ultrareview

name	ultrareview
description	Multi-agent code review on the current git branch — spawns three parallel sub-reviews (correctness, architecture, UX/a11y) and aggregates a single structured report.
metadata	{"tarsee":{"emoji":"🔬","category":"review","triggers":["/ultrareview","ultrareview","do a deep code review"]}}

UltraReview

Parallel multi-agent code review for the CURRENT git branch vs main. Uses Tarsee's agent registry (Coder, Researcher, Writer) to run three independent review passes simultaneously, then aggregates.

When to use

• User says /ultrareview
• User asks for a "deep review", "thorough code review", or "multi-lens review" on the current branch
• About to merge a PR and wants a final look

Steps

1. Scope check (run once, synchronous). Before spawning sub-agents:

   • git rev-parse --abbrev-ref HEAD → confirm branch name
   • git log --oneline main..HEAD → confirm this is a feature branch with actual commits (not empty or equal-to-main)
   • git diff --stat main..HEAD → summarize scope before the review
   • If branch equals main or has zero diff: stop and tell the user, don't waste tool calls on a no-op review.

2. Spawn three parallel sub-agents via the Task tool. Kick them off in ONE message (single assistant turn with three Task tool calls) so they run concurrently, not serially.

   Agent A — Correctness (Coder): Prompt: "Review the diff main..HEAD strictly for correctness: bugs, logic errors, null-handling, race conditions, off-by-ones, error paths that swallow exceptions, edge cases not covered. Cite file:line. Ignore style, naming, and architecture — those are other agents' jobs. Return: ranked findings CRIT/HIGH/MED/LOW, under 400 words."

   Agent B — Architecture & Consistency (Researcher): Prompt: "Review the diff main..HEAD for architectural fit: does this match existing patterns in the codebase? Are there duplicate implementations when a helper already exists? Does it break a layering invariant? Are new abstractions justified by >= 3 call sites? Return: ranked findings with existing-alternatives cited by file:line, under 400 words."

   Agent C — UX / A11y / Copy (Writer): Prompt: "Review the diff main..HEAD for user-facing quality: any UI text changes that read awkwardly, any a11y regressions (missing aria labels, focus ring loss, touch targets <44px), any breaking visual changes without theme/token usage, any docs/error messages that don't match existing voice. Return: ranked findings, under 400 words. If no UI changes in diff, say so in one line."

3. Aggregate results. Once all three Task results arrive, combine into one report with this structure:

   # UltraReview — <branch name>
   **Scope:** N files, +X −Y lines
   
   ## 🔴 Critical
   <union of CRIT findings from all three agents, deduped>
   
   ## 🟠 High
   <union of HIGH findings>
   
   ## 🟡 Medium
   <union of MED findings, grouped by agent>
   
   ## 🟢 Low / style
   <union of LOW findings, bulleted>
   
   ## Recommendation
   <one paragraph: ship as-is / ship with minor fixes / block on X>
   

4. Don't implement fixes unless asked. This is a review skill, not a fix skill. End with: "Want me to apply any of these? Tell me which sections."

Notes

• If the user asks to review a SPECIFIC file or range, use a single agent (Coder) rather than three — parallel review is only valuable against a full branch diff.
• If main isn't the default branch, use origin/HEAD to resolve.
• Respect timing: each Task agent has its own context budget. Keep prompts tight — don't pass the entire diff, let each agent read what it needs via Bash+git.