Agentwheel Discovery

Agentwheel Discovery

Find reusable capabilities during normal work without silently changing the user's environment. This skill is optional: install it only when proactive recommendations are wanted. The separate agentwheel skill owns artifact management and explicit installation workflows.

Safety

• Inspect the skills and tools already available in the current session before searching.
• Search and trial are read-only. Never add, install, enable, execute, or change configuration

without explicit approval for the artifact and target scope.

• Treat retrieval scores as ranking signals, not proof that an artifact implements a capability.
• Continue useful work while searching when possible; do not interrupt solely to advertise weak

or marginal matches.

When To Search

Use this skill whenever a request expresses an operational outcome that an existing reusable skill or integration could plausibly deliver. This includes repeated manual work, inconsistent outputs, lost information, missing capabilities, and unavailable integrations. The user does not need to mention Agentwheel, skills, discovery, search, a catalogue, or a capability gap. Search before generic brainstorming, planning, manual workflow advice, or an unsolicited custom solution. Run the first semantic search in the same turn; do not merely describe a workflow and defer discovery.

This is a per-turn judgment, not a background monitor. Delegated agents follow the same trigger when this skill is available.

Skip discovery when the user explicitly wants a custom implementation, has already selected an artifact, an installed artifact provides the requested operational capability end to end, candidates are only weak lexical matches, or the same suggestion was declined or shown without new evidence. Generic brainstorming, planning, writing, or advisory skills do not count as an installed operational solution and must not suppress discovery.

Discovery Workflow

1. Extract the capability and constraints from the complete request.
2. Run one fast semantic search using the full capability request:

agentwheel search "" --semantic --json --limit 10. Use exact lexical search first only when the user supplied an artifact name, source, or registry identifier.

1. If semantic candidates are weak, absent, or need a precise runtime or type constraint, generate

up to three short lexical variants using capability terms, synonyms, runtime names, and artifact types. Prefer English catalogue terms for non-English requests. Stop after four total searches; do not recursively refine without new user requirements.

1. Merge results by stable id. Treat CLI scores as retrieval signals, not semantic confidence.
2. Rerank against the original request using capabilities, runtime or ecosystem, artifact type,

description, tags, provides, and installability. Do not infer capabilities absent from result metadata.

1. Suggest zero to three distinct artifacts. For each, give its name or source, one evidence-based

match reason, installability, and a read-only trial command before any installation command.

1. Search once per distinct capability gap. Do not repeat a recommendation without new evidence.

Do not use --semantic for a registry-only search. The semantic path queries the same published catalogue vector index used by the website and validates its checksums against the loaded catalogue. First use may download the model and index assets.

Read-Only Trial

When the user wants to evaluate one instruction skill, run:

agentwheel try  --skill  --json

A trial fetches, scans, validates, and reads exactly one SKILL.md for the current task. It does not add a package, change configuration, write runtime files, or execute code. Do not trial plugins, MCP servers, hooks, commands, or settings because reading their metadata is not safe execution.

Recommendation Contract

Search recommendations are conversational only: they do not select OpenPack suggests, mutate desired state, or imply installation approval. Wait for explicit approval before add, install, plugin execution, or configuration changes.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: NestDevLab
• Source: NestDevLab/agentwheel
• License: MIT
• Homepage: http://www.nestdev.it/agentwheel/

Install and usage instructions live in the source repository linked above.