review-pr-ultra

name	review-pr-ultra
description	Deep PR review that runs `code-review:code-review` and `pr-review-toolkit` agents in parallel, then hands off to `review-pr` for fix or post.
when_to_use	Use when the user says 'ultra review', 'review-pr-ultra', 'deep review this PR', 'thorough PR review', 'comprehensive PR review', 'multi-agent PR review', 'kitchen-sink review', 'ultra PR review', or any variation of wanting an exhaustive pull-request review that combines multiple reviewers.
argument-hint	[findings|fix|post]
disable-model-invocation	true

Run two complementary upstream review skills in parallel, present their findings side by side, then hand off to review-pr for fix or post.

The two upstream skills cover non-overlapping ground:

• code-review:code-review — git history, prior PR comments, CLAUDE.md compliance, confidence-filtered output
• pr-review-toolkit agents — type design, comment accuracy, test coverage, silent failures

Modes

Same as review-pr:

• findings — display side-by-side findings only
• fix — display findings, then run review-pr's Fix Flow
• post — display findings, then run review-pr's Post Flow
• No argument: findings, then AskUserQuestion for fix / post / stop

Findings Phase

Identify the target PR (default: current branch's PR). Dispatch both passes in a single assistant message so they run concurrently — sequential dispatch defeats the purpose, since tool calls must complete before the main turn continues.

Pass A — code-review:code-review

Launch a general-purpose subagent and tell it to invoke the /code-review:code-review slash command via the Skill tool (skill name: code-review:code-review), skip the final GitHub posting step, and return findings as structured data (file, line, severity, description).

Pass B — pr-review-toolkit agents (selective)

Dispatch the toolkit agents directly based on what the diff touches:

• Always: pr-review-toolkit:code-reviewer
• If error handling changed: pr-review-toolkit:silent-failure-hunter
• If tests changed, or production code changed without tests: pr-review-toolkit:pr-test-analyzer
• If comments or docs changed: pr-review-toolkit:comment-analyzer
• If new or modified types: pr-review-toolkit:type-design-analyzer

Each agent must return findings as structured data and must not post to GitHub.

Present findings

Show one section per source. Don't merge — the two skills use different severity scales and filtering, so cross-source agreement is for the user to weigh, not the orchestrator.

If neither pass returned anything actionable, say so and stop regardless of mode.

After Findings

Behavior depends on the mode:

• findings mode: stop here.
• fix mode: follow the Fix Flow defined in the review-pr skill.
• post mode: follow the Post Flow defined in the review-pr skill.
• No mode argument: use AskUserQuestion with options fix, post, stop. Then run the corresponding flow (or stop).

When tracking per-issue answers in Fix Flow, also include the source (which pass flagged the issue) so the user has that context.