second-opinion

name	second-opinion
description	Get an independent review of completed work from OpenAI's gpt-5.6-sol with high reasoning effort — sends the work and test-tree contract through the Responses API with OPENAI_API_KEY and surfaces its critique. TRIGGER when: sync has just finished, the user asks for a second opinion, an independent review, or a sanity check on completed work before a PR or release.

Second Opinion

Sends the completed work to OpenAI's gpt-5.6-sol with high reasoning effort and surfaces its independent review. Once sync reports the trees and implementation are aligned, a second pair of eyes catches what a single perspective misses: bugs, drift from the test-tree contract, rule violations, gaps the author rationalised away.

When to Use

• Immediately after sync reports the project is in sync — the natural next step.
• Before a PR or release, as a final independent check on completed work.
• When the user asks for a second opinion, an independent review, or a sanity check.

Process

1. Gather the work to review

Work out what to review before gathering it. Rely on natural language, not on a fixed git boundary:

• If the user gave a natural-language indication of what to review — "the second-opinion change", "the last thing we did", "everything since the refactor" — review exactly that.
• Absent a clear indication, review the current worktree.

Gather the current worktree's tracked changes and untracked file contents:

CHANGE=$(git diff HEAD; git ls-files --others --exclude-standard | while read -r f; do printf '\nUNTRACKED FILE: %s\n' "$f"; cat -- "$f"; done)

Then read ## Test Trees (or TEST_TREES.md) — this is the contract the work must satisfy. If the current worktree contains no non-trivial work to review, say so and stop without calling the API.

2. Ask an independent model to review against the contract

Call OpenAI's Responses API at https://api.openai.com/v1/responses with OPENAI_API_KEY, the gpt-5.6-sol model, and high reasoning effort. Send the work and the test trees, and ask the model to review the work as an independent critic: does the change satisfy the test-tree contract, are there bugs or drift, does it honour the rules (KISS, YAGNI, no fake code, fail-fast, no comments), what database schema or API contract changes does it introduce, what does it impact in other systems, and what would it change?

REVIEW_INPUT=$(printf 'TEST TREES (the contract):\n\n%s\n\nCHANGE (the work to review):\n\n%s\n' \
  "$(cat TEST_TREES.md)" "$CHANGE")
REVIEW_BASE_URL="${OPENAI_BASE_URL:-https://api.openai.com/v1}"
REVIEW_MODEL="gpt-5.6-sol"

REVIEW_RESPONSE=$(curl -sS -f -X POST "$REVIEW_BASE_URL/responses" \
  -H "Authorization: Bearer ${OPENAI_API_KEY:?OPENAI_API_KEY is required}" \
  -H "Content-Type: application/json" \
  -d "$(jq -n --arg input "$REVIEW_INPUT" --arg model "$REVIEW_MODEL" '{
        model: $model,
        reasoning: { effort: "high" },
        instructions: "You are an independent code reviewer. Review the completed work against the test-tree contract. Review database schema changes, API contract changes, and impacts on other systems. Surface bugs, drift from the trees, rule violations, and gaps. Be specific and concrete.",
        input: $input
      }')")
REVIEW=$(printf '%s' "$REVIEW_RESPONSE" | jq -er '[.output[]? | select(.type == "message") | .content[]? | select(.type == "output_text") | .text] | join("\n") | select(length > 0)')
printf '%s\n' "$REVIEW"

3. Surface the review

Present the returned review to the user verbatim, attributed to gpt-5.6-sol so it is clear this is a second opinion, not your own. The user decides what to act on; where the review finds drift or gaps, route them back through change, sync, or tdd.

Failure is loud

If the review request fails — OPENAI_API_KEY is missing, the API returns an error or non-2xx response (curl -f), or the response contains no review (jq -e) — surface the failure as an error and stop. Never fabricate a review, summarise the diff yourself and pass it off as the independent model's review, or report a second opinion you did not get. A missing review is an honest outcome; a fake one is not.