Nebius × NVIDIA Global AI Hackathon • Coding & Agentic Engineering • Verified Coding Agent

MissionaryX: a governed coding agent that repairs real repositories — and proves the repair.

MissionaryX is not a code-completion wrapper. NVIDIA Nemotron Ultra proposes repairs; MissionaryX constrains authority, executes them in Nebius ConTree, runs real tests, and promotes only evidence-backed code.
ACCEPTED LIVE RUN REPLAY D6 evidence JSON DOMINO 5 DOMINO 4
AI proposes. MissionaryX verifies. Evidence decides.

The human delegates a bounded engineering mission. The system generates, validates, executes, tests, compares, and promotes. The model cannot declare success or grade its own work.

Technological Implementation
Nebius Token Factory + NVIDIA Nemotron Ultra + ConTree

Real generation, isolated execution, deterministic verification.

Design
Autonomy inside authority

Human sets bounds; the model cannot expand scope or promote itself.

Potential Impact
Safer autonomous maintenance

Test-backed execution for higher-trust repository work.

Quality of the Idea
Evidence has final authority

Verified effects outrank model confidence or self-reported success.

What happened in the accepted live run

REAL REPO • REAL MODEL • REAL SANDBOX
01 • Mission
Real Click bugReal Click #1921 on a pinned pre-fix commit.
02 • Generate
Nemotron UltraThree independent bounded repair candidates.
03 • Govern
D6 v2.1 authorityOne allowed file, bounded edit contract, syntax and policy gates.
04 • Execute
Nebius ConTreeValid candidates applied in isolated remote sandboxes.
05 • Verify
Tests, not confidenceTargeted regression plus bounded full-suite verification.
06 • Decide
PROMOTEDDeterministic evidence selected the accepted checkpoint.

Minimal human input • explicit human authority

GOVERNED AUTONOMY
HumanAuthorize the bounded mission and live model run. The human did not write the winning patch, choose the winner, or declare the tests successful.
MissionaryXGenerate candidates, enforce edit policy, execute valid patches, run verifiers, compare evidence, and decide promotion.

The real engineering target and result

pallets/click #1921

Real open-source repository

Repository
pallets/click
Issue
Click #1921
Authorized file
src/click/types.py
Model
NVIDIA Nemotron Ultra

Accepted result

PROMOTED
2 targeted PASS
Click #1921 regression verification
2 full-suite PASS
bounded regression verification

The repair Nemotron generated and MissionaryX proved

PROMOTED PATCH
--- a/src/click/types.py
+++ b/src/click/types.py
@@ -838,7 +838,7 @@
             if self.resolve_path:
                 # realpath on Windows Python < 3.8 doesn't resolve symlinks
                 if os.path.islink(rv):
-                    rv = os.readlink(rv)
+                    rv = os.path.realpath(rv)
 
                 rv = os.path.realpath(rv)
Three proposals • governance handled the difference BEST-OF-N
Candidate 0WINNER
Request
chatcmpl-87dca6a88fae9071
Contract
PASS
Policy
PASS
Syntax
PASS
Targeted
PASS
Full suite
PASS
Candidate 1VERIFIED
Request
chatcmpl-9b32bb4bdd6dc61c
Contract
PASS
Policy
PASS
Syntax
PASS
Targeted
PASS
Full suite
PASS
Candidate 2BLOCKED BEFORE EXECUTION
Request
chatcmpl-8a7b79e81467c16d
Contract
PASS
Policy
PASS
Syntax
SYNTAX_ERROR
Targeted
INDETERMINATE
Full suite
NOT_RUN

Candidates 0 and 1 independently produced the same passing repair. Candidate 2 failed syntax validation and was blocked before sandbox execution. Deterministic tie-breaking selected candidate 0.

Technical evidence and frozen provenance
Mission
m-test-25-domino6-click1921-live
Repository URL
https://github.com/pallets/click.git
Issue URL
https://github.com/pallets/click/issues/1921
Pinned base
76cc18d1f0bd5854f14526e3ccbce631d8344cce
Contract
d6-v2.1
Promotion rule
DETERMINISTIC_ORDER_TIEBREAK
Promoted candidate
r-0-248ab8dc-c0
ConTree checkpoint
contree:a1583e26-b578-4343-b407-10b8cacb5683
Promotion SHA-256
a9c16353a36e35f0a7ca3b5a2ca8e67a6554fdb7d1d04c460beb7b605c7fbfc9
Mission outcome SHA-256
0ffa35ff8d86de818820cfcc593049adecfcf93e72b815dd10b440779cbb022d
Finalized
2026-09-11T10:48:13.889094Z
Accepted live-run replay. No new model request or sandbox execution occurs when this page loads; it also makes no repository mutation or promotion decision.