Five reviewers, one install: parallel multi-perspective code review

A single reviewer — human or model — reads with one set of eyes. They catch what they’re primed to catch and slide past the rest. Security thinking crowds out naming nitpicks; architecture worries bury the missing test.

Zensu’s review chain splits the job. When the agent tries to call a feature record done, it fans out five read-only reviewers in parallel, each with exactly one lens:

• conventions — does it match the codebase’s established patterns?
• bugs — what breaks at the edges?
• architecture — does the shape hold as the system grows?
• tests — is the new behavior actually pinned down?
• security — what does an attacker do with this?

Each loads its own checklist and nothing else, so none of them dilutes its focus. Their findings merge in the main thread — deduplicated by file and line, sorted critical-first — into one report you actually read, instead of five you skim.

implementation complete  →  review chain (5 lenses, parallel)

  conventions   2 findings
  bugs          1 critical
  architecture  0
  tests         1 important
  security      1 critical

merged  →  deduped by file:line · critical-first

  CRITICAL  bugs       auth.ts:42   token compared with == (timing leak)
  CRITICAL  security   auth.ts:42   same line — confirms the leak
  IMPORTANT tests      auth.ts      no test for the expired-token path

Read-only on purpose

The reviewers don’t edit. They surface; you decide. Critical and important findings route back into the same TDD loop that built the code, so a fix is another red-green cycle — not an unreviewed patch bolted on at the end.

The result is the breadth of a review panel with the cost of a single pass, and a token budget that doesn’t balloon because each lens carries only what it needs.

None of it is something you request. The chain runs on the way out of the turn, and what it produces lands on the feature record — which is also what the release gate reads later. Install it for Claude Code or Kiro.