Merge PR #705: fix: detect, warn, and block file re-read/search loops after context compression

Authored by 0xbyt4. Adds read/search loop detection, file history injection after compression, and todo filtering for active items only.

tools/code_execution_tool.py

@@ -78,7 +78,7 @@ _TOOL_STUBS = {
     "web_extract": (
         "web_extract",
         "urls: list",
-        '"""Extract content from URLs. Returns dict with results list of {url, title, content, error}."""',
+        '"""Extract content from URLs. Returns dict with results list of {url, content, error}."""',
         '{"urls": urls}',
     ),
     "read_file": (
@@ -616,7 +616,7 @@ _TOOL_DOC_LINES = [
      "    Returns {\"data\": {\"web\": [{\"url\", \"title\", \"description\"}, ...]}}"),
     ("web_extract",
      "  web_extract(urls: list[str]) -> dict\n"
-     "    Returns {\"results\": [{\"url\", \"title\", \"content\", \"error\"}, ...]} where content is markdown"),
+     "    Returns {\"results\": [{\"url\", \"content\", \"error\"}, ...]} where content is markdown"),
     ("read_file",
      "  read_file(path: str, offset: int = 1, limit: int = 500) -> dict\n"
      "    Lines are 1-indexed. Returns {\"content\": \"...\", \"total_lines\": N}"),

tools/file_tools.py

@@ -14,6 +14,11 @@ logger = logging.getLogger(__name__)
 _file_ops_lock = threading.Lock()
 _file_ops_cache: dict = {}
 
+# Track files read per task to detect re-read loops after context compression.
+# Key: task_id, Value: dict mapping (path, offset, limit) -> read count
+_read_tracker_lock = threading.Lock()
+_read_tracker: dict = {}
+
 
 def _get_file_ops(task_id: str = "default") -> ShellFileOperations:
     """Get or create ShellFileOperations for a terminal environment.
@@ -132,11 +137,66 @@ def read_file_tool(path: str, offset: int = 1, limit: int = 500, task_id: str =
         result = file_ops.read_file(path, offset, limit)
         if result.content:
             result.content = redact_sensitive_text(result.content)
-        return json.dumps(result.to_dict(), ensure_ascii=False)
+        result_dict = result.to_dict()
+
+        # Track reads to detect re-read loops (e.g. after context compression)
+        read_key = (path, offset, limit)
+        with _read_tracker_lock:
+            task_reads = _read_tracker.setdefault(task_id, {})
+            task_reads[read_key] = task_reads.get(read_key, 0) + 1
+            count = task_reads[read_key]
+
+        if count >= 3:
+            # Hard block: stop returning content to break the loop
+            return json.dumps({
+                "error": (
+                    f"BLOCKED: You have read this exact file region {count} times. "
+                    "The content has NOT changed. You already have this information. "
+                    "STOP re-reading and proceed with your task."
+                ),
+                "path": path,
+                "already_read": count,
+            }, ensure_ascii=False)
+        elif count > 1:
+            result_dict["_warning"] = (
+                f"You have already read this exact file region {count} times in this session. "
+                "The content has not changed. Use the information you already have instead of re-reading. "
+                "If you are stuck in a loop, stop reading and proceed with writing or responding."
+            )
+
+        return json.dumps(result_dict, ensure_ascii=False)
     except Exception as e:
         return json.dumps({"error": str(e)}, ensure_ascii=False)
 
 
+def get_read_files_summary(task_id: str = "default") -> list:
+    """Return a list of files read in this session for the given task.
+
+    Used by context compression to preserve file-read history across
+    compression boundaries.
+    """
+    with _read_tracker_lock:
+        task_reads = _read_tracker.get(task_id, {})
+        seen_paths = {}
+        for (path, offset, limit), count in task_reads.items():
+            if path not in seen_paths:
+                seen_paths[path] = []
+            seen_paths[path].append(f"lines {offset}-{offset + limit - 1}")
+        return [
+            {"path": p, "regions": regions}
+            for p, regions in sorted(seen_paths.items())
+        ]
+
+
+def clear_read_tracker(task_id: str = None):
+    """Clear the read tracker. Called when starting a new conversation."""
+    with _read_tracker_lock:
+        if task_id:
+            _read_tracker.pop(task_id, None)
+        else:
+            _read_tracker.clear()
+
+
 def write_file_tool(path: str, content: str, task_id: str = "default") -> str:
     """Write content to a file."""
     try:
@@ -185,6 +245,24 @@ def search_tool(pattern: str, target: str = "content", path: str = ".",
                 task_id: str = "default") -> str:
     """Search for content or files."""
     try:
+        # Track searches to detect repeated search loops
+        search_key = ("search", pattern, target, path, file_glob or "")
+        with _read_tracker_lock:
+            task_reads = _read_tracker.setdefault(task_id, {})
+            task_reads[search_key] = task_reads.get(search_key, 0) + 1
+            count = task_reads[search_key]
+
+        if count >= 3:
+            return json.dumps({
+                "error": (
+                    f"BLOCKED: You have run this exact search {count} times. "
+                    "The results have NOT changed. You already have this information. "
+                    "STOP re-searching and proceed with your task."
+                ),
+                "pattern": pattern,
+                "already_searched": count,
+            }, ensure_ascii=False)
+
         file_ops = _get_file_ops(task_id)
         result = file_ops.search(
             pattern=pattern, path=path, target=target, file_glob=file_glob,
@@ -195,6 +273,13 @@ def search_tool(pattern: str, target: str = "content", path: str = ".",
                 if hasattr(m, 'content') and m.content:
                     m.content = redact_sensitive_text(m.content)
         result_dict = result.to_dict()
+
+        if count > 1:
+            result_dict["_warning"] = (
+                f"You have run this exact search {count} times in this session. "
+                "The results have not changed. Use the information you already have."
+            )
+
         result_json = json.dumps(result_dict, ensure_ascii=False)
         # Hint when results were truncated — explicit next offset is clearer
         # than relying on the model to infer it from total_count vs match count.

tools/todo_tool.py

@@ -105,8 +105,17 @@ class TodoStore:
             "cancelled": "[~]",
         }
 
-        lines = ["[Your task list was preserved across context compression]"]
-        for item in self._items:
+        # Only inject pending/in_progress items — completed/cancelled ones
+        # cause the model to re-do finished work after compression.
+        active_items = [
+            item for item in self._items
+            if item["status"] in ("pending", "in_progress")
+        ]
+        if not active_items:
+            return None
+
+        lines = ["[Your active task list was preserved across context compression]"]
+        for item in active_items:
             marker = markers.get(item["status"], "[?]")
             lines.append(f"- {marker} {item['id']}. {item['content']} ({item['status']})")