[BUG] Strange context bloat and truncation

[JNevrly]

Problem (one or two sentences)

Lately I'm getting problems with strange context bloat that is not related to actual prompt (I have experienced it even on empty project folder and in the beginning of the prompt session). I have only noticed it by increase token consumption (=costs) and context truncation messages, which look like the truncated context is larger than the original context:

Context (who is affected and when)

I have experienced it in Architect and Code roles, during sessions and even in the beginning of the prompt session. I haven't yet figured out the pattern.

Reproduction steps

I have not yet figured out clear reproduction pattern.

Expected result

Actual result

Variations tried (optional)

No response

App Version

3.51.1 (7c9722b)

API Provider (optional)

OpenRouter

Model Used (optional)

openai/gpt5.1-codex, openai/gpt5.3-codex

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote-v0]

I've been looking into this issue and reviewing the context management, truncation, and token-counting code. A few things stood out that could help narrow down the root cause, but I need some additional details to move forward effectively.

Clarifying questions:

1. Screenshot details: In the screenshot, the truncation row shows a "before -> after" token count (e.g. X -> Y tokens). When you say "truncated context is larger than the original context," do you mean the number on the right (after truncation) is larger than the number on the left (before)? Could you share exact numbers if possible?

2. Context management settings: What are your current settings for:

   • "Auto Condense Context" (enabled/disabled?)
   • "Context Management Threshold" percentage (if set)
   • Any per-profile threshold overrides?

3. Session start behavior: You mention this happens "even in the beginning of the prompt session." Does this mean the very first API request in a new task already shows unexpectedly high token counts? Or does it happen after the first few exchanges?

4. Model context window: You're using openai/gpt5.1-codex and openai/gpt5.3-codex via OpenRouter -- do you know what context window size is being reported for these models? (You can check in the model info or the API response metadata.)

5. System prompt size: Do you have lengthy custom instructions, custom mode prompts, or many MCP servers configured? These all contribute to the system prompt size, which is included in every API call but might not be immediately visible in the UI.

6. Workspace size: Even for "empty project folders," does your workspace have a .git directory or other hidden files/directories? The environment details include a recursive file listing (up to maxWorkspaceFiles, default 200) that gets appended to every user message. If git status is enabled, that also adds to context.

Areas of investigation I've identified so far:

• The contextTokens metric (displayed in the UI and used for truncation decisions) is calculated as tokensIn + tokensOut from the last API response. Meanwhile, newContextTokensAfterTruncation is calculated by re-counting message content via tiktoken + system prompt. These two different counting methods could produce inconsistent before/after values in the truncation UI.

• The environment details block (workspace files, git status, terminal output, reminders, mode info, etc.) is appended to every user message and can be substantial even for small projects.

• If condensation fails silently and falls back to sliding-window truncation, the error may not be clearly surfaced.

Please tag @roomote in your response with the answers, and I'll be happy to investigate further and propose a concrete fix plan.

[JNevrly]

@roomote

1. Yes, the indicator shows before 251367, after 334680
2. Auto-condense context: disabled; context management threshold - don't see this setting anywhere. No per-profile override.
3. Yes, at times even the first prompt already shows Context Truncation symbol
4. Context window is reported 400000 tokens
5. I have 0 MCP servers, no custom mode prompts and short custom instructions. Moreover, I havent changed those for some time and this has started happenning only recently.
6. The problem appears even when there is no hiden folder or .git folder.

[roomote-v0]

Thanks for the detailed answers, @JNevrly. The key data point here is that the "after" count (334,680) is larger than the "before" count (251,367) -- truncation should never increase the token count. This points to an inconsistency in how we calculate and display the before/after token metrics in the truncation UI.

I'll dig into the token counting and truncation display logic now and open a PR with a fix attempt.

[roomote-v0]

Opened PR #11991 with an attempt at fixing the inconsistent before/after token counts in the truncation display. The root cause was that "before" used API-reported token counts while "after" used local estimation, which can produce wildly different values depending on the provider. The fix makes both use the same local counting method. Feedback welcome!