Ω Agent Omega
Engineering demonstration

Autonomous AI repair system for open-source projects.

Agent Omega finds, localizes, fixes, and tests bugs in open-source projects. Its patches have been merged by maintainers of projects including Caddy, rclone, go-ethereum, and k3s. In one documented case, the complete repair-to-merge workflow ran without human intervention.

Explore the demo ↓ Merged PRs ↓

The workflow

From open issue to upstream pull request

Five stages. The model proposes; deterministic checks hold the authority to ship. Every stage links to the real artifact it produced — in production, or in the recorded demo below.

  1. Discover production · recorded

    A daily scout ranks every open candidate issue; the executor picks unprompted.

    the scout's ranking row →
  2. Investigate benchmarked

    A read-only session names the faulty files from the issue text alone — 78.4% top-1 on 37 ground-truth issues (29 of 37).

    the localization it declared →
  3. Fix production · recorded

    A bounded session — file tools, no shell — edits a fresh checkout; only the diff is kept.

    the merged production diff →
  4. Verify deterministic

    A fix-blind regression test must FAIL on base and PASS on patch; judge ×2; 16 fail-closed gates.

    inspect all 16 gates →
  5. Submit maintainer's call

    A pull request a maintainer reviews on its own merits. Omega only reads the public state afterward.

    the merged-PR record →
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.
One real run, end to end

A bug report went in. A merged fix came out.
No human in the loop.

  1. Aug 28 a user files an ordinary bug report
  2. Sept 1 Omega ships the fix — 18 minutes, selection to pull request
  3. Sept 25 the project's maintainer merges it
Read the full anatomy → every step, every verbatim artifact
Production record

Merged pull requests

Every link is the public record on GitHub. These are the 11 merged code changes in externally-owned repositories, from 74 pull requests submitted to 30 such repositories (GitHub API, 2026-10-24). A twelfth merged PR, naimkatiman/tradeclaw#48 (April 2026), is a documentation edit that predates this system and is not a repair.

RepositoryPRChangeStatusMerged
caddyserver/caddy#7898Register nested named-route invokes transitivelyMerged2026-07-30
caddyserver/caddy#7905Flush headers immediately for SSE responsesMerged2026-07-31
caddyserver/caddy#7904Isolate auth-provider responsesMerged2026-08-01
rclone/rclone#9639serve nfs: fix EIO creating symlinksMerged2026-08-01
caddyserver/caddy#7916Isolate health-check state per check configMerged2026-08-12
caddyserver/caddy#7920Shield hostnames from wildcard client authMerged2026-08-12
ethereum/go-ethereum#35433Synthesize finalized markers only with explicit sync targetMerged2026-08-12
rclone/rclone#9726drive: pool resumable-upload chunk buffersMerged2026-08-30
rclone/rclone#9847onedrive: manual drive-ID fallbackMerged2026-09-01
k3s-io/k3s#14552Restore Flatcar detection in install scriptMerged2026-09-03
esengine/DeepSeek-Reasonix#9707Merge live proxy-observed MCP tools into catalog — shipped with no human in the loop · full anatomy →Merged2026-09-25
danielgtaylor/huma#1101Surface the Body field's example tag as the request-body exampleMerged2026-10-25
Under maintainer review

Eight open pull requests — snapshot as of 2026-10-25

A static snapshot, not a live feed. An open pull request is pending the maintainer's judgment — it is not an endorsement by these projects, and it may close without merging. The work continues daily: a scout → executor cycle has run every morning since 2026-08-31 (one full cycle, published verbatim · production record).

RepositoryPRChangeStatusOpened
ethereum/go-ethereum#36015Add missing IsAmsterdam arms to precompile activation — opened 2026-10-24 by the unattended ship pathOpen2026-10-24
rclone/rclone#9851protondrive: fall back to decrypted-name search on name-hash missOpen2026-09-02
cosmos/cosmos-sdk#26800Take a read lock rather than an exclusive lock in getContextForTxOpen2026-09-03
caddyserver/caddy#7961caddytls: stop mutating the live cache's options on reloadOpen2026-08-26
go-task/task#2998Fall back to the include dir when no taskfile is specifiedOpen2026-08-26
helm/helm#32553Skip namespace creation in --create-namespace when it already existsOpen2026-08-17
encoredev/encore#2536Stop a 304 revalidation's Content-Length from truncating cached responsesOpen2026-08-16
go-git/go-git#2238Use git's funcname heuristic for diff hunk contextOpen2026-07-02
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 on uncertainty

DecisionUnknown or incomplete results block shipment.

WhyA retry can hide a real failure, create duplicate actions, or turn an ambiguous outcome into a false success.

Trade-offSome recoverable jobs require manual investigation. Safety takes priority over throughput.

Give the second judge neutralized issue text

DecisionRemove attacker-controlled framing before the second review.

WhyIssue descriptions are untrusted input. The second judge evaluates the fix without being steered by instructions embedded in the report.

Trade-offNeutralization is deliberately greedy and can strip harmless context; if the two verdicts then diverge, the candidate goes to human review instead of shipping. A lost candidate is cheaper than a steered patch.

Keep the public demo recorded

DecisionReplay verified evidence rather than execute live model sessions on the public website.

WhyVisitors inspect the repair workflow without exposed credentials, untrusted code execution, or model charges.

Trade-offVisitors cannot submit new bugs to the public site. Real execution remains a controlled, local capability.

Long-form rationale: autonomy & authority in the showcase documentation.

Under the hood

Full technical details

Pipeline architecture
  • Localize — a read-only model session names the faulty files from the issue text alone (78.4% top-1 on 37 ground-truth issues; keyword baseline 10.8%).
  • 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: 78.4% top-1 (29/37) and 89.2% top-3 (33/37) on 37 issues with merged human fixes; keyword baseline 10.8% (4/37).
  • Fixes: an LLM judge — not test execution — rated 45.9% of patches (17/37) equivalent to the merged human fix. Of 45 end-to-end validation runs against the projects' own test suites, 4 produced a test-verified fix, 9 demonstrably failed, and 32 were environmentally inconclusive.
  • 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

Go deeper

Pick your next step

Run it yourself The full pipeline on your machine in about a minute — no API key needed. git clone github.com/SillyZir/agent-omega-lab make demo the demonstration lab →
Read the anatomy One production fix, selection to merge, told through its verbatim artifacts — including the 24 days nothing happened. anatomy of an autonomous fix →
Inspect the evidence Diffs, judge verdicts, gate reports, transcripts — every claim on this page traces to a file you can read. technical details & raw artifacts →
Follow the open work Nine PRs are under maintainer review right now; the system code and its documentation are public. the review queue → system showcase →