From 94825316dc4e6f8a537ecce0261d1016bb290545 Mon Sep 17 00:00:00 2001
From: ZhenYi <434836402@qq.com>
Date: Fri, 24 Apr 2026 13:17:06 +0800
Subject: [PATCH] fix(agent): extract JSON from model output even with leading
 text prefix

ReAct loop was terminating early when the model returned:
  [Agent ran through N steps...]
  {"thought": "...", "action": {...}}

The extract_json function only checked the string start or code fences.
Now scans for { or [ at non-word positions and uses depth-counting
to strip trailing text, allowing JSON buried anywhere in the response.
---
 libs/agent/react/loop_core.rs | 51 +++++++++++++++++++++++++++++++++--
 1 file changed, 49 insertions(+), 2 deletions(-)

diff --git a/libs/agent/react/loop_core.rs b/libs/agent/react/loop_core.rs
index 747c19e..d11afff 100644
--- a/libs/agent/react/loop_core.rs
+++ b/libs/agent/react/loop_core.rs
@@ -378,15 +378,21 @@ fn parse_react_response(content: &str) -> ParsedReActResponse {
     }
 }
 
-/// Extract the first JSON object or array from a string, handling markdown fences.
+/// Extract a JSON object or array from a string, even when wrapped in non-JSON text.
+/// Handles: raw JSON at start, JSON in code fences, JSON buried in text (e.g. after
+/// a prefix line like "[Agent ran through N steps]").
 fn extract_json(s: &str) -> Option<String> {
     let trimmed = s.trim();
+
+    // Direct match — starts with { or [
     if trimmed.starts_with('{') || trimmed.starts_with('[') {
         return Some(trimmed.to_string());
     }
+
+    // Code fence handling
     for line in trimmed.lines() {
         let line = line.trim();
-        if line.starts_with("```json") || line.starts_with("```") {
+        if line.starts_with("```json") || line == "```" {
             let mut buf = String::new();
             let mut found_start = false;
             for l in trimmed.lines() {
@@ -409,6 +415,47 @@ fn extract_json(s: &str) -> Option<String> {
             }
         }
     }
+
+    // Scan for JSON object/array buried in text (common with prefix lines).
+    // Find the first '{' or '[' that is NOT preceded by a word character,
+    // then try to parse from there (stripping trailing non-JSON text).
+    let chars: Vec<char> = trimmed.chars().collect();
+    for i in 0..chars.len() {
+        let c = chars[i];
+        if (c == '{' || c == '[') && i > 0 {
+            // Skip if preceded by a word character (would be part of a string value)
+            let prev = chars[i - 1];
+            if prev.is_alphanumeric() || prev == '_' || prev == '"' || prev == '\'' {
+                continue;
+            }
+            let candidate: String = chars[i..].iter().collect();
+            // Try full candidate first
+            if serde_json::from_str::<serde_json::Value>(&candidate).is_ok() {
+                return Some(candidate.trim_end().to_string());
+            }
+            // Try stripping trailing text (text after the JSON closing brace/bracket)
+            let mut depth = 0isize;
+            let mut in_string = false;
+            let mut escaped = false;
+            for (j, c) in candidate.char_indices() {
+                if escaped { escaped = false; continue; }
+                if c == '\\' { escaped = true; continue; }
+                if c == '"' { in_string = !in_string; continue; }
+                if in_string { continue; }
+                if c == '{' || c == '[' { depth += 1; }
+                if c == '}' || c == ']' { depth -= 1; }
+                if depth == 0 {
+                    // Found the end of the JSON value
+                    let json_end = j + c.len_utf8();
+                    let trimmed_candidate = &candidate[..json_end];
+                    if serde_json::from_str::<serde_json::Value>(trimmed_candidate).is_ok() {
+                        return Some(trimmed_candidate.to_string());
+                    }
+                }
+            }
+        }
+    }
+
     None
 }