The wire

Claude Code 2.1.251 adds PreModelSwitch/PostModelSwitch hook events for blocking or annotating model switches, live streaming of foreground subagent tool calls to Remote Control clients, and new background-session CLI subcommands (attach/logs/stop/respawn/rm). SessionStart resume hooks now receive staleness and re-cache cost. Also includes /cost prompt-cache stats, spend-limit status fields, and many symlink/path-traversal security fixes.

The Anthropic Python SDK 1.2.0 moves the beta files and skills namespaces onto GA shapes (dropping dated beta header pins), fixes AWS Bedrock signing for binary file uploads, lets the read tool return a view_range for files over the size cap, preserves exact file bytes (no newline translation) in the agent toolset and memory tool, makes session event accumulators forward-compatible with new event types, and tightens webhook `unwrap()` header handling.

Hugging Face Transformers 5.16.1 lands support for GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series (320B total / 18B active MoE, hybrid sparse+linear attention, Manifold-Constrained Hyper-Connections, 30T-token multimodal pretraining). Reportedly outperforms GLM-5.2 and approaches Claude Opus 4.8 on coding/agentic benchmarks. Also restores BC for the tensor-parallel API and pins an ESMFold2 kernel commit for security.

vLLM 0.28.0 lands 584 commits from 270 contributors, headlined by a Kimi-K3 optimization push (DCP, fused FlashKDA kernels, GEMM-RS, adaptive spec budget), end-to-end DeepSeek V4 sparse MLA with DSpark, new DFlash2 speculative decoding, Model Runner V2 E/P/D disaggregation, tiered KV cache with disk offloading, a Rust gRPC frontend with multimodal image inference, higher default batched-token and CUDA-graph caps, plus a bitsandbytes plugin split-out.

DSPy 3.3.1 hardens the PythonInterpreter with an optional managed Deno/Pyodide runtime (`pip install dspy[deno]`), stronger sandbox isolation, and full callback lifecycle observability. It upgrades to GEPA 0.1.4 with multi-proposal sampling and objective-aware frontiers, adds MCP SDK v2 compatibility with opt-in `result_mode=\"structured\"`, applies default fallbacks for missing adapter fields, and deprecates CodeAct/ProgramOfThought in favor of `dspy.RLM`.

LlamaIndex 0.14.24 ships broad model allowlist updates (Claude Sonnet 5, Opus 5 with 1M context on Opus 4.6, GPT-5.6, Gemini 3.7 Flash as new default) across the anthropic, bedrock-converse, openai, and google-genai integrations. AG-UI protocol 0.4.0 now supports multimodal user input (images, audio, video, documents). Memory can accept any AsyncDBChatStore, LLMRerank gains async, and dozens of core bug fixes land for agents, ingestion, and retrieval.

The Model Context Protocol has published the stable 2026-07-28 revision of its specification. This is the versioned protocol contract that MCP clients (agents/hosts) and MCP servers implement to exchange tools, resources, prompts, and sampling requests. Full details are on modelcontextprotocol.io along with a changelog covering deltas from the prior revision.

Anthropic published version 1.30.0 of the official Model Context Protocol TypeScript SDK to npm. The SDK lets developers build MCP clients and servers in Node/TypeScript, exposing tools, resources, and prompts to MCP-compatible agents over stdio, SSE, and streamable HTTP transports.