The Hub for AI Coding Models, Tools & Workflows

n8n shipped [[email protected]](https://github.com/n8n-io/n8n/releases/tag/[email protected]) on 2026-09-29: Agents enabled by default (#39312); AI Agent Tool can use its own tools under a pre-v3 parent agent (#39160); MCP Client terminates sessions before close, MCP Registry flag removed from AI Assistant, editor adds Mistral Vibe to MCP picker; Azure OpenAI Chat Model renamed Azure AI Foundry with project/deployment listing and Entra app sign-in; plus shared Agent context and Agent Builder UX polish.

LangChain shipped `langchain-anthropic==1.7.5` and `langchain-core==1.6.6` on 2026-09-29, landing [#40882](https://github.com/langchain-ai/langchain/pull/40882): Claude Sonnet 5.5 model profile and capability discovery; reject unsupported forced tool_choice/thinking; allow mid-conversation system/tool changes; preserve computer-toolset namespaces, encrypted advisor blocks, and refusal details through streaming and replay. Invalid tool calls can fall back to unforced structured output or explicit `method="json_schema"`.

Charm's terminal coding agent Crush shipped v0.97.0/v0.97.1 on 2026-09-29 (npm `@charmland/[email protected]`): toggle MCP servers on/off in the UI (global or local); show the current Git branch in the TUI; continue Claude Channels session injection with exactly-once routing; toggle the sidebar with ctrl+b; fix Anthropic cache-creation token counting, clipboard false failures, and OpenAI OAuth refresh via Catwalk. ~36 commits vs v0.96.0.

Comet Opik 2.2.84 (2026-09-29) lets online LLM-as-a-Judge rules call TypeSafe Jev end-to-end via OpenRouter `POST /api/alpha/decisions` (not chat/completions). Each Boolean metric becomes a yes/no question answered in one call; score 1 if probability ≥0.5, with `Probability: 0.93` in the reason. Backend adds `OpenRouterDecisionsClient` (workspace OpenRouter key, 429/5xx retries) for `~typesafe/jev-latest` / `typesafe/jev-1.13`. Also: Annotation Queue automation + item provenance. ~8 commits / 104 files vs 2.2.83.

Cline shipped SDK v0.0.87 on 2026-09-29: Windows hub startup works when `HTTP(S)_PROXY` is set—`ensureLoopbackProxyBypass()` adds loopback to `NO_PROXY` so Bun fetch no longer routes health probes through Clash/v2ray/corp proxies; `normalizeUsage()` excludes reasoning from `outputTokens` to stop double-counting; Gateway models keep per-model `apiProtocol`; bare GPT-6/GPT-5.6 on Bedrock use inference profiles (`@ai-sdk/amazon-bedrock` ^5.0.96); catalog 209→211 providers, 6,386→6,447 models; recommended list adds Claude Sonnet/Opus 5.5.

OpenAI Codex shipped rust-v0.159.0 on 2026-09-29 (npm `@openai/[email protected]`): opt-in `instant_interrupt` steers mid-response or long code-mode calls; compact welcome/headers; warnings viewer dismisses reviewed items on close (`k` to keep); scroll transcript while reviewing a plan; richer native Mermaid; app-server can paginate thread history from an item anchor. ~89 commits / 300 files vs v0.158.0. Windows suppresses stray consoles for MCP/code-mode; `.aws` protected under writable roots; macOS sandbox TLS/proxy fixes.

Reinforcement learning for code agents typically relies on executable tests for binary rewards, leading algorithms like GRPO to assign identical scalar advantages to all passing rollouts regardless of code bloat, extraneous modifications, or architectural hygiene. Researchers from ByteDance, Peking University, and Tsinghua introduced GAGAR, a quality-aware advantage redistribution framework. By staging rollouts in a shared workspace where an SFT agentic grader ranks passing candidates and dynamically shifts credit toward clean solutions, GAGAR successfully stabilizes code agent RL on 310B Flash and 1.02T Pro models.

Addressing the fundamental limitation of benchmarks like SWE-bench that assess coding agents strictly within isolated repositories, researchers from Zhejiang University introduced WideSWE. Mining changes across 103 real-world software ecosystems, WideSWE establishes 120 cross-repository tasks (balanced between 60 bug fixes and 60 feature developments) with comprehensive regression test suites. Across seven frontier agent configurations, full task success tops out at just 42.50%, exposing systemic failures in cross-repo dependency tracking and coordinated verification.

To overcome the severe GPU starvation and explosive trajectory redundancy of extreme-long (xlong) horizon agents spanning hours, hundreds of environment steps, and nearly 1M tokens per rollout, Alibaba's Qwen team introduced QwenGyre. By dynamically reallocating GPUs between rollout and training without pausing live interactions and pruning non-linear trajectory branches, QwenGyre enables 700K-token RL on the 2.4T-parameter Qwen 3.8 flagship, driving a 6.0% absolute gain on NL2RepoBench in 48 steps while delivering a 1.85x speedup.

Multi-teacher on-policy distillation (MOPD) aims to synthesize diverse capabilities from specialized models (mathematics, coding, instruction-following) into a single versatile student. However, empirical studies reveal that standard MOPD students fail to beat single-specialist baselines and lose their math edge because instruction-following feedback is exponentially more dispersed and dominates student gradient updates. Researchers introduced Domain-Normalized MOPD (DN-MOPD), which rescales feedback by measured domain spread to recover math and coding expertise across six public benchmarks.

In collaborative multi-agent systems sharing a common foundation model, distinct system prefixes cause KV cache representations for the shared context to diverge, forcing each agent to redundantly re-prefill the entire conversation history. Researchers from Seoul National University and KAIST introduced KVCMAS, an online KV cache correction framework. By parameterizing cross-agent cache deviations as compact low-rank states and chaining corrections along agent pipelines without external reference prefills, KVCMAS accelerates TTFT by 2.0x while reducing peak GPU memory by up to 3.7x.

gh-aw v0.90.0 (2026-09-28) renders friction cost in step summaries, allows two `noop` safe-outputs by default, adds five trajectory graders with hardened generator contracts, and bumps `gh-aw-mcpg` to v0.4.27 / `gh-aw-firewall` to v0.28.27—shifting from 0.89.22’s sandbox focus toward observability and grading.

CrewAI 1.15.23 (2026-09-28) adds native Gemini 3.8 Flash support; `crewai eval` can score the last AMP-traced run (not just print it); task spans capture declared output format/results; LLM retries throttled providers and Bedrock `acall` falls back to sync. Tightens the observe→evaluate loop for multi-agent crews.

Official Ollama Python client v0.6.3 shipped 2026-09-29: adds `ollama.systemone()` matching server v0.35.0 `/v1/systemone` Decision Models (choice/noul/score); relaxes `think` to string levels; fixes multi-tool examples and case-insensitive image extensions. Requires Ollama ≥0.35.0 plus local decision models such as nimble/tev1.

While looped transformers achieve high parameter efficiency by iteratively recycling layers for latent compute, fixed-depth looping expends redundant iterations on simple tokens and suffers from performance saturation. Tsinghua University's NICS lab developed TaH2, an adaptive looped framework that co-trains the backbone and an iteration decider via lookahead depth supervision. Evaluated on the rigorous AIME competition benchmarks, TaH2 accelerates the accuracy-compute scaling slope by 53% (2.74 vs 1.79) and surpasses non-looped baselines by +3.4 to +3.9 points at matched test-time compute.

In reinforcement learning with verifiable rewards (RLVR), rollout trajectories are generated by high-throughput inference engines while backward gradients are computed by training engines, introducing subtle floating-point probability discrepancies that trigger unbounded variance under exact importance sampling. Researchers identified the invariance of logit displacement distributions and introduced Calibrated Importance Sampling (CIS), featuring confidence-aware truncation that replaces unbounded second moments with constant bounds, establishing top scores across five mathematical benchmarks.

Investigating how deeply an agent should revise an interrupted workflow, ControlScope systematically isolates three repair granularities: continuing generated code (KEEP), patching only tool call arguments (ARG), and rewriting the unfinished workflow (FULL). Across filesystem benchmarks, ALFWorld, and AppWorld, the study reveals that unconstrained full-workflow revisions regularly interrupt viable execution paths with self-doubt. Enforcing nested permissions and a 5-call budget cap saves 19.4% of model output tokens without degrading task success.

Addressing the limitation where synthetic training environments only generate tasks within isolated individual domains, researchers introduced CompoWorld. By scaling task spaces across 448 typed services exposing 10,130 tools and generating cross-service workflows via dependency graphs, paired with Completion-Focused Rubric Rewards in RL, a Qwen3.6-35B-A3B agent trained on just 3K SFT trajectories and 1K RL tasks gains +9.17 points across eight benchmarks and surpasses Claude Opus 4.6 on AutomationBench.

LangChain shipped `langchain==1.4.3` on 2026-09-28 (on PyPI): `init_chat_model` supports AWS Bedrock Mantle chat models via langchain-aws Mantle interfaces; recognizes GPT-6 structured output without profiles; `create_agent` emits error `ToolMessage`s for invalid tool calls missing results on replay; sanitizes cache settings for fallback models. Key PRs: #40837, #40844, #40530, #40886 vs 1.4.2.

Anthropic shipped Claude Code v2.1.284 on 2026-09-28 (`@anthropic-ai/[email protected]`): `claude-sonnet-5-5` is now the default Sonnet on the Anthropic API—1M context, $2/$10 per MTok, $0.20/MTok cache reads. Adds “Yes, but ask again next time” for auto-mode reads, `/mcp reconnect all`, dollar spend limits in `/usage` and the status line (`used_usd`/`limit_usd`/`period`), and rebindable effort-slider keybindings. Stability fixes: damaged streams writing undefined, overloaded errors after thinking blocks, and double-compact when Prompt-too-long persists.