Dreamer is an innovative tool designed to help software development teams effectively mine and visualize recurring feedback from their coding-agent sessions. This project aims to enhance the productivity of developers by systematically addressing common coding errors that arise during interactions with coding agents.

Most development teams frequently provide similar feedback to their coding agents, such as ensuring builds and tests are complete before declaring tasks finished, avoiding the weakening of tests for easy passes, and adhering to project naming conventions. Unfortunately, this valuable feedback often fails to make its way into AGENTS.md or CLAUDE.md, resulting in the persistence of the same mistakes.

Key Features

• Recurrence Mining: Dreamer analyzes recent coding-agent session history to identify and extract recurring corrections.
• Context File Enhancements: It transforms these corrections into actionable updates for context files, presenting the changes for review along with evidence from previous conversations, ensuring developer approval before implementation.
• Compatibility: Dreamer supports 20 coding harnesses, including Claude Code, Codex, Cursor, OpenCode, and Gemini CLI, making it versatile across different coding environments.

Workflow Overview

session history -> recurring corrections -> context file updates -> fewer repeated corrections

Local Execution

Dreamer operates entirely on local machines, requiring no provider keys. It utilizes an existing subscription and ensures that session history and context file edits remain confidential. The integration with cass is automated, making setup straightforward.

Dashboard Insights

For a comprehensive overview of patterns, trends, and evidence, Dreamer provides a dashboard. Users can review activity data and proposed context-file updates easily:

pipx install agent-dreamer
dreamer serve

Visualization

• Patterns illustrate repeated corrections with supporting evidence, transforming insights into candidate context improvements.
• Clusters analyze the concentration of coding agent activity, highlighting recurring task types.
• Statistics provide insights into activity trends, project volume, and workflow evolution over time.

Supported Coding Harnesses

Dreamer accommodates a range of coding harnesses via cass, including but not limited to:
Codex, Claude Code, Gemini CLI, Cline, OpenCode, Amp, Cursor, ChatGPT, Aider, Pi-Agent, GitHub Copilot Chat, Copilot CLI, OpenClaw, Clawdbot, Vibe, Crush, Hermes, Kimi Code, Qwen Code, and Factory (Droid).

Additional Resources

For more information on the Dreamer project, explore the following:

• Dream skill details
• Architecture
• Research papers

Contributions

Contributions are welcome. See the CONTRIBUTING.md for details on how to get involved.