Local evidence, PR context, and guarded cleanup lanes for evolving codebases.
AI-assisted projects can become working-but-sloppy fast: duplicated structure, unclear ownership, fragile boundaries, weak tests, and too many possible cleanup paths to hand to an agent or CI system safely.
A local-first maintainability system for codebases that already work but need disciplined follow-through before the next refactor.
It writes durable artifacts under `.deepclean/`: runs, feature maps, evidence, synthesis ledgers, candidates, clusters, reports, plans, handoffs, triage, lifecycle history, revalidations, and fix attempts.
Run `deepclean init`, `doctor`, `scan`, `report`, `status`, `cluster`, `next`, `show`, `plan`, and `handoff` to move from evidence to one bounded cleanup task.
`deepclean review-pr --json` gives controllers like OctoCheck source-safe PR context: changed files, related findings, architecture neighborhoods, risk summary, suggested verification, and prompt context.
The guarded fix lane requires one target, explicit mutation intent, current proof inputs, and verification before recording a fix attempt as useful.
The output is a reviewable cleanup queue with evidence IDs, stable candidate/theme identifiers, source-safe review packets, and proof-backed agent plans instead of vague smell lists.