Model Releases

  • GLM-5.3-CYBERSECURITY-FP8 🤖 A domain-specific GLM-MOE-DSA fine-tune for cybersecurity and red-team work, quantized to FP8 for more practical GPU deployments. The refusal-removal tuning is useful only if that is explicitly the workload.
  • Minimax-H3 Fine-Tuned Text-to-Video A Diffusers-integrated fine-tune for text-to-video, image-to-video, synchronized audio-video, and reference-based generation. Worth considering if your existing video pipeline already depends on Diffusers.

Open Source Releases

  • claude-code v2.1.266 Fixes a regression where CLAUDE_CODE_USE_GATEWAY forced Cloud sign-in even when configured with an API key. This is the version to use for proxy and LLM-gateway setups that relied on the previous behavior.
  • sst/opencode v1.18.30 Adds an Astra system-prompt adapter for GPT-6, giving clients a consistent way to format and invoke the model. Useful if you are integrating GPT-6 without maintaining your own prompt mapping.
  • router-maestro 0.9.6 Provides OpenAI-compatible routing, load balancing, and failover across multiple LLM backends. A reasonable fit for distributed inference systems that need backend diversity without rewriting client code.
  • clawmetry 0.12.847 🛠️ Monitors more than 30 AI-agent runtimes with real-time metrics and health visualization, including OpenClaw, NemoClaw, Claude Code, and Codex. Its main value is observability across agent runtimes that otherwise have fragmented tooling.
  • claude-code v2.1.265 Adds user.email and user.groups to gateway telemetry and supports --plugin-dir loading for dynamically added or removed plugin manifests. The plugin behavior is the operational win; the extra telemetry is a privacy consideration.
  • under-llm-governance 0.1.0b1 A least-privilege controller for local coding agents, enforcing scoped permissions and resource boundaries. It targets the security gap between autonomous code generation and controlled execution.

AI Dev Tools

  • PI-Desktop An Electron/Rust desktop coding agent built on the pi Agent Harness, with user-installable plugins and local-first operation. It is the option to inspect if offline assistance matters more than cloud integrations.
  • goose v1.50.0 Adds GPT-6 Astra support, tool calling for goose-agent, and Kotlin caller configuration for the Databricks AI Gateway path. Those are concrete integration gains rather than a reason to upgrade on their own.
  • cline v0.0.24 Fixes doubled streamed text and dropped messages caused by the sidecar opening two Hub sockets for session events. A straightforward correctness fix for CLI session rendering.

Today’s Synthesis

If you’re wiring up a distributed LLM pipeline today, router-maestro 0.9.6 gives you OpenAI-compatible routing, load balancing, and failover across multiple backends without touching client code—the backend diversity piece that usually means maintaining custom proxy logic. Pair that with under-llm-governance 0.1.0b1 , which enforces least-privilege permissions and resource boundaries for local coding agents, plugging the security gap between autonomous code generation and controlled execution. The missing link in many stacks used to be the gateway sign-in regression in claude-code v2.1.266 : it fixes the CLAUDE_CODE_USE_GATEWAY forced Cloud sign-in issue, so proxy and LLM-gateway setups work as intended without unnecessary Cloud dependencies. Together, these three form a concrete scaffolding: route traffic through router-maestro, constrain agents with under-llm-governance, and use the fixed claude-code version for reliable gateway integration. It isn’t a complete platform, but it is a practical stack for running distributed inference with actual operational control. If you’ve been burned by fragmented agent tooling or opaque backend failover, this combination is worth a serious look.