@dzhechkov/adapter-copilot

The GitHub Copilot platform adapter for DZ Harness Hub — compiles a canonical skill into a GitHub Copilot instruction file.

dz init --target copilot --select design-thinking

What it emits

GitHub Copilot does not scan a .../skills/ directory like Claude Code. It auto-reads repo instructions from .github/instructions/*.instructions.md (each with an applyTo glob). So a skill maps onto an instruction file:

.github/instructions/<id>.instructions.md   # always-on (applyTo: "**") + the skill body
.github/copilot-skills/<id>/…                # the skill's assets (scripts/, references/) — reference-only

Intentionally lossy

This is the harness's first lossy adapter (the five skill-tree adapters — claude/codex/opencode/hermes/openclaude — are lossless). Copilot loses:

• progressive disclosure — the instruction is always-on, not lazily loaded;
• script execution — scripts/ become reference-only (Copilot can't run them);
• skill invocation — Copilot has no per-skill activation command.

Per the @dzhechkov/core Adapter contract, the loss is surfaced as a warning; under strict the adapter throws instead. The canonical skill stays the lossless source of truth — recompiling to claude is still full fidelity.

Lint, not sandbox / runtime: the emitted instruction conveys the skill's methodology; it does not reproduce its tooling.

See docs/research/metaharness-analysis.md §8 and features/copilot-target/ for the design.