CI/CD failures follow a predictable loop: pipeline goes red, someone scrolls logs, finds the error, fixes it, pushes, waits. Steps two through five are repetitive. They map well to pattern matching plus a structured response.
PipelineHealer handles that loop. A GitHub webhook fires on a failed workflow run, the system analyzes the failure, and PipelineHealer either opens a fix PR or files a structured issue when confidence is low.
This is my entry for the AI Dev Days Hackathon, targeting the Agentic DevOps grand prize.
Why Multi-Agent
The single-prompt approach, "here are 500 lines of CI logs, fix it", produces inconsistent results. The context is too broad, the task is too vague, and the output oscillates between useful and hallucinated.
Breaking the problem into four focused agents improved consistency:
- Log Analyzer: parses raw output, extracts structured error patterns. Does not diagnose.
- Diagnosis Agent: maps extracted patterns to early failure categories: dependency issues, lint/format failures, test failures, build config errors, and timeouts. Does not fix.
- Remediation Agent: generates a targeted fix based on a confirmed diagnosis. Does not guess from raw logs.
- Orchestrator: coordinates the pipeline, manages state, opens the PR or issue.
Each agent has a narrow scope, structured inputs and outputs, and can be tested independently. The agents communicate through typed data models, not free-form text.
The Stack
- Agents: Microsoft Agent Framework workflows, Azure OpenAI
- Backend: Python, FastAPI
- Frontend: React, TypeScript, Tailwind, TanStack Query
- Storage: Azure Cosmos DB (serverless)
- Hosting: Azure Container Apps
- Trigger: GitHub
workflow_run.completedwebhook
The flow: GitHub sends the webhook → the backend validates and persists the failed run → the agent pipeline runs → a PR or issue is created on the source repo. The launch scope covered five core failure types.
What's Built
The scaffold is in place: agents, GitHub integration, webhook handler, dashboard, Docker. Next: Azure deployment, live webhook wiring, and validation against real failures.
What's Coming
This is post one of six. The series will cover:
- Architecture decisions behind the agent pipeline
- Deploying to Azure, failure modes discovered during deployment and mitigations
- Prompt engineering for CI log parsing
- End-to-end walkthrough from webhook to pull request
- Retrospective after submission
Post 1 of 6 in the AI Dev Days Hackathon series.