[BUG] Roo Code intelligent context condensing triggering on brand new conversation

[dempsey-ryan]

Problem (one or two sentences)

After only 1 user message, and a very brief description of the problem and next steps, Roo Code gets stuck in a loop of "context condensed"

Context (who is affected and when)

In a brand new task, after asking the AI to do a fairly simple task, while devising a plan and shortly after reading a few files, it gets stuck in a loop of repeatedly condensing the context and not being able to respond.

This is my current settings for intelligent context condensing:

Reproduction steps

1. Environment/setup

• running VSCode 1.104.1 on Windows 11
• connected via SSH to an AWS m5a.xlarge VM running Ubuntu 24.04.2 LTS

2. tell Roo to do a simple task, e.g., "change the welcome page so it says 'hello coder'". After a brief explanation it will get stuck condensing context for no reason.

Expected result

The task starts and shows progress

Actual result

After briefly summarizing the task, the AI gets caught in a loop of repeatedly condensing the context

Variations tried (optional)

Tried on latest version, tried on older version (3.27.0), and tried multiple anthropic models (4.0 Sonnet, 3.7 Sonnet w/ thinking)

App Version

3.38.3

API Provider (optional)

Anthropic

Model Used (optional)

Claude 4.0 Sonnet

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[hannesrudolph]

@dempsey-ryan Can you tell me how large that file translations.json is?

[dempsey-ryan]

@hannesrudolph The two translation.json files are about 3000 lines each, both < 250 KiB. I haven't had issues with them before but there may be 10-15 more lines than the last time I used Roo with these files.

[hannesrudolph]

@hannesrudolph The two translation.json files are about 3000 lines each, both < 250 KiB. I haven't had issues with them before but there may be 10-15 more lines than the last time I used Roo with these files.

How many tokens are the files?

[hannesrudolph]

Also, would you be able to expand the header at the top of your task so I can better se the entirety of the first prompt?

[hannesrudolph]

Basically what I think is happening here from what I can see is that the condensing is now containing the first turn and the the last 3 turns instead of just the last 3 turns. If your first turn contains an @ mention to the file and you have it in the most recent read then I am guessing your context is being overwhelmed. I would advise not including large files in your initial call and instead point to them without @ mentioning them.