@dzhechkov/adapter-claude

The Claude Code platform adapter for the DZ cross-platform harness.

It implements the Adapter contract from @dzhechkov/core: it compiles a CanonicalSkill into the file layout Claude Code expects — .claude/skills/<id>/SKILL.md plus any bundled scripts/, references/, and assets/ files.

Why this adapter is special

Claude Code's skill format is the Agent Skills open standard. So this adapter performs no lossy transformation — it is a faithful, lossless emitter. That makes it the byte-identity baseline: compiling a skill that was loaded from a .claude/skills tree reproduces the original SKILL.md byte-for-byte. The other adapters (adapter-codex, adapter-opencode, adapter-hermes) are measured against this one.

API

import { claudeAdapter } from '@dzhechkov/adapter-claude';

const emit = await claudeAdapter.compile(skill, { targetRoot: '.' });
// emit.files -> [{ path: '.claude/skills/<id>/SKILL.md', content, encoding }, ...]

const result = await claudeAdapter.verify(emit);
// result.ok -> boolean

compile is a pure function — it returns the files to write, it does not touch the filesystem. Applying an EmitResult to disk is the harness's job.

Companion integrations

claudeIntegrationAdapter.plan() is also pure. For a validated INTEGRATIONS.json it returns only the project-scoped .mcp.json fragment supported by the pinned evidence family; it does not write, merge, spawn, or contact a manifest URL. harness-core performs the ownership-safe transaction and the live registration probe. Hook intent is refused until an activation/negative-control receipt exists; it is never silently skipped.

Status

0.2.7 — staged, not published. Skill compilation remains lossless; the separate pure integration planner emits only the receipt-proven Claude project MCP fragment.