    let mut malformed_markup = false;
 
     // First, try to parse as OpenAI-style JSON response with tool_calls array
     // This handles providers like Minimax that return tool_calls in native JSON format
@@ -425,7 +426,7 @@ fn parse_tool_calls(response: &str) -> (String, Vec<ParsedToolCall>) {
                     text_parts.push(content.trim().to_string());
                 }
             }
-            return (text_parts.join("\n"), calls);
+            return (text_parts.join("\n"), calls, false);
         }
     }
 
@@ -456,10 +457,12 @@ fn parse_tool_calls(response: &str) -> (String, Vec<ParsedToolCall>) {
 
             if !parsed_any {
                 tracing::warn!("Malformed <tool_call> JSON: expected tool-call object in tag body");
+                malformed_markup = true;
             }
 
             remaining = &after_open[close_idx + close_tag.len()..];
         } else {
+            malformed_markup = true;
             break;
         }
     }
@@ -477,7 +480,7 @@ fn parse_tool_calls(response: &str) -> (String, Vec<ParsedToolCall>) {
         text_parts.push(remaining.trim().to_string());
     }
 
-    (text_parts.join("\n"), calls)
+    (text_parts.join("\n"), calls, malformed_markup)
 }
 
 fn parse_structured_tool_calls(tool_calls: &[ToolCall]) -> Vec<ParsedToolCall> {
@@ -593,7 +596,7 @@ pub(crate) async fn run_tool_call_loop(
         let llm_started_at = Instant::now();
 
         // Choose between native tool-call API and prompt-based tool use.
-        let (response_text, parsed_text, tool_calls, assistant_history_content) =
+        let (response_text, parsed_text, tool_calls, assistant_history_content, malformed_markup) =
             if use_native_tools {
                 match provider
                     .chat_with_tools(history, &tool_definitions, model, temperature)
@@ -610,13 +613,16 @@ pub(crate) async fn run_tool_call_loop(
                         let response_text = resp.text_or_empty().to_string();
                         let mut calls = parse_structured_tool_calls(&resp.tool_calls);
                         let mut parsed_text = String::new();
+                        let mut malformed_markup = false;
 
                         if calls.is_empty() {
-                            let (fallback_text, fallback_calls) = parse_tool_calls(&response_text);
+                            let (fallback_text, fallback_calls, fallback_malformed_markup) =
+                                parse_tool_calls(&response_text);
                             if !fallback_text.is_empty() {
                                 parsed_text = fallback_text;
                             }
                             calls = fallback_calls;
+                            malformed_markup = fallback_malformed_markup;
                         }
 
                         let assistant_history_content = if resp.tool_calls.is_empty() {
@@ -628,7 +634,13 @@ pub(crate) async fn run_tool_call_loop(
                             )
                         };
 
-                        (response_text, parsed_text, calls, assistant_history_content)
+                        (
+                            response_text,
+                            parsed_text,
+                            calls,
+                            assistant_history_content,
+                            malformed_markup,
+                        )
                     }
                     Err(e) => {
                         observer.record_event(&ObserverEvent::LlmResponse {
@@ -658,8 +670,15 @@ pub(crate) async fn run_tool_call_loop(
                         });
                         let response_text = resp;
                         let assistant_history_content = response_text.clone();
-                        let (parsed_text, calls) = parse_tool_calls(&response_text);
-                        (response_text, parsed_text, calls, assistant_history_content)
+                        let (parsed_text, calls, malformed_markup) =
+                            parse_tool_calls(&response_text);
+                        (
+                            response_text,
+                            parsed_text,
+                            calls,
+                            assistant_history_content,
+                            malformed_markup,
+                        )
                     }
                     Err(e) => {
                         observer.record_event(&ObserverEvent::LlmResponse {
@@ -684,7 +703,8 @@ pub(crate) async fn run_tool_call_loop(
         };
         let has_tool_call_markup =
             response_text.contains("<tool_call>") && response_text.contains("</tool_call>");
-        let malformed_tool_call_markup = looks_like_malformed_tool_call_markup(&response_text);
+        let malformed_tool_call_markup =
+            malformed_markup || looks_like_malformed_tool_call_markup(&response_text);
 
         if tool_calls.is_empty() {
             // Recovery path: the model attempted tool use but emitted malformed JSON.
@@ -1427,10 +1447,11 @@ mod tests {
 {"name": "shell", "arguments": {"command": "ls -la"}}
 </tool_call>"#;
 
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         assert_eq!(text, "Let me check that.");
         assert_eq!(calls.len(), 1);
         assert_eq!(calls[0].name, "shell");
+        assert!(!malformed);
         assert_eq!(
             calls[0].arguments.get("command").unwrap().as_str().unwrap(),
             "ls -la"
@@ -1446,18 +1467,20 @@ mod tests {
 {"name": "file_read", "arguments": {"path": "b.txt"}}
 </tool_call>"#;
 
-        let (_, calls) = parse_tool_calls(response);
+        let (_, calls, malformed) = parse_tool_calls(response);
         assert_eq!(calls.len(), 2);
         assert_eq!(calls[0].name, "file_read");
         assert_eq!(calls[1].name, "file_read");
+        assert!(!malformed);
     }
 
     #[test]
     fn parse_tool_calls_returns_text_only_when_no_calls() {
         let response = "Just a normal response with no tools.";
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         assert_eq!(text, "Just a normal response with no tools.");
         assert!(calls.is_empty());
+        assert!(!malformed);
     }
 
     #[test]
@@ -1467,9 +1490,23 @@ not valid json
 </tool_call>
 Some text after."#;
 
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         assert!(calls.is_empty());
         assert!(text.contains("Some text after."));
+        assert!(malformed);
+    }
+
+    #[test]
+    fn parse_tool_calls_marks_malformed_when_text_precedes_invalid_tool_call() {
+        let response = r#"I will schedule a 3AM update task. First, I will inspect existing tasks:
+<tool_call>
+{"action":"create","command":"nova update","expression":"0 3 * * *","id":"nova-self-update"}
+</tool_call>"#;
+
+        let (text, calls, malformed) = parse_tool_calls(response);
+        assert!(calls.is_empty());
+        assert!(text.contains("I will schedule a 3AM update task"));
+        assert!(malformed);
     }
 
     #[test]
@@ -1480,10 +1517,11 @@ Some text after."#;
 </tool_call>
 After text."#;
 
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         assert!(text.contains("Before text."));
         assert!(text.contains("After text."));
         assert_eq!(calls.len(), 1);
+        assert!(!malformed);
     }
 
     #[test]
@@ -1491,7 +1529,7 @@ After text."#;
         // OpenAI-style response with tool_calls array
         let response = r#"{"content": "Let me check that for you.", "tool_calls": [{"type": "function", "function": {"name": "shell", "arguments": "{\"command\": \"ls -la\"}"}}]}"#;
 
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         assert_eq!(text, "Let me check that for you.");
         assert_eq!(calls.len(), 1);
         assert_eq!(calls[0].name, "shell");
@@ -1499,16 +1537,18 @@ After text."#;
             calls[0].arguments.get("command").unwrap().as_str().unwrap(),
             "ls -la"
         );
+        assert!(!malformed);
     }
 
     #[test]
     fn parse_tool_calls_handles_openai_format_multiple_calls() {
         let response = r#"{"tool_calls": [{"type": "function", "function": {"name": "file_read", "arguments": "{\"path\": \"a.txt\"}"}}, {"type": "function", "function": {"name": "file_read", "arguments": "{\"path\": \"b.txt\"}"}}]}"#;
 
-        let (_, calls) = parse_tool_calls(response);
+        let (_, calls, malformed) = parse_tool_calls(response);
         assert_eq!(calls.len(), 2);
         assert_eq!(calls[0].name, "file_read");
         assert_eq!(calls[1].name, "file_read");
+        assert!(!malformed);
     }
 
     #[test]
@@ -1516,10 +1556,11 @@ After text."#;
         // Some providers don't include content field with tool_calls
         let response = r#"{"tool_calls": [{"type": "function", "function": {"name": "memory_recall", "arguments": "{}"}}]}"#;
 
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         assert!(text.is_empty()); // No content field
         assert_eq!(calls.len(), 1);
         assert_eq!(calls[0].name, "memory_recall");
+        assert!(!malformed);
     }
 
     #[test]
@@ -1530,7 +1571,7 @@ After text."#;
 ```
 </tool_call>"#;
 
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         assert!(text.is_empty());
         assert_eq!(calls.len(), 1);
         assert_eq!(calls[0].name, "file_write");
@@ -1538,6 +1579,7 @@ After text."#;
             calls[0].arguments.get("path").unwrap().as_str().unwrap(),
             "test.py"
         );
+        assert!(!malformed);
     }
 
     #[test]
@@ -1547,7 +1589,7 @@ I will now call the tool with this payload:
 {"name": "shell", "arguments": {"command": "pwd"}}
 </tool_call>"#;
 
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         assert!(text.is_empty());
         assert_eq!(calls.len(), 1);
         assert_eq!(calls[0].name, "shell");
@@ -1555,6 +1597,7 @@ I will now call the tool with this payload:
             calls[0].arguments.get("command").unwrap().as_str().unwrap(),
             "pwd"
         );
+        assert!(!malformed);
     }
 
     #[test]
@@ -1609,13 +1652,14 @@ I will now call the tool with this payload:
         let response = r#"Sure, creating the file now.
 {"name": "file_write", "arguments": {"path": "hello.py", "content": "print('hello')"}}"#;
 
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         assert!(text.contains("Sure, creating the file now."));
         assert_eq!(
             calls.len(),
             0,
             "Raw JSON without wrappers should not be parsed"
         );
+        assert!(!malformed);
     }
 
     #[test]
@@ -1776,9 +1820,10 @@ I will now call the tool with this payload:
 
 </tool_result>
 Done."#;
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         assert!(text.contains("Done."));
         assert!(calls.is_empty());
+        assert!(!malformed);
     }
 
     #[test]
@@ -1793,10 +1838,11 @@ Done."#;
     fn parse_tool_calls_handles_empty_tool_calls_array() {
         // Recovery: Empty tool_calls array returns original response (no tool parsing)
         let response = r#"{"content": "Hello", "tool_calls": []}"#;
-        let (text, calls) = parse_tool_calls(response);
+        let (text, calls, malformed) = parse_tool_calls(response);
         // When tool_calls is empty, the entire JSON is returned as text
         assert!(text.contains("Hello"));
         assert!(calls.is_empty());
+        assert!(!malformed);
     }
 
     #[test]
@@ -2086,6 +2132,86 @@ Done."#;
             .any(|m| m.role == "user" && m.content.contains("[Tool parser error]")));
     }
 
+    struct TextPrefixedMalformedThenValidToolProvider;
+
+    #[async_trait]
+    impl Provider for TextPrefixedMalformedThenValidToolProvider {
+        async fn chat_with_system(
+            &self,
+            _system_prompt: Option<&str>,
+            _message: &str,
+            _model: &str,
+            _temperature: f64,
+        ) -> anyhow::Result<String> {
+            anyhow::bail!("chat_with_system should not be called in this test");
+        }
+
+        async fn chat_with_history(
+            &self,
+            messages: &[ChatMessage],
+            _model: &str,
+            _temperature: f64,
+        ) -> anyhow::Result<String> {
+            if messages
+                .iter()
+                .any(|m| m.role == "user" && m.content.contains("[Tool results]"))
+            {
+                return Ok("Scheduled successfully.".to_string());
+            }
+
+            if messages
+                .iter()
+                .any(|m| m.role == "user" && m.content.contains("[Tool parser error]"))
+            {
+                return Ok(r#"<tool_call>
+{"name":"shell","arguments":{"command":"echo fixed"}}
+</tool_call>"#
+                    .to_string());
+            }
+
+            Ok(
+                r#"I will schedule a 3AM update task. First, I will inspect existing tasks:
+<tool_call>
+{"action":"create","command":"nova update","expression":"0 3 * * *","id":"nova-self-update"}
+</tool_call>"#
+                    .to_string(),
+            )
+        }
+    }
+
+    #[tokio::test]
+    async fn run_tool_call_loop_retries_text_prefixed_invalid_tool_call_markup() {
+        let runs = Arc::new(AtomicUsize::new(0));
+        let tools_registry: Vec<Box<dyn Tool>> = vec![Box::new(CountingShellTool {
+            runs: Arc::clone(&runs),
+        })];
+
+        let mut history = vec![
+            ChatMessage::system("sys"),
+            ChatMessage::user("set schedule"),
+        ];
+
+        let response = run_tool_call_loop(
+            &TextPrefixedMalformedThenValidToolProvider,
+            &mut history,
+            &tools_registry,
+            &NoopObserver,
+            "test-provider",
+            "test-model",
+            0.0,
+            true,
+        )
+        .await
+        .unwrap();
+
+        assert_eq!(response, "Scheduled successfully.");
+        assert_eq!(runs.load(Ordering::SeqCst), 1);
+        assert!(!response.contains("<tool_call>"));
+        assert!(history
+            .iter()
+            .any(|m| m.role == "user" && m.content.contains("[Tool parser error]")));
+    }
+
     struct PrefixMalformedThenValidToolProvider;
 
     #[async_trait]