[BUG] File Reads getting compressed

[harmony-ai-solutions]

Problem (one or two sentences)

When working on larger tasks which require a lot of read_file tool usage; Roo often complains about file contents being "compressed" by the read tool. I checked the config settings but couldn't find any option to enable read file compression. Also I couldn't find an Issue here for this problem. Sharing a Screenshot:

Context (who is affected and when)

Happens to me only on my dev setups, but occurs on both Windows and Linux machines. Usually when using Architect mode to plan features, but also during edit phases.

Reproduction steps

Environment: Linux, Windows, VSCode Extension
Mode: Architect; model-independent
Roo Version: 3.51.1
VSCode Version: 1.110.1

The issue just happens almost in every planning session for features when automatically fetching files into context. No explicit trigger recognizable

Expected result

Roo reads the full file or relevant parts via file tools

Actual result

Constant fetching of small snippets to workaround "compression"

Variations tried (optional)

Happened for the last 3 Weeks since I migrated from Cline to Roo on the regular.

App Version

3.51.1

API Provider (optional)

None

Model Used (optional)

Claude Sonnet 4.6, MiniMax M2.5, MiniMax M2.7

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote-v0]

I've analyzed this issue and here's what I found:

Root Cause

The "compressed" file content the model complains about is actually caused by the context condensation system. When a long conversation triggers condensation (automatic or manual), previously-read files are represented as "folded" summaries showing only function signatures and class declarations via tree-sitter, wrapped in <system-reminder> blocks like:

<system-reminder>
## File Context: src/utils/helpers.ts
1--15 | export function formatDate(...)
17--45 | export class DateHelper {...}
</system-reminder>

The model sees these condensed structural summaries and interprets them as "compressed" file reads from the read_file tool. It then tries to work around this perceived limitation by fetching small snippets repeatedly, which is inefficient and confusing.

Proposed Implementation Plan

Improve the folded file context blocks in src/core/condense/foldedFileContext.ts to include a clear explanatory note that helps the model understand what it is seeing:

1. Modify generateFoldedFileContext() in src/core/condense/foldedFileContext.ts to add a brief instruction at the top of each <system-reminder> block explaining:

   • This is a structural summary (signatures only) of a previously-read file, generated during context condensation
   • It is NOT compressed output from the read_file tool
   • The model should use read_file to get full file contents when implementation details are needed

2. The updated format would look like:

<system-reminder>
## File Context: src/utils/helpers.ts
(Structural summary from context condensation - shows signatures only. Use read_file to get full content when needed.)
1--15 | export function formatDate(...)
17--45 | export class DateHelper {...}
</system-reminder>

3. Update existing tests in src/core/condense/__tests__/foldedFileContext.spec.ts to verify the new explanatory text is present.

This approach is minimal and targeted -- it clarifies the nature of the folded content directly where the model encounters it, without changing any core condensation logic or read_file behavior.

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

[harmony-ai-solutions]

@roomote seems like a viable solution we could try, feel free to implement as proposed.

[roomote-v0]

I've seen the approval and will implement the proposed solution now -- adding explanatory text to folded file context blocks so models understand these are structural summaries from context condensation, not compressed read_file output. I'll open a PR to share progress shortly.

[roomote-v0]

Opened a draft PR with my attempt at addressing this: #12000

The change adds an explanatory note to each folded file context block so models understand these are structural summaries from context condensation, not compressed read_file output. Feedback welcome!