[Bug]: Read numbers lines with str.splitlines(), so a form feed or U+2028 shifts its line numbers away from Grep #3212

Prerequisites

• I have searched the existing issues and discussions, and this is not a duplicate.
• This is a bug, not a usage question. (For questions, please use Discussions instead.)

Background / Description

Read builds its numbered lines with str.splitlines(keepends=True) (src/agentscope/tool/_builtin/_read.py:522). After _normalize_newlines, splitlines() still breaks lines at \v, \f, \x1c, \x1d, \x1e, \x85, \u2028 and \u2029. Grep (rg -n), sed -n, wc -l and editors count only \n.

So in a file that contains one of those characters, every line after it has a different number in Read than in Grep:

• A form feed is a common page break in older Python, C and Emacs-style sources. Grep reports 5: return 2 # needle, while Read(offset=5) returns def second():.
• A \u2028 inside a string literal is legal in JS since ES2019 and also shows up in pasted text. Read shows that one line as two numbered lines. If the agent then copies those lines into Edit's old_string, joining them with \n, Edit answers old_string not found.

Expected: Read numbers lines the way Grep and the shell do, so an offset taken from Grep output, or a line copied from Read output, refers to the same text.

Write's line count (#2734) and the read cache that Write and Edit refresh use splitlines() too (_write.py:304,309, _edit.py:412), so they would need the same rule to stay consistent with Read.

Error Messages

Output of the script below on main (the form feed and \u2028 are shown escaped):

wc -l: 5
Grep: 5:    return 2  # needle
sed -n 5p:     return 2  # needle
Read offset=5: '     5\tdef second():'
Read whole file:
     1	def first():
     2	    return 1
     3	\f
     4	
     5	def second():
     6	    return 2  # needle
Read strings.js:
     1	const msg = "first\u2028
     2	second";
     3	const other = 1;
Edit with the lines as Read showed them: error | Error: old_string not found in /tmp/tmpXXXX/strings.js

Steps to Reproduce

1. Code:

import asyncio
import os
import subprocess
import tempfile

from agentscope.tool import Edit, Grep, Read


async def main() -> None:
    with tempfile.TemporaryDirectory() as tmp:
        path = os.path.join(tmp, "report.py")
        # A form feed (page break) between two sections.
        content = "def first():\n    return 1\n\f\ndef second():\n    return 2  # needle\n"
        with open(path, "w", encoding="utf-8", newline="") as f:
            f.write(content)
        print("wc -l:", subprocess.run(["wc", "-l", path], capture_output=True, text=True).stdout.split()[0])

        grep = await Grep()(pattern="needle", path=path, output_mode="content", n=True)
        print("Grep:", grep.content[0].text.strip().splitlines()[-1])
        print("sed -n 5p:", subprocess.run(["sed", "-n", "5p", path], capture_output=True, text=True).stdout.rstrip())

        read5 = await Read()(file_path=path, offset=5, limit=1)
        print("Read offset=5:", repr(read5.content[0].text))
        read_all = await Read()(file_path=path)
        print("Read whole file:")
        print(read_all.content[0].text)

        # A line holding U+2028 inside a string literal.
        path2 = os.path.join(tmp, "strings.js")
        with open(path2, "w", encoding="utf-8", newline="") as f:
            f.write('const msg = "first\u2028second";\nconst other = 1;\n')
        read_js = await Read()(file_path=path2)
        print("Read strings.js:")
        print(read_js.content[0].text)
        edit = await Edit()(file_path=path2, old_string='const msg = "first\nsecond";', new_string='const msg = "x";')
        print("Edit with the lines as Read showed them:", edit.state, "|", edit.content[0].text[:80])


asyncio.run(main())

2. Run: python repro.py
3. See the output above.

A possible fix would be a small helper next to _normalize_newlines that splits only on \n and keeps the line endings. Read, Write and Edit would use it in place of splitlines(). The existing CRLF/CR normalization would stay as it is. I'm happy to open a PR with tests if this direction works for you.

Environment

• AgentScope Version: 2.0.10dev (main at a833ab9)
• Python Version: 3.11.15
• OS: Linux