lazybrain-find

name	lazybrain-find
description	Choose, compare, visualize, and sequence installed local Skills, Plugins, MCP servers, agents, or commands from a natural-language task. Use in Codex Desktop when the user asks what capability to use, gives a vague prompt, cannot remember a tool name, or wants a Codex/Claude Code workflow. Do not install or execute a recommendation without confirmation.

LazyBrain capability router

Turn the user's task into one explainable local capability decision. Codex Desktop is the primary interaction surface; CLI and Claude Code are compatibility surfaces.

Workflow

1. Preserve the user's original task text. Do not invent missing scope.
2. Prefer the read-only lazybrain_recommend MCP tool.
3. If the MCP tool is unavailable but the CLI is installed, run lb desktop "<task>" --json.
4. If neither surface is available, explain that LazyBrain is not ready and suggest npm install -g lazybrain && lb quickstart; do not install or scan automatically.
5. Read the returned decision, source, score, alternatives, and suggested order.
6. If the decision is clarify, ask the returned question instead of guessing.
7. If the decision is compare, explain the user-visible tradeoff among at most three candidates.
8. If the decision is use, recommend the primary capability and explain why it fits.
9. Read desktopVisualization.shouldRender from the result. When it is true and @Visualize is exposed in the current Codex Desktop task because the user selected it in the composer, pass desktopVisualization.visualizePrompt to @Visualize without altering its data.
10. Installing @Visualize does not automatically expose it to every task. When it is not exposed, still rolling out, or fails to render, show the MCP Markdown/table fallback and provide the exact visualizePrompt for a new task where the user selects @Visualize. Do not route this desktop workflow to Codex CLI or claim a cross-plugin call occurred.
11. Treat any suggested workflow as a plan. Selecting a card is not execution. Never invoke, install, enable, or execute another capability unless the user has authorized that action.

Presentation

• Lead with one recommendation or one clarification question.
• Show source and confidence when they help the choice.
• Keep alternatives to three or fewer.
• For two or three close choices, use one compact decision explorer: recommendation first, visible scores and reasons, then optional workflow.
• Preserve the payload's candidate, kind, platform, and workflow controls. Keep essential values visible without hover and retain its keyboard/table fallback.
• When visualization rendering is unavailable, use the tool's Markdown fallback. Do not claim a visualization was rendered without Codex Desktop host readback.

Useful calls

lazybrain_recommend({ "query": "review this payment PR safely" })
lazybrain_catalog({})
lazybrain_orchestrate({ "query": "ship a database migration" })

Use lazybrain_catalog only when the user asks what is installed or the recommendation needs inventory context. Use lazybrain_orchestrate only when the task genuinely needs multiple capabilities.

Safety and privacy

• LazyBrain indexes metadata and names, not secrets or credential values.
• A recommendation is not permission to write files, change configuration, install software, publish, send messages, or perform destructive actions.
• If a candidate's side effects are unknown, say so and require confirmation before execution.