roo loads a file into context that is bigger than the model context window - followup error of autocondense

[deafsquad]

App Version

3.19.1

API Provider

Google Gemini

Model Used

Pro and flash for condensing

🔁 Steps to Reproduce

let the model request a file > than the contextwindow, in my case it was a logfile that roo requested after the process ran that it invoked. followup auto condensing then starts with a too big context and keeps looping.

💥 Outcome Summary

Expect roo to recognize if file > contextwindowsize then abort or log and cut off file or tail so it fits into the context window

📄 Relevant Logs or Errors (Optional)

[dosubot]

Hi @deafsquad! I'm Dosu and I’m helping the Roo-Code team.

Roo Code relies on large language models with a limited context window—the maximum amount of text the model can process at once. If a file (like a large logfile) exceeds this window, Roo may not handle it gracefully, which can lead to issues like the auto condensing loop you described. The context window includes the system prompt, conversation history, any files mentioned with @, and command/tool outputs. When the total context exceeds the model's limit, Roo's Intelligent Context Condensing feature tries to reduce the input, but if the initial file is too large, this process can get stuck or loop without resolving the overflow read more.

You can adjust how and when Intelligent Context Condensing triggers in the Roo Code settings under the "Context" section. There, you can set the threshold (as a percentage of the context window, default 100%), choose which API configuration to use for condensing, and customize the condensing prompt. Lowering the threshold or customizing the condensing prompt may help prevent runaway loops, but currently, there is no documented mechanism for Roo to automatically abort, truncate, or tail a file that is too large before it enters the context window details.

Recommended strategies include breaking large tasks into smaller ones, summarizing or only including essential parts of large files, and being selective with what you load into context. For logfiles or other large outputs, consider manually tailing or summarizing before referencing them in Roo see strategies.

If the context gets "poisoned" or stuck, starting a new chat session is the most reliable way to reset recovery advice.

I did not find documentation on specific codebase components that handle file loading or context window enforcement, nor any built-in feature that automatically truncates or aborts oversized files before they are loaded. If you need more granular control, manual intervention is currently the best approach.

To reply, just mention my name @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[deafsquad]

@dosu in case of a logfile you can not forsee if it will break the context window and the behaviour in this case is a never ending loop of the auto condensing which at least should have been avoided, it just needs an initial check if the context is already blowing the models capabilities and it stops after the first try. previous version had a note on this, i say it lacks a first check if its even worth starting if the context is already too big to send it.
[3.19.0] - 2025-05-29
Skip condense and show error if context grows during condensing

[dosubot]

You are correct that in version 3.19.0, Roo Code introduced smart error handling to skip condensing and show a clear error if the context grows during condensing. This feature is also present in version 3.19.1, as confirmed in the official release notes see release notes.

However, there is no documentation or release note indicating that Roo Code performs an initial check to prevent starting the condensing process if the context is already too large to fit in the model's context window. The current implementation only skips condensing and shows an error if the context grows during the condensing process itself, not before it starts details.

If you are running into a never-ending condensing loop with very large files (like logfiles), this means the initial pre-check you described is not yet implemented. The best available workaround is to manually limit or summarize large files before referencing them in Roo, as there is no built-in mechanism to automatically tail or truncate files to fit the context window.

To reply, just mention my name @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}   

[deafsquad]

@dosu that wont help in an autonomous application in case of logfiles, its also easy to just blow the context as you dont have a tokencounter in your view, you can achieve the same problem by manually adding a logfile, this was worse as it ran autonomous and if you dont have an eye on it it will run without any limits.

[dosubot]

You’re right—Roo Code does not currently have a built-in mechanism to automatically prevent context window overflows in autonomous scenarios, especially with unpredictable log file sizes. While Roo does have smarter context window management, chunked loading for large files, and a sophisticated token counting system, these features do not guarantee that an autonomous process won’t exceed the model’s context window or enter a condensing loop if a file is too large to begin with. The system only skips condensing and shows an error if the context grows during condensing, not before it starts, so a runaway loop is still possible if the initial context is oversized details.

There is a setting called “File read auto-truncate threshold” in Advanced Settings, which controls how many lines are read from a file in one batch. Lowering this value can help limit the amount of data loaded from large files like logs, but it’s not a foolproof solution for autonomous workflows, as it still requires some manual tuning and awareness of file sizes reference.

Roo Code does have a context window visibility feature (since v3.3.5) that shows token counts in the task header, but this may not be surfaced in all autonomous flows or UI contexts reference.

At this time, there’s no documented way to enforce a hard pre-check or automatic truncation/tailing of files before they are loaded into context in an autonomous process. The best available mitigation is to use the auto-truncate threshold and be as specific as possible with file references, but this does not fully solve the problem for unsupervised or automated runs.

If you need strict control, you may need to implement external log rotation, pre-processing, or file tailing outside of Roo, or use the Model Context Protocol (MCP) to integrate custom logic for file access reference.

If you have a specific workflow or integration in mind, I can help look for more targeted workarounds or settings. Let me know if you want details on MCP or other advanced options.

To reply, just mention my name @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}