Skip to content

Review

/mx:review is the starter dose. Run it on any changes and it bundles all eight review-grade agents into a single verdictPASS, PASS WITH WARNINGS, or REJECT — with per-agent findings and a recommendation you can act on.

It’s the refusal layer in one command: it interrogates code (yours or an AI’s) and pushes back on silent failures, hallucinated APIs, suppressed errors, weak types, missing tests, and unjustified shortcuts. It is read-only — it reports and recommends, it never modifies your code.

Review-grade verdict on your changes

/mx:review [--strict | --advisory] [--staged | --commit <sha> | --branch <name>] [--scope <path>]

By default it reviews everything that differs from trunk — both your committed branch work and your uncommitted changes (including brand-new untracked files, which is where AI-generated code most often hides). Because the scope is computed from the merge-base with main/master, it also does the right thing when you’re sitting on trunk with just uncommitted edits.

When to use it:

  • Before opening a PR — get a verdict on the whole branch
  • After Claude (or Cursor, Copilot, Claude Tag…) generates a chunk of code, to interrogate it
  • Any time you want a fast, opinionated quality read on your changes

Always: mx-code-reviewer, mx-silent-failure-hunter, mx-performance-auditor, mx-quality-keeper (structural checks). Conditionally: mx-type-design-analyzer (if types changed), mx-mr-test-analyzer (if tests changed), mx-comment-analyzer (if comments changed). Finally mx-code-simplifier runs in report-only mode — it surfaces simplification suggestions but never edits files here.

The verdict format is always Pass / Warnings / Reject plus per-agent findings. What flips it to REJECT is configurable:

ModeFlagREJECT when
Strict--strictany CRITICAL or HIGH finding
Balanced (default)(none)any CRITICAL finding (HIGH → warning)
Advisory--advisorynever blocks — labels findings by severity

Whatever the mode, the report always includes a recommendation with reasons — and in advisory mode it still tells you what the verdict would be under balanced rules, so the signal is never lost.

FlagEffect
(none)Committed branch work and uncommitted changes vs trunk
--stagedOnly staged changes
--commit <sha>A single commit’s diff
--branch <name>Compare against an explicit base branch
--scope <path>Limit to a directory or file

A terminal summary with the verdict, a per-agent findings table, and the top findings by severity. The full report is saved to .agents/reports/review-{YYYY-MM-DD}.md.

/mx:review vs /mx:qa/mx:review is the headline interrogation: all eight review agents, one verdict, tuned for “is this code good enough to merge?” /mx:qa is the deeper structural + spec-conformance audit (suppressions, dependencies, contract and requirement verification) best run after a feature or team build. Reach for /mx:review constantly; reach for /mx:qa --full at milestones.