[BUG] v3.43.0 not condensing context automatically

[srulyt]

Problem (one or two sentences)

Version v3.43.0 is not condensing context automatically.

I have run multiple tasks in multiple modes and the context overflows to well over 100%. Roo them fails with an error. Stopping, manually condensing the context, and continuing works some of the time.

Context (who is affected and when)

This happens in any task that runs long enough to fill the context.

Reproduction steps

Environment

VS Code v1.108.2
Roo: v2.43.0
Provider: VS Code LM Api (Github copilot)
Context size: 128k
Model: Opus 4.5

This repeats itself in multiple agent modes across multiple tasks.
This began to happen after I upgraded to the latest version.

Expected result

Context should auto condense

Actual result

Context is never auto condensed

Variations tried (optional)

I tried lowering the condense threshold to 80% and it did not improve the issue.

App Version

v3.43.0

API Provider (optional)

None

Model Used (optional)

Opus 4.5

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

Date/time: 2026-01-26T11:21:33.433Z
Extension version: 3.43.0
Provider: vscode-lm
Model: copilot/claude-opus-4.5

Request Failed: 413 Request Entity Too Large

[roomote]

Proposed Implementation Plan

Thank you for reporting this issue. I have investigated the auto-condense mechanism and the VS Code LM API handler. Here is my analysis and proposed fix:

Root Cause Analysis

The auto-condense mechanism in willManageContext() triggers when:

contextPercent >= effectiveThreshold || prevContextTokens > allowedTokens

Where contextPercent = (100 * prevContextTokens) / contextWindow.

After tracing the code flow, I identified the likely issue:

1. Token counting returns 0 on failure: In the VS Code LM handler (src/api/providers/vscode-lm.ts:274-290), the internalCountTokens method returns 0 on any error:

   catch (error) {
       // ...
       return 0 // Fallback to prevent stream interruption
   }

2. Token counting may fail silently: The countTokens method on the VS Code LM client may fail or return 0 for certain models (like Opus 4.5 via GitHub Copilot), resulting in totalInputTokens being 0 for the usage calculation.

3. Context window may be correct, but token count is 0: If prevContextTokens is 0 or very low (due to failed token counting), then contextPercent would be near 0%, and auto-condense would never trigger even when the actual context is very large.

Proposed Solution

Option A: Add fallback token estimation when counting fails

Add a fallback character-based token estimation when the VS Code LM client's countTokens returns 0 or fails:

// In internalCountTokens(), add fallback estimation
if (tokenCount === 0 && text.length > 0) {
    // Fallback: estimate ~4 chars per token (conservative)
    tokenCount = Math.ceil((typeof text === 'string' ? text.length : extractTextCountFromMessage(text).length) / 4)
    console.warn('Roo Code <Language Model API>: Using fallback token estimation')
}

Option B: Add diagnostic logging to help identify the specific failure

Before implementing a fix, add detailed logging to identify exactly where the token counting is failing:

console.debug('VS Code LM token count:', {
    inputTokens: totalInputTokens,
    outputTokens: totalOutputTokens,
    maxInputTokens: client.maxInputTokens,
    modelId: client.id
})

Recommendation

I recommend implementing Option A with diagnostic logging:

1. Add fallback token estimation in internalCountTokens() when the count is 0
2. Add a warning log when fallback is used so users know it's an estimation
3. This ensures auto-condense can trigger even when the VS Code LM API's countTokens method fails or returns inaccurate values

Questions for Confirmation

1. Can you confirm whether you see any warnings in the Roo Code output channel related to "Token counting failed" or similar messages?
2. What exact model ID is shown in the Roo Code interface for "Opus 4.5"? (This would help verify if the model is correctly identified)

Please approve this plan and I will implement the fix.

[daniel-lxs]

Hey, have you tried rolling back to the previous version to see if the same prompts/tasks work there? Would help confirm if this is a regression in v3.43.0 specifically.

[srulyt]

Yes, v3.43.0 seems to almost never compact. v3.42.0 also has issues but it compacts some of the time. I have gone back further and see a similar pattern. The issue reproduces in older version but less frequently.

[daniel-lxs]

Thanks for testing that. Since it's gotten progressively worse across versions rather than being a sudden break in v3.43.0, this points to an underlying issue with token counting reliability on the VS Code LM API. We're looking into it now.