Agent Core Toolkit Public

A public, reusable, runtime-neutral toolkit for agent instructions.

It provides a canonical AGENTS.md contract, generic role overlays, starter skills, and rule snippets. Everything in this repository is intended to be generic, synthetic, and publishable.

Contents

AGENTS.md                         # Base public agent contract
openpack.json                     # OpenPack package manifest
roles/<role>/AGENTS.md            # Generic role overlays
skills/<skill>/SKILL.md           # Reusable public skills
rules/<rule>.md                   # Runtime-neutral rule snippets
plugins/README.md                 # Plugin structure guidance
LICENSE                           # MIT license

Quick Start

Install agentwheel:

npm i -g agentwheel

Install from the GitHub source:

agentwheel init
agentwheel add github:NestDevLab/agent-core-toolkit-public --adapter openclaw --mode tracking
agentwheel update --dry-run
agentwheel update

After this package is available in the public registry, the short name works too:

agentwheel registry update
agentwheel add nestdev-core-toolkit --adapter openclaw --mode tracking
agentwheel update --dry-run
agentwheel update

For one-off preview/sync without saving a package entry:

agentwheel sync github:NestDevLab/agent-core-toolkit-public --adapter openclaw --dry-run

agentwheel supports bundled adapters such as openclaw, claude, codex, hermes, and copilot, plus custom/private adapters.

Codex Suggested Next Message Hook

The package includes a Codex-only Stop hook that can append a context-aware, copyable Suggested next message block after a response. It is installed inert and performs no model call until Codex is started with:

CODEX_SUGGESTED_NEXT_MESSAGE_ENABLED=1 codex

When enabled, it runs an ephemeral, read-only gpt-5.6-luna child at low reasoning effort. The child receives the final answer and the latest user message from a bounded transcript tail; it may also receive a compact, explicitly supplied context summary through CODEX_SUGGESTED_NEXT_MESSAGE_CONTEXT. It never loads an unbounded transcript or retains the inputs or generated suggestion after the hook finishes. Failures and timeouts leave the main response unchanged.

Review and trust the generated Codex hook before installation. See skills/codex-suggested-next-message/SKILL.md for prerequisites, context limits, and the optional script-path override.

agentwheel Package

openpack.json exposes:

• instructions -> AGENTS.md
• rules -> rules/
• skills -> skills/
• hooks -> opt-in Codex, Claude, and Copilot configurations

Skill Evolution

OpenPack v3 injects fragments/skill-evolution.md into every rendered skill when this toolkit is a configured graph root, except the evolution skill itself. The skill-evolution composite resolves authoritative source and ownership before preparing any improvement. Its runner validates, deduplicates, and classifies bounded events with explicit --dry-run and --apply modes. Deterministic candidates require an authoritative script, focused tests, policy-bounded paths, and second-run idempotence proof.

Selecting a Codex or Claude failure-observer hook activates it; omission is the opt-out. The standalone observer still requires SKILL_EVOLUTION_HOOKS_ENABLED=1. Observers retain fingerprints, not prompts, transcripts, commands, or tool output. The Copilot hook stays environment-gated; OpenClaw and Hermes observers are not part of v0.1.

The roles/ directory contains generic role overlays for humans or future adapter support. It is not currently mapped as an agentwheel subagents artifact because this repository stores roles as nested roles/<role>/AGENTS.md folders, while the current package manifest flow expects directly installable files or supported artifact directories.

Skills And Rules

Starter skills:

• code-review: findings-first technical review.
• conversation-handoff: verified context transfer between sessions or agents.
• plan-task: ordered planning with assumptions and stop points.
• decision-interview: focused questions before a confident answer, opinion, decision, or plan.
• write-docs: durable technical documentation.

Starter rules:

• engineering-standards.md
• public-audience-privacy.md
• safe-actions.md

The base instructions also include bounded capability discovery when substantive work exposes a missing reusable capability.

Public Boundary

This repository must stay generic and safe to publish. Do not add real identities, organizations, clients, hostnames, IP addresses, secrets, private workspace paths, transcripts, or operational details.

Contributions should use synthetic examples and runtime-neutral language. Deployment-specific overlays belong in separate private layers.