toolshed

My personal Agent Skills for Claude Code and Codex. Treat it like dotfiles: I update it when I choose, breaking changes are normal, and there are no stability or compatibility promises. You're welcome to clone it, fork it, install plugins, or copy any piece into your own setup — just don't expect anything in here to stay still.

The portable skill bodies live under plugins/<plugin>/skills/. Each harness loads them through its own marketplace metadata.

Installation

Claude Code

Add the marketplace:

claude plugin marketplace add richhaase/toolshed

Install a plugin:

claude plugin install memento@toolshed

Inside Claude Code, the equivalent commands are:

/plugin marketplace add richhaase/toolshed
/plugin install memento@toolshed

Codex

Add the marketplace:

codex plugin marketplace add richhaase/toolshed
codex plugin add memento@toolshed
codex plugin add actuary@toolshed
codex plugin add steward@toolshed
codex plugin add ostrich@toolshed

Codex reads .agents/plugins/marketplace.json and each plugin's .codex-plugin/plugin.json.

For local development:

codex plugin marketplace add /path/to/toolshed

Runtime requirements

The skill instructions themselves are Markdown. Optional bundled helpers use Node.js (dependency-free JavaScript), Bash, and Git. Install all three to use the complete Memento workflow and the repository validation/privacy gates. Memento also expects standard Unix utilities; its health check uses rg when available and falls back to grep.

The host agent may use its own configured tools for workflow steps, including network access described by a skill. Bundled helper scripts do not install packages or make network requests on their own.

Plugins

Memento — Personal Memory Base

Multi-layer cache memory base with automated compilation. Treats knowledge like a CPU cache hierarchy: L1 (AGENTS.md hot set) -> L2 (wiki, loaded on demand) -> L3 (sources, cold storage). Claude Code uses a thin CLAUDE.md entrypoint that imports AGENTS.md. Includes setup, compilation, read-only health checks, session capture, active interview, follow-up queue review, and gated skill/tool promotion.

Memento can be installed globally while the actual wiki lives in one configured data root. Set MEMENTO_ROOT=/path/to/memento or add a .memento-root file to a project repo; the bundled skills/_shared/scripts/memento-root and skills/_shared/scripts/memento-run helpers make each skill operate from the resolved root instead of assuming the current repo is the Memento.

Actuary — Skill Audit

Audit and evaluate Agent Skills against the agentskills.io specification and authoring best practices. The skill-audit skill separates portable L1 spec compliance from named harness profiles, then reports L2 structure and L3 craft. With --tier, it adds privacy/genericization checks and a static Gate-1 verdict used by Memento's promote flow; final behavioral readiness remains separate.

Steward — Intent and Assurance Contracts

Separate software intent and assurance from an interchangeable builder. frame freezes the minimum decision-complete intent delta, critique conditionally challenges material ambiguity without expanding scope, and assess binds proportionate post-build evidence to a frozen contract and immutable change identity. Any agent, bundled workflow, or human process may construct the change. The dependency-free CLI preserves local immutable revisions without a remote ticket store or builder adapter.

Ostrich — Deliberate Distraction

Break away from overwhelming or unwanted work through one agent-chosen tangent. Ostrich silently excludes the active work cluster, leads a maximally unrelated five-minute escape without presenting another decision, grounds factual tangents when needed, and learns broad preferences without retaining the work context that prompted the break.

Distribution layout

Harness	Marketplace metadata	Plugin manifest	Context
Claude Code	.claude-plugin/marketplace.json	plugins/*/.claude-plugin/plugin.json	CLAUDE.md
Codex	.agents/plugins/marketplace.json	plugins/*/.codex-plugin/plugin.json	AGENTS.md

Keep plugin behavior in SKILL.md; keep wrappers thin. For Memento consumer repos, AGENTS.md is the canonical shared context and vendor entrypoints should only adapt their harness to it.

License

MIT