What is code_review_agent?

This is a Go-based code review tool that uses a large language model with a very large context window to read through a codebase and report on what it finds. By default it is configured to use DeepSeek, a model that supports up to one million tokens of context, which allows it to hold a large portion of a codebase in memory at once. To use it, you create a config.yaml file in your project directory specifying the AI provider, API key, model name, and settings such as how many turns the agent can take before stopping, and at what ratio of context usage to compress earlier conversation history to free up space for more code. You then run the tool, enter the project path, and the agent begins reading through the code. A terminal interface lets you press the left arrow key to navigate output, scroll with the up and down arrow keys, press Escape to pause and have a conversation with the model, and type "go" to resume automated auditing. The README notes a relationship between the summary interval setting and how thoroughly the model works: a lower value causes DeepSeek to summarize its progress more frequently, which tends to keep it from stopping the review loop too early. However, a value that is too low can cause unproductive looping, so some tuning is needed depending on the codebase and model. The tool also supports an offline mode using a local model called duckgpt when no internet API is available. It currently runs on Windows only. The README is written in Chinese with configuration examples in YAML.