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

Autonomous Engineering With Proof

Agents that don’t just act.
They audit · verify · attack · and prove.

Jeffy Loop is an autonomous engineering system built on one principle: AI agents shouldn’t just produce work. They should produce evidence that the work is correct.

Every run is an Audit → Attack → Verify → Prove loop that ends in a receipt: what changed, why it changed, and how the result was checked.

Jeffy treats “done” as something that must be demonstrated, not declared.

The proof

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

Projects tested	Converged	Failed	PRs merged	PRs open	Issues filed
132	103	28	38	31	4

_{103 projects run to convergence across 13 languages with no language-specific analyzer or ruleset. Derived from the scorecard at render time.}

Independent Validation

A merged pull request is the one result Jeffy cannot award itself. It takes an independent maintainer, someone with no stake in this project, to review the patch and accept it into their own codebase. Maintainers across 30 open source projects have done exactly that, including those below.

Merged by	Finding	Merged in
Google	snappy #257 Every release build compressed a 4 GiB input into a stream whose header claimed 0 bytes	1 day
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
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
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
Oracle	macaron #1466 The build spec dropped the JDK version read from the JAR whenever the artifact recorded no language version	3 days
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
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
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
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, reported through their program and reproduced and triaged by their own security team. The details stay unpublished at their request until the report resolves.	14 days

See every project, every patch, and every failure

Quick Install

You need Claude Code, signed in once, and git. The installer handles everything else, including jq.

git clone https://github.com/lenamonj/jeffy-loop.git
cd jeffy-loop
./install.sh        # Windows PowerShell: .\install.ps1

If PowerShell refuses with "running scripts is disabled on this system", run in PowerShell:

powershell -ExecutionPolicy Bypass -File .\install.ps1

Or from PyPI, no clone needed; pipx install jeffy-loop and uv tool install jeffy-loop work the same way:

pip install jeffy-loop
jeffy install       # installs Jeffy as a Claude Code skill

Running Jeffy

Open Claude Code in the project you want to improve and type /jeffy 10. It is a slash command inside the session, not a shell command.

/jeffy                                     # 10 iterations, full-spectrum improvement
/jeffy 5                                   # 5 iterations
/jeffy 12 accessibility and performance    # 12 iterations with a focus directive
/jeffy 5 --highs                           # High hunt: find and fix only the Highs
/jeffy 10 --max-time 2h                    # 10 iterations, but stop after two hours either way

When the run ends, start a new session to run it again; the restart is doing real work. 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, rounds and budgets, scoped mode, and cancelling.

What the engine enforces

Each one is enforced by the iteration prompt, the state files, or the Stop hook. How.

1. It audits like an engineer, not a linter. A finding exists only when the loop can point at it and prove it with a runnable check.
2. It cannot wreck your repo. Every iteration is a local commit, a broken verify is reverted, and nothing is ever pushed.
3. "Done" is not the agent's opinion. An adversarial evaluator and a shell gate re-check every declaration.
4. It cannot declare convergence over code it never looked at. The loop maps the public surface into a checklist, every swept row records the commit it certified, and the Stop hook refuses the declaration while any row is unswept.
5. Lessons become machinery. A rule learned once binds every later iteration, and the engine itself passes 347 behavioural checks on Linux, Windows and macOS.

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

Important

Trust model. The engine is one shell script, skills/jeffy/hooks/stop-hook.sh, plus the small library beside it in skills/jeffy/hooks/lib/, registered as a Claude Code Stop hook. In a session with no live Jeffy state file it exits at once and does nothing. The installer writes two skill folders under ~/.claude/skills, one hook entry in ~/.claude/settings.json, and, only if you say yes when jq is missing, a jq install through your package manager.

License

MIT