FixPilot: Autonomous 5-Stage Code Remediation Agent

Stop prompt and pray coding. FixPilot uses an NVIDIA Nemotron-3 cascade & Tavily search to diagnose runtime bugs, write surgical diffs, and prove they pass in an AST sandbox.

Comment

Inspiration and The Problem

Developers spend over 35% of their working hours reproducing, searching, and diagnosing runtime exceptions and library migration errors. Traditional AI coding assistants are blind single shot generators that suffer from three critical flaws:

They hallucinate deprecated syntax because static training weights cannot track breaking API changes, such as Pydantic version 1 to version 2 or Node.js version 22 deprecations. They operate without feedback loops. If an AI generated patch crashes on an empty collection or introduces a secondary bug, standard tools leave the developer stranded. They risk executing dangerous payloads without static AST inspection.

FixPilot replaces blind generation with adversarial closed loop agentic engineering, combining specialized reasoning tiers with real time web retrieval and deterministic sandboxed verification.

What It Does

FixPilot is an autonomous five stage code remediation engine that takes broken source code, error tracebacks, or GitHub repositories and systematically fixes and validates them:

Stage 1 (Classify): Subsecond polyglot triage between Python and JavaScript, exception categorization, and reproduction test harness synthesis using NVIDIA Nemotron 3 Nano 30B.

Stage 2 (Search Docs): Real time web retrieval querying authoritative documentation across official Python docs, MDN, and StackOverflow with relevance scoring via Tavily Search API.

Stage 3 (Diagnose): Frontier execution tracing and deep root cause hypothesis formulation using NVIDIA Nemotron 3 Ultra 550B.

Stage 4 (Synthesize): High precision code repair generating surgical patches, guard clauses, and unified git diffs using NVIDIA Nemotron 3 Super 120B and Ultra 550B.

Stage 5 (AST Sandbox Verification): Static AST security gating and isolated subprocess execution. If an edge case fails, FixPilot triggers an Adaptive Tavily Delta Search and autonomously reprompts up to 3 self healing attempts until all assertions pass.

How We Built It

Multi Model LLM Orchestration: Direct integration with the NVIDIA NIM API and Nebius Token Factory routing tasks across the NVIDIA Nemotron 3 family. Autonomous Doc Grounding: Integrated Tavily Search API with polyglot domain filtering for official Python docs, MDN developer docs, and StackOverflow. Layered AST Security Sandbox: Custom security visitor that scans AST nodes before execution to prohibit dangerous calls including os, subprocess, socket, eval, and exec, paired with Node preflight syntax validation. Real Time Streaming Engine: FastAPI backend using an asyncio queue producer consumer architecture delivering live Server Sent Events with 10 second heartbeat keepalive signals. Frontend: Linear and Vercel inspired dark minimalist UI with live model telemetry, syntax highlighted diff viewer, and one click patch export.

Challenges We Ran Into

Multilayer Bug Cascades: Fixing an obvious KeyError often uncovers a hidden downstream ZeroDivisionError on empty batches. We engineered an Adaptive Delta Grounding Loop: when the sandbox fails, the agent extracts the new stderr, executes a targeted Tavily delta search, and injects fresh documentation into Nemotron 3 Ultra for Attempt 2. Concurrency Limits and Timeouts: Handled occasional HTTP 503 worker limits on cloud clusters by implementing an escalating waterfall timeout sequence with automatic fallbacks.

Accomplishments That We Are Proud Of

28 of 28 Passing Automated Tests: Full test suite verifying AST security gates, multilanguage sandbox execution, and end to end streaming workflows. Verified Pipeline Under 20 Seconds: Full five stage remediation cycle running in under 20 seconds on real world breaking library migrations. Zero Hallucination Guarantee: Code fixes are only presented to the developer after passing 100% of the sandboxed verification assertions.

What We Learned

Specialized multimodel cascades drastically outperform monolithic LLMs for software engineering. Decomposing tasks so that lightweight models handle classification, search APIs handle real time grounding, and frontier reasoning models handle diagnosis yields higher accuracy at a fraction of the token cost.

What Is Next for FixPilot

GitHub Pull Request Bot: Automatically detect failing CI runs and open verified PR patches with diffs and reproduction test cases. Expanded Sandbox Runtimes: Adding isolated execution environments for TypeScript, Go, and Rust. IDE Extensions: VS Code and JetBrains extension bringing FixPilot's verified self healing loop directly into the editor.

Built With

Share this project:

Updates

Submission history