second-opinion

Get a quick, independent second opinion from one agent on a decision, design choice, or approach.

Raw slash-command arguments: $ARGUMENTS

Agent selection:

• Use --agent gemini (default), --agent claude, --agent codex, --agent cursor, or --agent kilo.
• Default is Gemini — fastest responses, good for quick sanity checks.
• Use --agent claude when you want depth and edge-case analysis.
• Use --agent codex when you want a terse "is this over-engineered?" check.
• Use --agent cursor for integration and codebase-fit questions.
• Use --agent kilo for naming, readability, and maintainability questions.
• If the chosen agent is not installed, the companion automatically falls back to the next available agent and tells you which one it used. Default fallback order (no --agent flag): gemini → claude → codex → kilo → cursor. When --agent is specified and unavailable, fallback iterates agentDefs insertion order instead.

Execution mode:

• Default: foreground (quick — one agent only).
• If --background is in arguments, run background mode.
• If --json is in the arguments, the companion emits {"command":"second-opinion","results":[{name,output,error,exitCode}]} on stdout instead of delimited text. Warnings still go to stderr.
• Strip --background, --wait, and --json before passing to companion.

Pre-flight: The companion checks whether the chosen agent is installed and falls back to the next available one if not. If no agents are available at all, it exits non-zero — tell the user to install at least one agent.

Foreground execution:

node "$CLAUDE_PLUGIN_ROOT/scripts/companion.mjs" second-opinion $ARGUMENTS

Background execution:

Bash({
  command: `node "$CLAUDE_PLUGIN_ROOT/scripts/companion.mjs" second-opinion $ARGUMENTS`,
  description: "Second opinion",
  run_in_background: true
})

Tell the user: "Second opinion started in the background. You'll be notified when it completes."

After the companion exits:

Return the agent's output verbatim — do not add your own commentary unless the user asks. The agent is instructed to give a verdict (approve / approve-with-caveats / reject), so the output is self-contained.