GitHub - whistler/dreamer: Mine recurring feedback from coding-agent sessions and visualize it.

2 min read Original article ↗

PyPI License: MIT

Stop repeating the same coding-agent corrections.

Most teams give their coding agent the same feedback over and over: verify builds and tests before declaring done, don't weaken tests to make them pass, don't add comments nobody asked for, follow the project's naming conventions. That feedback rarely makes it into AGENTS.md or CLAUDE.md, so the agent keeps making the same mistakes.

Dreamer mines your recent coding-agent session history for the corrections that keep recurring, and turns them into concrete additions to your context files, with the original conversation as evidence and your approval required before anything is written.

Dreamer works with 20 coding harnesses including Claude Code, Codex, Cursor, OpenCode, Gemini CLI and more.

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

Get started

Requires Python 3.11-3.13.

npx skills add whistler/dreamer --skill dream -g -y

Then, in your agent:

Dreamer proposes context-file updates as a reviewable in-chat plan (instructions, supporting evidence, and destination files) and applies only what you approve.

No provider keys needed; Dreamer uses your existing subscription. Everything runs locally: session history is read on your machine, context files are edited on your machine, and nothing is uploaded. Session history comes from cass, which Dreamer installs automatically on first run.

The dashboard

For reviewing patterns, evidence, and trends across many sessions at once, use the dashboard:

pipx install agent-dreamer
dreamer serve

If you want the CLI to mine sessions without the /dream skill, run dreamer run --yes after configuring a provider key.

See it in action

Patterns turn repeated corrections into candidate context improvements with evidence attached.

Dreamer feedback view

Clusters show where your agent work is concentrated and which kinds of tasks keep recurring.

Dreamer clusters view

Stats show activity trends, project volume, and how your workflow changes over time.

Dreamer stats view

Supported coding harnesses

Dreamer works with 20 coding harnesses via cass:

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).

Learn more

Contributing

See CONTRIBUTING.md.