Quick Install  ·  Usage  ·  How it works  ·  The receipts  ·  Headless  ·  White paper

A Claude Code loop that finds bugs in your repo, fixes them, and proves every fix before it commits. Maintainers have merged 58 of its patches.

Jeffy Loop turns Claude Code into a self-correcting engineering loop. Type /jeffy and it audits your codebase, finds a problem, fixes it, checks the fix, and commits the result locally, one task at a time. It keeps going until the codebase passes a fresh audit, rather than stopping when the first patch looks good.

The loop is deliberately hard to fool. If a fix breaks your tests, it is undone. Nothing is ever pushed. A standard run counts as complete only when a fresh audit is clean, your tests pass again, and an adversarial review by a second AI agent with no part in the run signs off. The goal is changes that survive scrutiny.

There is a real-world test for that. Maintainers with no stake in this project have merged its patches, each filed as a pull request from a local clone, in 46 projects, including ones run by NVIDIA, Meta, Tesla, Google, Apple, Microsoft, Netflix, Apache, Oracle, IBM, Cisco, Square, Cloudflare, and more. See them all.

The record

Jeffy was run against 132 open-source projects with no connection to this repository, each judged by its own test suite. Every run is published, failures included.

Projects tested	Converged	Failed	PRs merged	PRs open	Issues filed
132	103	28	58	27	4

Converged: the closing audit came back clean and the loop's adversarial evaluator, a fresh-context sub-agent, countersigned it, a standard this repository set and checks itself. That happened in 103 projects across 13 languages with no language-specific analyzer. Failed: the run spent the budget declared before it started without converging, or, for libuv, was abandoned before it had one. That leaves PapaParse, an audit held to the same method rather than a loop run, which the receipts page counts as Fixed alongside the 103.

See every project, every patch, and every failure

Quick Install

You need Claude Code, signed in once, and git. Each installer checks for everything else and asks before installing jq.

Install Jeffy with pip

Step 1. Install the package.

pip install jeffy-loop

Step 2. Install Jeffy as a Claude Code skill.

jeffy install

---

Install Jeffy with uv

You need uv. If you do not have it, pip install uv gets it.

Step 1. Install the package.

uv tool install jeffy-loop

Step 2. Install Jeffy as a Claude Code skill.

jeffy install

---

Install Jeffy by cloning the repo locally

Mac and Linux

git clone https://github.com/lenamonj/jeffy-loop.git
cd jeffy-loop
./install.sh

Windows PowerShell

git clone https://github.com/lenamonj/jeffy-loop.git
cd jeffy-loop
powershell -ExecutionPolicy Bypass -File .\install.ps1

The -ExecutionPolicy Bypass form runs the installer on a machine where PowerShell scripts are disabled by default and changes no policy.

Running Jeffy

Open Claude Code in the project you want to improve and type /jeffy 10 into the session.

Start a new session for each run, so each run reads its state files with a clean context.

A High hunt fixes only the Highs and stops at the first audit that finds none, so it is usually the faster run.

Usage covers every flag.

What the engine enforces

Each rule is enforced by the iteration prompt, the state files or the Stop hook, and each can be checked in this repository. How.

Rule	What it means
A finding needs proof	The loop must point at the defect and prove it with a check that runs.
Three gates decide done	A fresh audit with no High or Medium findings, an adversarial evaluator's countersignature, and a shell re-run of your tests. In one greenfield build the evaluator was invoked 8 times and rejected 7. A High hunt skips the evaluator and never claims convergence.
No convergence over unswept surface	The Stop hook refuses to converge while any row of the public-surface checklist is unswept, and a row reopens when its code changes.
Lessons become checks	A rule learned once binds every later iteration. The engine passes at least 470 behavioural checks on each of Linux, Windows and macOS. How the loop improves itself.

Trust model

Question	Answer
What runs	One Stop hook, skills/jeffy/hooks/stop-hook.sh, plus the small library beside it in skills/jeffy/hooks/lib/. With no live Jeffy state file it exits at once and does nothing.
What it can touch	The loop acts through your Claude Code session with that session's permissions. Its no-push rule lives in the iteration prompt, so never allowlist push or force operations for it. Usage, Blast radius.
What the installer writes	Two skill folders under ~/.claude/skills and one hook entry in ~/.claude/settings.json. Upgrading and removing them.
How to stop it	/cancel-jeffy ends a run at any time.

---

Independent Validation - The Receipts

Each finding below was accepted upstream by the project's own maintainers.

Merged by	Finding	Merged in
NVIDIA	go-nvml #207 The buffer handed to dlinfo for a library's directory was allocated with zero bytes, so the first Path() on a library opened by soname wrote past it and the next dlclose crashed	12 days
	k8s-device-plugin #2002 An empty deviceListStrategy list in the config file passed validation, and the container started with no GPU access and no error	13 days
	nvidia-container-toolkit #2108 A GPU id repeated in the visible devices request left an empty device in the list, and in CDI mode the container failed to start with an unresolvable device	11 days
Meta	stylex #1850 stylex.positionTry emitted every declaration twice, one of them with the property name as its own value, and an RTL variant for every rule	17 days
	stylex #1849 Configuring a custom importSources on the rollup plugin or unplugin narrowed the compile gate, so ordinary @stylexjs/stylex modules stopped reaching the compiler and the custom source was never transformed	17 days
Tesla	vehicle-command #479 The proxy read every client request body with no size limit, so any client that reached the listener could drive memory allocation without bound	15 days
Google	snappy #257 Every release build compressed a 4 GiB input into a stream whose header claimed 0 bytes	1 day
	benchmark #2294 The complexity report gave its BigO coefficient in nanoseconds whatever time unit the benchmark declared	7 days
Apple	swift-log #504 A documented no-op setter asserted instead	2 days
	swift-log #503 A handler implementing only log(event:) overflowed the stack on the SwiftLog 1.0 entry point	5 days
	swift-protobuf #2164 The project's own CMake build of protoc-gen-swift had not compiled since June	15 hours
	swift-format #1286 Formatting in place replaced the file, so a 0600 source came back 0644 and a read-only one lost its bit	4 days
	swift-metrics #244 The package's own test kit crashed on a repeated dimension name	4 days
	swift-certificates #317 A /0 iPAddress name constraint matched nothing, so a CA barred from every IP address by 0.0.0.0/0 and ::/0 exclusions still verified a leaf with an IP SAN	16 days
Microsoft	mimalloc #1385 The zeroing allocator returned uninitialized memory above the small-size threshold	8 hours
	snmalloc #878 The header-only build recipe named a CMake target removed in 2021 and include paths that resolved nowhere	2 hours
	GSL #1271 The documented conversion from an iterator to its const_iterator was an access error at every use	4 days
	GSL #1272 dyn_array_iterator declared random access but had no relational operators, no -> and no n + it, so std::sort over a dyn_array did not compile	9 days
Netflix	zuul #2213 HttpQueryParams.get lower-cased the lookup key against names stored verbatim, so every query parameter with an uppercase letter in its name came back empty	17 days
Apache	commons-text #768 A StringMatcher overload forwarded the buffer end as its start	2 days
	commons-csv #633 The record counter's Javadoc said headers were not counted while the constructor's header was	3 days
	commons-lang #1784 Fraction.add and subtract overflowed on operands not in lowest terms, or returned them unreduced	1 day
	commons-lang #1783 MethodUtils.invokeMethod threw on an instance of any non-public class, every JDK collection factory result included	1 day
	commons-lang #1787 Fraction's zero shortcuts returned the other operand unreduced, and threw where the reduced result fits	6 hours
	commons-codec #443 The Git tree-id builder sorted entries by UTF-16 code units where Git sorts UTF-8 bytes, so a name outside the Basic Multilingual Plane gave a different id from git write-tree	9 days
Oracle	macaron #1466 The build spec dropped the JDK version read from the JAR whenever the artifact recorded no language version	3 days
IBM	sarama #3740 The round-robin balancer never returned when the topics map held a topic no consumer group member subscribed to	14 days
	fp-go #236 A provider graph that referred back to itself deadlocked the dependency injector forever with no diagnostic; it now reports the chain that closes the circle	10 hours
Cisco	libsrtp #821 Encrypted packets carrying no authentication tag failed to unprotect, because the key lookup stepped a full tag length back from the packet end to find the key identifier	4 days
	libsrtp #822 issue, not a patch The autotools configure aborted on stock Ubuntu because pkg-config was forced static, so no OpenSSL build was possible; reported here with the diagnosis and fixed by another contributor's #823	3 days
Square	kotlinpoet #2380 String literals it emitted turned CRLF into LF, and raw strings in a constant context let newlines pick up indentation	4 days
	kotlinpoet #2382 A class in the default package got a ClassName whose toString() threw	4 days
Cloudflare	circl #700 The PKI marshal functions panicked on the library's own post-quantum keys instead of returning an error	1 day
	circl #699 The hybrid KEM derived a different key pair from the same seed on a random subset of calls	1 day
JetBrains	kotlinx-datetime #650 Deprecation quick-fixes pointed developers at the wrong replacement	90 minutes
	kotlinx-datetime #649 The Unicode pattern parser dropped the escaped quote inside a literal	4 days
Node.js	ada #1244 The URL parser Node.js ships reported host_end one byte short	12 minutes
Canonical (Ubuntu)	pebble #938 With more than one service or check, the metrics endpoint repeated each # HELP and # TYPE family and the Prometheus parser rejected the whole response	16 days
natsort	natsort #196 The locale sentinel meant to sort last was three ASCII bytes, so PyICU keys sorted after it (19 million PyPI downloads a month)	8 days
uuid-rs	uuid #907 The UUIDv7 counter lost its top four bits to the version nibble (179 million crates.io downloads in the last 90 days)	6 days

Private security

Reported to	Outcome	Answered in
Anthropic	A security issue in claude-code-action, scored Low (2.3), reported through their program and reproduced and triaged by their own security team. The details stay unpublished at their request until the report resolves.	Scored Low (2.3) in 6 days

Contributor agreements signed

Documentation

Page	What it covers
Usage	Flags, rounds and budgets, High hunt, scoped mode, cancelling, upgrading, uninstalling, and what to know before a first run
How it works	The run lifecycle, what the engine enforces, the full rule set, what a converged stop looks like, and how the loop improves itself
Headless runs	Running budgeted rounds unattended from bash or PowerShell
The receipts	Every open-source target with its outcome, the merged patches, the greenfield builds
Contributing	The validator and the review bar
White paper	For readers new to agent loops: how loops got here, every rule from first principles, and what this method still cannot do

License

MIT