dreamwork

A heartbeat-driven autonomous dev loop for coding agents. Leave an agent dreaming on a project and come back to steady, safe, well-chosen progress.

The loop wakes itself every 4.75 minutes, works in small committed increments with deliberate reflection after each change, captures every idea as a task, and picks the next one by an explicit algorithm when idle. It knows what you want because you told it once, in a file it keeps current.

What makes it different

• Small increments are the error-catching mechanism. Each task is capped at 15-20 minutes and ends verifiable and committable, so mistakes surface while the agent is still on the path that made them.
• Nothing durable lives in the conversation. Goals, open questions, lessons, plans, and the task queue itself are files in the repo — the loop survives a restart, a compaction, or a fresh agent.
• It asks, and records the ask. Anything needing a human decision lands in .dreamwork/questions.md with enough context to answer cold, and can be answered from the dashboard instead of the terminal.
• It delegates to itself. Substantial work goes to dreamer subagents that share the loop's memory and write back what they learned.
• It watches itself dream. watch.py is a dependency-free local dashboard showing live status, questions, review artifacts, and a composer for steering the loop — no build step, no npm.
• It also does errands. ud-dreamtask is the same loop bounded to one task: it establishes what done looks like, works it in increments, and stops when the criteria verify.
• It scales past one project. dreamhub.py is a read-only page over every dreaming project on the machine: which loops are moving, which have gone quiet, which are waiting on you, and a link through to each project's own dashboard (dreamhub add <path>, then dreamhub serve).

Install

The skill is consumed in-harness, so installing is a symlink:

ln -s "$PWD" ~/.claude/skills/ud-dreamwork      # Claude Code
ln -s "$PWD" ~/.agents/skills/ud-dreamwork      # Codex

Then start it on a project: say dreamwork (or start dreaming) to the agent. On first run it interviews you briefly and writes DREAMWORK.md at the project root — your goals, philosophy, and working preferences. Everything after that is the loop.

Where to look

SKILL.md is the entry point and the product: philosophy, the tick flow, the selection algorithm, subagents, durable state, commands, guardrails. Beside it, initialization.md is the per-session startup, reflection.md the post-change checklist, watch-design.md the dashboard styleguide, and writing-plugins.md the contract for extending the loop.

---

Created by Max Kaye (XertroV) + AI.