---
name: code-eye
description: Run an evidence-backed AI review of current changes, then open persistent findings in the browser. Use when the user asks to review changes, find bugs, walk through risks, or leave diff comments.
---

Review the current changes (uncommitted + base..HEAD commits), then open the result in the user's browser via the code-eye MCP `review` tool.

Use `deep` mode by default. Use `quick` only when the user explicitly asks for a fast/high-signal scan.

1. Inspect the diff. In deep mode also inspect relevant callers, tests, configuration, and invariants.
2. For each candidate issue, try to disprove it against the current repository.
3. Keep only findings with a concrete impact and specific evidence. Do not report style preferences or speculative concerns.
4. For the initial review round, call `review` once with a structured `report`:
   - `mode`: `quick` or `deep`
   - `summary`
   - `findings`: each has a stable `identity` slug based on symbol/invariant/failure mode (never prose or line number), plus `title`, `claim`, `impact`, `evidence`, `severity` (`critical|high|medium|low|info`), `category`, `confidence` (`0..1`), exact `location` (`sha?`, `file`, `line`, `side?`), and optional `suggestion`
   - `checks`: commands actually run with `passed|failed|skipped` and a short summary
   - `coverage`: `reviewedFiles` and `skipped` files with reasons

Locations must point to lines present in the diff. The tool blocks until the user closes the panel.

For every returned user conversation:

1. Follow its `requested action` when present (`fix`, `explain`, `add_test`, or `show_alternative`).
2. Address it and run relevant checks.
3. Call `review` once more with an `addressedComments` entry containing its stable `commentId`, a concise `summary`, `changedLocations` with Before/After evidence where useful, and `checks`. Include the refreshed structured report when appropriate.

That follow-up records the agent response and marks the conversation agent-addressed; never claim or assume it is resolved. Only the user resolves it in the panel. Continue only for conversations returned as waiting for more agent work. Agent-addressed conversations remain visible without being returned again.