Tenkai Daily — June 16, 2026

Open Source Releases

• alibaba/zvec: A Lightweight In-Process Vector Database — C++ vector DB that runs in your process, no separate service needed. Low latency, minimal overhead. If you’ve been putting off vector search because you didn’t want to operate Milvus or Qdrant, this is worth a look 🤖
• turnstone 1.7.0a2 — Multi-node AI orchestration with tool use, agent routing, and cluster simulation. Still alpha, but if you’re building distributed agent systems that need actual coordination (not just “hey LLM, call this tool”), the simulation bits alone might save you weeks of debugging 🛠️
• novalab-adk 2.9.0 — Provider-agnostic agent kit supporting Ollama, Bedrock, Claude, GPT, Gemini with multimodal and training features. Useful if you’re tired of rewriting your agent layer every time your org switches LLM vendors 🔄

Research Worth Reading

• Dr-DCI: Scaling Direct Corpus Interaction via Dynamic Workspace Expansion — Lets agents reorganize and cross-reference documents dynamically instead of just getting ranked snippets back. Addresses the “I found the needle but lost the context” problem in large-corpus agentic search 📄
• Relational Structural Causal Models — Combines causal reasoning (interventions, counterfactuals) with combinatorial generalization. Formally characterizes when you can learn these models. Heavy theory, but if you’re doing causal ML with relational data, this is the paper to cite 🧠
• PrologMCP: A Standardized Prolog Tool Interface for LLM Agents — MCP-style interface for delegating deductive reasoning to Prolog. Because LLMs still can’t count to three reliably, and symbolic solvers exist for a reason. The standardization angle is the real contribution here 🔧
• OSGuard: A Benchmark for Safety in Computer-Use Agents — Dual-granularity benchmark catching agents that achieve goals through dangerous shortcuts. Task success ≠ safety. Finally a benchmark that asks “how did you get there?” not just “did you arrive?” ⚠️
• CONCORD: Asynchronous Sparse Aggregation for Device-Cloud RAG under Document Isolation — Device-cloud RAG where private docs never leave the device. Async sparse aggregation keeps latency down. Relevant if you’re building on-device AI with privacy constraints 📱
• QPILOTS: Efficient Test-Time Q-Steering for Flow Policies — Test-time Q-steering for flow-matching/diffusion policies without backpropagating through the whole denoising chain. Numerically stable, computationally cheaper. Robotics folks, this one’s for you 🤖

AI Dev Tools

• Claude Code v2.1.178 — Tool-level permission rules with parameter matching — Tool(param:value) syntax for granular permissions (e.g., Agent(model:opus) to block Opus subagents). Skills now load from nested .claude/skills with namespacing. Finally, permissions that don’t require an allowlist the size of a phonebook 🔐

Today’s Synthesis

If you’re building on-device RAG with privacy requirements, alibaba/zvec and CONCORD solve complementary halves of the problem. zvec gives you a C++ in-process vector DB — no sidecar, no network hop, sub-millisecond latency — while CONCORD’s async sparse aggregation lets you sync only embeddings (never raw docs) to a cloud index for cross-device retrieval. The architecture writes itself: ingest locally into zvec, generate sparse vectors on-device, push those to your CONCORD-compatible cloud aggregator, and query either tier depending on latency budget. Private corpus never leaves the machine; global knowledge stays queryable. For a concrete starting point, wrap zvec’s C++ API with pybind11, implement CONCORD’s sparse encoder (they reference SPLADE), and you’ve got a local-first RAG stack that respects document isolation without sacrificing recall. 📱🤖