Mine rules from your own fix history

Every repository accumulates a record of its own conventions: the commits where someone fixed an agent's (or a human's) mistake. Mining turns that record into rules your agents read before they write — with a verification gate before any rule ships.

method

Who does what, and which steps touch a model

Work Done by LLM?
Fix-commit classification, candidate extraction Bundled CLI, deterministic no
Invariant drafting from distilled candidates Your invoking agent, skill-driven your session
Detector generation CLI scaffolds; agent fills logic your session
Verification gate CLI, deterministic no

Candidate distillation keeps spelling-level detail — AST shapes, source excerpts, concat forms — so detectors aren't hollow. Verification samples are produced by an agent session independent of the drafting one, and every promoted rule carries that provenance.

trust

Your source never leaves your machine. Judgment runs in your own agent session, inside your own harness. Deterministic steps are local.

Everything deterministic is open and auditable. What stays ours: the judgment prompts and the verification corpus.

status

Currently validating on our own repositories.

Mining engagements open after the pipeline produces a rule we didn't hand-write, verified end-to-end on real history. We won't sell it before then. No projections here — when there are recorded runs to point at, this page will point at them.

Changelog/launch list only. Or email directly: ambus1989@gmail.com.