[Bug] Long-context agent (task) calls corrupt returned content: newlines/backslashes lost, mojibake injected

Summary

In a long-context session, an agent (task/subagent) call that asked it to inspect local files produced a corrupted multi-line code payload displayed directly in the chat: newlines and backslashes were dropped (content flattened to one line), and mojibake <課 was injected at the start. The file the agent tried to write (check_cfg.py) was never created.

Environment

• OS: Windows 10 (build 19045)
• Node: v24.18.0
• Locale: Chinese (code page 936 / GBK) - suspected relevant

Corruption pattern (shown natively in the chat, not a copy-paste artifact)

Expected:

import os, glob
d = r'C:\Users\Dell.config\opencode'
print(...) on separate lines...

Actually shown:

<課 import os, glob newline d = r'C:\Users\Dell.config\opencode' print(...) for root, dirs, files in os.walk(d): ... (all flattened, \ in path lost, <課 prefix).

1. \n newlines lost -> single line.
2. \ backslashes lost -> C:\Users\Dell.config\opencode became C:\Users\Dell.config\opencode.
3. Mojibake <課 injected before import.

check_cfg.py was never written (File not found), so failure is at content-generation/return stage, not runtime.

Correlation with context length

• Short-context earlier session: identical task (writing multi-line helper, e.g. mk_scpb.py generating scpb.ps1) succeeded with correct newlines/backslashes.
• Long-context session: failed; user observed interruptions become more frequent as context grows.

Log evidence

In ~/.local/share/opencode/log/opencode.log, the 課 codepoint is logged as U+FFFD replacement char (<�n), e.g.:

...pattern=sshpass|scpb|<�n|�n|...

That shows a UTF-8 decode failure on this codepoint in the logging/rendering pipeline. Combined with the dialog display, corruption appears to happen in streaming/assembly of agent tool-call arguments, likely interacting with the GBK (cp936) Windows locale.

Suggested investigation

• Check how task/subagent return payloads (especially large multi-line code with \n and ) are streamed/assembled into the displayed tool call.
• Verify UTF-8 handling for multibyte content on Windows with non-UTF-8 system codepage.
• The \n+\ byte loss + mojibake injection suggests byte-level truncation/transcoding, not a model artifact.