dcdpr / jp / 17465383542

58%

push

github

refactor(llm, anthropic): Improve message structure and caching (#239)

The Anthropic provider now properly uses system content for instructions
and attachments instead of injecting them as user messages with fake
assistant acknowledgments. This change aligns with the best-practices as
described in the Anthropic docs[1].

Additionally, the previous implementation sometimes caused more than
four cache points to be used, which resulted in an API error from
Anthropic. This is now fixed.

[1]:
https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview

Signed-off-by: Jean Mertz <git@jeanmertz.com>

18 of 41 new or added lines in 1 file covered. (43.9%)

5175 of 12736 relevant lines covered (40.63%)

5.6 hits per line