Back to timeline

Thu, October 1501:30ToolsMCP & SkillsAI codingAgentsMCP & Skills guide

AAS Core V19.6.0 released: an agent-facing local skill catalog and control plane

Decision Brief

What changedGitHub project agentic-awesome-skills releases AAS Core V19.6.0, giving Codex and Claude a local skill catalog, agent-driven skill selection, stack validation, and plan previews, with 2,689+ SKILL.md skills indexed.
Why it mattersIt hands skill selection to Codex or Claude locally and uses read-only compose_stack and aas stack plan to produce an immutable preview before anything is applied, fixing the old problem of manually browsing catalogs and only finding out if it works after the change.
Who should careAI coding tool users
Affected stackClaude CodeOpenAI
Builder actionUpgrade Claude Code to 19.6.0
Source confidenceHigh · Official release / blog / repo

The project is at version V19.6.0. AAS Core is a local, agent-first control plane: Codex or Claude can search the full catalog of 2,689+ SKILL.md skills, read each skill file, pick an exact skill set, and use the read-only compose_stack tool to validate selected IDs and structure in memory. The client or aas CLI can persist aas-stack.json plus selection credentials; aas stack validate checks the manifest, and aas stack plan generates an immutable preview for human review. The manifest has a technical cap of 128 skills. Core does not rank, recommend, or directly install skills along this preview path. The impact is concrete: developers using Codex CLI or Claude Code can now let the agent choose skills and then review the plan themselves, avoiding blind bulk installs that pollute context. The team warns that a structurally valid stack is not necessarily semantically suitable or safe; apply and recovery remain experimental and require explicit opt-in. Antigravity's skill-watch directory easily overwhelms context, so default targets need a specified skill set or filters. Separately, the direct installer's 19.6.0 commands support --dry-run previews and can narrow install scope with --path, --category, --risk, and similar flags. Workbench reviews stacks and plan artifacts only in browser memory and does not access the local filesystem.

Summary basis: full article readCompiled from the source scope noted above; the original remains authoritative.

Sources

Related intel

留言

登入后即可留言,和其他 builder 交换实测心得。

还没有留言,抢头香。