harness-multitool

0030 — Skills of the harness's own, the same on every adapter

Why

A skill written once should work whatever runs the conversation: the harness's own loop, Claude Code or Codex. Claude Code and Codex each have skills of their own, and those keep working inside them (CON-145). The harness's skills are the portable layer above them.

Decision (draft)

Where skills live

They are laid out as Claude Code lays out its own, with .harness in place of .claude:

<project>/.harness/skills/<folder>/SKILL.md     the project's
~/.harness/skills/<folder>/SKILL.md             the user's, global (MULTITOOL_HOME when set)

SKILL.md starts with YAML frontmatter with name and description, and the body is the instructions. Files beside it (scripts, templates) are the skill's own, read by path.

How an agent uses them

Two tools are served by the harness's MCP server to every adapter. The native loop gets them as tools; Claude Code and Codex get them through the harness server they are already given.

The specification of the agent workflow gains a row: when a task matches a skill's description, skill_load it first.

On the screen

Open questions for the owner

  1. Should the harness also offer the delegates' own skills (~/.claude/skills) to the native loop and Codex, read-only? The draft says no; each program keeps its own, and only .harness/skills is portable.
  2. Should a skill be able to declare which adapters it is for? The draft says no, a skill is for all.

Rejected

This page's source