fix: Context Overload in Read File Tool

[liwilliam2021]

Fix Context Overload in Read File Tool

Problem

The read_file tool could attempt to read files that exceed the available context window, causing API failures and poor user experience.

Solution

Added preemptive context validation that dynamically calculates available space and applies intelligent file size limits.

Changes

New contextValidator module:

• Calculates remaining context space based on current token usage
• Reads files incrementally in batches to count actual tokens
• Returns safe maximum line count with 25% safety buffer
  Enhanced readFileTool:
• Validates file size before reading when partial reads are off
• Provides inline instructions to model when files are truncated to use line_range syntax
• Respects user-specified limits when provided

Fix

Turn off partial reads, and open Roo-Code-Evals/javascript/lodash. Ask the following prompt from this task: https://app.roocode.com/share/fa7affcc-611d-46ff-89a4-d741d8593b2a

Remaining concerns:

• The model is now shown the tool line_range if a file cannot fit in context. This means that if it reads a very large file, then normal sized files, in theory it could continue to use the line_range tool to do unnecessary partial reads even when partial reads are deactivated.
• The 25% safety buffer is not guaranteed robust but empirically seems fine
• Fails if you try reading the E.coli document which is one big line
• Still fails if you try and force Roo to read the entirety of the bible. This seems to be a problem of inefficient context condensing.

Fixes #7744

---

Important

Introduces context validation in the Read File Tool to prevent context overload by dynamically calculating available space and applying file size limits.

• Behavior:
  • Adds contextValidator module to calculate available context space and apply file size limits.
  • Updates readFileTool to validate file size before reading and provide instructions for truncated files.
• Tests:
  • Adds contextValidator.test.ts to test context validation logic.
  • Updates readFileTool.spec.ts to test file reading with context validation.
  • Adds read-partial-content.spec.ts to test partial content reading.
• Localization:
  • Updates translation files for new context-related messages in multiple languages.
• Misc:
  • Adds providerFormat.spec.ts to test provider format logic.
  • Updates api.ts to include provider format mapping logic.

^{This description was created by for 2285015. You can customize this summary. It will automatically update as commits are pushed.}

[liwilliam2021]

@cte

[cte]

I think in general the strategy is good, but I left a few questions / comments. Separately, I'm going to test this in a few scenarios just to get a sense of how it behaves.

[daniel-lxs]

LGTM.
Adding a dynamic hard cap based on the context window would be a solid improvement for the future, but this is a good first step.

[mrubens]

I would really like to find a solution that doesn’t require special casing of minified files. Can we try reading lines in chunks instead of reading to the newline every time?

[liwilliam2021]

I would really like to find a solution that doesn’t require special casing of minified files. Can we try reading lines in chunks instead of reading to the newline every time?

resolved!

[Kilian-Collender]

Is there a plan to merge this soon?

[hannesrudolph]

Stale. Implementation not the one we want.