jul 16, 2026
7 links from the engineering internet.
llama.cpp b10043 adds cuda virtual devices support
the cuda backend can now expose virtual devices, splitting one physical gpu into multiple logical ones. the build disables the nccl multi-gpu path while virtual devices are in use to avoid conflicts.
openai codex 0.144.5 hardens dangerous-command detection
the codex cli stable release widens its blocklist for risky shell commands, catching more forced rm variants, and gives clearer feedback when it refuses to run a command.
ollama 0.32.1 improves gemma 4 tool calling and fixes an mlx cache leak
the pre-release sharpens gemma 4 tool calling and multi-turn reasoning with more reliable tool-response continuations, and fixes a recurring mlx model cache leak that grew memory use across requests.
n8n 2.31.2 requires an execution mode for mcp workflow runs
the patch makes mcp workflow executions specify an execution mode instead of defaulting silently, and fixes a regression that changed the width of every dropdown in the editor ui.
ruff 0.15.22 adds noqa-modernizing preview rules and an e402 autofix
the release ships preview rules ruf105 and ruf106 for cleaning up noqa comments plus ruf201 for readable config selectors, adds an autofix for pycodestyle's e402, and speeds up lexing and parsing.
tokio 1.52.4 fixes a runtime driver skip when before_park schedules work
the patch fixes a runtime bug where the io driver could skip a cycle if the before_park hook scheduled new work, so that work no longer stalls until the next wakeup. it also cleans up taskdump output.
pnpm 11.13.1 fixes pnpm pack ignoring package-level .npmignore
the patch makes pnpm pack respect a package's own .npmignore instead of the workspace root rules, improves visibility of the minimumreleaseage approval prompt, and fixes self-update linking native binaries.