unity-code-review-agent-v0

A minimal Unity C# static review tool for interview teaching and manual review practice.

Unity C# file -> static rule scan -> Markdown review report -> manual validation

Positioning

This project is a small, inspectable v0 tool. It is useful for practicing Unity client code review, preparing interview explanations, and turning review findings into reusable rules.

It is not a production CI gate, not a Roslyn analyzer, and not a semantic C# analyzer. Every finding is a suspicious signal that must be checked by a human reviewer.

What It Does

• Scans one .cs file or a directory of .cs files.
• Detects five Unity client review signals.
• Generates a Markdown report with line numbers and manual validation rows.
• Keeps v0 zero-dependency and standard-library-only.
• Avoids modifying Unity source files.

Rules

ID	Name	Severity	Notes
U001	Update 中疑似频繁 GetComponent	HIGH	Checks Update, LateUpdate, and FixedUpdate.
U002	帧循环中疑似使用 LINQ	WARNING	Checks frame-loop LINQ-like calls; excludes using R3; and using UniRx; files.
U003	疑似事件注册，需检查解绑	HIGH	Flags AddListener and event-style +=; ordinary arithmetic += is filtered.
U004	疑似 Resources.Load 使用	WARNING	Prompts a project resource-policy review.
U005	疑似 Addressables 加载，需检查释放策略	WARNING	Suppresses same-file hits when obvious Addressables.Release calls exist.

Quick Start

Run from the project directory:

python main.py --input examples/InventoryView.cs --output reports/

Expected output:

Scanned files: 1
Findings: 6
Report: reports\InventoryView_review.md

Scan a directory:

python main.py --input path\to\UnityProject\Assets\Scripts --output reports/

Print version:

python main.py --version

Report Workflow

1. Run the scanner.
2. Open the generated Markdown report.
3. Mark each finding as 有效, 误报, or 需上下文.
4. Record false positives in docs/false_positives.md.
5. Move stable lessons into docs/unity_review_rules.md.

The generated files under reports/ are local validation artifacts and are ignored by Git. A stable sample is available at docs/sample_inventory_review.md.

Verified Sample

The sample examples/InventoryView.cs is intentionally compact and teaching-oriented. It produces six findings:

U003 x2: event registration review
U001 x1: GetComponent in Update
U002 x1: LINQ-like call in Update
U004 x1: Resources.Load
U005 x1: Addressables load without visible release

Manual validation notes are kept in docs/inventory_view_manual_validation.md.

Real Project Validation

v0.1.1 was checked against completed Unity learning projects under projects/verified:

File	Result	Meaning
AddressablesDemo/.../PrefabSpawner.cs	0 findings	Correctly silent; release paths are visible.
UniTaskDemo/.../CooldownButton.cs	1 finding	U003 needs context; Button + UniTask lifetime is a softer signal.
R3Demo/.../ReactiveHealth.cs	0 findings	R3 .Where() false positives were fixed in v0.1.1.

Tests

Run:

python -m unittest discover -s tests

Expected:

Ran 16 tests
OK

The tests cover empty files, method-scope checks, event registration vs arithmetic +=, Addressables release suppression, R3/UniRx exclusions, and generated sample coverage.

Project Structure

unity-code-review-agent-v0/
├── main.py
├── src/
│   ├── scanner.py
│   ├── report_writer.py
│   └── llm_client.py
├── rules/
│   └── unity_rules.json
├── examples/
│   └── InventoryView.cs
├── docs/
│   ├── false_positives.md
│   ├── inventory_view_manual_validation.md
│   ├── sample_inventory_review.md
│   └── unity_review_rules.md
└── tests/

src/llm_client.py is intentionally a v1 placeholder. v0 does not call any LLM API.

Limitations

• Text-level heuristic scanning only.
• No C# AST or semantic analysis.
• No Unity Editor integration.
• No automatic fixes.
• No LLM, LangChain, RAG, LangGraph, or MCP integration.
• Findings should not be treated as confirmed bugs.

License

MIT License. See LICENSE.

Related

• hermes-harness — Agent 工作线，含 unity-gc-scanner MCP 工具
• unity-verification-projects — 历史 Unity 验证项目