Doc(Clarification): Document File Truncation Limits When Sending Content to LLM

[dariuszkowalski-com]

After testing a prompt like:

<task>create xxx based on rules from @[PROJECT_STANDARDS.md]</task>

I noticed that PROJECT_STANDARDS.md (2696 lines) was truncated to 2000 lines when sent to anthropic/claude-sonnet-4-20250514:

<file_content path="/mnt/...md" start_line="1" end_line="2000" total_lines="2696" ... >

Questions

• Is this 2000‑line cutoff hardcoded in Forge, or is it a model limitation (Claude‑Sonnet‑4)?
• If this is expected behavior, could it be documented?

Why this matters

• Users may include large files (e.g. via repomix or similar tools).
• Silent truncation means the LLM won’t see the full context, leading to incorrect or incomplete results.
• Clear documentation would help users understand limitations and plan around them.

[tusharmath]

There are two ways LLMs can read the files -

1. Using tools that truncate to 2000 lines
2. Using an attachment ie ,@[file path] in the prompt.

Though truncated in the first case, LLMs get enough information about the total number of lines and how to read more if required. In the second case file is not truncated and sent as a whole.

[dariuszkowalski-com]

Though truncated in the first case, LLMs get enough information about the total number of lines and how to read more if required. In the second case file is not truncated and sent as a whole.

What is a good practice for working with long files?
I’m considering splitting one PROJECT_STANDARDS.md file into several smaller ones, each focused on a specific domain, and leaving only links with brief summaries in the original file. Would that help?

[tusharmath]

Global rules should be embedded in custom_rules. Module-specific information can be in the form of comments in source files. Consider using a memory MCP if you have a lot of instructions that need to be pulled in on a need-basis.