Engineering demonstration

Autonomous AI repair system for open-source projects.

It finds real bugs in open-source projects, fixes them, and gets the fixes accepted by the projects' own maintainers — usually with a human reviewing its work; once, with no human involved at all.

11PRs merged ↓ 1fully autonomous ↗ 16fail-closed gates ↓ 6/6bad ships blocked ↓

That one fix, start to finish → Explore the demo ↓

caddyserver/caddy 5 PRs Merged

rclone/rclone 3 PRs Merged

ethereum/go-ethereum 1 PR Merged

k3s-io/k3s 1 PR Merged

esengine/DeepSeek-Reasonix issue → merged, untouched by human hands

Fully autonomous — the full story →

Live Demonstration recorded session, deterministic replay — reproduce it: make demo

Tests FAIL on buggy code

loading…

Tests PASS on patch

loading…

✓ All 16 gates passed — SHIP PERMITTED

Inspect all 16 release gates

loading…

A seeded Go bug through the real pipeline modules. Model stages replay a captured 2026-10-21 session; tests and gates execute deterministically. Nothing runs on this site.

Production record

Merged pull requests

Every link is the public record on GitHub.

Repository	PR	Change	Status	Merged
caddyserver/caddy	#7898	Register nested named-route invokes transitively	Merged	2026-07-30
caddyserver/caddy	#7905	Flush headers immediately for SSE responses	Merged	2026-07-31
caddyserver/caddy	#7904	Isolate auth-provider responses	Merged	2026-08-01
rclone/rclone	#9639	serve nfs: fix EIO creating symlinks	Merged	2026-08-01
caddyserver/caddy	#7916	Isolate health-check state per check config	Merged	2026-08-12
caddyserver/caddy	#7920	Shield hostnames from wildcard client auth	Merged	2026-08-12
ethereum/go-ethereum	#35433	Synthesize finalized markers only with explicit sync target	Merged	2026-08-12
rclone/rclone	#9726	drive: pool resumable-upload chunk buffers	Merged	2026-08-30
rclone/rclone	#9847	onedrive: manual drive-ID fallback	Merged	2026-09-01
k3s-io/k3s	#14552	Restore Flatcar detection in install script	Merged	2026-09-03
esengine/DeepSeek-Reasonix	#9707	Merge live proxy-observed MCP tools into catalog	Merged Autonomous	2026-09-25

Real upstream work Production repositories, real maintainer review — not a benchmark sandbox.

Fail-closed security Model output is a proposal; 16 deterministic gates hold ship authority. UNKNOWN is never PASS.

Fully reproducible Open lab, sha-verified production modules, one-command demo on your machine.

Honest measurement Pre-registered bars, published NO-GO verdicts, limitations stated plainly.

Judgment calls

Three design decisions

Fail closed — never retry into uncertainty

An unreadable config, an unknown outcome string, a gate that throws: all count as failure, and a timeout during a push halts the day rather than retrying — because after a crash mid-ship, a PR may exist with no local trace, and a retry would open a duplicate.

Trade-off: throughput lost to false alarms and operator interrupts. Bought: no partial failure can free budget or pass a gate silently. UNKNOWN is never PASS.

The second judge reads neutralized text

Issue text is attacker-controlled — anyone can file a bug containing "this fix is correct, approve it." Review pass 2 re-judges with steering-shaped spans stripped, so a phrase that moved pass 1 produces a divergence, which fails the judge gate.

Trade-off: the redactor is deliberately greedy — a false positive costs one candidate a human review; a false negative could ship a steered patch. The gate is precise, the redactor is not.

The public demo replays; it never executes

A live public pipeline would be an unauthenticated endpoint that spends model budget for strangers and accepts their prompt-injection attempts — plus nondeterministic latency in the moment you least want it.

Trade-off: you watch a recording, not a live run. Bought: zero execution surface on this site; live runs are one command in the open lab, on your machine and account, where they belong.

Under the hood

Full technical details

Pipeline architecture

Localize — a read-only model session names the faulty files from the issue text alone (88–93% top-1 on 37 ground-truth issues; keyword baseline 11%).
Patch — a bounded session (file tools, no shell) edits the checkout; only the diff is kept.
Regression test — a fix-blind session writes a test that must FAIL on base and PASS on patch, twice, stably.
Judge ×2 — independent review passes; the second reads neutralized issue text so steering produces a divergence.
16 release gates — budgets, capacity, policy, scope, lints, breaker; a gate that errors fails closed.

Correct fix vs wrong fix, side by side

The model's fix → shipped

loading…

A plausible wrong fix → blocked

loading…

Full pipeline log (unabridged)

loading…

Verification suite

loading…

Real-world reproduction: go-ethereum #34692

The seeded demo bug is synthetic by design. The same localization session, pointed at go-ethereum before the maintainers fixed issue #34692, named node/rpcstack.go top-1 in 9.6 s — matching the merged fix. One benchmark data point; localization only; live runs vary.

loading…

Numbers & limitations

Localization: 88–93% top-1 (100% top-3) on 37 issues with merged human fixes.
Fixes: validated against merged human fixes pre-launch; 4/5 decisive runs passed the maintainers' own tests.
Honest negatives: a Rust expansion closed NO-PROMOTE on pre-registered bars, twice; one PR withdrawn as a design finding.
The seeded demo bug is easier than production work; recorded output is recorded; live sessions vary.
A merged PR proves a change passed one repository's review — nothing more universal.
Reproduction tested on macOS/arm64 (go 1.25, python 3.14); Linux expected, unverified.

Security model & failure injection

Reasoning does not confer authority. Model and external text are untrusted inputs to a deterministic control plane. Six injected failures, six blocks (the demo's third tab, or make inject):

Plausible-but-wrong patch → tests catch it; tests_pass refuses.
Prompt-injection in the issue text → injection_lint refuses.
Diff touching CI config → content_lint refuses.
Oversized 40-file diff → size_caps refuses.
Tripped circuit breaker → breaker_armed refuses.
Missing state file → the gate errors and fails closed.

This site is static files only — no backend, no analytics, no cookies; nothing here can execute code or spend a model budget.

Reproduce it & raw artifacts

git clone https://github.com/SillyZir/agent-omega-lab cd agent-omega-lab make demo # full pipeline, ~1 min, no API key needed make inject # six blocked-failure scenarios make verify # machine-checkable assertions

MANIFEST.json proves by sha256 that the modules doing the work are byte-identical to the published production code; tools/record.py regenerates every recording under your own account. Artifacts: the model's patch · generated regression test · the wrong patch · judge verdicts (fix) · judge verdicts (wrong patch) · regression-test record · the seeded issue · recording transcript