fix(llm): make Responses API cache hits visible; default example to auto#40
Merged
Merged
Conversation
openai_responses parsed only input/output tokens, so gpt-5.x via the Responses API always logged cache_read=0 even when the prefix was cached. Split cached_tokens out of input_tokens (mirroring openai_compat) so token stats are honest and total_input_tokens isn't double-counted.
review mode hangs an unattended/headless run waiting for input; auto is the right default for a copy-paste first run.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Two small fixes surfaced while dogfooding algotune_knn end-to-end.
fix(llm): report Responses API cached_tokens —
openai_responsesparsed only input/output tokens, so gpt-5.x via the Responses API always loggedcache_read=0even when the prefix was cached. Now splitscached_tokensout ofinput_tokens(mirroringopenai_compat), so token stats are honest andtotal_input_tokensisn't double-counted. Verified by unit-checking the parse; KV cache itself confirmed ~99% read-hit on the anthropic path.docs(examples): default algotune_knn to auto —
reviewhangs an unattended/headless first run waiting for input;autois the right default for a copy-paste run.