RepoMedic
Autonomous GitHub Issue-to-PR Verification Agent
Powered by NVIDIA Nemotron 3.5 Lightning (30B MoE) via NVIDIA NIM API. Diagnoses issues, semantically retrieves code with Tree-sitter AST parsing, validates security constraints, and executes unit tests inside an isolated sandbox before opening 100% test-verified pull requests.
Zero Hallucination Gate
Strict pass/fail gate. Pull requests are never dispatched unless 100% of sandbox pytest tests pass.
Nemotron 30B MoE
30B parameter Mixture-of-Experts activating ~3.5B params per token with sub-second chain-of-thought.
Tree-sitter AST Slicing
Language-aware AST parsing indexes symbols, definitions, and callers without sending entire repos to LLM.
Tavily Deep Research
Selectively queries live documentation, GitHub issues, and PEP standards when API drift is detected.
Autonomous Issue-to-PR Walkthrough
Featuring real open-source case study python-humanize #466 and step-by-step breakdown (WHAT, HOW, WHY, WHY IT IS BEST).
Autonomous 12-Node Agent State Machine
Interact with RepoMedic's 3D state machine. Drag to orbit, click any node to inspect live telemetry, and sync audio narration.
Nemotron 30B MoE Plannerreasoning
NVIDIA Nemotron 3.5 Lightning (30B A3B) executes chain-of-thought to formulate repair plan.
Open-Source Bug Fix: python-humanize/humanize #466
Live authentic bug resolution in a 2,500+ star open-source repository verified in an isolated execution sandbox.
12-Node Pipeline Execution Summary
Reported Bug: Passing negative ndigits=-1 to intcomma() crashed with an obscure ValueError: Format specifier missing precision.
Nemotron 30B MoE Diagnosis: Located src/humanize/number.py, synthesized input validation if ndigits < 0: raise ValueError("ndigits must be non-negative"), added targeted regression tests in tests/test_number.py, and ran full pytest in the isolated sandbox with 241 tests passing (0 failures, 0 regressions).
python3 scripts/run_opensource_proof.py~210s end-to-endProof Execution Telemetry
Key Architectural Innovations
Architected specifically for the Nebius × NVIDIA Global AI Hackathon (Coding & Agentic Engineering Track).
NVIDIA Nemotron 30B MoE Engine
Sub-second Latency & Active Chain-of-Thought- Nemotron 3.5 Lightning (30B A3B): 30 billion total parameters with ~3.5 billion active per token served via NVIDIA NIM with streaming chain-of-thought tokens (
enable_thinking=True). - 12-Node Cyclic LangGraph: Full state machine with triager, planner, coder, test synthesizer, reflection, and gate nodes.
- Tree-sitter AST Analysis: Language-aware code chunking indexing symbols and cross-file dependencies.
- PatchValidator Security Guard: Hardened security layer blocking absolute paths, directory traversals, and non-whitelisted files.
Zero-Hallucination Verification Policy
Deterministic Sandbox Verification Gate- Zero Hallucinated PRs: Most coding assistants create unverified diffs that break CI. RepoMedic executes tests inside a sandbox and strictly rejects PR creation unless 100% of tests pass.
- Self-Healing 3-Iteration Cycle: If tests fail, the agent parses the pytest stack trace, reflects on the failure mode, and re-plans repairs automatically.
- Regression Test Synthesis: Synthesizes new targeted regression test cases before modifying production code.
Real-time Developer Observability
Full Telemetry Across All 12 Nodes- Live Node Stepper: Step-by-step visual tracker showing real-time agent progression across all 12 nodes.
- Unified Diff Viewer: Clear color-coded file diffs highlighting additions, deletions, and AST modifications.
- Terminal Test Execution Logs: Real-time streaming of sandbox pytest output, showing exact pass/fail counters.
- Automated PR Documentation: Formats clean PR markdown with reproduction steps, fix explanations, and test evidence.
Eliminating Maintainer Burnout
Hours of Triage Reduced to Minutes of Automated Execution- Velocity Shift: Open-source maintainers spend hours each week triaging issues, reproducing failures, and reviewing broken PRs. RepoMedic reduces resolution time from hours of manual triage down to 1-3 minutes of automated, test-verified execution.
- High-Confidence Automation: By enforcing strict sandbox verification, maintainers can safely enable automated merge for verified low-risk bug fixes.
- Drop-In GitHub App: Installs on any repository with zero configuration changes required.
Autonomous Knowledge Engine: Tavily Deep API Research
Real-time External Documentation & API Drift IntelligenceWhen bug reports involve third-party dependency updates or breaking API changes (e.g., deprecated function arguments or framework migrations), RepoMedic automatically invokes Tavily AI Search to fetch real-time documentation, changelogs, and GitHub migration discussions before drafting the code repair.
Interactive 5-Stage System Pipeline
Click through the stages or simulate live execution to inspect AST trees, chain-of-thought tokens, and sandbox verification.
End-to-End Autonomous Issue-to-PR Execution Pipeline
Webhook & Ingestion
GitHub Issue Ingestion & Stacktrace Extraction
Tree-sitter AST Slicing
Zero-Hallucination Symbol Call Graph Indexing
Nemotron 30B MoE Engine
NVIDIA NIM API with Streaming CoT Reasoning
Deterministic Sandbox
Isolated Docker Container with Full Pytest Verification
Security Gate & PR Dispatch
PatchValidator AST Guard & Signed GitHub PR Creation
Nemotron 30B MoE Engine
30B Mixture-of-Experts model activates ~3.5B parameters per token, delivering sub-second deep chain-of-thought code reasoning.
<thought>
1. Issue analysis: intcomma() crashes when ndigits is negative (-1).
2. Root cause: f".{ndigits}f" evaluates to f".-1f", invalid Python float format specifier.
3. Decision: Explicit input validation at entry point of intcomma().
4. Validation rule: if ndigits < 0: raise ValueError("ndigits must be non-negative").
5. Regression test: Create test_intcomma_negative_ndigits() asserting ValueError.
6. Target file: src/humanize/number.py (lines 58-62).
</thought>Live Verification Workbench & Diff Viewer
Switch between the unified syntax-highlighted diff, interactive pytest sandbox runner, and Nemotron 30B reasoning inspector.
Strict gate enforced: 241/241 unit tests verified in isolated sandbox before PR dispatch.
Previously, passing ndigits=-1 propagated into Python float formatting f".{ndigits}f", raising ValueError: Format specifier missing precision. The boundary check safely raises a descriptive ValueError and preserves backwards compatibility.