Implement tool-agnostic invocation guard for audit-pipeline sub-stage skills

[yvonnedevlinrh]

Problem

The audit-pipeline sub-stage skills (adherence.md, mapping.md, scoping.md) use user-invocable: false in their YAML frontmatter to prevent direct invocation. This is a Claude Code-specific field — OpenCode, Gemini CLI, and other tools do not recognize it.

Without a tool-agnostic guard, an agent can auto-trigger a sub-stage directly, bypassing the pipeline router's prerequisite checks:

• adherence requires delta-report.yaml (produced by mapping)
• mapping requires scoping.yaml (produced by scoping)

Skipping prerequisites could produce invalid or incomplete compliance artifacts.

This gap becomes more relevant now that skills are distributed to consumer repos via lola (#119, #126), where the consuming tool may not be Claude Code.

Current state

File	Guard mechanism	Honored by
skills/audit-pipeline/adherence.md	user-invocable: false	Claude Code only
skills/audit-pipeline/mapping.md	user-invocable: false	Claude Code only
skills/audit-pipeline/scoping.md	user-invocable: false	Claude Code only

Options to evaluate

1. Guard text in description — Prefix descriptions with NEVER AUTO-TRIGGER - or similar text that any LLM-based tool would interpret as a constraint. Simple but relies on the model respecting the instruction.

2. Move to agents/ — Restructure sub-stages as subagents (.opencode/agents/) instead of skills. OpenCode's agent system supports subagent delegation, which would give the pipeline router explicit control. Requires updating the dispatching logic in audit-pipeline/SKILL.md.

3. Convention-based approach — Define a frontmatter field or naming convention (e.g., _scoping.md prefix, or a dispatch-only: true field) that all tool integrations agree to honor. Requires cross-tool coordination.

Acceptance criteria

• Sub-stage skills cannot be directly invoked by any supported AI coding tool
• The pipeline router remains the sole entry point for stage dispatch
• Prerequisite validation (scoping before mapping, mapping before adherence) is enforced regardless of tool
• module/ copies are updated to match canonical sources
• Decision is documented (ADR or inline rationale)

Context

• Raised in PR feat: add lola AI Context Module for skill distribution #126 review by @trevor-vaughan
• Related: Add lola AI Context Module for OpenCode skill distribution in consumer repos #119 (lola module), feat: add lola AI Context Module for skill distribution #126 (module distribution)