Second O Pi Nion — AI skill for Claude Code

second-o-pi-nion

A [pi-coding-agent](https://github.com/mariozechner/pi) extension that lets the active LLM ask a **different** LLM for a quick second opinion when it gets stuck, goes in circles, or wants a sanity check on its reasoning.

Anecdotally, when an LLM is stuck in its own reasoning loop, being forced to articulate the problem to another model (and getting a fresh perspective back) unblocks it surprisingly often. This extension gives the model that escape hatch as a first-class tool.

🔒 Privacy / data flow — read this before installing

**TL;DR: this extension ships parts of your pi session to a third-party LLM. Don't paste secrets into your prompts.** The included regex redaction is best-effort hygiene, not a security control.

This extension sends parts of your pi session to a third-party LLM provider (OpenAI, Google, xAI, Anthropic, or whatever you have configured — including GitHub Copilot if you fall back to that). Specifically:

• second_opinion tool — sends the problem and context strings the active LLM passes in. Those strings are LLM-controlled. The model can put anything from your project into them.
• /second-opinion command — sends the literal question you typed.
• /second-opinion-on-last-turn command — sends a transcript of the last N user/assistant turns (default 5), including:
  • your previous user messages,
  • the assistant's previous responses,
  • summaries of tool calls (tool name + first 200 chars of arguments),
  • tool results are NOT included.

What this means in practice

• File paths, command lines, code snippets, and anything you've pasted into the conversation can be sent to the reviewer model.

• A best-effort regex layer redacts common token shapes before sending. **This is damage reduction, not protection** — treat it as a nudge, not a security control. Currently covered: