The loop
Each phase is a skill your agent loads. Each one stops at a checkpoint and waits for you. Decisions are kept in a reviewed record under .muda/ in your repository, so the work can stop and resume.
measure
muda measure reports where the delivery time goes over a window: run counts, p50 and p90 durations, queue time, re-runs, per-job and per-step times and the top time sinks. muda gates lists the merge and deploy gates a fix must keep.
muda.toml with the confirmed settings, baseline.json, gates.toml and scan.json.diagnose
The agent works through the time sinks and the signals from muda scan and muda notices. For each one it reads every matching recipe and runs its Detect steps against the real logs. Each finding names its waste class, its evidence and an estimate.
findings.toml, and any exemptions in exemptions.toml.fix
The agent first shows the plan for the change: the files it will edit, the expected effect, the risk and the gates it touches. Once the plan is approved, the agent makes a branch and opens a pull request. It merges after verify, once you say yes.
findings.toml, in a separate record commit.verify
muda compare checks the change on the PR's real pipeline, and again after merge. It reports before and after timings with the runs it used, and a gate result.
receipts/, and the finding's new status in findings.toml.Every write goes through muda record. Record changes arrive on a branch as a pull request, merged by you or by the agent once you say so. Never edit .muda/ by hand. The gate results are on the proof page.