[Bug]: Edit silently corrupts non-UTF-8 bytes outside the replacement target #3320

Prerequisites

• I searched existing issues and pull requests for encoding, UTF-8, Edit and corruption.
• This is a reproducible bug, not a usage question.

Background / Description

Editing an ASCII setting in a Latin-1 file silently corrupts unrelated text. Both direct Edit and Read followed by Edit with an AgentState report success, but replace the original 0xe9 byte with UTF-8 0xef 0xbf 0xbd (U+FFFD).

Read decodes with errors="replace" and caches that lossy text. Edit then writes the cache as UTF-8. Its no-state path also uses replacement decoding before writing.

Error Messages

There is no error. Both paths return state="running":

cached False edit running bytes b"name = 'caf\xef\xbf\xbd'\ncount = 2\n"
read running "     1\tname = 'caf\ufffd'\n     2\tcount = 1"
cached True edit running bytes b"name = 'caf\xef\xbf\xbd'\ncount = 2\n"

Steps to Reproduce

Run this against current main; it uses only temporary local files and the public tools, with no model or network requests:

import asyncio
import tempfile
from pathlib import Path

from agentscope.state import AgentState
from agentscope.tool import Edit, Read


async def main():
    with tempfile.TemporaryDirectory() as directory:
        path = Path(directory) / "config.py"
        original = b"name = 'caf\xe9'\ncount = 1\n"
        for cached in (False, True):
            path.write_bytes(original)
            state = AgentState() if cached else None
            if state is not None:
                read = await Read()(file_path=str(path), _agent_state=state)
                print("read", read.state, ascii(read.content[0].text))
            result = await Edit()(
                file_path=str(path),
                old_string="count = 1",
                new_string="count = 2",
                _agent_state=state,
            )
            print("cached", cached, "edit", result.state, "bytes", path.read_bytes())


asyncio.run(main())

Expected Behavior / Proposed Fix

An edit must not silently replace bytes outside the target. Without an explicit encoding option, rejecting non-UTF-8 text before caching/writing seems safer than guessing an encoding. A narrow fix would use strict UTF-8 decoding in Read and the no-state Edit path, letting their existing error handling return an error while preserving the original bytes. Alternatively, read-only previews could remain lossy, but those previews would need to be excluded from the edit cache.

Environment

• AgentScope: 2.0.11.dev0, main 9b94a63407ec59881fde19300b61e469a0a7d8b2
• Python: 3.12.8
• OS: Windows

Reproduced with Codex assistance using the actual local backend, not mocked I/O.