[BUG] Incremental file reading is broken across multiple providers

[pwilkin]

Problem (one or two sentences)

When introducing a context size limit to read_file, even very strong open models fail to read files correctly, instead often falling into a loop of calling read_file again and again.

Context (who is affected and when)

When there is a large file and there's a need to read it incrementally, models will often fail to properly use line ranges.

Reproduction steps

1. Create a task that requires understanding a large file
2. Set max size of file read to eg. 500 lines
3. Run the task

Expected result

The model reads the file with line ranges incrementally

Actual result

The model re-reads the start of the file again and again

Variations tried (optional)

No response

App Version

3.36.16

API Provider (optional)

Not Applicable / Other

Model Used (optional)

GLM 4.6, Minimax M2, Kimi K2-Instruct

Roo Code Task Links (optional)

https://app.roocode.com/share/3bd34071-609f-48fb-87bc-f9a2a353c5c6

Relevant logs or errors (optional)

[pwilkin]

The way I see it, one of two approaches would be worthwhile:
a) add a default "cursor" behavior to limited read_file, with consecutive reads providing consecutive fragments of the file, with an optional "reset" modifier
b) add stronger guidelines to the read_file tool when line limits are present, maybe with a one-shot example of how to proceed with reading the entire file with chunked reading

I've analyzed this issue and here's my proposed implementation plan:

Problem Analysis

When files are truncated (via maxReadFileLine setting or token budget), the truncation notices don't clearly tell models how to continue reading incrementally. The current messages are:

• "Showing only {X} of {Y} total lines. Use line_range if you need to read more lines"
• "File truncated: showing {X} lines ({N} tokens) due to context budget. Use line_range to read specific sections."

These messages don't specify:

1. Which line number to start from next (models don't know they should start at line X+1)
2. An example of the syntax to use for continuation

This causes models (especially non-Anthropic/OpenAI models like GLM 4.6, Minimax M2, Kimi K2-Instruct) to repeatedly re-read from line 1 instead of continuing from where they left off.

Proposed Solution

Enhance the truncation notice messages in src/core/tools/ReadFileTool.ts and src/core/tools/simpleReadFileTool.ts to include:

1. The exact next line number to continue from
2. A concrete example of the line_range syntax

Example of improved messages:

Before:

Showing only 500 of 2000 total lines. Use line_range if you need to read more lines

After:

Showing lines 1-500 of 2000 total lines. To continue reading, use line_range starting at line 501 (e.g., <line_range>501-1000</line_range>)

Files to modify:

1. src/core/tools/ReadFileTool.ts (lines 461, 496, 551)
2. src/core/tools/simpleReadFileTool.ts (line 237)
3. src/integrations/misc/extract-text.ts (line 101)

Implementation details:

• Include both the next start line and a syntax example in the truncation notice
• For native protocol, provide JSON-style example: line_ranges: [[501, 1000]]
• For XML protocol, provide XML-style example: <line_range>501-1000</line_range>

This approach directly addresses the issue by giving models explicit continuation instructions they can follow.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[pwilkin]

@roomote Looks reasonable, but the feedback should be even clearer, like "To continue reading, use the read_file tool again with the line_range parameter starting at line 501...".

Also, can you check if the simpleReadFileTool even HAS line-range support? Because if not, that could be the problem.

And for versatility, the tool should accept both a single range (with line_range) and multiple ranges (with line_ranges).

I've seen the feedback and will investigate. Here's what I've found so far:

Confirmed: simpleReadFileTool does NOT have line-range support

Looking at src/core/tools/simpleReadFileTool.ts, lines 30-34 explicitly state:

* This is a streamlined version of readFileTool that:
* - Only accepts a single path parameter
* - Does not support multiple files
* - Does not support line ranges  // <-- This is the problem!
* - Has simpler XML parsing

This is a significant issue since models using simpleReadFileTool receive truncation messages suggesting to "use line_range" but the tool cannot actually process that parameter.

I'll work on:

1. Adding line_range (single range) and line_ranges (multiple ranges) support to simpleReadFileTool
2. Improving truncation notices to include the exact next line number and clearer syntax examples (e.g., "To continue reading, use the read_file tool again with the line_range parameter starting at line 501...")

Will open a PR with my attempt at addressing this.

Opened a PR with my attempt at addressing this: #10240

Summary of changes:

1. Added line_range support to simpleReadFileTool - Models using the simple read_file tool can now use <line_range>start-end</line_range> for incremental reading
2. Improved truncation notices - Now include the exact next line number and a concrete syntax example (e.g., "To continue reading, use the read_file tool again with the line_range parameter starting at line 501...")

Feedback welcome!

[hannesrudolph]

@roomote the simple read file tool is only used for grok models isn't it?

Good catch! Looking at the code in packages/types/src/single-file-read-models.ts:

export function shouldUseSingleFileRead(modelId: string): boolean {
    return modelId.includes("grok-code-fast-1") || modelId.includes("code-supernova")
}

So simpleReadFileTool is used for:

1. Grok models (specifically grok-code-fast-1)
2. Code Supernova models

The original issue mentions GLM 4.6, Minimax M2, and Kimi K2-Instruct - none of which would use simpleReadFileTool. They use the regular ReadFileTool which already has line range support.

However, PR #10240 does include improvements to BOTH tools:

• simpleReadFileTool: Added line_range support (for Grok/Supernova models)
• ReadFileTool: Improved truncation notices with clearer continuation instructions

So the truncation notice improvements (showing the next line number and example syntax) should help all models, not just Grok/Supernova.

[pwilkin]

@roomote Which exactly parameter is used in native mode: line_range (singular) or line_ranges (plural)?

I feel like a lot of confusion stems from this and the differing syntax (XML: range string with format start-end, native: JSON array of JSON 2-element arrays of line ranges). The syntax should be unified, preferrably into the simple text-based one and only one parameter should be used. If needed, multiple line ranges can be specified with commas, eg. 100-500,700-1200, but I feel like the multiple line-range use case is very rare and shouldn't sabotage the much more typical cases.