alex/g3
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

to get anthropic provider more reliable with tokens

Browse Source
This commit is contained in:
Michael Neale
2025-10-22 09:47:24 +11:00
parent 758e255af8
commit 738c3ac53e
4 changed files with 323 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
crates/g3-core/src/lib.rs
View File

@@ -319,10 +319,26 @@ impl ContextWindow {
/// Update token usage from provider response
pub fn update_usage_from_response(&mut self, usage: &g3_providers::Usage) {
// Add the tokens from this response to our running total
// The usage.total_tokens represents tokens used in this single API call
self.used_tokens += usage.total_tokens;
self.cumulative_tokens += usage.total_tokens;
// The provider's usage represents the tokens used in the last API call
// We need to be smarter about how we update our running total
let old_used = self.used_tokens;
// If the provider's total is greater than our current count, use it as the authoritative value
// This handles cases where our estimation was off
if usage.total_tokens > self.used_tokens {
self.used_tokens = usage.total_tokens;
self.cumulative_tokens += usage.total_tokens - old_used;
} else {
// Otherwise, add the tokens from this response
self.used_tokens += usage.completion_tokens; // Add only the new completion tokens
self.cumulative_tokens += usage.completion_tokens;
}
info!(
"Updated token usage - was: {}, now: {} (provider reported: prompt={}, completion={}, total={})",
old_used, self.used_tokens, usage.prompt_tokens, usage.completion_tokens, usage.total_tokens
);
debug!(
"Added {} tokens from provider response (used: {}/{}, cumulative: {})",
@@ -429,8 +445,18 @@ Format this as a detailed but concise summary that can be used to resume the con
if current_percentage >= 50 {
let current_threshold = (current_percentage / 10) * 10; // Round down to nearest 10%
if current_threshold > self.last_thinning_percentage && current_threshold <= 80 {
info!(
"Context thinning triggered - usage: {}% ({}/{} tokens), threshold: {}%, last thinned at: {}%",
current_percentage,
self.used_tokens,
self.total_tokens,
current_threshold,
self.last_thinning_percentage
);
return true;
}
} else {
debug!("Context usage at {}% ({}/{} tokens) - no thinning needed", current_percentage, self.used_tokens, self.total_tokens);
}
false