Context Condensing is disabled, but the context is shrinking all the time

[alex29ua]

App Version

3.27.0

API Provider

OpenAI

Model Used

gpt-5-chat-latest

Roo Code Task Links (Optional)

Even though I have disabled the Context Condensing, it seem to be happening all the time. My context size grows up to 100k-200k, and then suddenly becomes 15k in the same chat.
This happens all the time, and it's impossible to continue because apparently it just picks up one or two latest messages.
I don't see the "Context condensed" in the chat log when this happens.
So maybe it's not condensing, but just losing some (almost all) messages somehow?

🔁 Steps to Reproduce

just keep chatting

💥 Outcome Summary

I would expect the context grow until it reaches some hard limit or produces and API error response

📄 Relevant Logs or Errors (Optional)

[roomote]

I'm exploring a potential fix for the context condensing issue. I'll investigate why the context is shrinking even when condensing is disabled and will open a PR to share progress.

[roomote]

✅ Fix implemented and PR created: #7812

Summary

Successfully fixed the issue where context was being truncated even when Context Condensing was disabled (threshold set to 100%).

Root Cause

The truncateConversationIfNeeded function was applying sliding window truncation regardless of the user's intent to disable context management.

Solution

• Modified the logic to skip both summarization and truncation when threshold is 100%
• Added comprehensive test coverage to verify the fix
• All CI checks have passed

The fix ensures that when users explicitly disable context condensing by setting the threshold to 100%, their full conversation history is preserved as expected.

[alex29ua]

that's a nice automation here :)

but is there someone else experiencing the same issues at least?

[daniel-lxs]

Hey @alex29ua, I think you are experiencing the simple truncation. When you disable context condensing, you fall back to a middle-out truncation to prevent errors on the API. However I don't think this feature should cut down context by that much so something else might be going on here, maybe related to our new responses API implementation.