Jun 28, 2026
3 links from the engineering internet.
llama.cpp adds a minicpm5 tool-call parser
the inference engine's b9833 build implements a peg parser for minicpm5 tool calls, alongside jinja min/max api fixes and xml tool-call grammar improvements. one of five server and backend builds the project tagged through the day.
litellm cuts v1.91.0-rc.1 with shared mcp oauth token handling
the llm gateway's release candidate reworks mcp oauth into a shared layer with challenge/response flows, expiry-aware caching, single-flight refresh and cross-replica sync, adds a requested_model label to prometheus spend metrics, and a headroom guardrail for message compression.
clickhouse ships a second 26.3 lts patch in five days as v26.3.16.16
the analytics database tags v26.3.16.16-lts, its second long-term-support patch on the 26.3 line within a week, with signed binaries, libraries, and source archives across architectures for teams pinned to the lts branch.