cmd-review

Command Code (cmd) Review

Use Command Code (cmd) to get a second opinion. Model is optional — cmd uses its default unless a model is given. All execution goes through review.js.

Locating review.js

Find the script with:

printf '%s\n' ~/.claude/plugins/cache/second-opinion-skill/second-opinion-skill/*/bin/review.js 2>/dev/null | sort -V | tail -1

Store the result as REVIEW_SCRIPT. Do not call cmd directly.

Determining what to review

Pass the appropriate flag to review.js (it handles fetching and embeds diff/file content inline before the prompt):

What to review	Flag
Unstaged changes	--diff=unstaged
Staged changes	--diff=staged
Last commit	--diff=last-commit
Branch vs main	--diff=branch
Custom range	--diff="HEAD~3..HEAD"
Specific file	--file=<absolute-path>
General question	(no flag)

Running

"$REVIEW_SCRIPT" --engine=cmd --cwd=<repo-path> [--diff=<spec>|--file=<path>] "<review template>"

Model is optional. Bare --engine=cmd lets cmd pick its default. To pin a model, list the choices and pass the id verbatim:

cmd --list-models   # grouped list, e.g. "claude-sonnet-4-6", "gpt-5.5", "deepseek/deepseek-v4-flash"
"$REVIEW_SCRIPT" --engine=cmd:claude-sonnet-4-6 --cwd=<repo-path> [...] "<review template>"

review.js splits the engine spec on the first : only, so any model id survives intact. The runner launches it as cmd --print --permission-mode plan --skip-onboarding [-m <model>] "<prompt>", so all calls run non-interactively in read-only plan mode with taste-onboarding skipped (no interactive hang in an automated run).

Composing the prompt

For the full guidance on how to compose the prompt — when to embed the user's ongoing task / context (Tier A), when to fall back to default templates (Tier B), and when to use --no-embed for very large diffs (Tier C) — read the second-opinion skill. The default templates live there too.

Safety toggle

By default review.js applies this engine's read-only / plan-mode flags (--permission-mode plan). Pass --unrestricted only when the engine genuinely needs to edit files or run commands; review.js will drop the safety flags and log a stderr warning. (--skip-onboarding is functional, not a safety flag, so it stays on even with --unrestricted.)

Output envelope

review.js wraps every prompt with <<<SECOND_OPINION_START>>> / <<<SECOND_OPINION_END>>> markers and asks the engine to emit its real answer between them. After reading the log file, extract the text between the markers — that is the clean payload, free of reasoning traces and tool noise. Pass --no-wrap to disable.

Capturing output

When review.js runs from an agent harness (non-TTY stdout), engine output is not streamed to stdout. It's written to a temp file, and stdout receives only a banner with the log path. After the command exits, use the Read tool on the path shown after LOG FILE: to get the full Command Code review. Don't pipe to | tail -N / | head -N — the engine output isn't on stdout in this mode. Pass --log=<path> for a known location; pass --log=- to restore tee-to-stdout behavior.

Presenting results

Show Command Code's full response under a ## Command Code's Take heading. Don't filter or summarize — let the raw review speak. If Command Code raises issues that need fixing, address them and note what changed.