---
description: Run a deep evidence-backed review and walk the user through the findings in a browser panel. Use when the user asks for a deep review, risk walkthrough, or explanation of what needs attention.
---

Inspect uncommitted changes and the base..HEAD commits. Explore relevant callers, tests, configuration, and invariants outside the diff. For every candidate issue, try to disprove it before reporting it.

For the initial review round, call the code-eye MCP `review` tool (`mcp__plugin_code-eye_code-eye__review` when installed as a plugin) once with a structured `report` in `deep` mode:

- Each finding has a stable identity slug based on symbol/invariant/failure mode (not prose or line number), title, claim, impact, specific evidence, severity (`critical|high|medium|low|info`), category, confidence (`0..1`), exact diff location (`sha?`, `file`, `line`, `side?`), and optional suggestion.
- Include checks actually run with their result.
- Include reviewed files and skipped files with reasons.
- Do not turn style preferences or speculative concerns into findings.

The tool opens the findings in the browser and blocks until the user closes it. For every returned conversation, follow its requested action when present, address it, run relevant checks, then call `review` again with an `addressedComments` entry containing its stable `commentId`, summary, changed locations with Before/After evidence where useful, and checks. This marks it agent-addressed; only the user can resolve it. Continue only for conversations returned as waiting for more agent work; agent-addressed conversations remain visible without being returned again.