Skip to main content
Jeffy Loop

Validate Claude Code Platform License: MIT

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 60 of its patches.

The model writes the patch. The harness around it decides whether that patch can be trusted. Jeffy Loop is a verification harness for Claude Code: 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.

The harness 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 fixes that survive scrutiny.

There is a real-world test for that. Maintainers with no stake in this project have merged 60 of its pull requests, each filed from a local clone, into projects run by NVIDIA, Arm, Meta, Tesla, Google, Apple, Microsoft, Netflix, Apache, Oracle, IBM, Cisco, Square, Cloudflare, and more. See them all.

A run: /jeffy audits the codebase, then fixes one finding per iteration with a check that ran and passed and a local commit, until a fresh audit is clean, the evaluator agrees and the tests pass.

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 60 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.

Five ways to start a run: /jeffy for ten iterations, /jeffy 5, /jeffy 12 with a focus directive, /jeffy 5 --highs for a High hunt, and /jeffy 10 --max-time 2h.

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.

A receipt on an obsidian desk, sealed in terracotta wax pressed with the Jeffy Loop mark.

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
go-nvml #206
Shutdown() called nvmlShutdown() whether or not the library had loaded, so on a host without libnvidia-ml.so.1 the shutdown paired with a failed Init() took the process down with a symbol lookup error instead of returning ERROR_UNINITIALIZED
30 days
Arm astc-encoder #668
On F16C builds float16_to_float(vint4) packed the half-float codes with signed saturation, so every negative half decoded as NaN and a half-float normal map compressed differently on AVX2 than on SSE4.1
28 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

Metadata

Release files for jeffy-loop 1.24.7

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for jeffy-loop 1.24.7
File Size Uploaded
jeffy_loop-1.24.7.tar.gz 137.0 kB Details

Built distribution (wheel)

Table of built distributions (wheels) for jeffy-loop 1.24.7
File Interpreter ABI Platform
jeffy_loop-1.24.7-py3-none-any.whl Python 3 none any Details

Total release size: 284.8 kB

Release files / jeffy_loop-1.24.7.tar.gz

Download URL jeffy_loop-1.24.7.tar.gz
Size 137.0 kB
Tags Source
SHA-256 checksum
How to use checksums
e34dd3c8580bbb679f5ac43889c4a0cc59f21f1ef9a25a90ce5d3e6e769d180a
BLAKE2b-256 checksum
How to use checksums
f36cd5080251be98cbdb10cce1dd1a74cf0c995291c04b750583276b60abb7ef
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 6, 2026.

Transparency log

Release files / jeffy_loop-1.24.7-py3-none-any.whl

Download URL jeffy_loop-1.24.7-py3-none-any.whl
Size 147.7 kB
Tags Python 3
SHA-256 checksum
How to use checksums
da73c203fc534481f723070fe345202600901a8850461286a06e2acc53233b61
BLAKE2b-256 checksum
How to use checksums
9d9005aeba08e1b449b819365487953efa4234c11f0125c847bc7d7ff7de3c2a
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Oct 6, 2026.

Transparency log

Release history Release notifications | RSS feed

1.25.0

2 release files

This release

1.24.7 This release

2 release files

1.24.3

2 release files

1.24.2

2 release files

1.24.1

2 release files

1.24.0

2 release files

1.23.3

2 release files

1.23.2

2 release files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page