A reproducible, testable system-building agent that applies repo changes via hunk-level diffs
An autonomous coding agent you can actually test.
A deterministic system-building agent that takes a YAML mission, produces a multi-phase plan, and scaffolds and applies repository changes via hunk-level unified-diff patching, with full dry-run previews and a mock model for offline runs.
A look at the interface.
Representative UI — abstracted, never a client's real data.- •Enterprise +$22k (renewals)
- •Self-serve flat MoM
- Python
- PyYAML
- pytest
- GitHub Actions
Autonomous coding agents are usually nondeterministic and hard to trust. Vacant Lot makes the build loop reproducible and inspectable: every change is a parseable diff hunk, applied atomically, so plans can be previewed and regression-tested before anything touches disk.
How it
works
A YAML mission produces a multi-phase plan. Every change is applied as a parseable unified-diff hunk with atomic writes and rejection files, not full-file overwrites, so plans can be previewed and regression-tested before touching disk. A deterministic mock model adapter lets the whole loop run with no remote model, validated by a CI smoke test on every push.
What I built
(03) Inside the buildMulti-phase planner
Generates plan, scaffold, implement, test, and refine phases from a mission file, and can author per-feature plans with spec, implementation, and test steps.
Hunk-level diff applier
Parses and applies unified diffs hunk-by-hunk with atomic writes and rejection files, exposing pretty and machine-readable dry-run previews.
Offline-capable
A deterministic mock model adapter lets the whole loop run with no remote model configured, validated by a CI smoke test on every push.
The
receipts
- Fully deterministic agent loop: dry-runs are reproducible and unit-tested
- Hunk-level patching with atomic writes, not full-file overwrites
- Runs offline via a deterministic mock adapter
- Bundled tool layer for filesystem, git, http, and shell