wtf is dreamer?

Dreamer is a tool built to stop coding agents from repeating the same mistakes. Many teams find themselves giving their AI coding assistant the same corrections over and over, things like verifying that tests actually pass before saying a task is done, not weakening tests just to make them pass, or following naming conventions the project already has. That kind of feedback often never makes it into a project's shared instruction files, so the agent keeps making the same errors in future sessions. Dreamer solves this by mining a user's recent coding agent session history for corrections that keep showing up again and again, then turning those patterns into concrete suggested additions to instruction files like AGENTS.md or CLAUDE.md. It presents each suggestion along with the original conversation as evidence, and nothing is written to a file until the user reviews and approves it. The tool supports 20 different coding agent tools, including Claude Code, Codex, Cursor, OpenCode, and Gemini CLI, by reading session history through a separate tool called cass, which Dreamer installs automatically the first time it runs. Getting started requires Python 3.11 through 3.13. It is installed as a skill inside a coding agent using a single command, after which running a slash command like /dream inside that agent triggers the analysis and proposes an in chat plan of changes for review. No external API keys are needed since Dreamer uses the coding agent's existing subscription, and everything runs locally: session history is read and context files are edited entirely on the user's own machine, with nothing uploaded elsewhere. For a broader view across many sessions, Dreamer also offers a dashboard that can be installed separately and run as a local web server. This dashboard shows recurring correction patterns with supporting evidence, clusters of where agent work concentrates, and statistics on activity trends over time. The project is released under the MIT license.