crapkit

crapkit

crapkit scores every function in your repo on complexity times uncovered risk, ranks the worst ones by how often the file changes, and blocks commits that add more. It reads Python, TypeScript, TSX, JavaScript, Swift, Go, Rust, shell, PowerShell, C and C++, Objective-C, Vue, Java and Zig through lizard, and joins per-function branch coverage from the istanbul or coverage.py artifact your own test command already writes. JSON commands use sorted keys and a versioned schema for scripts, coding agents and the optional MCP server.

CRAP = ccn^2 * (1 - cov)^3 + ccn

The name is not ours: C.R.A.P. (Change Risk Anti-Patterns) was coined for crap4j by Alberto Savoia and Bob Evans in 2007.

ccn is the smaller of standard and modified cyclomatic complexity, both read off one lizard pass. cov is branch coverage inside the function's span; with no branches it falls back to statement coverage, and with no statements to invoked-or-not, so a half-executed straight-line function never reads as fully covered.

Above the ceiling, coverage cannot save you. Decompose. At the default ceiling of 6, a function at ccn 7 with 100% coverage still scores 7 and still fails the gate. The only move that clears it is splitting the function.

Why 6 and not 30. crap4j's conventional threshold of 30 is a CRAP score: it lets an untested ccn 5 through (25 + 5 = 30) and a fully covered ccn 30 too. crapkit's default is a complexity ceiling, because coverage can at best collapse CRAP to ccn, and a function you cannot cover past ccn 6 is one you decompose. Set target = 30 in crapkit.toml if you want the crap4j number. A repo with existing debt does not need to: ratchet seed marks today's over-ceiling functions at today's score, the gate then judges only the functions a change touches, and marks may only fall, so adoption never starts with a wall of red. Next to crap4py, radon, xenon, wily and SonarQube: docs/comparison.md.

crapkit scores git-tracked files only. Source you have not git added is invisible to it; init and verify name the first three such files they find.

Start with	When
Install and the 60-second start	You want the first score in an existing Git repository.
Python or TypeScript quickstart	You want a worked example from setup through a passing verify.
Adoption	You need to choose scopes, wire tests or introduce a ratchet to existing debt.
Upgrading	You already have saved runs, ratchet marks or an installed plugin.
Subcommands and JSON/MCP reference	You are scripting commands or connecting a coding agent.

---

The 60-second start

pip install crapkit
cd your-repo
crapkit init        # write crapkit.toml and ignore measurement output
crapkit doctor      # check scopes, test commands and coverage dependencies
crapkit coverage    # runs the lane, joins coverage, stores a scored run
crapkit worklist    # the ranked risk map
crapkit ratchet seed
git add crapkit.toml crapkit-ratchet.tsv .gitignore
git commit -m "adopt crapkit"
crapkit verify      # the first passing verdict

Not a Python repo? uvx crapkit init runs the same commands and adds nothing to your manifest: see A repo that is not Python. A Python repo skips uvx and installs crapkit as above, or with uv tool install or pipx install; that section says why. pip stops with externally-managed-environment? See When pip refuses.

coverage measures a commit, so a fresh git init repo needs its first commit before it. Until then it exits 4 and says so:

$ crapkit coverage
crapkit: the git repository at /repo has no commit yet: crapkit measures a commit, so make the first one (git add, then git commit) and run it again

init detects pytest, Vitest and Jest from the repository's own files. Review the generated config before running its commands. When detection leaves a commented lane, fill it in using the lane recipes. Commit the adoption files, then run crapkit verify to establish a passing verdict. Install the commit gate when the config and ratchet are ready. Each step prints the next one: init names coverage, coverage names worklist, worklist names ratchet seed, and ratchet seed names the commit and verify.

coverage scores, worklist ranks:

$ crapkit coverage
run 1 @ fae4db93108: 2 functions scored: 2 measured, 1 over ceiling 6, CRAP load 41.0, grade F
-> next: crapkit worklist

$ crapkit worklist
worklist @ fae4db93108 (run 1, floor ccn>=5, churn 12mo) - 1 of 1 active (worklist_top 50), 0 dormant
  risk     14.0  ccn  14  crap    38.5  cov  50%    1c/1a  calc/grade.py:7  classify( score , attempts , late , bonus )
no crapkit-ratchet.tsv yet: seed marks each function over its ceiling at today's score, and from then on a mark may only fall
-> next: crapkit ratchet seed

risk 14.0 is ccn times a churn weight of one: a one-commit repo has no spread of commits to weight, so each commit counts once and the ranking is complexity order until the history grows (Risk). crap 38.5 and cov 50% are the score and the coverage behind it.

ratchet seed signs today's debt at today's score. From then on marks only ever fall, so the repo can get better and never worse while you burn it down.

One thing stops most first runs: the coverage plugin. init writes a lane that shells out to your own test runner, and the runner needs its coverage package installed: pytest-cov for pytest, @vitest/coverage-v8 (pinned to your vitest major) for vitest. Without it the lane produces no artifact and coverage exits 5 quoting the runner's own error. A pytest lane also needs coverage.py 7.13.1 or newer, which writes the function start lines crapkit reads; an older one fails the lane at exit 5. For pytest, init probes the python its lane will run and prints the install command when pytest_cov is missing; pip install "crapkit[py]" pulls the plugin alongside crapkit when the two share a venv. On a Windows PATH holding only the py launcher it writes py, not a python3 the lane could never run, and when cmd.exe cannot start the interpreter at all (exit 9009, the Store alias) it names that instead of guessing at pytest-cov. A repo that pins no lockfile and carries its own .venv gets that venv's interpreter in the lane, when that interpreter can import pytest, rather than whichever python the shell answers with. The two quickstarts below walk a real repo end to end.

On Windows a lane command is read by cmd.exe, the shell that will run it, not by sh. Double quotes are the portable quoting. A single-quoted value is refused at config load with exit 3, because cmd.exe would hand pytest five words and the lane would write no artifact:

# the lane in crapkit.toml
command = "python -m pytest -m 'not live and not perf' --cov=calc --cov-branch --cov-report=json:.crapkit/cov/py.json"

$ crapkit doctor
crapkit: lane 'py': positional argument 'live' narrows a full-suite coverage run; drop it, attach it to the flag it belongs to (-n8, --numprocesses=8), or set full_suite = false deliberately (cmd.exe does not treat ' as a quote: write the value in double quotes); a suite whose testpaths cannot be collected in one process needs one lane per testpath, each with full_suite = false and its own artifact

Write it -m "not live and not perf". Carets, && and | segments, redirections and empty quoted arguments all read the way the shell reads them, so a chained lane (cd tests && python -m pytest --cov ...) is checked one segment at a time. doctor reads a lane the same way, and FAILs one whose runner will not start.

Install

pip install crapkit

That is the release on PyPI. For the unreleased tip of main, or from a local clone (run at the clone root):

pip install git+https://github.com/JeanFrancoisGagne/crapkit.git
pip install .

Commits on main between two releases carry the same version string, and pip keeps an installed crapkit whose version matches, so running the git line again after a new commit leaves the old code in place. This line replaces it with the tip:

pip install --force-reinstall --no-deps git+https://github.com/JeanFrancoisGagne/crapkit.git

pip install . in a clone rebuilds on every run and needs no flag.

When pip refuses

Debian 12, Ubuntu 23.04 and later, Homebrew and uv mark the Python they install as theirs (PEP 668), and pip refuses to write into it with error: externally-managed-environment:

$ pip install crapkit
error: externally-managed-environment

Give crapkit an environment of its own. Either line installs it into one and puts a crapkit command in ~/.local/bin (%USERPROFILE%\.local\bin on Windows), the command the agent plugins start:

pipx install crapkit
uv tool install crapkit

When that directory is not on PATH yet, the installer says so and names the fix: pipx ensurepath, or uv tool update-shell. Open a new shell after either. The python on PATH does not hold this install, and the commit hook runs the crapkit command before it tries python, so it reaches either install as printed.

A Python repo can carry crapkit in its own venv instead, beside the test dependencies its lane runs:

python3 -m venv .venv
. .venv/bin/activate
pip install "crapkit[py]"

The same from cmd.exe on Windows:

python -m venv .venv
.venv\Scripts\activate
pip install "crapkit[py]"

--break-system-packages makes pip write into the system Python anyway, where the next OS package update can break crapkit or the OS's own tools. None of the routes above needs it.

A repo that is not Python

crapkit is a command-line tool, never a dependency of the code it scores. A TypeScript, Go or Rust repo adds nothing to its own manifest. With uv on the machine, uvx fetches crapkit into a cache of its own and runs it:

$ uvx crapkit init
wrote crapkit.toml with 1 scope(s): src
detected 1 lane(s) from this repo's own files: js - next: run `uvx crapkit coverage`
added to .gitignore: .crapkit/

$ uvx crapkit coverage
$ uvx crapkit worklist

uvx puts no crapkit on PATH, so every next step crapkit prints under uvx starts with uvx crapkit. The lane still runs your own test runner, so Vitest or Jest and its coverage package come from the repo's node_modules as they do today. uv tool install crapkit or pipx install crapkit puts a crapkit command on PATH once, which is what the commit gate and the Claude Code plugin call. The ratchet's merge driver takes either form; the git merge driver gives the uvx line. uv brings its own Python when the machine has none.

A Python repo installs crapkit instead: in the environment its suite runs in (pip install "crapkit[py]"), or with uv tool install crapkit or pipx install crapkit. uvx puts its own environment first on the PATH the lane inherits, so the lane's python starts uv's cached interpreter, which holds neither the suite's packages nor pytest-cov. init then says that interpreter cannot import pytest_cov, and installing pytest-cov into uv's cache fixes nothing the suite needs.

Requires Python 3.11 or newer and Git on PATH. On an older Python pip stops with No matching distribution found for crapkit, and uvx crapkit runs crapkit on a Python 3.11 or newer that uv finds or downloads. The CLI has one runtime dependency, lizard>=1.24.0,<1.25; a package mirror needs both distributions. Install into the environment you intend to use, then check crapkit --version. The pip install -e ".[dev,accuracy-push]" under Development is a different thing: it adds the test extras, for people changing crapkit.

pip install --user crapkit puts the crapkit command in your user scripts directory: ~/.local/bin on Linux, ~/Library/Python/3.12/bin for a python.org Python 3.12 on macOS, %APPDATA%\Python\Python312\Scripts for Python 3.12 on Windows. When that directory is not on PATH, pip ends its install with WARNING: The script crapkit is installed in '...' which is not on PATH. Add the directory it names to PATH and open a new shell. Until then the shell answers crapkit with command not found (is not recognized on Windows), and the Claude Code plugin lists its server as Failed to connect.

Python projects can install pip install "crapkit[py]" in their test environment to include pytest-cov and subprocess-capable coverage.py. A separate tool installation still needs the coverage plugin in the environment that runs the suite.

Analysis and scoring run locally and send no telemetry. Configured lane, mutation and alert commands run with your permissions and can contact services or change files. Review those commands before running crapkit in a repository you do not trust (SECURITY.md).

$ crapkit --version
crapkit 0.8.1

python -m crapkit works identically to the console script and is what to use from a source checkout. A next step crapkit prints names crapkit when PATH finds this installation's console script, and otherwise the interpreter running it, spelled with forward slashes (C:/venv/Scripts/python.exe -m crapkit coverage) so Git Bash, cmd.exe and PowerShell all run it as printed. A segment that holds a space is quoted on its own (C:/"Program Files"/Python312/python.exe -m crapkit coverage), so the line never opens with a quote and runs in PowerShell too. One case loses cmd.exe: a venv's python.exe in a spaced directory with no 8.3 short name, where you put the whole path in double quotes by hand (docs/adr/0003). Every subcommand accepts --repo PATH (default: the nearest crapkit.toml at or above the current directory, so a monorepo workspace finds the root's), and with it you never have to cd into the repo you are scoring; Subcommands shows where the flag goes. A leading ~ in PATH is your home directory, also in cmd.exe and in an MCP client's args, where no shell expands it.

Upgrading

Keep the CLI and plugin versions aligned, measure fresh coverage after upgrading, and review any ratchet identity refusal before reseeding. The current reader is analysis version 13. It reads a coverage.py function's span from the start_line coverage.py 7.13.1 and newer write, and refuses a report without one. It scores a source saved as UTF-16 and keys an identifier that holds one of the five bytes cp1252 leaves undefined by its name. It moves numbers in every language but Python and some of Python's cognitive and nesting, puts JavaScript and TypeScript coverage on the function that owns it, and computes CRAP alike on every platform. So every marks file re-seeds once. Version 11, in 0.8.0, renamed once each Python def with a PEP 695 type parameter list and each def nested three or more deep, and listed a def whose body sits on its colon line; crapkit ratchet prune drops the marks left under the old names. Older JavaScript and TypeScript callback marks can require a reviewed mapping. Follow the upgrade guide for the upgrade line each installer takes (pip, pip --user, pipx, uv tool, uvx and a git install), saved state, portable records and Windows launcher locks. CHANGELOG.md says what each release changed.

Upgrading from 0.8.0 to 0.8.1

1. Install coverage.py 7.13.1 or newer where each Python lane runs: pip install -U "coverage>=7.13.1", or pip install -U "crapkit[py]" when crapkit shares the suite's venv.
2. Upgrade crapkit in every clone, then re-seed each repo once: crapkit coverage, crapkit ratchet prune, crapkit ratchet seed. When a failed verify pins the baseline, pass --baseline N to prune and to seed.
3. Commit the re-seed together with every pin that runs crapkit: the CI install pin, the Action's uses: pin and the pre-commit rev. 0.8.1's verify refuses marks stamped with analysis version 11, and 0.8.0's refuses marks stamped 13.
4. Keep {python} and {python:DIR} out of a committed crapkit.toml until every clone runs 0.8.1: 0.8.0 hands the token to the shell as written, and the lane fails.

Until the re-seed, crapkit verify refuses the marks 0.8.0 stamped, at exit 3:

crapkit: ratchet marks were recorded under [crapkit-analysis=11 lizard=1.24.0] but this run measures [crapkit-analysis=13 lizard=1.24.0] - CRAP scores are not comparable across metric versions; run `crapkit coverage`, then `crapkit ratchet prune`, then re-baseline with `crapkit ratchet seed`

The upgrade guide lists every step in order, and analysis version 13 says what moves. A metric-stamp refusal from an older upgrade is quoted under analysis version 8.

The exe lock on Windows

A running crapkit.exe mcp holds its launcher open, and each installer meets that lock its own way. From pip 23.3 on, pip succeeds: it moves the busy crapkit.exe aside and installs the new release, and the running server keeps serving the old code until you restart its client. pipx does the same when it installs through pip. An older pip, such as the 22.3.1 a Python 3.11.2 venv ships, fails with ERROR: Could not install packages due to an OSError: [WinError 5] Access is denied and puts the old release back: run python -m pip install --upgrade pip first, or stop the server as for uv. uv tool upgrade crapkit fails with os error 32. uv tool install crapkit@latest, and pipx upgrade crapkit when pipx installs through uv, fail with Access is denied. (os error 5); after that uv tool install the crapkit command fails with ModuleNotFoundError: No module named 'crapkit' until the install runs again. After any of those failures, stop the server or its agent session, rerun the same command, then restart the client. The Windows upgrade procedure lists what each installer printed.

Removing crapkit

Take the commit hook and the merge driver out before the package: both call crapkit. After pip uninstall crapkit alone, the sh hook Route 1 and Route 2 write keeps judging every commit through uvx crapkit on a machine with uv, with whichever crapkit release uv has cached or can download, and on a machine without uv stops every commit on No module named crapkit, or on exec: python: not found where there is no python at all, as on a Debian, Ubuntu or macOS that has only python3. The PowerShell hook Route 1 and the handbook write names the launcher that pip uninstall crapkit deletes, so it stops every commit on No such file or directory, with uv or without it. Every merge that touches crapkit-ratchet.tsv conflicts after the driver's crapkit: not found. The upgrade guide's Removing crapkit lists each piece, in a repo and on the machine.

The Claude Code plugin

claude plugin marketplace add JeanFrancoisGagne/crapkit --sparse .claude-plugin plugin
claude plugin install crapkit@crapkit

Two commands, installed once per user, and every repo on the machine gets it. The plugin ships three skills, the read-side MCP server, and one advisory PostToolUse hook that names any function an edit pushed over its ceiling. Claude reaches two of the skills by itself, crapkit and crapkit-recover; the third you type, as /crapkit:crapkit-onboard, because wiring a repo up happens once and its description has no business in every turn's window. It adds no files to your repo, and it needs the crapkit CLI on PATH.

The hook is one shell command, crapkit claude-hook --protocol 1, so it runs as written in any Claude Code version with plugin support, and in the other agents that load Claude Code plugins: Cursor (which imports them), GitHub Copilot CLI and VS Code. Claude Code gets the advisory on stderr with exit 2 and wakes the model with it. Those three read exit 2 as a deny, a message for the user alone, or a blocking error, so there the hook exits 0 and hands the model the same lines as added context (other harnesses).

--sparse .claude-plugin plugin checks out the two small directories the plugin ships from. Without it Claude Code clones the whole repository under a 120-second clone timeout, which one measured add ran out of. A marketplace added without --sparse keeps that full clone; claude plugin marketplace remove crapkit drops it and uninstalls the plugin, and the two lines above put both back.

A repo with no crapkit.toml costs a silent no-op per edit: 68 ms on Windows through the crapkit.exe launcher, where a bare python -c pass took 32 ms on the same machine. Every edit starts it, whatever the file type; an edit to a type crapkit does not measure stops at a suffix check before any config is read. On Windows, Claude Code starts the hook through Git Bash, whose own startup comes on top, and runs it async, so no edit waits for it. After upgrading the CLI, refresh the marketplace before updating the installed plugin:

claude plugin marketplace update crapkit
claude plugin update crapkit@crapkit --scope user
crapkit doctor --plugin-root

The installed plugin moves only at a release. claude plugin update compares version strings, and main carries the last release's version until the next one, so between releases it reports the plugin up to date. Restart existing Claude Code sessions to apply the plugin update. The check above compares installed files with the CLI on PATH; it does not reload a running session. When they disagree it prints the commands for the side that is behind, and when claude plugin update answered "already at the latest version" over files main has moved past, it prints the uninstall and install lines that replace them. The update above is for a user-scope install. For one made with --scope project or --scope local, doctor names that scope and the project directory to run it in; for a plugin installed from a marketplace you added as a local directory, which Claude Code loads in place, it names git -C <that directory> pull.

A headless claude -p has no one to grant a tool call, so it refuses every crapkit call and the model reads Claude requested permissions to use ..., but you haven't granted it yet. Pass --allowedTools mcp__plugin_crapkit_crapkit for the plugin's server, or --allowedTools mcp__crapkit for one you added yourself.

The hook registers on Edit|Write, which is every write that names a file. An agent that writes its source through a shell heredoc names none, so a Bash event is judged off the working tree instead. That half is yours to register, because it costs two git spawns per shell call. Add a second PostToolUse entry to your own settings, same command, matcher Bash:

{
  "hooks": {
    "PostToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          { "type": "command", "command": "crapkit claude-hook --protocol 1", "timeout": 20 }
        ]
      }
    ]
  }
}

The cost is one git rev-parse --show-toplevel and one git status --porcelain -z -uall per shell call in any git repo, whether or not crapkit measures it: about 50 ms together on crapkit's own checkout on Windows, and more on a bigger tree. What comes back is the dirty or untracked *.py files written in the last 12 seconds, 25 at most, each judged the way an edit is. Python only, so a TypeScript or Go repo pays the two spawns and hears nothing.

The hook remembers, per Claude Code session, the bytes each judgement read: one small record per file under .git/crapkit/claude-hook/<session_id>/, and a session idle for 7 days is pruned when another one starts. A recent file whose bytes match its record is skipped, so a touch, a rewrite of the same bytes, or a test run right after an Edit does not repeat an advisory the session already heard. Content that arrives with an old mtime is never judged by the Bash half: a file moved or copied in (mv, cp -p, an unpacked archive), or one a long command wrote well before it returned. The commit gate still judges those.

Codex

Codex 0.131.0 or newer installs the same marketplace's plugin through its own manager; 0.130.0 has no codex plugin add:

codex plugin marketplace add https://github.com/JeanFrancoisGagne/crapkit.git --ref v0.8.1 --sparse .claude-plugin --sparse plugin
codex plugin add crapkit@crapkit

Use the three skills and MCP server in Codex. Codex loads no crapkit hook: the plugin's Codex manifest leaves hooks out, because Codex reports an edit as apply_patch patch text, which the advisory does not read. Codex offers crapkit and crapkit-recover to the model by itself; crapkit-onboard stays out of the model's list until you type $crapkit:crapkit-onboard, as Claude Code waits for /crapkit:crapkit-onboard. A plugin from 0.8.0 or earlier ships no Codex manifest: Codex 0.156.1 lists its hooks as untrusted PostToolUse hooks that run a bare crapkit, and they should stay untrusted.

--ref pins the marketplace to this release's tag, and crapkit's release step rewrites it to the tag it cuts. Codex 0.156.1 checks every Git marketplace each time it starts and reinstalls the marketplace's plugins when it moved. Added without a ref, the marketplace follows main, and a push to main moves the plugin past the CLI you installed from PyPI with no command from you. --sparse fetches the two small directories the plugin ships from rather than the whole repository.

A marketplace added at a tag stays there: codex plugin marketplace upgrade keeps it at that tag. After upgrading the CLI, remove the marketplace, add it at the new tag, and install the plugin again. The listing's --json needs Codex 0.137.0; drop it on an older Codex.

codex plugin marketplace remove crapkit
codex plugin marketplace add https://github.com/JeanFrancoisGagne/crapkit.git --ref v0.8.1 --sparse .claude-plugin --sparse plugin
codex plugin add crapkit@crapkit
codex plugin list --marketplace crapkit --json

The installed plugin moves only at a release. Removing the marketplace keeps it, and it stays at the old version until codex plugin add installs the new one. The same four lines move a marketplace added without --ref onto the tag.

Check the installed Codex plugin with crapkit doctor --plugin-root: with no PATH it reads Codex's plugin cache when Claude Code has no install, and PATH names one install under ~/.codex/plugins/cache/crapkit/crapkit/. A gap there names these refresh lines, never a claude command. See plugin upgrades for choosing that path and starting a fresh MCP session. A runtime with a skills directory but no compatible marketplace can copy plugin/skills/* into it instead: ~/.claude/skills for Claude Code, $CODEX_HOME/skills (~/.codex/skills by default) for Codex, ~/.gemini/skills for Gemini CLI. Gemini CLI lists crapkit-onboard to its model with the other two; gemini skills disable crapkit-onboard --scope user takes it out once the repo is adopted.

Other agents

Every other agent starts crapkit mcp from its own config file, and each one reads its own key and fields. The mcpServers block Cursor takes starts nothing in OpenCode, Amp or VS Code's .vscode/mcp.json, and none of them says so; in Gemini CLI it connects, but a headless gemini -p offers the model none of the tools without "trust": true. Wiring crapkit into your agent gives the block for each of 27 agents, Gemini CLI, Goose, Zed and Continue among them, with the file it goes in and where the agent starts the server.

Languages

14 languages, two coverage parsers. Coverage joins where a parser exists; everything else scores on complexity alone.

Language	Files	Coverage
Python	.py	coverage.py
TypeScript	.ts	istanbul
TSX	.tsx	istanbul
JavaScript	.js .jsx .mjs .cjs	istanbul
Vue	.vue	istanbul, when your vitest run reports on .vue files
Swift	.swift	none: cc-only
Go	.go	none: cc-only
Rust	.rs	none: cc-only
shell	.sh .bash	none: cc-only
PowerShell	.ps1 .psm1	none: cc-only
C and C++	.c .cc .cpp .cxx .h .hpp	none: cc-only
Objective-C	.m .mm	none: cc-only
Java	.java	none: cc-only
Zig	.zig	none: cc-only

A cc-only scope declares coverage_optional = true, scores crap = ccn, and needs no lane. Nothing about it is provisional: the ceiling still binds and the gate still refuses a function over it. Add a coverage lane the day a parser exists and the same scope starts joining coverage.

crapkit init writes that key itself, on every scope whose languages all lack a parser, and leaves it off any scope a lane could still measure. So the 60-second start above runs unchanged on a Go, Rust or shell repo: crapkit coverage scores it with no lane at all, and that run is the baseline worklist, next-item, ratchet seed and verify read.

Shell, PowerShell and Rust run on crapkit's own readers. lizard ships none for shell or PowerShell, so crapkit counts their functions itself, and reads the command inside a quoted "$(...)" as code: its &&, ||, -and and -or count as they do written bare. lizard's Rust reader scores a 7-arm match as ccn 2 (filed as lizard #494), so crapkit counts each non-wildcard arm like a C case. It also reads a Rust signature, a closure's empty ||, a let-else and a for that is no loop the way Rust means them (see per-language gotchas), and retires each correction the day upstream fixes it. It lists #[inline] fn f() { written on one line, which lizard took for a C preprocessor line. The cognitive column charges a match once, the way Sonar charges a switch. The Rust, shell and PowerShell readers count each match, case or switch arm in the modified column too, so both columns agree and the arms are gated: a seven-arm match gates at ccn 8, where lizard's modified count, and so the gated ccn, of a C switch with seven cases is 2.

Go and Zig read through crapkit's subclasses of lizard's readers, which end a signature where the language does. A function type such as var cb func(int) error opens no function, a result type's braces are not the body, and a parameter of function type counts once.

Python and Swift read through crapkit's subclasses of lizard's readers too. lizard's Python reader ends a def at the first ) of a signature that runs past it, so crapkit reads the signature to the body's colon. lizard's Swift reader takes super.init(...), r.get() and Socket(protocol: p) for declarations and #fileID for the start of a preprocessor line, which hid the functions after them, and lists no function named by a raw identifier with a space in it (func `keeps onboarding if offline`()), so crapkit reads each one as the name or literal it is. It also reads the expression in a string's \( ) as code, which lizard read as part of the string.

C, C++, Objective-C and Java run on lizard's readers with crapkit's fixes on top. lizard hid every function after some constructs, named rows after an attribute or a macro, and left unnamed and array parameters out of params; the per-language notes list what crapkit reads differently.

Expression arrows in arrays and argument lists are measured separately. In TypeScript, wrap an arrow body in parentheses when it contains < before a comma, such as x => (pair<T,U>(x)) or x => (x < 0). Without that delimiter, analysis refuses the file because this reader cannot distinguish type arguments from an expression separator. A run scores the refused file as zero functions and names it, and every gate refuses it once a change touches it (The gate). Generic arrow parameter declarations remain supported.

Functions on the same line have separate occurrence identifiers. Existing ratchet marks with ambiguous old identities require a reviewed mapping; see same-line function identity.

The gate

Use the advisory while editing, the gate when committing, and verify for the full verdict. The preview and hooks differ in what their available evidence can prove:

Surface	Fires	Power
crapkit claude-hook	after an agent's edit lands	advisory. Names the breach to the agent: on stderr in Claude Code, as added context in Cursor, Copilot CLI and VS Code. Blocks nothing, because PostToolUse runs after the write
crapkit rescore FILE --gate	when you ask, after the first coverage run	preview. A stricter preview of the commit gate, sub-second, before you stage: a ratchet mark pardons a function only while its CRAP is at or under the mark. With no run behind it, exit 1 and no snapshot
crapkit hook-precommit	git commit, or by hand	blocks. The hook exits 6; git reports 1. Inside a commit it judges the staged blobs only, so it costs the size of the commit and needs no coverage. Run outside a commit with nothing staged, as pre-commit run --all-files does, it judges every tracked file and fails on committed debt no ratchet mark covers; crapkit coverage, then crapkit ratchet seed, records that debt
crapkit verify	before you push, and in CI	the verdict. Gate, ratchet, new test failures, diff coverage, against the trusted baseline

Both hooks pardon a function that the marks file in the working tree carries a mark for, whether or not that mark is staged or committed, so touching signed debt never refuses a commit. In CI the file on disk is the committed one, so verify there fails a function whose mark never reached a commit. verify is also what fails a mark that rises. Since 0.4.5 its gate pardons a touched function whose fresh CRAP sits at or under its mark, the rule rescore --gate already applied; push it past the mark and the gate fires again. The pre-commit hook still pardons on the mark's existence alone, on purpose: a staged blob has no coverage, so there is no fresh CRAP to compare against. It reports how many it pardoned on stderr (staged function(s) carry a ratchet mark and were not gated), and says the same about a staged file no [[scope]] claims, so a new top-level directory cannot go ungated in silence.

Every gate refuses a changed file no reader could read, exit 6, because it judged none of the functions in it: crapkit gate: 1 staged file(s) could not be read, so no function in them was judged:, then UNREAD src/a.ts: ... with the reader's reason, then what to do. A TypeScript expression-arrow body with < before a comma is the common case; wrap the body in parentheses or a block. claude-hook names the same file after the edit, and an override grants nothing while one is in the change. crapkit coverage scores such a file as zero functions and names it on stderr, and crapkit doctor WARNs about each one the newest run could not read, so you meet the file before the commit gate refuses it. A file the change never touched still passes, so an old unanalyzable file blocks nothing until someone edits it.

The crapkit root can sit below the Git top. A config in packages/api gates that package's staged files as project-relative paths such as app/m.py. A CI step starts at the top, where no crapkit.toml is, so Route 4 and the GitHub Action take working-directory: packages/api. Any command other than the hook that runs at the top exits 3, and the refusal names the root below it:

crapkit: no crapkit.toml at /repo - nothing to analyze; crapkit.toml sits below it in packages/api: pass --repo packages/api

Path and root rules also cover absolute arguments, literal filenames and Git diff settings. Routes 1 to 3 below need no --repo in such a monorepo when you run them at the git top. Git runs the hook there, and with no crapkit.toml at the top the gate runs in each crapkit root below that owns a staged file and names paths from the top (packages/api/app/m.py). A commit that stages nothing under any crapkit.toml, a docs-only commit or any commit in a repo armed before crapkit init, passes with one note on stderr. To pin the gate to one root instead, end each hook-precommit line of the hook below with --repo packages/api, as in exec python -m crapkit hook-precommit --repo packages/api, and Route 3 adds args: [--repo, packages/api] under id: crapkit-gate. Route 4's crapkit verify takes --repo packages/api.

Route 1 and Route 2 write this hook body. It runs the first of three that the hook's PATH offers: the crapkit command a pipx, uv tool or venv install puts there, then uvx crapkit, then python -m crapkit:

command -v crapkit >/dev/null 2>&1 && exec crapkit hook-precommit
command -v uvx >/dev/null 2>&1 && exec uvx crapkit hook-precommit
exec python -m crapkit hook-precommit

The uvx line is what gates a machine that runs crapkit only through uvx. It runs the release uv fetched first, or the newest one when uv has none, which need not be the release the rest of your team runs. It also keeps the gate running after pip uninstall crapkit on any machine with uv, so take the hook out before the package (Removing crapkit).

Route 1's PowerShell form writes the launcher's own path instead, as that route explains. Route 3's framework writes a hook of its own that runs crapkit hook-precommit from the environment it installs crapkit into, and Route 4 gates in CI with no hook.

Git hands the hook the PATH of whatever ran git commit. A terminal with your venv activated passes the venv on; an IDE or GUI client you did not start from that terminal does not. When that PATH reaches no crapkit command, no uvx and no python that imports crapkit, git refuses every commit. With no python at all, as on a Debian, Ubuntu or macOS that has only python3, the hook exits 127 with exec: python: not found. With a python that does not import crapkit, such as a system python, it exits 1 with No module named crapkit. Spell the path out for that client: exec /path/to/venv/bin/crapkit hook-precommit, or Scripts/crapkit.exe on Windows.

Run git config core.hooksPath before you pick a route. When it prints a directory, set globally or by husky, lefthook or another hook manager, git runs hooks from there and never reads .git/hooks: Route 1 arms nothing, and husky sets the path back over Route 2's on the next npm install. Add the gate to the pre-commit file that directory already runs. Under husky, that is crapkit hook-precommit as a line of its own in .husky/pre-commit, with no exec: husky runs the file under sh -e, so the gate's exit 6 stops the commit, and the lines after it still run when it passes.

Route 1: .git/hooks/pre-commit (local, not committed)

hook="$(git rev-parse --git-common-dir)/hooks/pre-commit"
cat > "$hook" <<'EOF'
#!/bin/sh
command -v crapkit >/dev/null 2>&1 && exec crapkit hook-precommit
command -v uvx >/dev/null 2>&1 && exec uvx crapkit hook-precommit
exec python -m crapkit hook-precommit
EOF
chmod +x "$hook"

git rev-parse --git-common-dir finds the repository's .git directory from any subdirectory and from a linked worktree, where .git is a file and cat > .git/hooks/pre-commit fails with Directory nonexistent. Every worktree runs the hook kept there. It ignores core.hooksPath on purpose: --git-path hooks follows a global hooks path, and would write this repo's gate into every repo on the machine.

The same hook from PowerShell. Windows PowerShell 5.1 puts a byte-order mark in front of the shebang, UTF-16 from > and Out-File and UTF-8 from Out-File -Encoding utf8, and git then refuses every commit with error: cannot spawn .git/hooks/pre-commit: No such file or directory (measured on git 2.43 for Windows) without ever running the gate. PowerShell 7 writes no mark, and Set-Content -Encoding ascii writes none in either. Git runs the hook with its own sh, so the launcher's path is spelled with forward slashes and quoted, and no chmod is needed on Windows:

$crapkit = (Get-Command crapkit -ErrorAction Stop).Source -replace '\\', '/'
$hook = "$(git rev-parse --git-common-dir)/hooks/pre-commit"
Set-Content -Path $hook -Encoding ascii -NoNewline -Value "#!/bin/sh`nexec '$crapkit' hook-precommit`n"

This hook names the crapkit your shell finds today, so a GUI client that never saw your venv runs the same one. It has no uvx line: while that path is gone, as after you move the install or pip uninstall crapkit, every commit stops on .git/hooks/pre-commit: line 2: .../crapkit.exe: No such file or directory. Paste the block again after you move or rebuild that install.

Get-Command stops the block when no crapkit is on PATH (a uvx-only machine): install one with uv tool install crapkit or pipx install crapkit, or write the sh form above from Git Bash.

crapkit doctor warns when the hook file git would spawn starts with a byte-order mark, and when a core.hooksPath set elsewhere (a global one, or husky's .husky/_) sends git away from .git/hooks, where git then runs no hook at all.

PowerShell 5.1's > writes UTF-16 behind the same mark wherever it saves a file, and crapkit reads each such file as far as the program that owns it can:

• A source file saved that way scores its functions as its UTF-8 twin does. git diffs it as binary, so crapkit doctor names it in a note.
• A .gitignore in UTF-16 is one git cannot read. init names it on stderr with the fix and leaves it as it was.
• A root package.json in UTF-16 is one init will not guess a lane from: it exits 3 before it writes any file, naming the file and the fix.
• The marks file (crapkit-ratchet.tsv) saved that way reads as its rows, and a write puts it back in UTF-16 behind its mark and in its own line endings. A mark whose name holds a cp1252 byte reads with that byte as U+FFFD, and ratchet seed, prune, move, merge, verify's tighten and both overrides refuse at exit 3, naming the byte, rather than save U+FFFD in place of the name: how the marks file is read.

Set-Content -Encoding utf8 saves any of them as UTF-8.

Route 2: a committed hooks directory

The whole route, from a repo that has no githooks/ yet:

mkdir -p githooks
cat > githooks/pre-commit <<'EOF'
#!/bin/sh
command -v crapkit >/dev/null 2>&1 && exec crapkit hook-precommit
command -v uvx >/dev/null 2>&1 && exec uvx crapkit hook-precommit
exec python -m crapkit hook-precommit
EOF
chmod +x githooks/pre-commit
printf 'githooks/pre-commit text eol=lf\n' >> .gitattributes
git add .gitattributes githooks/pre-commit
git update-index --chmod=+x githooks/pre-commit
git commit -m "add crapkit gate hook"
git config core.hooksPath githooks

PowerShell stops that block at its heredoc, before any hook is written, and the next commit goes through ungated. From PowerShell, the same route:

New-Item -ItemType Directory -Force githooks | Out-Null
Set-Content -Path githooks/pre-commit -Encoding ascii -NoNewline -Value "#!/bin/sh`ncommand -v crapkit >/dev/null 2>&1 && exec crapkit hook-precommit`ncommand -v uvx >/dev/null 2>&1 && exec uvx crapkit hook-precommit`nexec python -m crapkit hook-precommit`n"
Add-Content -Path .gitattributes -Encoding ascii -Value 'githooks/pre-commit text eol=lf'
git add .gitattributes githooks/pre-commit
git update-index --chmod=+x githooks/pre-commit
git commit -m "add crapkit gate hook"
git config core.hooksPath githooks

The committed hook keeps the PATH lookup rather than one machine's launcher path, since every clone runs it.

The --chmod goes between the add and the commit. It writes the executable bit to the index, so a commit that already happened does not carry it: run it after and git ls-tree HEAD still says 100644, which is a hook Unix checkouts silently skip. The .gitattributes line is the harder half of the same failure: under Windows' default core.autocrlf the hook checks out CRLF and #!/bin/sh\r dies on Linux and macOS with a bad-interpreter error. crapkit doctor warns when a file under core.hooksPath is not 100755 in the index and prints the update-index line for it.

Git will not read a hooks path out of a committed file, so that git config line belongs in your CONTRIBUTING setup steps. Every clone arms the gate with it. crapkit doctor names a clone that skipped it:

WARN githooks/pre-commit runs crapkit's gate, but core.hooksPath is unset and git runs .git/hooks/pre-commit, so every commit here skips the gate without a word; run `git config core.hooksPath githooks` in this repo, or call crapkit hook-precommit from .git/hooks/pre-commit

After husky's npm install takes core.hooksPath back to .husky/_, the same WARN names .husky/pre-commit as the file to call crapkit from.

Route 3: the pre-commit framework

crapkit ships a .pre-commit-hooks.yaml declaring id: crapkit-gate. In your .pre-commit-config.yaml:

repos:
  - repo: https://github.com/JeanFrancoisGagne/crapkit
    # crapkit's release step rewrites this line to the tag it just cut
    rev: v0.8.1
    hooks:
      - id: crapkit-gate

That file arms nothing on its own. The framework writes .git/hooks/pre-commit when you tell it to, and until then git commit runs no gate and says nothing:

pip install pre-commit
pre-commit install

pre-commit install is the line every clone needs, the way Route 2 needs its git config core.hooksPath line. prek reads the same .pre-commit-config.yaml: prek install arms the hook in its place, and the gate refuses and accepts the same commits. crapkit doctor WARNs on a config naming crapkit-gate while the hook git runs is not the one the framework writes.

pre-commit run --all-files, the form pre-commit.ci and pre-commit/action run, starts no commit and stages nothing. Outside a commit, with nothing staged, the hook judges every tracked file instead of a staged diff that is empty: a function over its ceiling fails unless the marks file in the working tree marks it, which in CI is the committed file. So a breach committed from a clone that never ran pre-commit install fails the CI job. Inside a commit the hook judges only what is staged, as before. It tells the two apart by GIT_INDEX_FILE, which git sets for the hooks a commit runs. For a CI gate that also weighs coverage and the marks, use Route 4.

rev is a git ref pre-commit resolves against that remote. Pin a release tag, not a branch: pre-commit autoupdate only moves between tags, and a moving main would change your gate under you.

Route 4: CI

A CI job runs on a fresh clone, which has no .crapkit/ store, so bare crapkit verify exits 1. Running coverage first would make the PR's own tree the baseline, a gate that can never fail. The portable baseline is the mechanism:

# on the default branch, after a passing verify: commit this file
crapkit verify --emit-baseline crapkit-baseline.tsv

# in the PR job, against the committed baseline
crapkit verify --baseline-tsv crapkit-baseline.tsv --github

--github emits ::error file=... annotations that land on the PR diff; --sarif PATH writes SARIF 2.1.0 for code-scanning upload. Refresh the committed baseline whenever the default branch's verify passes.

The file carries each lane's test count and failure list on its first line, so the PR job forgives a test that already failed on the default branch and warns when a lane runs fewer tests. A file written by crapkit 0.8.0 or older carries neither: against it every failing test is a new failure, exit 8, and verify says so and names the line that rewrites it:

warning: the baseline file crapkit-baseline.tsv holds no test results, so every test failure counts as new and no suite size is compared; write it again with `crapkit verify --emit-baseline crapkit-baseline.tsv` to carry them

Two things the job has to do before those lines run. Install crapkit, pip install crapkit, and pin the version the way Route 3 pins rev: an unpinned install moves your gate on whatever day a release lands. Fetch the whole history. actions/checkout clones one commit by default, verify reads the diff against the baseline's commit out of git, and a shallow clone does not have that commit:

$ crapkit verify --baseline-tsv crapkit-baseline.tsv
crapkit: baseline commit a74260f321f is not an ancestor of HEAD in this shallow clone, which does not hold it; set fetch-depth: 0 on the checkout or run git fetch --unshallow

That is exit 4 on a git clone --depth 1 of a repo whose baseline verifies at full depth. On a full clone that holds the commit, the same exit names the branch that holds it, or, when no branch does, blames a rebase or an amend that rewrote history, and asks for a fresh baseline instead. A clone that does not hold it at all, because .crapkit/ came from another clone or from a CI cache keyed on a branch, says so and names the fetch:

$ crapkit verify
crapkit: baseline commit a74260f321f is not in this clone, so git cannot say whether it is behind HEAD; fetch it with `git fetch origin a74260f321f4e0b9d2c61a8f3e57d0c1b2a9e8f7`, or run `crapkit coverage` here for a baseline this clone holds

verify --base and hook-precommit --base look up the fork point with git merge-base, and in the shallow clone above they refuse with exit 4 and the same fix:

$ crapkit hook-precommit --base c47a37b1df69c434ba42eec5979ddad03d2bf1e4
crapkit: git merge-base c47a37b1df69c434ba42eec5979ddad03d2bf1e4 HEAD failed in /home/runner/work/app/app: fatal: Not a valid commit name c47a37b1df69c434ba42eec5979ddad03d2bf1e4; this shallow clone does not hold every commit: set fetch-depth: 0 on the checkout or run git fetch --unshallow

When the clone holds both commits but not the one they fork from, the line reads no merge base between REF and HEAD in ROOT, followed by the same fix. ratchet report --enforce with debt_max_age_months or repayment_min_per_30d set refuses in a shallow clone, exit 4 with the same fix: every mark reads 0 days old there and no repayment shows, so an age limit a full clone fails would pass and a repayment quota a full clone passes would fail. Set fetch-depth: 0 on the checkout step, which is what crapkit's own .github/workflows/ci.yml does.

The whole PR job, on GitHub Actions:

on: pull_request
jobs:
  crapkit:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0        # verify needs the baseline's commit
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"
      - run: pip install "crapkit==0.8.1"
      - run: pip install -e ".[dev]"   # your own test dependencies
      - run: crapkit verify --baseline-tsv crapkit-baseline.tsv --github

The second install is the one people leave out. verify reruns your lanes, so the job needs whatever your test command needs: the coverage plugin, npm ci, a database, all of it. Without them the lane writes no artifact and verify exits 5 quoting the runner's own error, which is a broken job and not a verdict.

What a refusal looks like

$ git commit -m "add route"
crapkit gate: 1 staged function(s) exceed the complexity ceiling of 6:
  ccn   7  app/m.py:9  route( a , b , c , d )
decompose before committing (coverage cannot save a function above the target).

That commit exited 1, not 6. Git collapses any failed hook to 1, so 6 is a code you only ever see by running the hook yourself: crapkit hook-precommit exits 6 on a violation and 0 otherwise. The stderr block above is the same either way.

CRAPKIT_OVERRIDE_REASON is not a bypass. Setting it routes the commit through the full three-record audit: an alert line through alert_command, a ratchet entry staged into the commit, and a row in the override log. All three land or nothing does, and an unset alert_command refuses the override outright. See docs/ratchet.md.

The GitHub Action

action.yml at this repository's root is a composite action, so a reviewer sees crapkit's numbers on the pull request without installing anything. Four lines add it to a workflow, and every input has a default:

      - uses: actions/checkout@v4
        with:
          fetch-depth: 0
      - uses: JeanFrancoisGagne/crapkit@v0.8.1

The whole job those four lines sit in:

on: pull_request
jobs:
  crapkit:
    runs-on: ubuntu-latest
    permissions:
      contents: read                   # actions/checkout clones the repository
      pull-requests: write             # the comment
    steps:
      - uses: actions/checkout@v4
        with:
          fetch-depth: 0               # the diff, and verify's baseline commit
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"       # the interpreter the install below lands in
      - run: pip install -e ".[dev]"   # whatever your lanes need to run
      - uses: JeanFrancoisGagne/crapkit@v0.8.1
        with:
          gate: "false"

That pip install step is the one people leave out, and it is the same one Route 4 above names: the action installs crapkit and nothing else, so your lanes still need whatever your test command needs. Without it the lane writes no artifact and the comment says so.

fetch-depth: 0 is the other one. actions/checkout clones a single commit; the action reads the pull request's changed files out of git and verify reads the diff against the baseline's commit. A shallow clone lacks the base commit, so git refuses the diff: the step logs git's error, and the comment ranks the whole repository and says why under its heading, quoting git's first line and naming fetch-depth: 0. A shallow clone also flattens every number counted from history: churn sees one commit per file, so the worklist ranks on ccn alone, and a ratchet mark's age reads as 0 days. worklist, next-item, brief and ratchet report print one line on stderr and carry shallow: true in their JSON (false in a full clone), and the comment repeats the worklist's line above its table:

warning: churn counts read only the commits this clone holds; this shallow clone does not hold every commit: set fetch-depth: 0 on the checkout or run git fetch --unshallow

brief names churn counts and mark ages in that line, and ratchet report mark ages and repayments.

The action installs crapkit from $GITHUB_ACTION_PATH, which is its own checkout of the ref you pinned in uses:. So a pin left at last month's tag scores your tree with last month's crapkit rather than with whatever released since, and pinning a tag is the whole version policy; the snippets above name the current release. Move the pin in the same commit that re-seeds the marks under a new analysis version: an Action one release behind the marks refuses them with exit 3 on every pull request (a team's upgrade).

What the comment looks like

One comment per pull request, edited in place on every push. A hidden <!-- crapkit-action --> line opens it, and the next run edits the first comment that starts with that line, so a fifteen-push branch carries one comment and not fifteen. A reviewer's reply that quotes the comment does not start with it and is left alone. When GitHub answers a page of the comment list with an error, the step edits the crapkit comment it found before the error. When it found none, it lists the comments once more before it posts fresh; if that listing fails too, it posts a fresh comment, so the pull request still gets this push's verdict, and the job log says the listing failed twice. On a push event there is no pull request to carry it, and the same text goes to the job log instead.

Rendered from three saved payloads: a pull request that adds an untested route() (ccn 8) beside a ratchet-marked legacy_router(), in a repository whose diff_uncovered_max is 3. The payloads are under tests/fixtures/action_comment/, and the unit suite pins this block to their render:

<!-- crapkit-action -->

## crapkit

4 functions in 2 files, 2 over ceiling 6, CRAP load 149.59, grade F.

**verify failed, exit 6: complexity gate.**

- gate: `app/calc.py:34` `route( a , b , c , d )` ccn 8, cov 0%, crap 72.0 -> decompose
- uncovered lines in `app/calc.py`: 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45

Run 3 against baseline 1, 1 changed file (`app/calc.py`): 1 gate violation, 0 ratchet regressions, 0 new test failures, 11 uncovered changed lines.

### Worklist: 1 changed file

| File | Function | ccn | risk | remedy |
|---|---|---:|---:|---|
| `app/calc.py:34` | `route( a , b , c , d )` | 8 | 4.0 | decompose |
| `app/calc.py:19` | `legacy_router( a , b , c , d , e )` | 8 | 4.0 | decompose (accepted debt) |

The first line is the run crapkit coverage wrote: functions and files, how many sit over the ceiling (over ceiling 6, or over their ceilings (6; reports 12, util 4) when scopes set their own), CRAP load and grade, then ; scope `src` scored no function: it claims no file for each scope that scored nothing, and a failed lane's first error line appended as ; lane 'js' failed: .... When coverage --json died before a summary, the line quotes the error object it printed instead: `crapkit coverage` exited 5: lane 'py' cannot import pytest-cov; pip install pytest-cov.

The verdict opens with the exit code and the rule it stands for (complexity gate, ratchet regressions, new test failures, diff-coverage ceiling N), then one bullet per finding: each gate violation with its function, ccn, coverage, CRAP and remedy; each changed file no reader could read, as - unread: `src/a.ts`, so the gate judged none of its functions: and the reader's reason; each ratchet regression as recorded -> fresh; each new test failure by id; and the first twenty uncovered changed lines, one bullet per file, with a count of the rest. Each kind prints fifty bullets at most, then - and 1450 more new test failures; crapkit verify lists them all. The counts line closes it, naming up to three of the files verify judged and counting the rest: 1 changed file (`app/calc.py`). An unread file fails the gate as a function over the ceiling does, so the counts line counts it among the gate violations and says how many were files: 1 gate violation (1 unread file). A verify that passed is one line: **verify passed.** Run 2 against baseline 1, 7 changed files (a.py, b.py, c.py and 4 more). A verify older than 0.8.1 lists no changed_paths, and its line gives the count alone. A lane that recorded no test results (it declares no results_artifact) adds New test failures went unchecked in lane `py`: it recorded no test results. to that line, and a new failure in a lane whose baseline recorded no failure list gets a bullet of its own, - lane `py`: the baseline recorded no failure list, so its new failures may predate this change: the fork point's lane declared no results_artifact, so nothing can tell the pull request's failures from older ones, and verify still counts them.

GitHub refuses a comment over 65,536 characters, and a pull request then gets no comment at all. A body that still runs over (long names under a large top) is cut at the last line that fits, the marker line still first, and ends with the comment stopped at GitHub's 65,536-character limit; the job log above holds the whole text. The step prints the whole comment to the job log before it posts.

The rows are the ranked worklist for the files the pull request changed, worst first, top of them, with the rows a finding names listed first. risk is ccn times churn weight, the number crapkit worklist ranks on, and remedy is the run's own verdict for that function: decompose, split-lines, add-tests or ok. (accepted debt) marks a function the committed ratchet carries a mark for, so an untouched legacy_router does not read like the pull request's own new function. A pull request that touches no ranked function gets the heading and No ranked function in these files. When crapkit worklist wrote no ranking, the table's place says why rather than reading as an empty one: the error it printed (`crapkit worklist` exited 1: no snapshot in ...: run `crapkit coverage` first.), or, when it printed nothing, `crapkit worklist` printed no ranking; its error is in the job log.

The two file counts describe the same diff, counted twice. 39 changed files is git diff --name-only --relative base.sha...HEAD run in working-directory, the branch's own commits under it, and it is what the table is filtered to; the step's log names up to three of them. The count on the verdict line is what verify measured from the same fork point. With delta: "false" the second one is 0, because there is nothing behind the checkout to measure from.

The inputs

Input	Default	What it does
gate	"false"	"true" exits with crapkit verify's own code, so a finding fails the check. On a pull request with delta on it also exits 1 when the base run was not made, because a verdict with no base run judged no changed function. Anything else exits 0 and the comment is the whole output
delta	"true"	scores the pull request's base commit first, so the verdict covers the commits the pull request adds. Costs a second lane run; "false" scores the checkout alone, and the verdict then judges no changed function
top	"5"	worklist rows rendered in the table. An empty value renders 5. A value that is not a whole number ("ten", "5.0", "-1") renders 5 and puts a warning naming the input on the run's summary page
python-version	"3.12"	the interpreter actions/setup-python installs crapkit into. Match it to the version your own setup-python step named, or the lanes run on an interpreter your dependencies never reached
working-directory	"."	the directory that holds crapkit.toml, relative to the checkout. Every step that runs crapkit or lists the changed files runs there, and the base run scores the same directory at the fork point. Set it when crapkit.toml sits below the repository top: at the top, crapkit coverage finds no crapkit.toml and exits 3

On a monorepo whose crapkit.toml sits in packages/api, the crapkit step takes working-directory: packages/api under with:, and the job's own pip install -e ".[dev]" step takes the same key at step level, because .[dev] names that package's pyproject.toml and the top has none. The changed files are then listed from packages/api, the way the worklist names them, and a change elsewhere in the repository is not part of the table. Point it at the directory that holds crapkit.toml itself, not at one below it. The input is new in 0.8.1: an action pinned to an older tag warns Unexpected input(s) 'working-directory' and scores the top.

gate: "false" is the default on purpose. A team adopts the action before it has decided which findings should stop a merge, and a check that fails on day one gets turned off on day two.

What the verdict line covers

On a pull request, the commits the pull request adds. The action scores the fork point first, then the checkout, then runs crapkit verify --base <fork>, which measures the diff from there and takes the fork point's run as its baseline. So the gate judges the functions in the diff a reviewer is reading, and a repository that was already over its ceiling before the branch started does not fail every pull request that touches it.

The fork point is git merge-base of base.sha and HEAD, not base.sha itself. base.sha is the base branch's tip when the event fired, so a base branch that moved after the branch forked carries commits HEAD never saw, and a run there would be neither the baseline verify wants nor a diff anyone is reviewing.

The base run happens in a detached worktree under RUNNER_TEMP, in the same working-directory inside it, and its store is copied over the checkout's so both runs sit in one place. The cost is two lane runs on a pull request: your suite runs once at the fork point and once on the checkout. Set delta: "false" to skip the base run, and the verdict falls back to the checkout against its own run, which reports the tree's own health and judges no changed function. The comment says so in place of verify passed:

**verify judged no changed function:** the base run was not made (no base commit). Run 2 against baseline 2, 0 changed files.

Three things leave the base run unmade: a shallow clone that does not hold the fork point, a fork point older than your crapkit.toml, and a lane that will not run against that tree. The step logs crapkit base scoring exited N and writes the reason to crapkit-base.reason in the words the comment then quotes, shallow clone does not hold the fork point of <sha>; set fetch-depth: 0 on the checkout, no usable crapkit.toml at the fork point <sha>: ..., or lane failed at the fork point <sha>: ... with the first lane '<name>' FAILED: line crapkit printed, so a warning from a lane that passed never stands in for the lane that failed. A crapkit that died printing nothing reads crapkit coverage exited <code> and printed nothing. The verdict falls back the same way delta: "false" does, and the ratchet still runs, so exit 7 there is a finding. What differs is the job's status. With gate: "true" on a pull request whose base run was attempted and not made, the exit step exits 1 and prints the reason, because actions/checkout's default depth-1 clone would otherwise turn every pull request into a green check that judged nothing. A push event and delta: "false" never attempt the base run, so they keep verify's own code; a push has no base commit and no pull request to comment on.

One requirement the base run adds: the lane has to measure the tree it runs in. A lane that reaches an installed copy of your package instead of the checkout will measure the pull request's code while standing on the base commit, and the two runs then describe the same tree. crapkit verify refuses a run whose artifact names files outside the tree (exit 5), which catches the loud version of this; a lane pinned to a path outside the worktree is the quiet one. Point the lane at the tree, or set delta: "false".

--reuse-artifacts is what keeps each of those runs to one pass of your suite. coverage ran the lanes moments earlier on that tree, and verify parses those artifacts rather than running the whole suite a second time for the same numbers.

Which is why crapkit coverage has to exit 0 for there to be a verdict. When it exits anything else (a lane that failed, an artifact it refused), the action does not run verify: on a runner that keeps its workspace between jobs (clean: false), a verify --reuse-artifacts over a failed measurement read the artifact the dead lane had left from an earlier run, passed over it, and made that run the trusted baseline. The comment then carries coverage's failure in place of the verdict:

**no verdict: `crapkit coverage` exited 5 (lane 'py' failed: lane 'py' wrote no artifact on its last attempt; the .crapkit/cov/py.json on disk predates it); verify did not run.**

The parenthesis is the first line of the lane failure the summary carries. When every lane failed, coverage prints an error object instead of a summary and the lane errors are only in the job log, and the line says so: (every lane failed (1 of 1); the lane errors are in the job log). When coverage printed nothing at all, it crashed or was killed before it could: (it printed no run summary, so it crashed or was killed before scoring; its error is in the job log). When coverage refused files a scope takes whose names are not UTF-8, the parenthesis names the first and counts the rest, and one bullet under the line names each file the error object lists in unread_files: - refused: `src/caf\xe9.py`: its name is not UTF-8, .... When the directory coverage ran in holds no crapkit.toml, the parenthesis names that directory and ends set the action's `working-directory` input to the directory that holds crapkit.toml. With gate: "true" the job exits with coverage's code.

The other gate that judges a delta is the portable baseline in Route 4: commit crapkit-baseline.tsv on the default branch and run crapkit verify --baseline-tsv crapkit-baseline.tsv in a step of your own. It needs no second lane run, and it needs someone to keep that file current.

The comment is posted with gh api and the job's own GITHUB_TOKEN, which needs pull-requests: write, and contents: read for the checkout once the job names any permission: naming one sets every other to none. Two things it cannot do: a pull request from a fork gets a read-only token, so the POST is a 403 there, and a self-hosted runner without the gh CLI on PATH posts nothing. In both cases the step stays green, logs gh's exit code, and leaves the rendered text in the job log.

Subcommands

crapkit clean --dry-run --json previews abandoned temporary mutation recovery. See resource policies for shared analysis workers, process lifetime and bounded logs.

Every subcommand takes --repo PATH, and the flag goes after the subcommand. Without it the root is the nearest crapkit.toml at or above the current directory (ADR 0002): from a monorepo workspace crapkit worklist reads the root configuration that claims the workspace, says crapkit: using crapkit.toml at /repo on stderr, and reads a relative path argument from where you stand. claude-hook is the one exception: it has no --repo, because it takes its root from the file named in the hook payload it reads.

$ crapkit worklist --repo /path/to/repo --scope util --top 1
worklist @ a7c5c85ac37 (run 1, floor ccn>=5, churn 12mo) - 1 of 3 active (--top 1), 0 dormant
  risk      5.4  ccn   5  crap    30.0  cov   0%    5c/1a  util/stats.py:1  bucket( value , low , high )
-> next: crapkit next-item

Before it, argparse reads the path as the subcommand name and exits 2 without ever mentioning --repo:

$ crapkit --repo /path/to/repo worklist --top 1
crapkit: error: argument command: invalid choice: '/path/to/repo' (choose from 'inventory', 'coverage', ...)

--json prints one sorted-keys JSON object on stdout, always carrying a schema field.

Command	What it does
clean [--dry-run] [--json]	Recovers abandoned temporary mutation worktrees. Preserves active runs and intentional mutation pools. --dry-run reports planned recoveries. It leaves test evidence alone: crapkit's own development runner, tools/testing/run.py, prunes .crapkit/test-runs/ with --retention-days and --retention-count. --json keeps test_runs with every array empty.
init	Sniffs tracked source into per-directory scopes, writes a self-validated starter crapkit.toml whose lanes report into .crapkit/cov/, and appends .crapkit/ plus each runner's own droppings to .gitignore. Writes a live [[lane]] when it can detect the test runner, otherwise a commented template. Never rewrites an existing config: over one, it adds only the .gitignore entries an earlier run left out.
doctor [--show-files] [--json] [--tune] [--plugin-root [PATH]]	Checks the config still describes the repo: unknown keys (with the accepted spellings), zero-file scopes, tracked source no scope claims, scopes no lane covers, lane cwds and commands that no longer resolve, lizard importable, oversized files. It FAILs a lane whose runner does not resolve or that the shell cannot start, a pytest --cov lane whose coverage.py is older than 7.13.1, and a marks file a newer crapkit or lizard stamped. It WARNs on lane, hook and marks-file settings a later command refuses or skips, such as a scope a lane measures with no [crapkit.scoped_tests] template behind it, and names each crapkit launcher on PATH with its version when there are two or more. --show-files lists every file each scope matched. --json gives each lane a refusal, the sentence --reuse-artifacts would refuse its artifact with, or null. --tune prints suggested parallelism knobs and writes nothing. --plugin-root PATH reads no repo at all: it checks an installed plugin against the crapkit on PATH on both version and hook --protocol, one line per disagreement, each naming the command that closes it. Every check and warning: docs/commands.md: doctor. The JSON form: docs/agent-json.md.
inventory [--db PATH] [--export PATH] [--json]	One lizard pass over every in-scope file into a SQLite snapshot run, cached by content hash. --db is the only way to point crapkit at a store outside .crapkit/, and only this command accepts it.
coverage [--lane NAME] [--reuse-artifacts] [--reuse-unchanged] [--export PATH] [--sarif PATH] [--github] [--json]	Runs the lanes, joins branch coverage onto a fresh inventory, writes a scored run. A failed lane is recorded, not fatal: its scopes fall back to no-lane and the run is typed partial, so it can never serve as a baseline. --reuse-unchanged reuses a lane whose proof still holds and says what that proof leaves out; --reuse-artifacts reads the saved files and warns, naming up to three files whose bytes moved since the lane measured them. A failed attempt's leftover stays refused while it holds the same bytes: a touch does not lift the refusal, new bytes do. See docs/lanes.md.
verify [--baseline ID | --base REF | --baseline-tsv PATH] [--emit-baseline PATH] [--override REASON] [--reuse-artifacts] [--reuse-unchanged] [--no-tighten] [--sarif PATH] [--github] [--json]	The full verdict against the trusted baseline: gate on touched functions, ratchet, no new test failures, optional diff-coverage ceiling. The three baseline selectors are mutually exclusive; --baseline ID also bypasses the taint rule (The trusted baseline), and --baseline-tsv reads a commit-stamped file so a fresh clone verifies with no store. --no-tighten passes the verdict without rewriting the ratchet. Findings a dirty tree produced are tagged dirty and counted apart. A line under the verdict names the first three changed files, --json lists them all as changed_paths, and an untracked source file inside a scope is named on stderr as not judged. A changed file no reader could read fails the gate, exit 6, as an UNREAD finding (unread_files in --json), and --override grants nothing while one is present. It reads each istanbul artifact once for coverage, dead lines and its digest, and skips the artifact walk on an empty diff; skipping the whole run on an unchanged tree was measured and rejected, because a key made of HEAD plus the dirty names cannot see a second edit to a file that was already dirty.
worklist [--top N] [--scope NAME] [--batches N] [--json]	The risk map: every admitted function ranked by ccn * churn weight, floored by worklist_floor, with hot simple code and anything over its ceiling admitted past that floor. It ranks finished rows and no-lane rows too, marked ok and no-lane, so it never empties; next-item carries the stop condition. Every row carries the function's crap and cov off the ranked run and its ratchet_mark when the marks file in the working tree signs for it, and the header counts the active rows the cap hid: 50 of 3980 active (worklist_top 50). --scope NAME (repeatable) is exact, not a substring; a name no [[scope]] declares is a configuration error, exit 3, naming the declared scopes. --batches N adds a batches[] view cutting the active list into at most N file-disjoint batches with co-changing files kept together, off the same cached pairs coupling reads; the normal keys stay. Files go out largest summed risk first, each to the batch with the least risk so far (LPT scheduling). The text ends with the command to run next: coverage when the ranked run cannot serve as a baseline, ratchet seed while the repo has no marks file, next-item after that.
next-item [--top N] [--exclude FRAG] [--scope NAME] [--claim]	The actionable queue as JSON, with churn, budget estimates and uncovered lines. Same run and same admission floor as worklist, a different view of it: no-lane rows are skipped and counted in skipped_no_lane, and what is left is ranked by crap descending rather than by risk, so the item it hands out is often not the worklist's first row. --exclude FRAG (repeatable) skips items whose path or function name contains FRAG, and reads a path fragment the way a file argument is read, so ./pkg/legacy is pkg/legacy, while a function name keeps its case. --scope NAME (repeatable) is exact, not a substring, and a name no [[scope]] declares is a configuration error, exit 3, naming the declared scopes. --claim holds what it hands out so a second session skips it. Every item carries a handle, the name form that survives the edit the item asks for. docs/agent-json.md: next-item gives the tie order, stale, scored_changes and every field.
claims [list | release PATH NAME | release --all] [--json]	The open claims, and the way to hand one back without waiting for a verify. release takes the bare identifier, the whole long name, or the handle the claim was taken under, which is the only one that picks out a single (anonymous) claim. A claim taken before analysis version 11 on a nested Python def also answers to the name that version gives the def.
brief FILE NAME [--batch N] [--json]	The start-editing packet for one function: its own source text, every function in the file, the scored row and the scope ceiling, the ratchet mark and what the gate will bind on, uncovered lines, duplication twins, file churn, coupling partners, the config's notes, and the literal commands for the rest of the loop. Plus handle, remedy and the same est_splits / est_uncovered_paths the queue prints, and a commands.refresh that writes a run (refresh_writes_run) rather than re-reading the stale one. NAME takes the bare identifier, the long name next-item printed, the function's start line, (anonymous)#N for a function printed (anonymous) counting the file's anonymous functions from the top, or NAME#2 for the second of several functions a file gives one name to. --batch N drops the positionals and emits packets[] instead: the top N of the queue, built from one read of the store and one duplication pass over the snapshot for the whole batch (batch of 5: 11.8 s to 5.2 s, output byte-identical to five separate calls).
explain FILE NAME [--history] [--tests] [--json]	A function's score across runs plus its mark. NAME resolves exact first: a function whose bare identifier or long name is exactly NAME wins, and only when nothing matches exactly does it fall back to a prefix match, so route explains route rather than every route_* beside it. It also takes the function's start line, the form brief takes, which is how you open one printed (anonymous). --history adds the commits that touched it (git log -L over the span the run measured, carried through your uncommitted edits onto HEAD's lines), each carrying its message body as committed, and says so when the span holds only uncommitted lines or git cannot answer; --tests adds the tests that covered it, which needs coverage.py contexts turned on (recipe), and withholds them with the same note whenever the file's dark lines are withheld. --json emits the same content as one schema 1 object.
rescore FILE ... [--gate] [--json]	Fresh complexity for named files over the latest run's stale coverage, joined by name. A function on a line span another one shares, and a Python def whose body starts on the line its signature ends, scores untested, as the coverage run scores it. A function the run holds no row for (added or renamed since) or one in a scope no lane measures prints - for its cov and ends (coverage not measured); --json marks it unmeasured: true and keeps its cov, crap and remedy. Advisory: it writes no run. --gate applies the pre-commit hook's policy to the same selection the hook uses (functions the tree changed since HEAD), minus functions whose CRAP sits at or under their ratchet mark, and exits 6. It also exits 6 on a changed file no reader could read, listed under gate.unread_files in --json. A marked function past its mark is gated; the pre-commit hook pardons on the mark's existence instead, because a staged blob has no coverage to score.
ratchet seed | prune | merge | move | report [--baseline ID] [--enforce] [--json]	The mark lifecycle: seed new debt, prune gone code (a mark whose file git renamed follows it, and the prune line names up to three renames it followed; when this clone lacks the commit the renames start from and a marked file left before the oldest commit it holds, prune exits 4 and writes nothing), merge as a git driver, move re-paths marks, report reads burn-down from the file's own git history. seed ends by naming the commit and verify that follow it. seed and prune take --baseline ID to read a named run instead of verify's pick, refused for the reasons verify --baseline refuses one. See docs/ratchet.md.
runs [list | prune [--keep N]] [--json]	Run history, and retention. list marks the run verify compares against today baseline, and prints verdict=- for a run that produces no verdict rather than one that failed. See The trusted baseline. --keep (default 5) is a floor on the newest trusted runs, not a cap: the digest pair, every passing verify baseline, every run an override names, and the newest non-hook run are kept too. prune VACUUMs afterwards.
overrides [--json]	The override audit trail: who granted what, when, and why.
trend [--json]	Totals per trusted run: functions, over-target count, CRAP load, average, per-scope rollup. It reads a per-run rollup table rather than rescanning every scored row, and fills that table for any run missing one, so it writes to the store (best effort: a read-only .crapkit/ costs the speed, not the command).
digest [--alert]	The delta between the two newest runs with identical lane sets. Silent when nothing changed. Past the totals line it names up to five functions of each kind: those whose CRAP rose by more than 0.01, largest rise first; new functions over their ceiling, highest CRAP first; and functions that were over their ceiling and dropped by more than 0.01, largest drop first. Moves and scores equal to 4 decimal places list by path. An over-ceiling function the older run holds no row for reads new over ceiling when that run scored its scope, and newly scored over ceiling in scope NAME when it scored nothing there: a scope added to crapkit.toml between the two runs brings old code in, and that is not new debt. --alert pipes the body to alert_command on stdin. Plain lines, never JSON.
report [--out PATH]	One self-contained HTML page written to .crapkit/report.html (or --out PATH, repo-relative, or an absolute path you name), with the path printed on stdout. It renders what worklist --json and trend --json already answer at their defaults: the ranked worklist capped at worklist_top, the per-scope grades off the newest run, the trend series, and a banner naming every stale lane and, from scored_changes, how many files the run scored changed since. It measures nothing and opens no network connection. Every row carries the function's CRAP and coverage, and prints the crapkit explain call for the rest: dark lines, history, the mark. It reads the same per-run rollups trend does, and writes them on the same terms.
duplication [--min-lines N] [--similarity F] [--top N] [--json]	Near-duplicate functions by normalized line shingles with containment scoring. Defaults: --min-lines 8, --similarity 0.8, --top 50. Ties have a stable order across hash seeds. A positive --top bounds retained candidates and output; dense inputs still require pair comparisons. Each function is shingled from its own lines: the lines of a function nested in it, past that function's first line, are the nested function's. A function and its nested closure never pair. Blank lines and comment lines stay out, and so does a Python line starting with three quotes. Code after a block comment's closer is code (/* tag */ acc += 1;, */ x = a). s = "http://x"; /* note opens a block comment and x = 1; // see /* here does not. The reader knows plain strings and character literals only, so it misreads r"C:\", r#"a"b"#, R"(a"b)", /"/, re = /a /* b/; and echo a#b <# note: docs/commands.md: duplication gives the comment reader's rules and what each of these lines does.
coupling [--min-support N] [--min-confidence F] [--top N] [--json]	File pairs that keep landing in the same commits. Defaults: --min-support 5 shared commits, --min-confidence 0.5 max-direction ratio, --top 50. Bulk commits never couple pairs, and a young repo returns nothing at the default support. Pairs rank by support times confidence, highest first, and pairs that tie rank by their paths. The ranked pairs are cached in .crapkit/coupling-cache-v2.json, keyed on HEAD, the churn window, today's UTC date, the path format, the clone's history depth and a digest of the tracked set, and shared with brief and worklist --batches (warm: 1.05 s to 0.11 s on a 72k-commit repo). The date is part of that key, so the first run after midnight UTC rebuilds the pairs on an unchanged HEAD, and gets the same pairs: the window ends at HEAD's commit date, not today's. The depth is part of it too, so git fetch --unshallow rebuilds them the same day. --top reads the cache, because it truncates that same order; --min-support or --min-confidence off their defaults ask a wider question than the file answers, so they bypass it and recompute.
mutate [--files F ...] [--max-mutants N] [--drop-pool] [--json]	Diff-scoped mutation testing: flips comparisons, boundary shifts, boolean connectives and boolean literals on changed lines, runs mutation_command per mutant, lists survivors. --files replaces diff scope with the whole file. Both lists pass through the scored corpus first, the same predicate coverage uses: a file outside it is named on stderr and never mutated, --json lists it under outside_corpus, and when nothing is left stdout says nothing to mutate at exit 0. --max-mutants (default 100) caps the run, and mutants in --json is the capped count. A mutant whose suite timed out, or exited 5 and ran no test, counts as killed, and --json also counts it under timed_out or no_verdict, each a count inside killed. The summary counts the no-verdict ones on a line of its own, no verdict: N of the K killed ran no test (exit 5), so no test caught them. Shell and PowerShell files are refused by name on stderr, and some operators make no mutant in some languages: docs/commands.md: mutate. Every worker uses a kept worktree, including one; see mutation worktrees. --drop-pool removes them and exits.
test-scoped FILE ...	Runs each owning scope's [crapkit.scoped_tests] template on the files (quoted, longest-prefix scope wins). A file outside every scope runs only from a test, tests or __tests__ directory, under the one scope that declares a template. A template with no {files} runs as written, which is how a scope whose tests live outside its own paths runs its whole suite. Exit code only; a nonzero runner exits 1.
hook-precommit [--base REF]	The cc-only gate on staged blobs. No coverage, no snapshot, no repo-wide cache. Exit 6 on a violation, and on a staged file no reader could read, named on an UNREAD line with the reader's reason; CRAPKIT_OVERRIDE_REASON grants nothing while such a file is staged. --base REF compares the index with the merge base of REF and HEAD, the form a CI checkout runs. Outside a commit with nothing staged, as under pre-commit run --all-files, it judges every tracked file. With no crapkit.toml at or above where it runs, it gates in each root below that owns a staged file.
claude-hook [--protocol N]	Reads one PostToolUse payload from stdin, as Claude Code, Copilot CLI, Cursor or VS Code sends it, and judges each file it edited: ccn against the scope ceiling, on functions the edit changed, minus functions a ratchet mark already covers. Advisory only: the edit has landed, and hook-precommit stays the enforcement point. The advisory goes on stderr with exit 2 for Claude Code, and as one JSON object on stdout with exit 0 for Copilot CLI, Cursor and VS Code, which read exit 2 otherwise. An unscoped file, or one with no crapkit.toml above it, exits 0 in silence. A Bash event judges the working tree instead, where you register a Bash matcher (The Claude Code plugin). It opens no snapshot, and the one thing it writes is that session's record of judged bytes, under the git directory. What it prints and when it stays silent: docs/commands.md: claude-hook.
watch [--interval SECONDS] [--cycles N]	Rescores the files your scopes claim when their content changes (default every 2s, subprocess-isolated so a half-saved syntax error never kills the watcher). Each poll lists the tracked and untracked files under the scope paths again, so a file created while it runs is rescored too. A file whose mtime moved is read and rescored only when its bytes differ, so a touch or an editor saving the same bytes rescores nothing; an edit written under the file's old mtime (cp -p, touch -r) is not seen, the limit the analysis cache shares. --cycles N polls exactly N times and exits 0; without it the loop runs until ctrl-c. After a pip install -U under it, the next rescore exits 1 and says to restart it.
help [TOPIC]	The help git, npm and docker answer to. With no TOPIC it prints the command list; with one it prints that subcommand's own help, the same page as crapkit TOPIC --help. A TOPIC that names no subcommand exits 3.
mcp	A stdio MCP server with no extra dependency, exposing twelve read-side tools named verb_noun, each with a title and output schema. Tools call the CLI to inspect current scores, source and edited-file gates. They take no claims and run no verification; calls can write caches or store metadata. See the MCP contract, and Wiring crapkit into your agent for the block each agent's config file takes.

Reading the output

Every line crapkit writes from its own words is ASCII, so $x = crapkit doctor or $x = crapkit worklist 2>&1 in Windows PowerShell 5.1 captures it intact under any console code page. A path or function name crapkit quotes keeps its own characters and goes out as UTF-8; to capture one of those, set [Console]::OutputEncoding = [Text.Encoding]::UTF8 first.

Flags: why a coverage number is missing

Flag	Meaning	Scored
measured	A lane artifact spoke about this function.	Real cov.
untested	A lane covers the scope, but its artifact is silent on this function, which normally means no test imports the file.	cov = 0. A testing gap, and uncovered_lines comes back null because no artifact can name lines it never saw.
excluded	A lane's artifact measured the file and was told to leave this function out: # pragma: no cover, an exclude_lines pattern that takes every statement in it, or from coverage.py 7.10.1 a stub whose body is ..., under coverage.py; /* istanbul ignore next */ or /* v8 ignore next */ under istanbul. istanbul drops such a function from the file's fnMap, so a function missing from an instrumented file's fnMap (one written beside its statementMap) that lists others reads excluded too.	crap = ccn, and remedy can only be ok or decompose: no test can move its number. uncovered_lines comes back []. To have it measured, remove the exclusion.
no-lane	No lane's scopes list names this function's scope.	cov = 0. A tooling gap, not a testing gap. next-item never hands one out and counts them in skipped_no_lane; worklist ranks them and marks the row no-lane, because a wiring gap is a risk you have to see.
cc-only	The scope sets coverage_optional = true, so no coverage number can exist.	crap = ccn, and remedy can only be ok or decompose. uncovered_lines comes back null with a note naming that setting.

The coverage summary counts all five as measured / untested / excluded / no_lane / cc_only.

Remedy: what to do about it

Remedy	Condition	Action
decompose	ccn > ceiling	Split it. No amount of coverage clears this.
split-lines	ccn <= ceiling, crap > ceiling, and another function shares its source lines, or a Python def's body starts on the line its signature ends	Put each definition on its own lines, then measure again. Coverage cannot tell functions on one line apart, so the score stays at uncovered whatever the tests do. A one-line Python def shares its line with its def statement, which runs at import, so coverage.py cannot show a call. So does a body on the last line of a signature that spans several lines, or one that goes on from the colon's line inside brackets or after a backslash: move the body to its own line after the signature.
add-tests	ccn <= ceiling and crap > ceiling	Cover the branches.
ok	crap <= ceiling	Nothing.

The comparison is made on the exact CRAP, not on the float that computes it. CRAP(18, 2/3) is exactly 30, and its float is 30.000000000000004; at target = 30 it reads ok. The over-target count, the grade, ratchet seed and verify's gate judge it the same way.

Grade and CRAP load

The grade is the share of functions over their ceiling: A+ at exactly zero, A under 2%, B under 5%, C under 10%, D under 20%, F at 20% or more. crap_load beside it is the sum of every function's CRAP score, so it moves when a function gets better even if the letter does not. It is added exactly (math.fsum) and rounded once to 2 dp, so the order the rows come in and the Python or SQLite version never move it. Both are taken over the scopes the run measured: a scope whose lane failed or was left out by --lane scores at the cov-0 stand-in and reports its load under by_scope only. trend, coverage, the digest and brief's file totals print the same load for the same rows.

Risk: what ranks the worklist

risk = ccn * churn weight. The weight is a time-weighted sum over the file's commits in the churn window: each commit contributes a logistic weight rising to 0.5 for the newest commit in the log and falling to near zero for the oldest, so five edits last month outrank fifty from two years ago. The window reaches churn_window_months back from HEAD's commit date, never from the wall clock, so a fixed tree ranks identically forever.

Age is not the input, position in the log is. A log whose commits all share one timestamp has no range to weight against, so each commit counts once: a one-commit repo weighs every file 1.0, ranks by ccn, and promotes nothing under the floor, because a top 10% of equal weights would be every file. Commits minutes apart already rank. This repo was eight commits old, all made the same day:

$ crapkit worklist --scope util
worklist @ a7c5c85ac37 (run 1, floor ccn>=5, churn 12mo) - 3 of 3 active (worklist_top 50), 0 dormant
  risk      5.4  ccn   5  crap    30.0  cov   0%    5c/1a  util/stats.py:1  bucket( value , low , high )
  risk      4.5  ccn   9  crap    90.0  cov   0%    1c/1a  util/curve.py:1  curve( scores , mode , floor , ceiling , skip_none )
  risk      4.3  ccn   4  crap     4.2  cov  75%    5c/1a  util/stats.py:13  spread( values , cap )  ok
-> next: crapkit next-item

bucket at ccn 5 outranks curve at ccn 9 because five commits touched it and one touched curve. That is the whole point of weighting by churn. spread carries the ok marker: already at or under its ceiling, listed anyway, and next-item would not hand it out.

The list splits in two: active (files with commits in the window) and dormant (zero churn, kept out of the queue but counted). Two rules reach under the worklist_floor. A file whose churn weight sits in the top 10% is promoted down to ccn 3, which is why spread appears above at ccn 4. And a function over its ceiling is admitted whatever its ccn, so the floor can never hold back debt.

The trusted baseline

Every verify measures the working tree against one earlier run, the trusted baseline. crapkit runs list marks which one that is today.

Which runs qualify. A coverage run, or a verify that passed. A failed verify never qualifies, and neither does a partial run (a lane failed, so some scope fell back to no-lane) nor a hook override record, which carries no scored rows at all. In runs list, verdict=- marks a run that produces no verdict rather than one that failed: only verify renders a verdict. Four readers ask this one question and get this one answer: the baseline pick here, ratchet seed, prune, and the tighten damping that compares a mark against the same commit's previous run. A mark can no longer be signed off a run verify refused.

What advances it. Any qualifying run whose commit is at or behind HEAD. coverage writes one wherever HEAD is, so a dashboard cron advances the baseline exactly as CI does. A passing verify advances it and tightens the ratchet on the way.

A run on another branch never serves. The store keeps every branch's runs, and after git checkout main the newest qualifying run can be a feature branch's. verify, ratchet seed, ratchet prune and runs list all skip it and read the newest qualifying run in main's own history. When no qualifying run sits behind HEAD, verify exits 4 and says why the newest one does not: it was made on a branch HEAD does not contain (run coverage on this branch), or a rebase or an amend rewrote its commit.

The taint rule. A failed verify recorded findings against a tree. Until some verify passes, runs made after that failure do not become the baseline: choosing one would move the comparison point past the findings, the flagged function would stop counting as touched, and nothing would look at it again. verify says which run it refused and falls back to the newest run in front of the failure.

$ crapkit runs list
run   1 @ 88012a148f6 2026-08-23T09:27:46Z coverage  verdict=-      lanes=py  baseline
run   2 @ 803bdde8556 2026-08-23T09:27:53Z verify    verdict=FAILED lanes=py
run   3 @ 803bdde8556 2026-08-23T09:28:02Z coverage  verdict=-      lanes=py

$ crapkit verify
warning: run 3 is not the baseline: verify run 2 FAILED with 1 finding(s) and no passing verify has cleared it since - measuring against run 1 @ 88012a148f6 instead, so those findings stay visible. Fix them, or pass `--baseline 3` to accept the newer run deliberately.
verify FAILED @ d89068de7f3 vs baseline 88012a148f6 (2 changed files)
  changed files: calc/legacy.py, tests/test_legacy.py
  GATE  crap     72.0  ccn   8 cov 0%  calc/legacy.py:7  legacy_router( a , b , c , d , e )  -> decompose
  findings: 1 committed / 0 dirty (uncommitted edits and untracked files)

Run 3 is a coverage run somebody took on the tree run 2 refused, and it scores the same ccn-8 function. Without the rule it would have become the baseline, legacy_router would have stopped being a touched function, and that gate line would never print again.

A baseline with no test results for a lane. A trusted run can hold a lane with no test count and no failure list: the lane declared no results_artifact then, coverage --reuse-artifacts found its junit gone or unreadable, or a crapkit older than 0.8.1 stored a verify --reuse-artifacts run over such a junit. verify does not read that as a suite of zero tests that failed nothing. It compares that lane with the newest trusted run at or behind the baseline's commit that recorded it, and the line names that run:

warning: lane 'py': baseline run 2 recorded no test results, so its failures are compared with run 1's
warning: lane 'py' runs 8 fewer tests than run 1 (baseline run 2 recorded no test count for it)

When no run recorded a failure list for the lane, each of its failures still counts as new, exit 8, and a line says it may predate the change; verify --json lists the lane under lanes_without_baseline_results.

A junit this verify cannot read. verify --reuse-artifacts over a lane that declares a results_artifact it finds missing or unreadable exits 5 and stores no run, the exit a real run of that lane gets for the same file. The refusal names the lane and the file and ends run verify without --reuse-artifacts so the lane writes it again. Before 0.8.1 that verify passed at exit 0 with no test checked and stored a passing run, which became the next baseline. The Action never reaches this case: its coverage step runs the lanes and refuses the same junit first. coverage --reuse-artifacts over that junit still warns and scores on. A lane that declares no results_artifact passes with a stderr line, and verify --json lists it under lanes_without_results.

The escape, twice. Fix the findings and let a verify pass, which clears the taint for good. Or accept the newer run on purpose with verify --baseline 3: an explicit id bypasses the rule, and the run history records which run the verdict used. Nothing here touches a repo that has never run verify: with no failure to protect, coverage alone always advances the baseline.

ratchet seed --baseline ID and ratchet prune --baseline ID take the same name as verify --baseline ID, admitted by the same rule. When a failed verify pins seed to a run it cannot read or sign, that is the way out (docs/ratchet.md).

When the id you pass cannot serve. A --baseline ID naming a real run that is not a candidate says which run it is, why, and which ones can:

$ crapkit verify --baseline 3
crapkit: run 3 is an inventory run (no coverage was measured) and cannot serve as a baseline; trusted runs: 1, 2; pass `--baseline 2` for the newest

Exit codes

Code	Meaning
0	OK. For verify and hook-precommit: the gate passed.
1	Overloaded. Three unrelated things, listed below the table.
2	Usage error from argparse: unknown flag, missing positional. Raised before crapkit's own error handling.
3	Config error: crapkit.toml missing or unparseable, an unknown language or parser, a lane command the shell that runs it reads as a narrowed suite, a ratchet metric-stamp mismatch (Upgrading), a test-scoped file under no scope or under a scope with no template, a scoped file whose name is not UTF-8 or a path argument naming a file whose name is not UTF-8 (both end rename it (git mv) to a UTF-8 name), a root package.json that init cannot read as one UTF-8 JSON object (init wrote no file: ..., before it writes any file), a root on a Windows network share (the line gives the net use command that maps it to a drive letter).
4	Git error: not a repository, a repository with no commit yet, one git refuses to open (the refusal quotes git's own fix, such as a safe.directory exception), a baseline commit rewritten out of the history, made on a branch HEAD does not contain or missing from this clone, a baseline commit or fork point a shallow clone does not hold, ratchet report --enforce with a debt key set in a shallow clone (mark ages and repayments need the whole history), a ratchet prune that cannot tell whether a marked file was renamed because this clone lacks the commit its renames start from. The shallow refusals end with set fetch-depth: 0 on the checkout or run git fetch --unshallow.
5	Tool error: lizard not importable, a lane that produced no artifact, one that measured a different tree, one that measured this tree and reported it in absolute paths (the join is root-relative, so those match nothing either; the refusal names the runner's own switch, relative_files = true under [tool.coverage.run] for a coveragepy lane, the reporter's cwd/root option for an istanbul one), a lane that timed out past its retries, verify --reuse-artifacts over a lane whose declared results_artifact is missing or unreadable (it stores no run), an override alert command that failed, a process with no home directory (USERPROFILE on Windows or HOME on POSIX unset, and the operating system names none; the message names the variable to set). A timeout_seconds kills the whole process tree, so no orphan suite keeps running behind the failure.
6	Gate violation. A function the diff touched is over its ceiling and past any ratchet mark it carries: an edit that leaves a marked function at or under its mark is the debt the repo signed for and is pardoned. Also rescore --gate, which applies the same rule, and hook-precommit, which pardons on the mark's existence instead. All three also refuse a changed file no reader could read (UNREAD lines), since they judged none of its functions.
7	Ratchet regression the diff never touched. A marked function scores worse than its recorded high-water mark; a touched one past its mark reports 6.
8	New test failures against the baseline run. Failures the baseline already had do not count, against a --baseline-tsv file too: its stamp line lists each lane's failures (portable records).
9	Diff-coverage ceiling breached: diff_uncovered_max is set and more changed lines than that never ran. A changed file no lane artifact mentions counts every line of its functions.

Exit 1 means one of three things

CI cannot tell a crash from a clean policy verdict on the code alone. Which one you got depends on the command:

Command	What exit 1 means
doctor	A FAIL finding. This is a verdict, not a crash. A WARN (an unmeasured directory, or a lane writing its artifact at the repo root) and a note (a file over max_file_bytes, or no lanes declared) both exit 0.
ratchet report --enforce	The debt policy was breached. Also a verdict. In a shallow clone, with a debt key set, it judges nothing and exits 4.
anything else	An unexpected error: "no snapshot yet, run crapkit coverage first", a brief name that matches no function, a test-scoped runner that exited non-zero.

verify reports the first of 6, 7, 8, 9 that fires, in that order. A gate violation and a ratchet regression together report 6. A run that takes any of them fails, so it neither advances the baseline nor tightens the ratchet, exit 9 included.

Quickstart: Python

A repo with calc/grade.py, tests/test_grade.py, and a pyproject.toml. Commit first; crapkit reads git ls-files. Install the coverage plugin first, because the lane init writes runs pytest --cov and those flags come from pytest-cov:

pip install pytest-cov "coverage>=7.13.1"

Quote the coverage requirement, or the shell reads > as a redirect. Without it, pip keeps an older coverage.py the venv already holds, since pytest-cov accepts it, and the lane fails at exit 5 on its report.

(pip install "crapkit[py]" pulls both at once when crapkit shares the suite's venv.)

If your suite drives its own CLI through subprocess.run, add [tool.coverage.run] patch = ["subprocess"] to pyproject.toml and keep coverage>=7.13.1: pytest-cov 7.0.0 dropped subprocess measurement, so without that key every entry point scores 0% and nothing warns. 7.13.1 is also the oldest coverage whose report crapkit reads; an older one fails the lane at exit 5. docs/lanes.md has the whole rule.

1. Scaffold the config

$ crapkit init
wrote crapkit.toml with 1 scope(s): calc
detected 1 lane(s) from this repo's own files: py - next: run `crapkit coverage`
added to .gitignore: .crapkit/, .coverage, __pycache__/

init sniffs tracked source into one scope per top-level source directory, and detects a coverage lane from what the repo already has: a pytest marker file (pyproject.toml, pytest.ini, setup.cfg) writes a live [[lane]], and so does a test script or vitest/jest in package.json. A lockfile beside them names the environment: uv.lock, poetry.lock, pdm.lock or Pipfile.lock makes the lane uv run python -m pytest … (and the matching run for the rest), because a bare python binds to whichever venv the shell has active rather than the one the repo pins; see The interpreter a lane binds to. Whatever it detects, it also leaves commented templates for the runners it did not find, and those carry the same launcher, so uncommenting one cannot hand the bare python back. Every lane it writes reports into .crapkit/cov/, which is why the .gitignore list is so short: see Where artifacts live. With no lockfile the lane names its interpreter with a launcher token, so one committed line runs on every OS: {python} here, which crapkit reads as python on Windows and python3 on Linux and macOS, or {python:.venv} when a .venv in the repo holds pytest. The launcher token lists each form.

init reads each package.json past a UTF-8 byte-order mark, as npm does. A root package.json that is not one UTF-8 JSON object stops init at exit 3 before it writes any file, naming the file and the fix, since a lane read off such a file would be a guess: one in UTF-16 (what PowerShell 5.1's Out-File writes), one holding a byte that is not UTF-8, one that does not parse and one that holds an array or another value, as in init wrote no file: package.json is not UTF-8 (byte e9 at offset 36); save it as UTF-8. A nested one, a test fixture say, is skipped with one warning line naming it.

init writes .gitignore before crapkit.toml, and appends in that file's own line ending without touching a byte already there. It reads the lines past a UTF-8 byte-order mark, as git does, so an entry already there is not added twice. A UTF-16 .gitignore, which git cannot read either, is named on stderr with the fix and left as it was. Run init again over an existing crapkit.toml and it adds the .gitignore entries its lanes need, says what it finished, exits 0 and leaves crapkit.toml byte for byte. Over a UTF-16 .gitignore it exits 3 with the line that names that file and the entries to add; with nothing missing it refuses with already exists.

[crapkit]
target = 6

[[scope]]
name = "calc"
paths = ["calc"]
languages = ["python"]

[exclude]
# A leading **/ matches zero or more directories, so each glob below reaches the
# repo root and every nested copy. Test directories leave the corpus on their own.
globs = [
  "**/node_modules/**",
  "**/dist/**",
  "**/build/**",
  "**/vendor/**",
  "**/generated/**",
  "**/__generated__/**",
  "**/*.generated.*",
  "**/*.test.*",
  "**/*.spec.*",
  "**/test_*.py",
  "**/*_test.py",
  "**/conftest.py",
  "**/*_test.go",
  "**/*.config.ts",
  "**/*.config.js",
  "**/*.config.mts",
]

[[lane]]
name = "py"
command = "{python} -m pytest --cov --cov-branch --cov-report=json:.crapkit/cov/py.json --junitxml=.crapkit/cov/junit-py.xml --continue-on-collection-errors"
artifact = ".crapkit/cov/py.json"
results_artifact = ".crapkit/cov/junit-py.xml"
parser = "coveragepy"
scopes = ["calc"]

# Declare one [[lane]] per coverage command, then run `crapkit coverage`.
# [[lane]]
# name = "js"
# command = "npx vitest run --coverage --coverage.reportsDirectory=.crapkit/cov/js --coverage.reportOnFailure --reporter=default --reporter=junit --outputFile=.crapkit/cov/js/junit.xml"
# artifact = ".crapkit/cov/js/coverage-final.json"
# results_artifact = ".crapkit/cov/js/junit.xml"
# parser = "istanbul"
# scopes = ["<your-scope>"]

# `crapkit test-scoped FILES` runs one command per scope, with {files}
# replaced by that scope's files, each quoted; a template with no {files}
# runs as written, which is how a scope whose tests live elsewhere runs them.
[crapkit.scoped_tests]
# calc: no test file under calc/, so the whole suite runs, from tests/
calc = "{python} -m pytest tests -q -p no:cacheprovider"

The last block is the one an agent loop needs. crapkit test-scoped exits 3 for a file whose scope declares no template, and AGENTS.md makes it step 4 of the burn-down loop. Every key is in docs/configuration.md.

2. Check the config against the repo

$ crapkit doctor
resources: up to 8 analysis worker(s) per pool, 8 shared slot(s); lane log limit 16777216 bytes per file
ok   config keys all recognized
ok   scope 'calc': 1 file
ok   every tracked source file belongs to a scope
ok   1 lane(s) declared
ok   lane 'py': python3 -> /home/you/.venvs/ledger/bin/python3 (pytest 8.3.3, pytest-cov 7.1.0, coverage 7.13.1)
ok   lizard 1.24.0
doctor: no problems found

doctor prints one line per check and exits 1 only on a FAIL. WARN and note report and exit 0, and the closing line counts the WARNs above it (doctor: no problems found, 1 warning above).

3. Score the repo, and read the queue

$ crapkit coverage
run 1 @ fae4db93108: 2 functions scored: 2 measured, 1 over ceiling 6, CRAP load 41.0, grade F
-> next: crapkit worklist

$ crapkit worklist
worklist @ fae4db93108 (run 1, floor ccn>=5, churn 12mo) - 1 of 1 active (worklist_top 50), 0 dormant
  risk     14.0  ccn  14  crap    38.5  cov  50%    1c/1a  calc/grade.py:7  classify( score , attempts , late , bonus )
no crapkit-ratchet.tsv yet: seed marks each function over its ceiling at today's score, and from then on a mark may only fall
-> next: crapkit ratchet seed

Columns: risk, ccn, the function's crap and cov off the ranked run (both - on an inventory-only run, and cov - on a no-lane or cc-only row, which no lane measured), <commits>c/<authors>a in the churn window, path:line, the function's long name, then a marker on rows the burn-down queue will not hand out (ok, no-lane). The header counts the active rows against their total, so 50 of 3980 active (worklist_top 50) says what the cap hid, and reads (--top N) when the flag set the cap. --json also carries ccn_std, weight and ratchet_mark.

The last line names the step after this one. It is ratchet seed while the repo has no marks file, next-item once it has one, and coverage when the ranked run cannot serve as a baseline: an inventory run, a partial run or a failed verify. This walk looks at the top item first, in step 4, and seeds in step 5.

worklist is the risk map, not a to-do list. It ranks finished rows too, so it does not empty when the burn-down does. next-item is the other view of that run: it drops the no-lane rows, ranks by crap, and its empty: true is the stop condition.

4. Take the top item

$ crapkit next-item
{"commands": {"refresh": "crapkit coverage --reuse-unchanged"}, "commit": "fae4db93108b4841a00959f9117430679e7250ca", "empty": false, "item": {"authors": 1, "ccn": 14, "ccn_std": 14, "cognitive": 13, "commits": 1, "cov": 0.5, "crap": 38.5, "end": 28, "est_splits": 3, "est_uncovered_paths": 7, "flag": "measured", "function": "classify( score , attempts , late , bonus )", "handle": "classify", "nesting": 3, "nloc": 22, "occurrence": 1, "path": "calc/grade.py", "remedy": "decompose", "scope": "calc", "start": 7, "target": 6, "uncovered_lines": [9, 11, 15, 17, 19, 24, 25, 26, 27, 28], "unmeasured": false}, "run_id": 1, "schema": 1, "scored_changes": 0, "shallow": false, "skipped_no_lane": 0, "stale": false}

remedy: "decompose", est_splits: 3 (this needs roughly three pieces to fit under 6), and uncovered_lines naming the ten lines no test walks. handle is the name form to pass back. scored_changes: 0 says no file the run scored holds other content now, so these numbers describe the files on disk; anything but 0, null included, means run commands.refresh before acting on them. stale: false says HEAD is still the run's commit, a question about the commit and not the files. Every field is in docs/agent-json.md.

5. Seed the ratchet

Arm the debt gate before fixing anything. ratchet seed records every over-target function at its current score, and from then on nothing may get worse.

$ crapkit ratchet seed
crapkit-ratchet.tsv: added 1, tightened 0 - 1 mark(s) vs run 1 (fae4db93108)
-> next: commit crapkit-ratchet.tsv, then run `crapkit verify`

$ git add crapkit.toml crapkit-ratchet.tsv .gitignore && git commit -m "adopt crapkit"

A verify here would pass and give the repo its first passing verdict. This walk runs it in step 6, after the fix.

6. Fix it and verify

Extract until every piece sits at or under the ceiling. Here classify became _validate, _adjusted, _band and a classify that only sequences them, with the table of cases pushed into parametrized tests. Commit the fix, then:

$ crapkit verify
verify OK @ 8d10c13303d vs baseline fae4db93108 (5 changed files) ratchet: 1 dropped, 0 tightened -> git add crapkit-ratchet.tsv
  changed files: calc/grade.py, .gitignore, crapkit-ratchet.tsv and 2 more

$ crapkit coverage
run 3 @ 8d10c13303d: 5 functions scored: 5 measured, 0 over ceiling 6, CRAP load 19.0, grade A+
-> next: crapkit worklist

CRAP load 41.0 to 19.0, grade F to A+. verify reruns the lanes and checks three things against the trusted baseline: every function the diff touched sits at or under its ceiling, no marked function got worse, and no test that passed in the baseline fails now. Exit 0 advances the baseline and tightens crapkit-ratchet.tsv in place, so the repaid mark leaves the file: follow up with git commit -am "ratchet: classify repaid". The full mark lifecycle is in docs/ratchet.md.

crapkit next-item now comes back empty: true with a reasons object saying which ending you got. That is most of the stop condition, not all of it: AGENTS.md states the whole rule and reads the rest of reasons.

Quickstart: TypeScript

A vitest repo with src/grade.ts and test/grade.test.ts.

1. Scaffold the config

$ crapkit init
wrote crapkit.toml with 1 scope(s): src
detected 1 lane(s) from this repo's own files: js - next: run `crapkit coverage`
added to .gitignore: .crapkit/

The lane init wrote is npm run test -- --coverage --coverage.reportsDirectory=.crapkit/cov/js --coverage.reportOnFailure --reporter=default --reporter=junit --outputFile=.crapkit/cov/js/junit.xml. It reads vitest's json reporter from .crapkit/cov/js/coverage-final.json; the reportsDirectory flag is what keeps that report out of your root. The junit half is the lane's results_artifact, which the crashed-worker and no-new-failures checks read; both reporters are named because --reporter=junit alone would replace the console output you watch the suite through. Anything that produces an istanbul coverage-final.json works; see docs/lanes.md for the jest and pytest recipes, a package one directory down, and a crapkit root below the repo top.

init writes the vitest line under [crapkit.scoped_tests] commented out, because it cannot see which vitest config a file-scoped run needs. Until you uncomment that line or write your own, crapkit doctor WARNs scope 'src' has a lane but no [crapkit.scoped_tests] template. The WARN leaves doctor's exit at 0; it means crapkit test-scoped exits 3 on files under src/.

2. Install a coverage provider

This is the step that stops most TypeScript users. vitest ships no coverage provider by default. Without one, init and doctor are both happy and coverage dies with exit 5:

$ crapkit coverage
crapkit: lane 'js' FAILED: lane 'js' produced no artifact at .crapkit/cov/js/coverage-final.json (command exit 1); lane log: /repo/.crapkit/lane-js.log; last output: $ npm run test -- --coverage --coverage.reportsDirectory=.crapkit/cov/js --coverage.reportOnFailure --reporter=default --reporter=junit --outputFile=.crapkit/cov/js/junit.xml

> app@1.0.0 test
> vitest run --coverage --coverage.reportsDirectory=.crapkit/cov/js --coverage.reportOnFailure --reporter=default --reporter=junit --outputFile=.crapkit/cov/js/junit.xml

 MISSING DEPENDENCY  Cannot find dependency '@vitest/coverage-v8'

(exit 1)
crapkit: every lane failed (1 of 1); the errors are above

The two > lines are npm's banner: > <name>@<version> test from your package.json (> test when it has no name or version), then the command the test script runs.

That failure writes no run. Every lane failed, so coverage exits before it opens a store: there is no .crapkit/crap.sqlite yet and the run ids below still start at 1.

Install the provider, and pin the major yourself. Unpinned, npm resolves the newest provider against your older vitest and refuses the tree:

npm i -D "@vitest/coverage-v8@<your vitest major>"

Question	Answer
Which provider?	Either works. @vitest/coverage-v8 is vitest's default and needs no config. @vitest/coverage-istanbul also works and needs coverage.provider = "istanbul" in your vitest config.
Which crapkit parser?	Both feed parser = "istanbul". The provider name and the parser name are unrelated: v8 output is remapped to the istanbul JSON schema before it is written.
Which version?	The provider's major has to match vitest's. Read your vitest major with npm ls vitest, then install the provider at that major: on vitest 5, npm i -D "@vitest/coverage-v8@5". Drop the pin and npm answers ERESOLVE unable to resolve dependency tree, naming the peer it could not satisfy.

The artifact crapkit wants is coverage-final.json, written by vitest's json coverage reporter, which is on by default. If your vitest config sets coverage.reporter explicitly, keep "json" in the list.

vitest writes no coverage report at all when the run fails. The lane init wrote already carries --coverage.reportOnFailure, so a red test still produces the artifact. If you write the lane by hand, or you would rather keep the switch beside your other coverage settings, coverage.reportOnFailure = true in the vitest config does the same job; either one is enough. The full block is in docs/lanes.md.

3. Score the repo

$ crapkit coverage
run 1 @ 8bfbe613fcd: 2 functions scored: 2 measured, 1 over ceiling 6, CRAP load 56.68, grade F
-> next: crapkit worklist

$ crapkit worklist
worklist @ 8bfbe613fcd (run 1, floor ccn>=5, churn 12mo) - 1 of 1 active (worklist_top 50), 0 dormant
  risk     15.0  ccn  15  crap    52.4  cov  45%    1c/1a  src/grade.ts:8  classify ( row Row )
no crapkit-ratchet.tsv yet: seed marks each function over its ceiling at today's score, and from then on a mark may only fall
-> next: crapkit ratchet seed

classify is ccn 15 against a ceiling of 6: one function holding the late-and-retry penalty, the letter bands, the demotion rule and the null case.

4. Seed the ratchet and commit

ratchet seed records every over-target function at the score it has today, so nothing can get worse while you burn this one down.

$ crapkit ratchet seed
crapkit-ratchet.tsv: added 1, tightened 0 - 1 mark(s) vs run 1 (8bfbe613fcd)
-> next: commit crapkit-ratchet.tsv, then run `crapkit verify`

$ git add crapkit.toml crapkit-ratchet.tsv .gitignore && git commit -m "adopt crapkit"

5. Fix it

Above the ceiling, coverage cannot help, so classify gets split rather than tested. penalty, band and demote come out as their own exported functions, and classify keeps the null case and the bonus:

export function classify(row: Row): string {
  if (row.score === null) {
    return "N/A";
  }
  let score = row.score - penalty(row.attempts, row.late);
  if (row.bonus && score < 90) {
    score += 3;
  }
  return demote(band(score), row);
}

rescore --gate judges that edit on complexity alone, before the slow step:

$ crapkit rescore src/grade.ts --gate
rescore vs run 1 @ 8bfbe613fcd (coverage STALE, complexity fresh)
   ccn   cov     crap  remedy      function
     5     -     30.0  add-tests   src/grade.ts:22  band ( score )  (coverage not measured)
     5     -     30.0  add-tests   src/grade.ts:38  demote ( letter , row Row )  (coverage not measured)
     4     -     20.0  add-tests   src/grade.ts:8  penalty ( attempts , late )  (coverage not measured)
     4   45%      6.7  add-tests   src/grade.ts:48  classify ( row Row )
     4   75%      4.2  ok          src/grade.ts:59  average ( scores Array )
gate: 4 changed function(s) judged, 0 over ceiling 6

Exit 0 and the gate: line: every changed piece is at or under 6. The crap column is loud because its coverage half is still run 1's, from before three of those functions existed. Run 1 has no row for band, demote or penalty, so each prints - for cov, ends (coverage not measured) and scores as if no test ran it. add-tests is the literal instruction for step 6.

6. Cover the new pieces

rescore --gate passed on complexity, not on coverage. penalty, band and demote are three functions no test has ever called, so each gets a table test:

describe("band", () => {
  it.each([[95, "A"], [85, "B"], [75, "C"], [65, "D"], [10, "F"]])(
    "scores %i as %s", (score, expected) => expect(band(score)).toBe(expected));
});

Run the suite once before the slow step:

$ npx vitest run
 Test Files  1 passed (1)
      Tests  21 passed (21)

Skip this step and step 7 fails rather than passes. Run on a copy of this repo with steps 1 to 5 as written, step 5 left uncommitted and step 6 left out, verify reruns the lanes against the real tree and three functions the old suite never called come back over the ceiling. Each GATE line ends [dirty] because the edit is not committed, and the last line counts the findings that way. The block leaves out the two warnings verify prints above its verdict, one for changed lines no test covers and one for debt no mark signs:

$ crapkit verify
verify FAILED @ 0296156ff21 vs baseline 8bfbe613fcd (6 changed files)
  changed files: src/grade.ts, .gitignore, crapkit-ratchet.tsv and 3 more
  GATE  crap     17.8  ccn   5 cov 20%  src/grade.ts:38  demote ( letter , row Row )  -> add-tests  [dirty]
  GATE  crap     12.4  ccn   5 cov 33%  src/grade.ts:22  band ( score )  -> add-tests  [dirty]
  GATE  crap     10.8  ccn   4 cov 25%  src/grade.ts:8  penalty ( attempts , late )  -> add-tests  [dirty]
  findings: 0 committed / 3 dirty (uncommitted edits and untracked files)

7. Verify

$ crapkit verify
verify OK @ 2af3433d979 vs baseline 8bfbe613fcd (7 changed files) ratchet: 1 dropped, 0 tightened -> git add crapkit-ratchet.tsv
  changed files: src/grade.ts, .gitignore, crapkit-ratchet.tsv and 4 more

$ crapkit coverage
run 3 @ 2af3433d979: 5 functions scored: 5 measured, 0 over ceiling 6, CRAP load 22.0, grade A+
-> next: crapkit worklist

CRAP load 56.68 to 22.0, grade F to A+, and the mark seeded in step 4 is gone: verify dropped it once classify scored under the ceiling, rewriting the tracked crapkit-ratchet.tsv in place. Commit it with your change. Marks only ever fall.

The seven changed files are everything since run 1's commit: the three files step 4 committed, the package.json and package-lock.json that step 2's install changed, and the two files steps 5 and 6 edited.

A verify may also print warning: N changed line(s) have no coverage above its verdict; that block is advisory unless diff_uncovered_max is set (docs/configuration.md). It prints warning: N function(s) over the ceiling carry no ratchet mark and names the first three when the tree holds debt ratchet seed never signed: the gate judges touched functions only and the ratchet check compares marks only, so coverage loss on such a function would pass unseen. The count is unmarked_over_target in --json, fires no exit code, and is zero on a repo with no debt.

Documentation

Page	Covers
The handbook	Start here for anything deeper. The illustrated handbook: what crapkit is, how every piece works, and where each command earns its keep. The same page ships in the repository as docs/handbook.html, self-contained, so it opens straight from a clone.
docs/adoption.md	The judgment layer over the quickstarts: scope granularity, exclude vs lane, scoped_tests wiring, the first-verify taint hazard.
docs/configuration.md	Every crapkit.toml key: type, default, and what it does.
docs/lanes.md	The lane model, vitest and jest and pytest recipes, artifact reuse, flake retest, containers.
docs/resources.md	Worker budgets, command cleanup, log rotation and safe cleanup.
docs/ratchet.md	Seeding, pruning, the git merge driver, metric stamps, debt policy, overrides.
docs/upgrading.md	Existing installations: analysis and key versions, saved state, plugin alignment and Windows upgrades.
docs/portable-records.md	Lossless exports, portable baselines and ratchets, including filenames with delimiters.
docs/commands.md	What doctor, mutate and claude-hook check, print and refuse, and which lines duplication compares, past what a Subcommands row holds.
docs/agent-json.md	The machine surface: schema, every payload field, real captured examples.
docs/harnesses.md	Wiring crapkit into 27 agents: each one's MCP config, whether it runs the advisory hook, and how it restarts after an upgrade.
docs/comparison.md	Where crapkit sits next to radon, xenon, wily, coverage.py and SonarQube, and how they run together.
docs/accuracy.md	How crapkit checks its own numbers: each calculation against outside tools, hand tables and models, the tiers that run the checks, and every place crapkit reads a construct differently from an oracle on purpose.
AGENTS.md	The burn-down loop an agent runs, and the rules for changing crapkit itself.
CHANGELOG.md	What each release changed and how to upgrade to it.
plugin/	Three skills and the MCP server for Claude Code and Codex, and the advisory PostToolUse hook that Claude Code, Cursor, Copilot CLI and VS Code run.

crapkit.schema.json is the authority on the config file shape.

Development

pip install -e ".[dev,accuracy-push]"
git config core.hooksPath git-hooks
git config merge.crapkit-ratchet.driver "python -m crapkit ratchet merge %O %A %B"
python tools/testing/run.py

The dev extra includes pytest, pytest-cov, pytest-xdist and coverage.py; the accuracy-push extra holds the pinned oracles of the calculation-accuracy suite. The shared runner owns the unit and E2E schedule; use --unit-workers 1 for serial unit reproduction or --coverage for combined branch coverage and JUnit. The core.hooksPath line arms the complexity gate on commits and change control on pushes, and the merge.crapkit-ratchet.driver line merges crapkit-ratchet.tsv through crapkit ratchet merge instead of as text. See CONTRIBUTING.md for development, docs/accuracy.md for how crapkit checks its own numbers, and tools/deploy/README.md for the install tests each release passes.

Maintainer and project background

crapkit is created and maintained by Jean-François Gagné. Read the project background for the problem it addresses and how it fits into his work on software and AI.

License

MIT. See LICENSE.