leehow/pipiui/create-subagent

Trust tier 3
  • community
Source check pending

Checking that the SKILL.md file is still in the repository.

Create a least-privilege schema v1 PipiUI subagent package through the bundled subagent_manage tool.

Install

npx vskill@latest install leehow/pipiui/create-subagent

SKILL.md

Create a PipiUI subagent

Agent, Skill, and Prompt

Use this only in the main session. This Skill is guidance for the current main session; it is not an Agent and grants no tools. Dispatched workers run skill-free and must not load this Skill. A Prompt is only task text, not a reusable definition or permission grant.

The bundled subagent_manage tool is the authority for agent package scaffolding, validation, permission summaries, and installation. Bundled Skill names are retained and bundled-first: create-subagent is the built-in workflow for this task, so a same-named user Skill cannot replace it. Do not copy a schema by hand or write AGENT.md directly when that tool is available. The template beside this Skill is documentation/test material only; runtime rules come from the bundled subagent extension's agents.ts parser.

Collect only what is needed

Ask only for information absent from the request:

  1. Scope — user or project.
  2. Name — concise lower-kebab-case name.
  3. Description — one clear sentence about role and outcome.
  4. Mode / optional model — read-only or worker; omit model unless the

user needs one.

  1. Minimum capabilities — filesystem, shell, web, exact MCP tool names,

desktop, and delegation.

Do not ask for accounts, login, credentials, marketplaces, remote installs, or infrastructure details. The least-privilege baseline is read-only with no filesystem/shell/web/MCP/desktop/delegation, worktree: none, and a report.

Tool-first workflow

  1. If the user wants to inspect existing roles, call

subagent_manage({action:"list", scope:"user"|"project"|"both"}) or inspect. Read its diagnostics; a shadowed or malformed definition is not a reason to silently invent a replacement.

  1. Call `subagent_manage({action:"scaffold", scope, name, description, mode?,

model?, capabilities?, worktree?, deliverable?, tools?, prompt?})`. It returns schema v1 content, the exact target path, parser diagnostics, and the capability/tools permission summary.

  1. Call subagent_manage({action:"validate", scope, name, draft}) on the

proposed content. Use the returned structured errors/warnings and permissionSummary; do not imitate the parser in prose.

  1. Before any install, show the user the proposed scope/path, mode/model,

worktree/deliverable, declared capabilities, explicit tools narrowing, and the tool's runtime caveats. In particular: global disabled-tools and missing extensions can still remove tools; desktop: requestable needs the existing global Computer Use host gate plus a per-task grant; delegation is depth and runtime-role limited.

  1. After the user approves that preview, call

subagent_manage({action:"install", scope, name, draft, overwrite?}). Omit overwrite by default. Set overwrite:true only when the user explicitly approved replacing that exact existing package. Report the tool's final install path and verification result.

install only writes a validated standard package under the selected user or project agent root. A project install writes <project>/.pi/agents/<name>/AGENT.md; it still requires the existing runtime project-agent confirmation when that role is run. It rejects bundled targets, traversal, symlink escape, and arbitrary output paths.

Capability choices

  • read-only must use worktree: none; it cannot use

filesystem: workspace-write.

  • Use worktree: isolated only for a worker that actually needs isolation.
  • MCP is false by default. If needed, give only exact registered

mcp_<server>_<tool> names. Never use true, *, a server-wide grant, or a guessed name.

  • desktop: requestable is only a request for per-task desktop access. Never

put computer or open_application in tools.

  • web: true is retrieval capability, not a browser grant. Never put

browser in tools.

  • Set delegation only for a real orchestration role. It does not let a worker

load Skills or self-grant desktop/search access.

  • tools is optional and only narrows capability-derived tools; it cannot add

a tool or bypass a capability.

Keep the flow local, direct, and preview-first. No remote market, account, or pairing ceremony is needed.

Security

  • Security scanPassed, 90 of 100Pattern checks on the skill's files.
  • Model reviewPassed

Quality evaluation

  • Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

Source

Popularity

7-day trend score
0

Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Codex
  • Gemini CLI
  • Cline
  • Amp
  • Roo Code
  • Goose
  • Continue
  • Kilo Code
  • Junie
  • Trae
  • OpenHands
  • Qwen Code
  • AI Agent
  • Tabnine Chat
  • Sourcegraph Cody
  • Amazon Q
  • Firebase Genkit
  • Sweep
  • Devin
  • Phind
  • Mentat
  • Claude Desktop
  • Aider
  • GPT Pilot
  • Plandex
  • bolt.new
  • v0
  • Replit Agent
  • Anse
  • TypingMind
  • BetterChatGPT
  • LibreChat
  • Warp AI
  • Pieces
  • Fig AI
Embed badge
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/leehow/pipiui/create-subagent/badge)](https://verified-skill.com/skills/leehow/pipiui/create-subagent)

Something wrong with this skill or its page? Tell us.