Begin at the work.
The next agent can trust what is current, understand what happened, and see the next useful action without a forensic warm-up lap.
Repository closeout + clean-commit control
Mister Clean pays inherited repository debt, reconciles code, planning, and Git into one trustworthy reality, then guards the commit boundary so the clean codebase stays clean.
Every agent inherits the repository before it inherits your intent. When the code, plan, and Git history disagree, every new session pays to reconstruct reality before useful work can begin.
The repository is the team’s shared memory. Make that memory coherent and every session begins with leverage.
The next agent can trust what is current, understand what happened, and see the next useful action without a forensic warm-up lap.
Context windows go toward product judgment and implementation—not repeated reconstruction of stale plans, mystery branches, and half-finished procedures.
A clean baseline teaches good patterns. GUARD protects it at every accepted commit, making coherence easier for each team to preserve than to erode.
An unfinished review or stale plan is not a one-time inconvenience. Every later agent must rediscover it, reinterpret it, and decide whether the repository can be trusted. Mister Clean pays that debt once, at the source.
Implementation says done. QA, holdout, or planning truth says otherwise.
Agents spend context and judgment reconstructing what should already be explicit.
The repository teaches the team that partial completion and drift are normal.
Finish the procedure, reconcile every projection, and leave one trustworthy reality.
“Clean” is not a feeling and it is not a green formatter. It is a successor-ready end state with no hidden work, no conflicting projections, and no archaeology required before the next useful change.
A clean target, only genuinely owned working branches and worktrees, no stale stashes, no accidental untracked files, and no unreviewed commit waiting off to the side.
Stories, slices, holdouts, epics, rollups, maintenance records, and even historical archives conform to their standards and occupy the lifecycle lane their current state requires.
Folders explain themselves, names are consistent, generated and local evidence live where they belong, and one thoughtful root-cause repair replaces a spray of symptom patches.
Builds, tests, acceptance, independent review, and required holdouts bind the actual result. The next safe action is obvious—or there is honestly no remaining action.
CLOSE pays accumulated repository debt before a handoff. GUARD moves the same standard into daily development so debt never reaches accepted history in the first place.
Invocation is standing authority to finish owed work, repair planning and Git drift, remove debris, run the applicable gates, synchronize the result, and keep going until the repository is actually successor-ready or a real hard boundary is named.
Development produces an immutable candidate tree. QA, Mister Clean, and a final holdout independently attest to that exact tree. Any byte-changing repair invalidates the old receipts and restarts the gate.
Every action is compared before and after. Exact result bytes are preserved and fingerprinted. A new regression must be repaired inside the action or the action is rolled back; net improvement never buys permission to hand new debt forward.
Isolated worktrees, explicit write custody, versioned semantic conflict domains, short fenced integration leases, and compare-and-swap checks prevent a clean textual merge from hiding a stale plan or invariant collision.
Raw findings remain visible, but payable work is grouped by independently repairable root cause. The preferred change is the smallest coherent adjustment that solves the class of problem and leaves the whole codebase simpler.
The live instrument distinguishes repository truth, planning coherence, root completion debt, raw observations, holdouts, action-caused regressions, and successor-readiness evidence. It reports what is known; it never manufactures a CLEAN verdict.
Open dashboard →The MIT-licensed release includes CLOSE and GUARD workflows, the executable closeout CLI, a read-only MCP server, zero-harm evidence, semantic-boundary probes, and collision-safe multi-agent coordination.
pnpm add --global @bradheitmann/mister-clean@6.3.0
pnpm dlx --package @bradheitmann/mister-clean@6.3.0 mister-clean-mcp
Register the included SKILL.md with the agent harness, then invoke it bare. Mister Clean does not stop at an audit, a queue, or a good explanation of the debt: it performs the ordinary in-scope cleanup it is authorized and able to perform. For long campaigns, the package also includes an optional orchestration-goal template that keeps the loop alive until the clean contract is satisfied. Hard safety, ownership, secrets, and production-effect boundaries remain hard.
$mister-clean