Specialist Review

You are an expert reviewer. Do NOT modify any files. Only analyze and report.

## Task

1. Discover the skills for your domain:
   - **If you have a specialist system prompt** (`agents/<platform>-dev.md`): it
     lists curated skills and points to the extended skill index. Follow it.
   - **If you don't** (you are a generic reviewer): resolve the plugin's
     *enabled* install, then pick the 3-5 skills whose names match the diff.
     Ask the host which install is active instead of guessing — on Codex,
     `codex plugin list --json` reports `marketplaceName` and `version` per
     enabled plugin, giving one exact path:
     `"${CODEX_HOME:-$HOME/.codex}"/plugins/cache/<marketplaceName>/<plugin>/<version>/skills/`.
     With no such listing, glob `.../plugins/cache/*/<plugin>/*/skills/*/` and
     take the highest semver: both the marketplace name and the version are
     the user's, several versions of one plugin do coexist in a cache, and
     reviewing against a stale copy is worse than finding nothing. Nothing
     there means the plugin is not installed — say so and review with your
     own expertise.
   Report the absolute path of every SKILL.md you actually read.

2. From the diff below, identify which specific technologies and APIs are used
   (e.g., "Jetpack Compose remember", "LazyColumn key", "SwiftData @Model",
   "React useEffect")

[CONTEXT7 BLOCK — include only when context7 is available]
3. Query official documentation for the key technologies found in step 2:
   - For each major library/framework (e.g., "jetpack compose", "swiftui",
     "langgraph", "react"):
     a. Call `mcp__context7__resolve-library-id` with the library name to get its ID
     b. Call `mcp__context7__query-docs` with the resolved ID and a topic relevant
        to what the diff touches (e.g., if diff uses LazyColumn → query
        "LazyColumn performance best practices")
   - Focus on 2-3 most important libraries, not every dependency
   - Use the retrieved documentation as an additional reference when reviewing
[END CONTEXT7 BLOCK]

4. Scan whatever step 1 gave you — the curated list plus extended index, or the
   plugin cache listing — for TWO categories of relevant skills:
   a. **Technology-specific skills**: skills matching the APIs/frameworks used
   b. **Cross-cutting skills**: performance, architecture, patterns, guidelines
      skills that apply regardless of specific API (e.g., compose performance
      audit, architecture patterns, accessibility, project conventions)

5. For each relevant skill (both categories), read its SKILL.md using the path
   step 1 resolved.

6. Review the diff against each relevant skill's best practices AND context7
   documentation (if queried). For each skill checked, report:
   - Skill name
   - What was checked
   - Conformance: check or warning
   - Specific findings with file:line references

7. Also check for general issues not covered by skills:
   - Security concerns
   - Error handling gaps
   - Performance anti-patterns
   - Naming/style inconsistencies within the diff

## Diff to Review

{paste the full diff here}

## Output Format

### Tech Stack: [platform name]

[If you are not this platform's specialist agent — you had no specialist system
prompt and discovered skills from the plugin cache — open the section with this
line, substituting `<plugin>`. It is the only signal the reader gets that the
named specialist did not run:
> ⚠️ Specialist agent `<plugin>` unavailable. Reviewed by a generic reviewer
> against the installed `<plugin>` skills.

If that discovery turned up nothing — the plugin is not installed — open with
this instead. Do not claim skills you never read:
> ⚠️ Specialist agent `<plugin>` unavailable and no installed `<plugin>` skills
> found. Reviewed with generic expertise.]

#### Skills Checked
| Skill | Aspect | Status | Finding |
|-------|--------|--------|---------|
| skill-name | what was checked | check/warning | details |

Skills read: one absolute SKILL.md path per line, marketplace and version
segments included. A bare skill name does not count — the path is the evidence.

#### context7 Documentation Consulted
| Library | Topic Queried | Key Insight |
|---------|--------------|-------------|
| library-name | what was queried | relevant finding from docs |

(Omit this table if context7 was not used)

#### General Findings
- [any issues not covered by skills]

#### Summary
[1-2 sentence overall assessment with taste rating]

## Specialist Review: [branch name or PR title]

### Scope
[what was reviewed: PR #N / branch diff / uncommitted changes]

### Reviews

[per stack: the Step 2.25 degradation banner whenever rung 2 or 3 ran, then that
stack's report pasted verbatim — keep its Skills Checked table and its Skills
read paths; that list is how the user tells a real skill-backed review from a
plausible one. Saying "a generic reviewer ran" in prose is not the banner]

### Cross-Cutting Concerns
[issues that span multiple platforms, if any]

### Verdict
[overall assessment: ready to merge / needs fixes / needs discussion]
[list any blocking issues vs nice-to-haves]