cds-second-opinion

name	cds-second-opinion
description	Get a second opinion on your current conversation from a different AI model. Routes to Claude from Codex and Codex from Claude using a fresh local shell session.
argument-hint	[optional review focus]

Second Opinion

Get a second opinion on the current conversation from a different AI model by invoking the other CLI with a pointer to this session's transcript on disk.

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:

Step 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. If the user asked for the reviewer to see the reasoning/chain-of-thought, note that for Step 4.

Step 2: Identify the Current Session

The transcript must be the current thread — never guess the session from file modification times.

• Inside Claude Code: the session ID is ${CLAUDE_SESSION_ID}.
• Inside Codex: the session ID is ${CODEX_THREAD_ID}.

If the variable for the current environment is unavailable, stop and explain that deterministic transcript selection is not possible here.

Step 3: Resolve the Transcript Path

• Claude Code session: ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl, where <encoded-cwd> is the absolute working directory with / replaced by -. If that exact path does not exist, search ~/.claude/projects/*/<session-id>.jsonl.
• Codex session: search ~/.codex/sessions/ (dated subdirectories) for rollout-*<session-id>*.jsonl.

Verify the file exists before proceeding; if it cannot be found, stop and report the paths checked.

Step 4: Render the Reviewer Prompt

Create a scratch directory (mktemp -d). Render reviewer.md into <scratch>/prompt.md, filling every placeholder:

• {{REVIEWER_NAME}} — the reviewing model's CLI (Codex when running from Claude Code, Claude when running from Codex)
• {{SOURCE_TOOL}} — Claude Code or Codex
• {{TRANSCRIPT_PATH}} — absolute path from Step 3
• {{REVIEW_FOCUS}} — from Step 1
• {{OUTPUT_PATH}} — absolute path to <scratch>/review.md

By default the template tells the reviewer to skip reasoning/thinking content so the opinion stays independent; if the user explicitly asked for reasoning to be included, remove that instruction from the rendered prompt.

Step 5: Invoke the Other CLI

Run the opposite CLI in the background (a review can take several minutes):

• From Claude Code:

  codex exec --dangerously-bypass-approvals-and-sandbox \
    --output-last-message "<scratch>/last-message.txt" \
    "$(cat "<scratch>/prompt.md")" > "<scratch>/reviewer.log" 2>&1
  

• From Codex:

  claude --dangerously-skip-permissions -p \
    "$(cat "<scratch>/prompt.md")" > "<scratch>/reviewer.log" 2>&1
  

If the reviewer CLI is not on PATH, stop and tell the user which CLI is required.

Poll for <scratch>/review.md roughly every 30 seconds rather than blocking indefinitely. If the process is still running with no review after ~15 minutes, tell the user and let them decide whether to keep waiting — do not kill the reviewer and do not retry silently.

The run succeeded when <scratch>/review.md exists; reviewer.log is diagnostic only. If the process exits without writing review.md, fall back to the reviewer's own final message before declaring failure:

• From Claude Code: use <scratch>/last-message.txt if it is non-empty — --output-last-message is Codex's designed handoff artifact and carries the complete final message. The non-empty check matters: codex exec has a known regression class of exiting 0 with empty output in detached contexts, so an empty file means the run failed, not that the review was empty.
• From Codex: reviewer.log itself is the fallback — claude -p prints exactly the final message to stdout.

If the fallback is also empty, report the failure with the tail of reviewer.log — do not retry silently.

Step 6: Present the Review

Relay review.md to the user, noting which model produced it and printing the file's absolute path. The scratch directory is temporary — if the review is worth keeping, offer to copy it somewhere durable (e.g. docs/reviews/second-opinion-<date>.md, matching where cds-review-council writes its runs). Do not soften or re-adjudicate the reviewer's opinion — disagreements are signal.

How It Works

• The harness provides the current session ID; this skill only turns it into a transcript path.
• The reviewer is a full agent with its own file tools: it reads the on-disk transcript itself (unaffected by context compaction) and selectively, instead of receiving a pre-truncated export.
• There is no companion CLI or build step. The only subprocess is the other vendor's CLI, spawned directly.