OSS Release Recap: Sept 15, 2026 - AI, Infra, and Runtime Highlights

Frameworks

• vercel/next.js, v16.3.5: Backport release fixing disk LRU cache corruption for next/image, adding CSP nonces to loading and template script tags, and correcting use cache prerender signal retention. Release
• vitejs/vite, v8.3.0: Performance-focused minor adding pre-compiled proxy context matchers at server creation, avoiding redundant preload dependency settling during builds, and fixing CRLF code-frame position handling. Release

Infra

• hashicorp/terraform, v1.17.0-beta1: Variables and locals are now supported inside provider requirement blocks; a new -minimal-refresh planning flag limits refresh to resources with proposed changes; and Terraform Policy graduates to GA, dropping the --allow-experimental-features prerequisite. The JSON output of version gains a format_version field for safer future evolution. Release
• hashicorp/terraform, v1.16.2: Patch fixing a panic in module installation when encountering invalid module calls. Release

Data

• supabase/supabase, self-hosted/v0.8.1: Self-hosted release adds a PgBouncer override option alongside component version bumps. Release
• ClickHouse/ClickHouse: Maintenance patches shipped across the v26.8.4.11-lts, v26.7.8.15-stable, and v26.6.6.15-stable branches with no published changelogs; skipping until notes are available.

AI

• vllm-project/vllm, v0.29.0: Major release with 594 commits from 277 contributors. Model Runner V2 is now the default for all models, bringing CUDA graph memory profiling, batch-sharded sampling cutting per-step logits memory by 1/TP, and prompt embeds. New model support includes Hy4-preview (Tencent 770B MoE), Qwen3.8-Flash-Next, GraniteSWA, NemotronH_Omni_Reasoning_V3, and Kimi K3 NVFP4. Significant latency improvements for Kimi-K3 and DeepSeek V4 via fused MXFP4, tuned Hopper GEMMs, and MLA gate fusion. Release
• openai/openai-python, v3.14.0: Normalizes errors raised while reading streams, fixes vector store file polling, and adds standard OPENAI_LOG level support. Release
• openai/openai-python, v3.13.0: Adds the Agents API. Release
• openai/openai-python, v3.12.0: Adds the Live API and fixes AsyncStream missing aclose() for standard async cleanup. Release
• anthropics/anthropic-sdk-typescript, sdk-v0.125.0: Adds auto mode tool permissions for Managed Agents, a content_too_large web_fetch error code, external user details via the user-profiles-2026-09-04 beta, and support for mounting public GitHub repositories without an authorization token in Managed Agents sessions. Release
• BerriAI/litellm, v1.101.0: Emits timing and overhead headers for /v1/messages and /v1/responses; large typing cleanup clearing over 4,500 basedpyright Any errors across 63 files. Release

Runtime

• cloudflare/workerd, v1.20260915.1: Exposes client IP and port as remoteAddress in TCP connect() handlers, makes the isolate the port namespace for Node.js http and net servers, replaces string Pyodide patches with AST patches for robustness, and fixes memory snapshots under Python 3.14. Release
• cloudflare/workerd, v1.20260913.1: Introduces kj-rs-tokio, a kj::EventPort driven by a per-thread Tokio runtime, and adds node:net net.Server plus a cloudflare:node connect handler. Release
• vercel/turborepo, v2.10.13: Adds support for Python virtual environments via uv workspace metadata, Cargo root package support, and expanded lockfile package metadata for repository metrics. Release

Devtools

• microsoft/vscode, 1.137.0: Full release notes at the VS Code update page; no inline changelog provided in the tag. Release notes
• neovim/neovim, v0.13.0-dev nightly: Prerelease build shipping the week's accumulated fixes and features; stable release cadence unchanged. Release

The throughline this week is maturity: AI inference (vLLM, openai-python, LiteLLM) is consolidating around production-hardened defaults and typed interfaces, while infra tooling (Terraform Policy GA, workerd Tokio integration, Turborepo Python support) is quietly expanding the surface area of what monorepo and cloud-native builds can coordinate out of the box.