light-code-architectureunit

name	light-code-ArchitectureUnit
description	Audit a Delphi project's own units (files) for SHALLOW modules that should merge into one deep module — information leakage, temporal decomposition, pass-through layers, always-paired usage. Reports ranked merge candidates; never merges. Say "find shallow units", "what units should be merged". For CLASS shapes use light-code-ArchitectureClass.

/light-code-ArchitectureUnit — Find shallow units (launcher)

This skill is a thin launcher. The audit is done by the light-code-ArchitectureUnit agent in its own context window, so the whole-project read does not clog the main context — only the ranked report comes back. Do not map units or analyse them yourself; resolve the scope, then launch the agent.

When to run

• User invokes /light-code-ArchitectureUnit (with or without args).
• User says variants: "find shallow units", "what units should be merged", "which files should be folded together", "audit the module structure".

If the real question is about a class shape (god class, anemic class, misplaced responsibility), this is the wrong skill — point at /light-code-ArchitectureClass.

Steps

1. Resolve scope. From $args: a path, a folder, or several files. With no argument, the scope is the project's own app source under the project root. Pass --include-lightsaber through if the user gave it. You do not need to ask anything — the agent applies the skip rules (LightSaber, third-party, tests, forms) itself.
2. Launch the light-code-ArchitectureUnit agent via the Agent tool with subagent_type: light-code-ArchitectureUnit. Put the resolved scope (and any flags) in the prompt.
3. Print the agent's returned report verbatim. It is already self-contained (ranked candidates + verdict line). Do not re-analyse or second-guess its verdicts.

That's it. No mapping, no signal-walking, no edits in the skill itself.

Where the rules live (for the agent, not for you)

• references/architecture-principles.md — the shared deep/shallow model, counter-analysis discipline, LightSaber boundary, risk flags (read by both this and the ArchitectureClass agent).
• The agent's own file (.claude/agents/light-code-ArchitectureUnit.md) — the unit-specific signals, anti-signals, procedure, report format, and hard rules.