second-opinion

Second Opinion

Get a second opinion on your current conversation from a different AI model by invoking the other CLI in a fresh local shell session.

When to use

Invoke with /second-opinion mid-conversation when you want another model to review the current conversation's accuracy, approach, or any specific concern.

Usage

/second-opinion [optional review focus]

Examples

/second-opinion
/second-opinion is this database schema correct?
/second-opinion check the security implications
/second-opinion are there edge cases I'm missing?

Instructions

When this skill is invoked:

1. Parse arguments: If the user provided text after /second-opinion, use it as the review focus. If no arguments, the default focus is accuracy, approach, and completeness.

2. Run the review script:

npx @crown-dev-studios/skill-issue second-opinion --cwd "$PWD" --source SOURCE --session-id SESSION_ID [--focus "REVIEW_FOCUS"]

• The --focus flag is only needed if the user specified a custom focus.
• This command requires @crown-dev-studios/skill-issue to be resolvable by npx. Copying the skill directory alone is not enough.
• When this skill runs inside Claude, pass --source claude --session-id "${CLAUDE_SESSION_ID}" so Codex reviews the current Claude session deterministically.
• When this skill runs inside Codex, pass --source codex --session-id "${CODEX_THREAD_ID}" so Claude reviews the current Codex thread deterministically.
• The caller must pass both --source and --session-id explicitly. The CLI does not auto-detect them.
• The script calls the other CLI directly as a subprocess and parses its structured JSONL output.
• It reads the full session file from disk (not affected by context compaction).
• Timeout is 5 minutes by default. Override with --timeout-ms if needed.
• For local development from a source checkout, anchor the command to that checkout, for example pnpm --dir /absolute/path/to/skill-issue run second-opinion -- --cwd "$PWD".

3. Present the review: Show the reviewer's output to the user. Add a brief note about which model reviewed it.

Options

The script supports these flags for advanced use:

• --source claude|codex — Source conversation to review
• --session-id <id> — Session or thread ID to review
• --reviewer claude|codex — Force which CLI does the review (defaults to the opposite of source)
• --timeout-ms <n> — Reviewer timeout in milliseconds (default: 300000)
• --include-thinking — Include chain-of-thought reasoning in the review context
• --extract-only — Just print the extracted conversation without calling a reviewer
• --max-chars N — Max conversation characters to send (default: 200000)

Defaults:

• Claude: claude --dangerously-skip-permissions --verbose --output-format stream-json --include-partial-messages -p
• Codex: codex exec --json --dangerously-bypass-approvals-and-sandbox

The review prompt is written to reviewer stdin, and the CLI parses the structured JSONL response to return the final assistant review text.

Deterministic Session Selection

The caller must pass the live session ID for the current tool instead of asking the CLI to infer it.

Claude

When this skill runs inside Claude, the source transcript must be the current Claude session.

• Use the live Claude session ID provided to the skill as ${CLAUDE_SESSION_ID}.
• Pass that value explicitly to the review script with --source claude --session-id.
• If ${CLAUDE_SESSION_ID} is unavailable, stop and explain that deterministic Claude transcript selection is not possible in the current environment.

Example:

npx @crown-dev-studios/skill-issue second-opinion --cwd "$PWD" --source claude --session-id "${CLAUDE_SESSION_ID}" [--focus "REVIEW_FOCUS"]

Codex

When this skill runs inside Codex, the source transcript must be the current Codex thread.

• Use the live Codex session ID provided to the skill as ${CODEX_THREAD_ID}.
• Pass that value explicitly to the review script with --source codex --session-id.
• If ${CODEX_THREAD_ID} is unavailable, stop and explain that deterministic Codex transcript selection is not possible in the current environment.

Implementation Requirements

The review script should resolve sessions from the explicit --source and --session-id pair provided by the caller. It must not guess from runtime env vars or modification-time heuristics.

How it works

1. Finds the current session JSONL file on disk:
   • Claude Code: ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl
   • Codex: ~/.codex/sessions/<year>/<month>/<day>/rollout-<id>.jsonl
2. Parses all user messages, assistant responses, chain-of-thought, and tool usage
3. Strips system prompts, permission blocks, and injection artifacts
4. Truncates intelligently if the conversation exceeds the context limit (keeps beginning + end)
5. Sends to the other CLI as a direct subprocess:
   • If in Claude Code → codex exec for review
   • If in Codex → claude -p for review
6. Parses the structured JSONL response and returns the final review text