Weaviate v1.36.15 text2vec-google batch fix; llama.cpp Vulkan+OpenCL drops; Goose hooks system

ApiDelta · 2026-05-23 · 384 words · apidelta.maxiaworld.app

🚨 Breaking

Weaviate v1.36.15 ships a fix to text2vec-google batch logic (release notes). The changelog lists no formal breaking changes, but a behavioral fix inside a production embedding module's batch path warrants a staging smoke test before you promote — batch throughput and error rates may shift.

🗑️ Dépréciations

None in today's brief.

💰 Pricing

No pricing changes in today's brief.

🆕 Nouveautés

Weaviate v1.36.15 also adds a text2vec-digitalocean module — first-class DigitalOcean embedding support for vectorizer pipelines (release notes).

llama.cpp shipped multiple builds in the past 24 hours: - b9292: integer overflow fix in perplexity computation — relevant if you run eval harnesses (b9292) - b9279: Vulkan backend now fuses the snake activation (5-op decomposition emitted by BigVGAN/Vocos audio decoders) into a single shader, reducing dispatch overhead (b9279) - b9295: Vulkan find_package(SPIRV-Headers) broken on Windows — fixed (b9295) - b9294: OpenCL MoE kernels generalized for Adreno GPUs (b9294)

Goose v1.35.0 lands a hooks system for pre/post tool execution, a PreToolUse denial hook, and a /goal command for agent self-evaluation before task completion (v1.35.0). PreToolUse denial is the one to watch — it lets you intercept and block tool calls at the framework level without forking the agent.

Crush v0.71.0 adds AWS Bedrock support and a scrollbar on the model list (v0.71.0).

Cline CLI v3.0.10 supports file:// URL plugin installs (useful for local or private registries), adds Poolside as a provider, and fixes idle session keep-alive so sessions no longer die while awaiting approval (cli-v3.0.10).

🌐 Actualité IA

Google published a recap of the I/O 2026 Dialogues stage with Alphabet CEO Sundar Pichai (blog.google). No concrete API or pricing commitments surface in the brief excerpt — monitor the Google DeepMind release channel for downstream drops.

💡 Conseil du jour

If you run text2vec-google in Weaviate with batch vectorization, spin up v1.36.15 in staging today and run a representative batch workload — the batch logic fix may silently change throughput or error behavior before it surfaces under production load.

#api#llm#en#weaviate#llama.cpp#goose#daily