AutoDebug — 5-Agent Bug Detection Pipeline
AutoDebug is a multi-agent pipeline built entirely on Band that autonomously detects, analyses, and reports bugs in GitHub repositories — with a human always in the loop for the final fix. Five independent agents run as real Band-registered processes, coordinating through a shared Band chat room using @mentions. Each agent hands off accumulated context as JSON to the next, forming a clean pipeline: 1. DetectorAgent — clones the repo, finds failed CI runs via GitHub Actions API or runs Python files directly to surface errors 2. AnalyserAgent — uses AI to identify root cause, severity (low/medium/high/critical), and affected files 3. FixerAgent — generates a precise, minimal code fix with explanation using AI 4. TesterAgent — runs pytest to confirm the bug is real and reproducible 5. ResolverAgent — writes a clear, actionable fix report and delivers it by email The system never auto-merges code. Every fix is a report, not a commit — a human always reviews and applies the change. This human-in-the-loop design makes AutoDebug safe to run on production repositories. Built with Band SDK, NVIDIA API (MiniMax M2.7), GitHub API, Gmail SMTP, and pytest. Includes a CLI tool (autodebug start/stop/status/doctor) and a local fallback pipeline for offline testing.
Category tags:
"Application of technology: Five Band-registered debugging agents (Detector/GitHub Actions API + CI clone, Analyser/root cause + severity, Fixer/minimal code fix, Tester/pytest reproducibility confirmation, Resolver/email report delivery) with JSON context handoffs via @mentions. NVIDIA MiniMax M2.7 + Claude + OpenAI. CLI tool with start/stop/status/doctor commands. Local fallback pipeline. AgentOps monitoring. Presentation: "Never auto-merges code — every fix is a report, not a commit" is a precise and responsible HITL statement for production repository safety. CLI with autodebug doctor for system health checking is a practical developer UX addition. Student submission with polished engineering. Business value: CI/CD bug detection and fix report generation with pytest reproducibility confirmation and Gmail email delivery creates a developer-friendly workflow. GitHub Actions API integration enables automatic triggering on CI failures. Originality: CLI interface with doctor command for pipeline health checking (autodebug doctor) is a practical operational tool. Tester Agent confirming bug reproducibility via pytest before FIxerAgent generates a fix prevents generating fixes for false positives."
Sanem Avcil
"Hi Ekaanksh — thank you for a genuinely useful tool; this is the kind of thing developers would actually reach for. What I liked: It genuinely runs on Band — the five agents are real Band agents, set up with the Band SDK and coordinating through @mention handoffs. A real, working pipeline: clone a repo, find failing tests, propose a fix, run the tests again, and report back. A responsible, safe design — it never auto-merges; a human always reviews before anything is applied. Clear, well-separated agent roles and a handy command-line tool. A few small things to tidy: The Band dependency isn't pinned in the installation file, so a fresh setup won't build cleanly. Some earlier code (a local message-queue path) is still in the repo alongside the real Band path — removing it would avoid confusion. The submission video has no narration — a short voiced walkthrough would show the pipeline off better. Thank you — genuinely solid work and great to see real Band coordination under the hood."
.png&w=96&q=75){kind=small}
Andrew Kozlov
Principal Agentic Engineer
For Innovators & Creators