Agent Kit

A curated collection of skills, Pi extensions, BB plugins, and instructions for AI coding agents. Skills sync to a shared ~/.agents/skills directory and are symlinked into Claude Code, Pi, Codex, and other runtimes; _AGENTS.md is the single instruction file linked to all of them.

Setup

Requires mise.

git clone https://github.com/vaayne/agent-kit.git
cd agent-kit

mise run sync             # everything: skills + instructions + extensions
mise run sync:skills      # link local skills + install remote skills in ~/.agents/skills
mise run sync:agents      # _AGENTS.md → CLAUDE.md / AGENTS.md symlinks for every framework
mise run sync:extensions  # Pi extensions → ~/.pi/agent/extensions symlinks

Development workflow skills

The core of the kit is a development lifecycle where every step is an independently invocable skill, and spec-dev is the thin orchestrator over them:

scout → grill → blueprint → mason → code-review → teach
探地形    拷问     画图纸      施工      审计         内化

Skill	Role
scout	Find your unknowns before they get expensive — blindspot pass, prototypes, references, quadrant diagnostic
grill	Stress-test an idea through structured interrogation — a decision tree worked frontier-first, in rounds
blueprint	Write a decision-first plan.md — decisions with tradeoffs, phased tasks with acceptance blocks
mason	Execute a plan.md phase by phase: catch up → implement → verify → commit → handoff
code-review	Multi-perspective adversarial review with verifier subagents and near-zero false positives
teach	Socratic quiz loop — merge only what you can pass a quiz on
spec-dev	The orchestrator: sequence, review gates, and skip rules over the skills above
refine-code	Improve existing code without changing behavior — Code, Architecture, and Entropy modes: sharpen, deepen, or prove-and-delete
handoff	Transfer context to a fresh focused session

Tool & service skills

Skill	Description
curator	Maintain the nmem knowledge base — lint + synthesis passes
humanizer	Strip AI writing patterns from prose
python-script	Robust Python automation with logging and safety checks
cf-email	Send email through the Cloudflare Email Sending API
gws	Google Workspace operations via the gws CLI
lark-cli	Lark/Feishu workspace operations via lark-cli
openlist	Manage files on OpenList/AList cloud storage

Remote skills installed during sync (see skills/remote-skills.txt): skill-creator, gh-stack, herdr, bento-slides, tailscale, diagram-design.

Extensions (Pi)

Extension	Description
auto-continue-after-compact	Continue the task automatically after a threshold compaction
generic-provider	Register configurable API providers from auth.json, enriched with models.dev
codex-usage-status	Show Codex usage windows in the status line
model-context	Tell the agent its active Pi model without changing the prompt

BB extensions

Extension	Description
workspace-navigator	Browse projects, worktrees, and sessions in a compact sidebar

Install the Workspace Navigator from this repository with:

npm --prefix bb-extensions/workspace-navigator install
bb plugin install ./bb-extensions/workspace-navigator --yes

Project structure

agent-kit/
├── _AGENTS.md    # Shared agent instructions, symlinked to every framework
├── skills/       # Local skills + remote-skills.txt registry
├── pi-extensions/   # Pi extensions
└── bb-extensions/   # BB plugins

Generic Pi providers

generic-provider.ts reads all provider configuration from ~/.pi/agent/generic-provider.json; matching entries in auth.json are neither required nor consulted. Keep this file private because it contains API keys.

{
  "providers": {
    "my-gateway": {
      "baseUrl": "https://gateway.example.com/v1",
      "apiKey": "sk-...",
      "api": "openai-responses",
      "models": {
        "include": ["gpt-*", "claude-sonnet-*"],
        "exclude": ["*-preview"],
        "overrides": {
          "gpt-custom": {
            "contextWindow": 128000,
            "maxTokens": 32000,
            "reasoning": true,
            "thinkingLevelMap": {
              "minimal": "low",
              "xhigh": "xhigh"
            }
          }
        }
      }
    }
  }
}

api supports openai-responses, openai-completions, and anthropic-messages, and defaults to openai-responses. OpenAI providers use {baseUrl}/models; Anthropic providers use {baseUrl}/v1/models. The provider model-list response remains the availability source, then pricing, context limits, and modalities are enriched from models.dev. include and exclude accept * and ? globs. Overrides are keyed by exact model ID and apply last, including when the discovered list came from the local cache.

License

MIT