fix(sdk): read_file pagination skipping lines after wrapping (#3641)

Closes #2453 --- `read_file` applies row-based truncation twice: once when the backend slices raw lines by `offset`/`limit`, then again after `format_content_with_line_numbers` wraps long lines into continuation rows. When a wrapped line inflates the row count past `limit`, real source lines fall off the page — and the next page resumes from `offset + limit`, skipping them entirely. `_truncate` now takes the row-based limit as optional. The new-style `ReadResult` path (the backend has already paginated by source lines) drops the row chop and keeps only the byte-budget cap. The legacy `str` backend path keeps the row-based behavior, since those backends still format with line numbers inline. `limit` now bounds *source* lines, not formatted rows — a wrapped 15k-char line with `limit=3` renders all of its continuation chunks alongside the surrounding source lines instead of silently dropping the tail. The `read_file` tool description is updated to match. Carries forward the fix originally proposed in #2452 by @farrimoh; that PR was closed and could not be reopened, so the patch was rewritten against current `main` with farrimoh credited as a co-author on the commit. --------- Co-authored-by: farrimoh <farri.mohajerani@gmail.com>
2026-07-22 01:35:28 -04:00 · 2026-05-27 17:02:16 -04:00
parent f87d61f01a
commit 390551d61d
7 changed files with 104 additions and 78 deletions
@@ -350,7 +350,7 @@ Usage:
    - Only omit limit (read full file) when necessary for editing
 - Specify offset and limit: read_file(file_path="...", offset=0, limit=100) reads first 100 lines
 - Results are returned using cat -n format, with line numbers starting at 1
- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). When you specify a limit, these continuation lines count towards the limit.
+- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). `limit` applies to source lines, so continuation rows do not consume the budget.
 - You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.
 - If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.
 - Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).
@@ -844,11 +844,11 @@ class FilesystemMiddleware(AgentMiddleware[FilesystemState, ContextT, ResponseT]
        tool_description = self._custom_tool_descriptions.get("read_file") or READ_FILE_TOOL_DESCRIPTION
        token_limit = self._tool_token_limit_before_evict

-        def _truncate(content: str, file_path: str, limit: int) -> str:
-            lines = content.splitlines(keepends=True)
-            if len(lines) > limit:
-                lines = lines[:limit]
-                content = "".join(lines)
+        def _truncate(content: str, file_path: str, *, line_limit: int | None = None) -> str:
+            if line_limit is not None:
+                lines = content.splitlines(keepends=True)
+                if len(lines) > line_limit:
+                    content = "".join(lines[:line_limit])

            if token_limit and len(content) >= NUM_CHARS_PER_TOKEN * token_limit:
                truncation_msg = READ_FILE_TRUNCATION_MSG.format(file_path=file_path)
@@ -877,7 +877,7 @@ class FilesystemMiddleware(AgentMiddleware[FilesystemState, ContextT, ResponseT]
                )
                # Legacy backends already format with line numbers
                return ToolMessage(
-                    content=_truncate(read_result, validated_path, limit),
+                    content=_truncate(read_result, validated_path, line_limit=limit),
                    name="read_file",
                    tool_call_id=tool_call_id,
                    status="success",
@@ -922,10 +922,11 @@ class FilesystemMiddleware(AgentMiddleware[FilesystemState, ContextT, ResponseT]
                )

            content = format_content_with_line_numbers(content, start_line=offset + 1)
-            # We apply truncation again after formatting content as continuation lines
-            # can increase line count
+            # `limit` already bounded raw source lines at the backend; do not
+            # re-truncate by row count here, or wrapped continuation rows would
+            # push real source lines off the end of the page (#2453).
            return ToolMessage(
-                content=_truncate(content, validated_path, limit),
+                content=_truncate(content, validated_path),
                name="read_file",
                tool_call_id=tool_call_id,
                status="success",
@@ -59,7 +59,7 @@
  },
  {
    "function": {
-      "description": "Reads a file from the filesystem.\n\nAssume this tool is able to read all files. If the User provides a path to a file assume that path is valid. It is okay to read a file that does not exist; an error will be returned.\n\nUsage:\n- By default, it reads up to 100 lines starting from the beginning of the file\n- **IMPORTANT for large files and codebase exploration**: Use pagination with offset and limit parameters to avoid context overflow\n    - First scan: read_file(file_path=\"...\", limit=100) to see file structure\n    - Read more sections: read_file(file_path=\"...\", offset=100, limit=200) for next 200 lines\n    - Only omit limit (read full file) when necessary for editing\n- Specify offset and limit: read_file(file_path=\"...\", offset=0, limit=100) reads first 100 lines\n- Results are returned using cat -n format, with line numbers starting at 1\n- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). When you specify a limit, these continuation lines count towards the limit.\n- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.\n- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.\n- Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).\n\nFor multimodal reads (image, audio, video, PDF, etc.):\n- Use `read_file(file_path=...)`\n- Do NOT use `offset`/`limit` for images (pagination is text-only)\n- If file details were compacted from history, call `read_file` again on the same path\n\n- You should ALWAYS make sure a file has been read before editing it.",
+      "description": "Reads a file from the filesystem.\n\nAssume this tool is able to read all files. If the User provides a path to a file assume that path is valid. It is okay to read a file that does not exist; an error will be returned.\n\nUsage:\n- By default, it reads up to 100 lines starting from the beginning of the file\n- **IMPORTANT for large files and codebase exploration**: Use pagination with offset and limit parameters to avoid context overflow\n    - First scan: read_file(file_path=\"...\", limit=100) to see file structure\n    - Read more sections: read_file(file_path=\"...\", offset=100, limit=200) for next 200 lines\n    - Only omit limit (read full file) when necessary for editing\n- Specify offset and limit: read_file(file_path=\"...\", offset=0, limit=100) reads first 100 lines\n- Results are returned using cat -n format, with line numbers starting at 1\n- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). `limit` applies to source lines, so continuation rows do not consume the budget.\n- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.\n- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.\n- Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).\n\nFor multimodal reads (image, audio, video, PDF, etc.):\n- Use `read_file(file_path=...)`\n- Do NOT use `offset`/`limit` for images (pagination is text-only)\n- If file details were compacted from history, call `read_file` again on the same path\n\n- You should ALWAYS make sure a file has been read before editing it.",
      "name": "read_file",
      "parameters": {
        "properties": {
@@ -59,7 +59,7 @@
  },
  {
    "function": {
-      "description": "Reads a file from the filesystem.\n\nAssume this tool is able to read all files. If the User provides a path to a file assume that path is valid. It is okay to read a file that does not exist; an error will be returned.\n\nUsage:\n- By default, it reads up to 100 lines starting from the beginning of the file\n- **IMPORTANT for large files and codebase exploration**: Use pagination with offset and limit parameters to avoid context overflow\n    - First scan: read_file(file_path=\"...\", limit=100) to see file structure\n    - Read more sections: read_file(file_path=\"...\", offset=100, limit=200) for next 200 lines\n    - Only omit limit (read full file) when necessary for editing\n- Specify offset and limit: read_file(file_path=\"...\", offset=0, limit=100) reads first 100 lines\n- Results are returned using cat -n format, with line numbers starting at 1\n- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). When you specify a limit, these continuation lines count towards the limit.\n- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.\n- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.\n- Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).\n\nFor multimodal reads (image, audio, video, PDF, etc.):\n- Use `read_file(file_path=...)`\n- Do NOT use `offset`/`limit` for images (pagination is text-only)\n- If file details were compacted from history, call `read_file` again on the same path\n\n- You should ALWAYS make sure a file has been read before editing it.",
+      "description": "Reads a file from the filesystem.\n\nAssume this tool is able to read all files. If the User provides a path to a file assume that path is valid. It is okay to read a file that does not exist; an error will be returned.\n\nUsage:\n- By default, it reads up to 100 lines starting from the beginning of the file\n- **IMPORTANT for large files and codebase exploration**: Use pagination with offset and limit parameters to avoid context overflow\n    - First scan: read_file(file_path=\"...\", limit=100) to see file structure\n    - Read more sections: read_file(file_path=\"...\", offset=100, limit=200) for next 200 lines\n    - Only omit limit (read full file) when necessary for editing\n- Specify offset and limit: read_file(file_path=\"...\", offset=0, limit=100) reads first 100 lines\n- Results are returned using cat -n format, with line numbers starting at 1\n- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). `limit` applies to source lines, so continuation rows do not consume the budget.\n- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.\n- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.\n- Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).\n\nFor multimodal reads (image, audio, video, PDF, etc.):\n- Use `read_file(file_path=...)`\n- Do NOT use `offset`/`limit` for images (pagination is text-only)\n- If file details were compacted from history, call `read_file` again on the same path\n\n- You should ALWAYS make sure a file has been read before editing it.",
      "name": "read_file",
      "parameters": {
        "properties": {
@@ -59,7 +59,7 @@
  },
  {
    "function": {
-      "description": "Reads a file from the filesystem.\n\nAssume this tool is able to read all files. If the User provides a path to a file assume that path is valid. It is okay to read a file that does not exist; an error will be returned.\n\nUsage:\n- By default, it reads up to 100 lines starting from the beginning of the file\n- **IMPORTANT for large files and codebase exploration**: Use pagination with offset and limit parameters to avoid context overflow\n    - First scan: read_file(file_path=\"...\", limit=100) to see file structure\n    - Read more sections: read_file(file_path=\"...\", offset=100, limit=200) for next 200 lines\n    - Only omit limit (read full file) when necessary for editing\n- Specify offset and limit: read_file(file_path=\"...\", offset=0, limit=100) reads first 100 lines\n- Results are returned using cat -n format, with line numbers starting at 1\n- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). When you specify a limit, these continuation lines count towards the limit.\n- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.\n- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.\n- Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).\n\nFor multimodal reads (image, audio, video, PDF, etc.):\n- Use `read_file(file_path=...)`\n- Do NOT use `offset`/`limit` for images (pagination is text-only)\n- If file details were compacted from history, call `read_file` again on the same path\n\n- You should ALWAYS make sure a file has been read before editing it.",
+      "description": "Reads a file from the filesystem.\n\nAssume this tool is able to read all files. If the User provides a path to a file assume that path is valid. It is okay to read a file that does not exist; an error will be returned.\n\nUsage:\n- By default, it reads up to 100 lines starting from the beginning of the file\n- **IMPORTANT for large files and codebase exploration**: Use pagination with offset and limit parameters to avoid context overflow\n    - First scan: read_file(file_path=\"...\", limit=100) to see file structure\n    - Read more sections: read_file(file_path=\"...\", offset=100, limit=200) for next 200 lines\n    - Only omit limit (read full file) when necessary for editing\n- Specify offset and limit: read_file(file_path=\"...\", offset=0, limit=100) reads first 100 lines\n- Results are returned using cat -n format, with line numbers starting at 1\n- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). `limit` applies to source lines, so continuation rows do not consume the budget.\n- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.\n- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.\n- Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).\n\nFor multimodal reads (image, audio, video, PDF, etc.):\n- Use `read_file(file_path=...)`\n- Do NOT use `offset`/`limit` for images (pagination is text-only)\n- If file details were compacted from history, call `read_file` again on the same path\n\n- You should ALWAYS make sure a file has been read before editing it.",
      "name": "read_file",
      "parameters": {
        "properties": {
@@ -59,7 +59,7 @@
  },
  {
    "function": {
-      "description": "Reads a file from the filesystem.\n\nAssume this tool is able to read all files. If the User provides a path to a file assume that path is valid. It is okay to read a file that does not exist; an error will be returned.\n\nUsage:\n- By default, it reads up to 100 lines starting from the beginning of the file\n- **IMPORTANT for large files and codebase exploration**: Use pagination with offset and limit parameters to avoid context overflow\n    - First scan: read_file(file_path=\"...\", limit=100) to see file structure\n    - Read more sections: read_file(file_path=\"...\", offset=100, limit=200) for next 200 lines\n    - Only omit limit (read full file) when necessary for editing\n- Specify offset and limit: read_file(file_path=\"...\", offset=0, limit=100) reads first 100 lines\n- Results are returned using cat -n format, with line numbers starting at 1\n- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). When you specify a limit, these continuation lines count towards the limit.\n- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.\n- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.\n- Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).\n\nFor multimodal reads (image, audio, video, PDF, etc.):\n- Use `read_file(file_path=...)`\n- Do NOT use `offset`/`limit` for images (pagination is text-only)\n- If file details were compacted from history, call `read_file` again on the same path\n\n- You should ALWAYS make sure a file has been read before editing it.",
+      "description": "Reads a file from the filesystem.\n\nAssume this tool is able to read all files. If the User provides a path to a file assume that path is valid. It is okay to read a file that does not exist; an error will be returned.\n\nUsage:\n- By default, it reads up to 100 lines starting from the beginning of the file\n- **IMPORTANT for large files and codebase exploration**: Use pagination with offset and limit parameters to avoid context overflow\n    - First scan: read_file(file_path=\"...\", limit=100) to see file structure\n    - Read more sections: read_file(file_path=\"...\", offset=100, limit=200) for next 200 lines\n    - Only omit limit (read full file) when necessary for editing\n- Specify offset and limit: read_file(file_path=\"...\", offset=0, limit=100) reads first 100 lines\n- Results are returned using cat -n format, with line numbers starting at 1\n- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). `limit` applies to source lines, so continuation rows do not consume the budget.\n- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.\n- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.\n- Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).\n\nFor multimodal reads (image, audio, video, PDF, etc.):\n- Use `read_file(file_path=...)`\n- Do NOT use `offset`/`limit` for images (pagination is text-only)\n- If file details were compacted from history, call `read_file` again on the same path\n\n- You should ALWAYS make sure a file has been read before editing it.",
      "name": "read_file",
      "parameters": {
        "properties": {
@@ -59,7 +59,7 @@
  },
  {
    "function": {
-      "description": "Reads a file from the filesystem.\n\nAssume this tool is able to read all files. If the User provides a path to a file assume that path is valid. It is okay to read a file that does not exist; an error will be returned.\n\nUsage:\n- By default, it reads up to 100 lines starting from the beginning of the file\n- **IMPORTANT for large files and codebase exploration**: Use pagination with offset and limit parameters to avoid context overflow\n    - First scan: read_file(file_path=\"...\", limit=100) to see file structure\n    - Read more sections: read_file(file_path=\"...\", offset=100, limit=200) for next 200 lines\n    - Only omit limit (read full file) when necessary for editing\n- Specify offset and limit: read_file(file_path=\"...\", offset=0, limit=100) reads first 100 lines\n- Results are returned using cat -n format, with line numbers starting at 1\n- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). When you specify a limit, these continuation lines count towards the limit.\n- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.\n- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.\n- Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).\n\nFor multimodal reads (image, audio, video, PDF, etc.):\n- Use `read_file(file_path=...)`\n- Do NOT use `offset`/`limit` for images (pagination is text-only)\n- If file details were compacted from history, call `read_file` again on the same path\n\n- You should ALWAYS make sure a file has been read before editing it.",
+      "description": "Reads a file from the filesystem.\n\nAssume this tool is able to read all files. If the User provides a path to a file assume that path is valid. It is okay to read a file that does not exist; an error will be returned.\n\nUsage:\n- By default, it reads up to 100 lines starting from the beginning of the file\n- **IMPORTANT for large files and codebase exploration**: Use pagination with offset and limit parameters to avoid context overflow\n    - First scan: read_file(file_path=\"...\", limit=100) to see file structure\n    - Read more sections: read_file(file_path=\"...\", offset=100, limit=200) for next 200 lines\n    - Only omit limit (read full file) when necessary for editing\n- Specify offset and limit: read_file(file_path=\"...\", offset=0, limit=100) reads first 100 lines\n- Results are returned using cat -n format, with line numbers starting at 1\n- Lines longer than 5,000 characters will be split into multiple lines with continuation markers (e.g., 5.1, 5.2, etc.). `limit` applies to source lines, so continuation rows do not consume the budget.\n- You have the capability to call multiple tools in a single response. It is always better to speculatively read multiple files as a batch that are potentially useful.\n- If you read a file that exists but has empty contents you will receive a system reminder warning in place of file contents.\n- Image files (`.png`, `.jpg`, `.jpeg`, `.gif`, `.webp`, etc.), audio and video files, and PDFs are returned as multimodal content blocks (see https://docs.langchain.com/oss/python/langchain/messages#multimodal).\n\nFor multimodal reads (image, audio, video, PDF, etc.):\n- Use `read_file(file_path=...)`\n- Do NOT use `offset`/`limit` for images (pagination is text-only)\n- If file details were compacted from history, call `read_file` again on the same path\n\n- You should ALWAYS make sure a file has been read before editing it.",
      "name": "read_file",
      "parameters": {
        "properties": {
@@ -417,28 +417,23 @@ class TestDeepAgentEndToEnd:
            assert len(result["messages"]) > 0

    def test_deep_agent_truncate_lines(self, tmp_path: Path, backend: BackendProtocol) -> None:
-        """Test line count limiting in read_file tool with very long lines."""
-        # Create a file with a very long line (18,000 chars) that will be split into continuation lines
-        # With MAX_LINE_LENGTH=5000, this becomes line 2, 2.1, 2.2, 2.3 (4 output lines for 1 logical line)
-        very_long_line = "x" * 18000  # 18,000 characters -> will split into 4 continuation lines (5k each)
-
-        # Add some normal lines before and after
+        """`limit` bounds source lines; wrapped continuations don't displace later lines."""
+        # 18k chars wraps into 4 rows (2, 2.1, 2.2, 2.3) but still counts as one
+        # source line against `limit`.
+        very_long_line = "x" * 18000
        lines = [
            "short line 0",
-            very_long_line,  # This becomes lines 2, 2.1, 2.2, 2.3 (4 output lines)
+            very_long_line,
            "short line 2",
            "short line 3",
            "short line 4",
        ]
        content = "\n".join(lines)

-        # Create backend and write file
-
        file_path = "/my_file"
        starter_files = prepopulate_file(backend, file_path, content)

-        # Create a fake model that calls read_file with limit=3
-        # This should return: line 1 (short line 0), line 2 (first chunk of very_long_line), line 2.1 (second chunk)
+        # `limit=3` source lines → lines 1, 2 (all 4 wrapped chunks), 3.
        model = FixedGenericFakeChatModel(
            messages=iter(
                [
@@ -460,42 +455,29 @@ class TestDeepAgentEndToEnd:
            )
        )

-        # Create agent with backend
        agent = create_deep_agent(model=model, backend=backend)

-        # Invoke the agent
        invoke_input: dict[str, Any] = {"messages": [HumanMessage(content=f"Read {file_path}")]}
        if starter_files:
            invoke_input["files"] = starter_files
        result = agent.invoke(invoke_input)

-        # Verify the agent executed correctly
        assert "messages" in result
-
-        # Get the tool message containing the file content
        tool_messages = [msg for msg in result["messages"] if msg.type == "tool"]
        assert len(tool_messages) > 0
-
        file_content = tool_messages[0].content

-        # Should have the first short line
        assert "short line 0" in file_content
-
-        # Should have the beginning of the very long line (line 2 with continuation)
-        assert "xxx" in file_content  # The very long line should be present
-
-        # Should NOT have the later short lines because the limit cuts off after 3 output lines
-        # (line 1, line 2, line 2.1)
-        assert "short line 2" not in file_content
+        assert "xxx" in file_content
+        # All four wrapped chunks of source line 2 render in order.
+        for marker in ("2\t", "2.1\t", "2.2\t", "2.3\t"):
+            assert marker in file_content, f"missing continuation marker {marker!r}"
+        # Source line 3 is the third source line and must be included.
+        assert "short line 2" in file_content
+        # Source lines 4 and 5 fall outside `limit=3`.
        assert "short line 3" not in file_content
        assert "short line 4" not in file_content

-        # Count actual lines in the output (excluding empty lines from formatting)
-        output_lines = [line for line in file_content.split("\n") if line.strip()]
-        # Should be at most 3 lines (the limit we specified)
-        # This includes continuation lines as separate lines
-        assert len(output_lines) <= 3
-
    def test_deep_agent_read_empty_file(self, tmp_path: Path, backend: BackendProtocol) -> None:
        """Test reading an empty file through the agent."""
        # Create backend and write empty file
@@ -1115,32 +1097,19 @@ class TestDeepAgentEndToEnd:
        assert len(file_content) < 85000

    def test_deep_agent_read_file_single_long_line_behavior(self, tmp_path: Path, backend: BackendProtocol) -> None:
-        """Test the behavior with a single very long line.
+        """`limit` bounds source lines, not formatted rows.

-        When a file has a single very long line (e.g., 85,000 chars), it gets split
-        into continuation markers (1, 1.1, 1.2, etc.) by format_content_with_line_numbers.
-
-        The current behavior:
-        - offset works on logical lines (before formatting)
-        - limit applies to formatted output lines (after continuation markers)
-        - This allows pagination through long lines by increasing limit
-        - Limitation: cannot use offset to skip within a long line
-
-        This test verifies:
-        1. A single long line with limit=1 returns only the first chunk (respects limit on formatted lines)
-        2. Size-based truncation applies if the formatted output exceeds threshold
+        When a source line is wider than `MAX_LINE_LENGTH`, every continuation
+        chunk for that line is rendered — `limit=1` returns the full set of
+        chunks rather than just the first one. The byte-budget guard still
+        clamps the result when the formatted output exceeds the size cap.
        """
-        # Create a file with a SINGLE very long line (no newlines)
-        # This will be split into ~17 continuation chunks (85000 / 5000)
+        # 85k characters in one line → 17 continuation chunks at 5k each.
        single_long_line = "x" * 85000

-        # Create backend and write file
-
        file_path = "/single_long_line.txt"
        starter_files = prepopulate_file(backend, file_path, single_long_line)

-        # Create a fake model that calls read_file with limit=1
-        # This should return just 1 formatted line (the first chunk of the long line)
        model = FixedGenericFakeChatModel(
            messages=iter(
                [
@@ -1162,32 +1131,88 @@ class TestDeepAgentEndToEnd:
            )
        )

-        # Create agent with backend
        agent = create_deep_agent(model=model, backend=backend)

-        # Invoke the agent
        invoke_input: dict[str, Any] = {"messages": [HumanMessage(content=f"Read {file_path}")]}
        if starter_files:
            invoke_input["files"] = starter_files
        result = agent.invoke(invoke_input)

-        # Verify the agent executed correctly
        assert "messages" in result
-
-        # Get the tool message containing the file content
        tool_messages = [msg for msg in result["messages"] if msg.type == "tool"]
        assert len(tool_messages) > 0
-
        file_content = tool_messages[0].content

-        # Verify behavior: with limit=1, we get only the first formatted line
-        # (not all continuation markers)
-        assert len(file_content) < 10000  # Only got first chunk (~5000 chars)
-        assert len(file_content.splitlines()) == 1  # Only 1 formatted line
-        assert "1.1" not in file_content  # No continuation markers (would need higher limit)
+        # `limit=1` (one source line) renders the wrapped chunks; size cap
+        # still trims when the formatted result exceeds the byte budget.
+        assert "1.1" in file_content
+        assert "Output was truncated due to size limits" in file_content
+        assert len(file_content) <= 80000

-        # To get more of the line, the model would need to increase limit, not offset
-        # E.g., read_file(offset=0, limit=20) would get first 20 formatted lines
+    def test_deep_agent_read_file_pagination_does_not_skip_wrapped_lines(self, tmp_path: Path, backend: BackendProtocol) -> None:
+        """Wrapped long lines must not displace later source lines across pagination.
+
+        Regression for #2453: previously `limit` re-truncated formatted output
+        after wrapping, so a 15k-char line on page 1 pushed `important
+        instruction` off the page, and page 2 resumed past it.
+        """
+        long_line = "x" * 15000
+        content = f"line1\n{long_line}\nimportant instruction\nline4"
+        file_path = "/wrapped.txt"
+        starter_files = prepopulate_file(backend, file_path, content)
+
+        model = FixedGenericFakeChatModel(
+            messages=iter(
+                [
+                    AIMessage(
+                        content="",
+                        tool_calls=[
+                            {
+                                "name": "read_file",
+                                "args": {"file_path": file_path, "offset": 0, "limit": 3},
+                                "id": "call_1",
+                                "type": "tool_call",
+                            }
+                        ],
+                    ),
+                    AIMessage(
+                        content="",
+                        tool_calls=[
+                            {
+                                "name": "read_file",
+                                "args": {"file_path": file_path, "offset": 3, "limit": 3},
+                                "id": "call_2",
+                                "type": "tool_call",
+                            }
+                        ],
+                    ),
+                    AIMessage(content="done"),
+                ]
+            )
+        )
+
+        agent = create_deep_agent(model=model, backend=backend)
+
+        invoke_input: dict[str, Any] = {"messages": [HumanMessage(content=f"Read {file_path}")]}
+        if starter_files:
+            invoke_input["files"] = starter_files
+        result = agent.invoke(invoke_input)
+
+        tool_messages = [msg for msg in result["messages"] if msg.type == "tool"]
+        assert len(tool_messages) == 2
+        combined = tool_messages[0].content + tool_messages[1].content
+        assert "important instruction" in combined
+        assert "line4" in combined
+        # All three continuation chunks of the wrapped line 2 must render in
+        # order, before `important instruction`, with nothing dropped at the
+        # page boundary.
+        for marker in ("2\t", "2.1\t", "2.2\t"):
+            assert marker in combined, f"missing continuation marker {marker!r}"
+        idx_first = combined.index("2\t")
+        idx_cont1 = combined.index("2.1\t")
+        idx_cont2 = combined.index("2.2\t")
+        idx_next = combined.index("important instruction")
+        assert idx_first < idx_cont1 < idx_cont2 < idx_next

    def test_read_large_single_line_file_returns_reasonable_size(self) -> None:
        """Test that read_file doesn't return excessive chars for a single-line file.