preetamnath/second-opinion

Second Opinion

Steps

Step 1 — Brief the subagent

If there's no concrete anchor (specific edit, chosen option, phrasing under review), stop and ask the user for one.

Otherwise, spawn one subagent with a briefing containing exactly:

• Context — file path(s), surrounding code/prose, constraints.
• Problem statement — what the proposal is trying to solve and why this approach.
• The proposed fix — verbatim.
• What's already ruled out — alternatives the user has considered and rejected, with the reason.
• The 3-part ask. The subagent returns:
  1. Rate the proposal. Pros, cons, and a confidence score (0.00–1.00) that it improves on the current state.
  2. Generate 2–4 alternatives. Each with proposed option, reasoning, and a confidence score (0.00–1.00) that it beats the user's proposal.
  3. Flag additional issues. Anything the proposal does not address — scope ambiguity, missing trigger classes, downstream impact.

Step 2 — Synthesize back

• Table:

| # | Option | Tradeoff | Reviewer conf. | |---|---|---|---| | 1 | User's proposal — [content] | … | 0.xx (vs. status quo) | | 2 | Alt 1 — [content] | … | 0.xx (vs. proposal) | | … | up to 4 alternatives, ordered by reviewer confidence | | |

• Additional issues flagged: reviewer-surfaced blind spots not addressed by any alternative (omit if none).
• Recommendation: your synthesized pick with brief reasoning and your own confidence (0.xx, distinct from the reviewer scores above).
• Checkpoint: call AskUserQuestion with the top 2–3 options.