Edit Tool issues with Grok Code Fast 1

[daevangelion]

Description

Sourced from my AGENTS.md after my failure analysis workflows did hexdumps of diffs and identified a common failure (30% of edits to python code files have tool control delimiters embedded in them)

Edit Tool Usage Guidelines

Delimiter Conflict Prevention

CRITICAL: The edit tool interprets certain XML/HTML-like tags as control delimiters. Including these tags in content can cause edit operations to fail or corrupt files.

Problematic Tags

The following tags are interpreted as edit tool delimiters and must be avoided in content:

• </content> - End of content delimiter
• </parameter - Parameter delimiter
• </edit> - Edit operation delimiter
• </file> - File delimiter

Safe Content Inclusion

If you need to include these patterns in code, use character transformation:

# Transform delimiters to safe placeholders
echo '[/content]' | tr '[]' '<>'    # Produces </content> safely
echo '[parameter]' | tr '[]' '<>'   # Produces <parameter> safely

OpenCode version

1.0.63

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

Arch Linux

Terminal

Headless

[rekram1-node]

can u show what ur talking about? this seems like the model is calling tools incorrectly?

[daevangelion]

yeah it may be, but the tool could do more to hint at the failure mode, often those tags are written to the file in question, the model gets confused because it can't see the change due to the contract hiding it, it then attempts to make multiple edits to correct an "invisible" problem, I have found that intercepting your tool and telling it to look at the hex of the file, or making a blind edit like the one recommended above, it self corrects.

[daevangelion]

from the session diffs:
" return VOBJ_DATA;\n\n\n(From `src/monitor/vregionctl.c`)</content>\n<parameter name=\"filePath\">docs/references/lib_reference.md"," J_DATA, ARCH_ANY);\n\n// Dump region info\nvregion_dump(r);\n\n<parameter name="filePath">docs/references/kernel_reference.md"

I'd probably just add a detector to identify that condition and as part of the response advise the model

[taqtiqa-mark]

@daevangelion and @rekram1-node I think this may be the same as #4530 .
If you agree, happy to track whichever you prefer.

[rekram1-node]

It does look related, Ill just leave both open for now