🤖 fix: prevent double newlines in file_edit_insert tool

[ammar-agent]

Problem

The file_edit_insert tool was adding an extra newline when content already ended with \n, causing failures on simple tasks like terminal-bench's hello-world where file content precision matters.

Example failure:

• Task: Create hello.txt with "Hello, world!\n"
• Agent: file_edit_insert(content="Hello, world!\n")
• Result: File contains "Hello, world!\n\n" (double newline!)
• Test: ❌ FAILED - Expected exactly one newline

Root Cause

The tool's implementation (lines 96-97):

const newLines = [...lines.slice(0, line_offset), content, ...lines.slice(line_offset)];
const newContent = newLines.join("\n");

When content is inserted as an array element and joined with "\n", it always gets a trailing newline. When content already includes "\n", this produces double newlines.

The tool was designed to add a trailing newline, but this behavior was:

1. Undocumented in the tool description
2. Unexpected by the agent (naturally includes \n when instructed to "make sure it ends in a newline")
3. Causing subtle bugs where exact file content matters

Solution

Smart newline detection - If content already ends with \n, strip it before joining to prevent doubling:

const contentEndsWithNewline = content.endsWith("\n");
const normalizedContent = contentEndsWithNewline ? content.slice(0, -1) : content;

This matches the agent's natural expectation - if they explicitly add \n, it should be preserved as-is without doubling.

Testing

Added 2 regression tests:

• ✅ Content with trailing newline (the hello-world case)
• ✅ Multiline content with trailing newline

All 14 tests pass (12 existing + 2 new).

Impact

Estimated: +5-10% accuracy on terminal-bench tasks

Direct fixes:

• hello-world task (was failing 1/2 tests, now should pass both)
• Likely fixes other tasks where exact file content matters

Indirect improvements:

• Agent won't avoid tool after encountering unexpected behavior
• Fewer edge case bugs in config files, data formats, etc.

Analysis Source

Full analysis: terminal-bench-results/analysis_run_18894357631.md

From run 18894357631:

• 40% accuracy (16/40 tasks)
• This bug identified as Priority 1 critical issue
• Several other tasks showed partial failures (5/7 tests, 4/5 tests) that may also benefit from this fix

Generated with cmux

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".