rework-rate

Measure how much of your git history is rework. One command. No config. No signup.

rework-rate
rework-rate 34
rework-rate 34 --minutes 20 --hourly 5000

--minutes and --hourly have no defaults. Both must be given, or neither. The printed yen line repeats those inputs. They are yours, not a measurement.

Definition (public, breakable)

A commit counts as rework when:

1. its subject starts with fix or revert (case-insensitive), and
2. it touches a file that this same history already touched less than 24 hours earlier.

rework = a fix touching a file you changed less than 24h earlier

That is the whole definition. If you disagree with it, you can show a counterexample on a real repository and we will treat that as a successful break. This definition is not the only valid one. A claim that it is "the right" definition is not part of the tool.

Limits of this definition

• Fixes that are not labeled fix / revert are not counted.
• Operators who commit corrections (instead of amending silently, squashing, or force-pushing) will score higher. The number rewards leaving a trail, not "worse work."
• 24 hours is an arbitrary threshold. It is not a natural unit of software. A 25-hour fix of the same file is invisible to this tool. A 23-hour rename that happens to start with fix is visible.
• Merge commits, empty commits, and files whose names contain whitespace follow the same git show --name-only split the prototype used. That is a limitation, not a feature.

If you have a definition that is stricter, looser, or labeled differently, write it down and run it. A definition that cannot be broken is not a definition.

When this tool cannot measure

If the window contains no commit whose subject starts with fix or revert, the tool prints not measurable here and exits 0. It does not print 0%. A repository that never labels fixes is outside the definition, not a healthy score.

If at least one such subject exists, the rate and the weekly table are printed. The published reference line is printed only when there are 5 or more fix/revert subjects. Fewer than that is too thin to sit next to another person's number.

Published reference (one operator, one repo)

window     2026-07-21 … 2026-08-30 (34 days)
commits    817
overall    181 / 816 = 22.2%
Cursor     170 / 593 = 28.7%
reference  28.7%  (n=1 repository, 1 operator, 6 complete weeks)
reference r = +0.87  (n=6 complete weeks)

The n of the rate is not 817 commits. It is one repository and one operator. The n of r = +0.87 is six complete ISO weeks. This is a published run, not a benchmark. The tool prints yours next to reference and does not say which is better.

Relation to change failure rate

This number is a pre-merge cousin of change failure rate (CFR), the DORA / Accelerate term for the share of production deployments that fail.

They are not the same quantity. This tool is not part of the DORA research program, and it is not an AI-era stand-in for those four measures. The vocabulary is borrowed. The authority is not.

Weekly correlation is Pearson's r of weekly commit count vs weekly rework rate. It is printed only when at least 3 complete ISO weeks remain after dropping incomplete weeks at both ends of the window. Two weeks can always produce r = ±1.00. We do not print that number.

An ISO week is incomplete when:

• it has not finished yet (the following Monday 00:00 UTC is still after the window end), or
• the window opens two or more UTC days after that week's Monday.

The first case is why a Monday run and a Sunday run used to disagree: one in-progress week (2 commits, 0% rework) moved r by 0.22. That week is still listed in the table, marked (partial), and is not fed into r. The second case is why N days ago landing on a Tuesday does not, by itself, drop the first week. A two-day-or-more cut does.

Display: every ISO week that has commits in the window appears in the weekly table. Incomplete weeks are marked (partial). Exclusion applies to the correlation only.

What this tool does not do

• It does not send telemetry, usage data, crash reports, or version-check requests.
• It does not import urllib, requests, http, or socket.
• It does not write to your repository. The only git verbs it runs are log, show, and rev-parse.
• It does not create temporary files.
• It does not fetch, checkout, or change branches.

If it is not a git repository, or the window has no commits, it exits with a clear error.

Install

pip install rework-rate

Changelog

• 0.1.1 — README の導入手順を PyPI 名に修正（機能変更なし）

License

MIT.