LLMSkillHub
mirroredCopied from ShenSeanChen/waku-agent under MIT. ShenSeanChen has not claimed this page, and is not affiliated with LLM SkillHub.

shenseanchen / new-tool

Add a new tool to Waku the right way — schema, safe execution, deterministic eval, honest output. Use when adding or modifying agent tools or when the user asks for a new capability.

0
SKILL.md
1.3 KB

Procedure

  1. One file per tool in waku/tools/, exposing make_tool(...) -> Tool.
    Copy the shape of waku/tools/calendar.py (the reference implementation).

  2. The tool function must be:

    • Idempotent where repeat calls could occur (check-before-insert, like
      create_event's title+start guard).

    • Honest in its return string: say exactly where the artifact went
      (file path, table) and how the user can see it. Never imply effects the
      tool doesn't have.

    • Local-first by default: write to .waku/ or state.db; real external
      services go behind env-gated adapters.

  3. Register it in waku/tools/__init__.py:build_registry.

  4. Add deterministic eval coverage in evals/deterministic/:

    • offline: scripted client fires the tool → assert the artifact (DB row/file)
    • a dataset case in evals/dataset.jsonl for the live tier

  5. If the tool needs usage guidance, add a rule to DEFAULT_SOUL
    (waku/runtime/session.py) or a skill in skills/.

  6. Check scope first: the repo ships flagship-task tools only. New capabilities
    beyond scheduling/notes/messages should be discussed before building.