fix: use proportional safety reserve for context compression trigger - #2051
Open
xielixing wants to merge 1 commit into
Open
fix: use proportional safety reserve for context compression trigger#2051xielixing wants to merge 1 commit into
xielixing wants to merge 1 commit into
Conversation
Replace fixed 10K safety reserve with a 10% ratio of context window (minimum 10K floor for small windows). This makes compression trigger fire at a consistent ~65% ratio for large-context models (128K-256K) instead of the previous 67-71%, addressing the 'triggers too late' complaint in GCWing#1529. Closes GCWing#1529
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
Issue #1529 reports that context compression triggers too late, especially for large-context models. The root cause is a fixed 10,000-token safety reserve constant that doesn't scale with context window size.
For a 64K context window, compression fires at ~59% — reasonable. But for 128K it fires at 67%, 200K at 70%, and 256K at 71%. The larger the context, the later compression triggers, degrading conversation quality.
Fix
Replace the fixed AUTO_COMPRESSION_SAFETY_RESERVE_TOKENS = 10_000 constant with:
This makes compression trigger at a consistent ~65% ratio across all context window sizes (128K-256K), while keeping 64K behavior unchanged (the floor applies).
Validation
Closes #1529