What problem does it solve? Same-model self-review confirms its own blind spots, so AL development artifacts like test specifications, verification plans, and refactor checklists can ship with gaps the authoring model never notices. This Skill sends the artifact to a structurally different model family for an independent read before the caller reconciles its work. ## Core Features & Use Cases - Cross-family review dispatch: Shells to the GitHub Copilot CLI pinned to a GPT model (-m gpt-5.5), guaranteeing the reviewer is a different model family than the Claude-based caller. - Read-only safety envelope: Assembles sandboxing from --deny-tool=write/shell/url, an isolated COPILOT_HOME, disabled MCP servers, and a 600-second timeout so the reviewer can read but never modify the codebase. - Fail-closed skip lines: Returns verbatim gap bullets or an explicit Second opinion skipped: <reason> line, with no retries, no flag widening, and no fabricated output. - Use Case: While running /al-implement on a non-trivial Test Specification, invoke this Skill to have a GPT model list missing concerns as markdown bullets, then reconcile each bullet before writing code. ## Quick Start Ask the agent to get a second opinion on the current test specification or verification plan before reconciling it.
