AI Agent GitHub Digest — 2026-09-26

🌏 中文版

Today's Highlight

The word popping up everywhere this week is Jev — the "silent" decision model TypeSafe shipped last week. Instead of text, it outputs structured decisions: Choice, Score, or Noul (a true/false probability). Today shows exactly how fast it's spreading: Pydantic AI v2.50.0 promoted last week's TypeSafeModel into a formal DecisionModel base class, DSPy 3.4.0 added a TypeSafe client integration the same day, and Nokia's applied research team open-sourced AnyJev, letting teams still stuck on the waitlist wrap any open LLM into the same calibrated decision layer without training it. A primitive that was public for barely a week is now infrastructure for three independent camps.

Trending Repos

nokia-applied-research/AnyJev ⭐ 662 (created 09-21, within 5 days, ~132/day average)

GitHub　·　Python　·　Apache-2.0

• What it is: A Python library from Nokia's applied research team that turns any open LLM into a calibrated decision model without additional training, using cyclic shifts (showing the option list in every rotation and taking the geometric mean to cancel out position bias) plus batch prior correction.
• Why it matters: Reading decisions straight off next-token logits has two known flaws — the answer flips when you reorder the options, and the raw probabilities aren't calibrated. AnyJev fixes both in two tiers: L0 needs no labeled data, L1 adds temperature scaling from 100–500 labeled examples. On Qwen3-8B/BANKING77, Nokia's own benchmark raised the share of traffic that's "auto-decidable within 5% error" from 7.7% to 52.0%, and cut the order-flip rate from 0.230 to 0.073. It's a rare case of a telecom-equipment R&D team shipping an open-source alternative for a paradigm that's barely a week old.
• Tech stack: Python + Hugging Face Transformers / vLLM inference backends + shared-prefix batch scoring
• Getting started: Medium — pip install anyjev[hf] runs out of the box, but getting correct probabilities requires understanding the L0/L1 calibration tiers and picking the right question type

mikehasa/golive-skill ⭐ 942 (created 09-23, within 3 days, ~314/day average)

GitHub　·　TypeScript　·　MIT

• What it is: An Agent Skill usable by Claude Code or Codex that picks up where most coding agents stop — actually taking the app they just built live.
• Why it matters: Most coding agents stop once the code compiles; wiring up hosting, a database, a domain, email, and payments is still manual. golive-skill runs a detect→plan→approve→apply→verify pipeline straight against your own Vercel/Netlify/Supabase/Neon/Cloudflare accounts, with no backend of its own and no telemetry — the deployment still happens on your own accounts, not through another managed layer in between.
• Tech stack: TypeScript + zero-dependency Node CLI + each hosting/DB/DNS provider's own API
• Getting started: Low — drop it into your agent's skill directory and the agent decides which provider to call

yetone/magpie ⭐ 791 (created 09-23, within 3 days, ~264/day average)

GitHub　·　Go　·　MIT

• What it is: A macOS menu-bar utility that turns "run Codex on DeepSeek, run Claude Code on Kimi" into a two-click operation.
• Why it matters: Coding-agent shells (Claude Code, Codex) are decoupling from their backend models fast, but switching providers today usually means editing config files or reconciling API formats. magpie sits in as a routing layer that absorbs those protocol differences, addressing a very concrete pain point developers are already solving by hand.
• Tech stack: Go + local routing daemon + compatibility adapters for each LLM API
• Getting started: Low — pick the model combination from the menu bar; no changes needed to the agent itself

sno-ai/sno-station ⭐ 212 (created 09-19, within 7 days, ~30/day average)

GitHub　·　TypeScript　·　Apache-2.0

• What it is: A local multi-agent coordination layer that lets Claude Code and Codex work as a squad on the same machine, sharing encrypted memory and messaging each other.
• Why it matters: Most multi-agent frameworks are built for agents in the cloud calling each other. sno-station instead targets the much more concrete case of one person's two different coding-agent brands dividing up work, and adds a nightly loop — gated on human approval — that lets the agents rewrite their own skills. That "self-rewriting" step is rarer than the usual "self-remembering" one, and deserves more caution.
• Tech stack: TypeScript + a local encrypted messaging layer (Reach) + no daemon, no cloud required
• Getting started: Medium — the concept is novel enough that you'll want to understand "squad skills" and the approval flow before letting it run unattended

mitkox/esf ⭐ 148 (created 09-19, within 7 days, ~21/day average)

GitHub　·　Go　·　MIT

• What it is: A self-hosted "engineering software factory" that wraps sandboxed coding agents in Temporal workflows, leaving a verifiable audit trail at every step.
• Why it matters: Most coding-agent projects care whether the code got written; esf cares how it got written — whether the run can be replayed and audited. That's the direction that pushes coding agents from personal toys toward a pipeline a team can actually trust in production, with Temporal guaranteeing failed steps retry instead of silently vanishing.
• Tech stack: Go + Temporal workflow engine + the CubeSandbox execution sandbox
• Getting started: High — you need to self-host a Temporal cluster; this is aimed at teams building an internal agent-development pipeline, not individual developers

Notable Releases

DSPy 3.4.0

Release Notes

• What changed: Added a TypeSafe client integration that calls Jev directly for typed Noul/Choice/Score decision outputs; added a ReAnchor optimizer that recalibrates decision thresholds against your own scoring function; the native dspy.lm15 LM engine replaces the execution path that previously depended entirely on LiteLLM, and opens up registering custom HTTP providers; added a persistent, non-sandboxed LocalInterpreter (for RLM/Flex) and an async ReActV2.
• Breaking Changes: RLM's interpreter_factory parameter is now keyword-only; the old experimental LM types from 3.3 have been removed entirely.
• What it means for you: If your code calls interpreter_factory positionally, or still relies on 3.3's old experimental LM types, upgrading will break immediately until you rewrite against the new interface.

Pydantic AI v2.50.0

Release Notes

• What changed: Added a DecisionModel base class, promoting last week's TypeSafeModel (from v2.49.0) into a formal "Decisions protocol" model family that routes by named labels; added gemini-3.8-live/gemini-3.8-live-extended-thinking realtime voice support; every DecisionModel request now emits its own decide tracing span; fixed several billing bugs around Anthropic's one-hour cache pricing, realtime-voice image pricing, and audio-duration billing.
• Breaking Changes: None — this release is mostly new features and billing/streaming bug fixes.
• What it means for you: If you're using Anthropic's one-hour cache or Gemini Live voice features, this release fixes billing bugs that directly affect your invoice — worth upgrading to check.

Today's Takeaway

I used to think "model capability" was mostly a race over whose text generation was more accurate and faster. Jev offers a different answer this week: strip language out entirely and leave only structured decisions and probabilities that ordinary code can consume directly. What's more striking is how fast it spread — a model that went public barely a week ago, still on a waitlist, already has native support in two mainstream frameworks (DSPy and Pydantic AI), and Nokia's applied research team has already shipped a training-free compatibility layer for it. Framework maintainers are now treating "decision models" as a category parallel to text-generation models, rather than squeezing them into the existing LLM call interface.

References

• nokia-applied-research/AnyJev
• mikehasa/golive-skill
• yetone/magpie
• sno-ai/sno-station
• mitkox/esf
• DSPy 3.4.0 — Release Notes
• Pydantic AI v2.50.0 — Release Notes
• Nokia Open-Sources AnyJev — MarkTechPost