Overview

Relevant source files

• README.ja-JP.md
• README.ko-KR.md
• README.md
• README.ru-RU.md
• README.zh-CN.md
• internal/config/rules/system_rules.go
• internal/viewer/static/session.js
• pages/src/content/docs/en/architecture.md
• pages/src/content/docs/ja/architecture.md
• pages/src/content/docs/ru/architecture.md
• pages/src/content/docs/zh/architecture.md

Open Code Review (OCR) is an AI-powered code review CLI tool designed to identify code defects with high precision README.md35-37 Originally developed as an internal tool within Alibaba Group, it has been validated at scale across tens of thousands of developers and millions of code review tasks README.md37-40

OCR differentiates itself from general-purpose AI agents by utilizing a Hybrid Deterministic + Agent architecture. This design combines rigid engineering constraints—for file selection, positioning, and rule matching—with the dynamic reasoning capabilities of Large Language Models (LLMs) README.md75-77

Sources: README.md35-77

Key Capabilities

• Deep Contextual Review: Unlike surface-level diff analysis, OCR agents can read full file contents, search the codebase, and inspect related files to provide deep architectural feedback README.md41-42
• Precision-First Design: OCR favors precision over recall to minimize "noise" and false alarms, achieving significantly higher F1 scores than general-purpose agents in real-world benchmarks README.md49-51
• Multi-Mode Operation: Supports reviewing local workspace changes, specific git branch diffs, individual commits, or performing full-file audits via ocr scan README.md132-151
• High Efficiency: Consumes approximately 1/9th of the tokens compared to general-purpose agents by using smart file bundling and scenario-tuned toolsets README.md49-51
• Extensive Language Support: Includes built-in rules for 10+ languages, covering specific risks like NPE, thread safety, and SQL injection README.md51 internal/config/rules/system_rules.go89-115
• Flexible Integration: Can be used as a standalone CLI, a CI/CD plugin, or in "Delegate Mode" where a host agent uses OCR's engineering logic while providing its own LLM reasoning README.md153-160

Sources: README.md41-160 internal/config/rules/system_rules.go89-115

System Architecture

The core philosophy of OCR is "Deterministic Engineering × Agent Hybrid" README.md75-77

1. Deterministic Engineering (Hard Constraints)

Engineering logic handles tasks where correctness is non-negotiable:

• Precise File Selection: Logic determines exactly which files require review and which should be filtered based on allowlists, glob patterns, and binary detection README.md81-83
• Smart Bundling: Related files are grouped into a single review unit to maintain context while enabling concurrent processing README.md84
• Rule Matching: Matches fine-grained review rules to each file's characteristics using a multi-tier resolution hierarchy README.md85
• Positioning & Reflection: Independent modules systematically improve location accuracy (line numbers) and content accuracy to reduce hallucinations README.md86

2. LLM Agent (Dynamic Decision-Making)

The agent layer handles semantic understanding and exploration:

• Scenario-Tuned Toolset: A purpose-built set of tools (FileRead, CodeSearch, etc.) distilled from large-scale production data README.md93-95
• Iterative Reasoning: An LLM loop that dispatches tool calls, manages memory compression, and aggregates feedback README.md41-42

Sources: README.md75-95

Architecture & Code Entity Mapping

Title: "OCR High-Level Pipeline"

Sources: pages/src/content/docs/ru/architecture.md14-25 pages/src/content/docs/ja/architecture.md11-22 pages/src/content/docs/zh/architecture.md12-23 pages/src/content/docs/en/architecture.md14-25

Title: "OCR Subsystem Relationship"

Sources: README.md75-95 internal/config/rules/system_rules.go89-115 internal/agent/selection.go1-50 pages/src/content/docs/en/architecture.md27-35

Core Subsystems

Subsystem	Responsibility	Key Code Entities
Review Agent	Orchestrates the end-to-end review pipeline, including task splitting and concurrency README.md84	internal/agent.Agent, internal/agent.Agent.Run, internal/agent.subTask
LLM Layer	Abstracts different providers (OpenAI, Anthropic, etc.) and handles protocol translation README.md122-124	internal/llm/client.Client, internal/llm/provider.Registry, internal/llm/client.ResolveEndpoint
Rule Engine	Matches files to language-specific best practices using glob patterns and brace expansion README.md85	internal/config/rules.Resolver, internal/config/rules.SystemRule, internal/agent/selection.FileFilter
Tool Registry	Provides the Agent with capabilities to "see" the codebase beyond the diff README.md93-95	internal/tool.Registry, internal/tool.FileRead, internal/tool.CodeSearch, internal/tool.FileReadDiff
Session Manager	Persists review results to JSONL for later viewing and resumption README.md147-148	internal/session.History, internal/session.TaskRecord, internal/session.FileSession

Sources: README.md84-148

Title: "Data Flow From Git Diff to Review Comment"

Sources: README.md41-42 internal/config/rules/system_rules.go89-115 pages/src/content/docs/en/architecture.md27-35

Further Exploration

For detailed technical information, please refer to the following child pages:

• Getting Started: Covers installation via NPM or binary, initial configuration via interactive TUI, and connectivity verification README.md108-128 For details, see Getting Started.
• CLI Command Reference: Comprehensive guide to all subcommands including review, scan, config, llm, rules, delegate, and viewer README.md132-159 For details, see CLI Command Reference.

Sources: README.md108-160