daydream
What it is
Daydream runs automated code review against a Git diff or pull request. It uses stack-specific Beagle review skills, verifies findings against the repository, applies fixes when configured to do so, and gates those fixes on the project's test suite. Each run is recorded as an ATIF (Agent Trajectory Interchange Format) v1.6 trajectory, then indexed by a corpus pipeline that can project accepted runs into JSONL records for supervised fine-tuning (SFT) or reinforcement learning (RL).
A structural evidence gate drops any finding that lacks a grounded
path:line citation before it reaches the merged report, the fix stage, or PR
posting. Dropped items are logged to a dropped-speculative.json audit sidecar.
Documentation
| Page | Use it for |
|---|---|
| Architecture | Deep review stages, shallow mode, trajectory recording, archive schema |
| CLI reference | Commands, flags, defaults, and examples |
| Configuration | Config files, backend/model resolution, review skills |
| Corpus and training pipeline | Harvest, reward scoring, corpus build, labels, lineage, training roadmap |
| Benchmarking | Martian benchmark setup, held-out PR set, scoring outputs |
| Self-hosted bot setup | GitHub App, secrets, workflow templates, verify doctor |
| Integration patterns | CI/CD, Phase A/B split, PR feedback loop, corpus production |
| Extensions | Fork extension seam: flow-step registry, skill slots, prompt overrides, ext validate |
Quick start
daydream /path/to/project
For a pull-request comment-only workflow:
daydream --non-interactive --comment
For a self-hosted GitHub bot:
daydream setup /path/to/project --repo owner/repo