Docket Tool Creates Per-Commit Evidence Records for AI-Generated Code

Addressing AI Code Review Challenges

Coding agents can generate code at a speed that outpaces human verification capabilities. Traditional code review processes receive a final diff without insight into the agent's development journey, including initial attempts, verification steps, or lines that received no scrutiny. Docket aims to bridge this gap by providing transparency into the AI's coding process.

How Docket Works

Docket captures the agent's development journey, which is typically discarded at commit time. It then integrates this information with the code diff to generate a per-hunk evidence record. This record highlights areas where human review attention is most needed, specifically where evidence from the agent's process is absent.

The tool displays details such as the originating agent (e.g., 'claude-code/main'), the task, the agent's intent, superseded attempts, test failures, and evidence coverage (e.g., lines executed by tests). It also indicates if there was no recorded human interaction with specific lines of code.

Local and Self-Contained Operation

Docket operates without external accounts or network dependencies. All evidence records are stored directly within the repository on an orphan ref. The tool is distributed as a single static binary, requiring no Go toolchain or runtime for execution. It supports macOS, Linux (arm64 and x86-64), and Windows platforms.

Installation and Usage

Users can install Docket via a shell script or by downloading platform-specific archives from the releases page. SHA256SUMS are provided for integrity verification. Building from source using `go install` is also an option. A demo script is available to create a temporary repository, simulate an agent's coding session, and display the resulting Docket record without affecting the user's existing repositories.