AI Agent GitHub Digest — 2026-10-30

🌏 中文版

Today's Highlight

Yesterday Agno v3.1.11 made Claude Code and Codex full members of a Team; today Agno v3.1.12 follows up with "Native Approvals" — permission requests these external coding agents raise on their own now surface directly in the AgentOS approval UI (see yesterday's deep dive). The same day, Mastra 1.80.0 lets end users bind their own OAuth connections, and metaharness bakes governance policy, approval gates, and verifiable signing straight into the default output of generating an agent harness. Three different layers — runtime permissions, service connections, and harness generation itself — are converging on the same idea the same day: lock down the checkpoints first, then let agents do more on their own. On the trending side, where agents get applied keeps spreading: claude-mem crossing 100k stars shows that "memory" is now infrastructure-grade demand, DeepTutor applies it to education, and Claude-Code-Game-Studios imports an entire game studio's org chart into Claude Code.

Trending Repos

thedotmack/claude-mem ⭐ 104,523 (+459)

GitHub　·　TypeScript　·　Apache-2.0

• What it is: A hook-layer tool that gives Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, and OpenCode persistent memory across sessions — at the end of each session it compresses tool-use history into AI-generated summaries, then injects the relevant memories back in when a new session starts.
• Why it matters: This repo made it onto the site once before, back in late August, at 92k stars; two months later it's broken 100k with no obvious slowdown, which says cross-session memory is a far more universal pain point than it first looked. More notably, the latest release extends that memory timeline to Codex subagents — a new SubagentStart hook gives a freshly spawned subagent the same memory context the parent session has, while skipping subagents forked from a parent (which already inherit its history) to avoid double-injecting.
• Tech stack: hooks into each CLI's own lifecycle events (SessionStart / UserPromptSubmit / PreToolUse / PostToolUse / Stop / SessionEnd / SubagentStart) + a Node.js worker for compression and retrieval
• Getting started: Easy — npx claude-mem install for Claude Code; other CLIs have their own install flags or marketplace plugin versions

HKUDS/DeepTutor ⭐ 49,631 (+1,926)

GitHub　·　Python　·　Apache-2.0

• What it is: A "lifelong personalized tutoring" system from HKU's Data Science Lab — it pulls books, videos, PDFs, and even GitHub repos into a continuously-synced knowledge base, then uses a multi-agent setup to help you read, get quizzed, take notes, and track learning progress.
• Why it matters: Most education-flavored AI tools stop at "chatbot that answers questions." DeepTutor turns the whole learning workflow into agents: immersive reading cites sources page by page, and a Mastery Path adjusts difficulty dynamically based on how you're doing, rather than following a fixed curriculum. The knowledge base even syncs bidirectionally with existing reading apps like MarginNote and IMA instead of becoming its own silo. The release cadence is unusually fast — nearly 20 releases in the past two months — which itself signals a team iterating hard on real usage feedback.
• Tech stack: Python backend + first-party LightRAG for retrieval + multi-provider model routing (local models included) + browser-based reader frontend
• Getting started: Medium — self-hosting needs at least one model provider key configured; the official hosted version is also available to try directly

Donchitos/Claude-Code-Game-Studios ⭐ 28,954 (+538)

GitHub　·　Shell　·　MIT

• What it is: A template that packages Claude Code into a full game development studio — 49 AI agents with clearly divided roles (design, programming, art, QA, and more) paired with 72 workflow skills, mirroring how a real game studio's org chart collaborates.
• Why it matters: Unlike the single-skill packages that have been trending lately, this project hardcodes an entire development team's division of labor directly into agent roles and skill combinations, rather than relying on one all-purpose agent to decide what role to play moment to moment. Set against claude-mem (a memory layer) and DeepTutor (an education application) on the same day, this repo represents a different path: lift an existing industry's organizational structure wholesale into an agent-collaboration framework. GitHub flags it as a template repo (is_template: true), so it's meant to be forked directly as the starting point for your own project.
• Tech stack: Shell scripts for orchestration + Claude Code plugin/skill mechanisms defining 49 agent roles and 72 skills
• Getting started: Medium — you'll want to be familiar with Claude Code's skill/subagent mechanism first; given the sheer number of roles and skills, read the official collaboration-flow docs before diving in

ruvnet/metaharness ⭐ 729

GitHub　·　TypeScript　·　MIT

• What it is: A "harness for making harnesses" — it generates a branded agent harness for any GitHub repo, or a blank slate: your own npx CLI, a local MCP server, a memory namespace scoped to that project, governance policy, and a signature-verified release pipeline, pluggable into Claude Code, Codex, pi.dev, Hermes, OpenClaw, and more.
• Why it matters: Most "give your agent more capability" tools today solve one dimension at a time — memory, skills, review. metaharness goes after a layer above that: every serious repo deserves a harness built for it specifically, not a generic one, but hand-crafting a harness from scratch is too high a bar for most people. It positions what it generates as "your own product" (your name, your branding) rather than another shared framework to depend on; the built-in governance policy and witness-signed provenance echo the same approval- and connection-scoping convergence showing up in today's Agno and Mastra releases. Star count is still small — this is early-stage — but the timing lines up with its GitHub Universe 2026 launch, so it's worth watching how it grows from here.
• Tech stack: TypeScript CLI + publishable npm package output + MCP server scaffolding + a browser-based Studio (zero-upload, nothing leaves your machine)
• Getting started: Medium — npx metaharness in the browser gets you the fastest full run, but actually customizing governance policy and memory namespaces means reading the USERGUIDE and ARCHITECTURE docs first

Notable Releases

Mastra @mastra/core 1.80.0

Release Notes

• Key changes: @mastra/connect adds end-user (per-resource) OAuth connections — tools({ resourceId }) plus a new userConnectRoutes() let each end user authorize with their own account, with hard enforcement that one user's connection can't be reused by another user's agent; a new request_connection tool lets an agent surface a structured "connect" card to the chat UI when it's missing an integration; a new span aggregation API (ClickHouse, Postgres, and DuckDB all supported) returns grouped stats — run counts, latency, error rate, tokens, cost — in one query; streaming error classification is now unified, with provider status codes, request IDs, and model IDs carried through the stream to the client.
• Breaking changes: Yes, four of them — queryTraceThreads() thread rows no longer include model or provider, and page.limit is now capped at 100 (default 25); @mastra/connect's connect-session creation API now returns launchUrl instead of connectUrl; @mastra/code-sdk no longer registers built-in subagents by default, so keeping the old behavior means passing subagents: [...builtinSubagents] yourself; @mastra/playground-ui's TracesExplorer sorting API changed to a single { field, direction } object.
• What it means for you: If your app has multiple end users who each need to connect their own third-party accounts — rather than the whole app sharing one service account — the per-resource connection feature is worth prioritizing. If your code calls queryTraceThreads(), @mastra/connect's session creation API, or the TracesExplorer sorting API, update it to match the four changes above before upgrading, or it will break outright.

crewAI v1.15.33

Release Notes

• Key changes: A new crewai skill search command lists an organization's available skills; crewai run --inputs can now pass parameters straight into a Python flow; deploy create --zip now prints which commit the resulting zip was built from; a batch of bug fixes landed too — LiteLLM tool-call and structured-output calls now correctly report token usage, some backwards-compatible import paths were restored, JSON crew runners now work against older crewai-core versions, crewai chat can load JSON crews, and crewai update now only prints upgrade commands instead of rewriting your uv project config directly.
• Breaking changes: None stated — this is mostly bug fixes correcting inconsistent existing behavior.
• What it means for you: If you rely on crewai update to manage dependency versions, note that it now only suggests commands rather than automatically rewriting pyproject.toml for you. Everything else here is a fix-forward update that shouldn't require code changes to adopt.

Today's Takeaway

I used to think "governance policy plus approval gates" was something only enterprise-scale frameworks could afford — infrastructure you needed real scale to justify. But metaharness treats that as the default output of generating an agent for any repo, not an add-on bolted on afterward. Governance is shifting from "something big companies retrofit after something goes wrong" to "built in the moment an agent gets generated" — and that bar is dropping fast enough that a single person's side project can have it with one command.

References

• thedotmack/claude-mem — GitHub repo
• claude-mem Releases (Codex subagent memory timeline)
• HKUDS/DeepTutor — GitHub repo
• Donchitos/Claude-Code-Game-Studios — GitHub repo
• ruvnet/metaharness — GitHub repo
• Mastra @mastra/core@1.80.0 Release Notes
• crewAI v1.15.33 Release Notes
• Agno v3.1.11 deep dive (Harness Agents)
• GitHub Trending (daily snapshot, 2026-10-30) — source of today's trending snapshot, including each repo's daily star delta