Repository navigation
Conversation
This branch has not been deployed
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Term2 marks later message prefixes for caching, so a changed system message can invalidate reuse of the unchanged tool definitions before it. Add a five-minute breakpoint to the last eligible function tool. Preserve caller-owned cache controls, the four-breakpoint allowance, TTL ordering and all tool schemas/behavior.
A native Claude Haiku 5.5 test used one cold, warm and changed-system request per arm with a synthetic ~20k-token tool prefix. On the changed-system request, the candidate read 19,934 cached tokens versus zero for control; estimated cost was $0.000216 versus $0.002508. This demonstrates conditional prefix reuse, not general task savings or Pi parity. Provider cache state and minimum prefix size still govern eligibility.
Validation: registry → installed SDK → loopback HTTP coverage; full provider suite 273 passed/5 skipped; fake-Codex E2E 18 passed; typecheck, lint, static and baseline gates pass. Independent verification accepted cache controls, TTL/breakpoint protections, failure/recovery and all six retained live request/usage captures. Original experimental worktree preserved.