Project Phoenix — Autonomous CI/CD Self-Healing

Most AI tools tell you WHY a pipeline failed. Phoenix actually FIXES it. Three specialized agents diagnose, repair, and validate CI/CD failures automatically, before you even notice.

Comment

Inspiration

Developers waste 30% of their time debugging CI/CD failures instead of writing features. Every broken pipeline means lost focus, wasted compute, and frustrated engineers. We asked: what if an AI could fix it automatically, before a human even notices?

What it does

Project Phoenix is a multi-agent autonomous CI/CD recovery system built on GitLab Duo Agent Platform. When a pipeline fails, three specialized agents activate automatically:

  • Agent 1 — Diagnostic: Pulls CI/CD job logs, strips noise, identifies the exact error type, affected file, and line number
  • Agent 2 — Architect: Fetches the failing source file, analyzes the root cause, and generates a precise fix with a confidence score
  • Agent 3 — Validator: Creates a shadow branch, commits the fix, triggers a micro-pipeline, and posts a full recovery report to the original MR

The developer wakes up to a comment with the diagnosis, fix, confidence score, and a branch ready to merge. in under 60 seconds.

How we built it

  • GitLab Duo Agent Platform - Three custom agents defined in YAML under .gitlab/duo/agents/, orchestrated by a DAP Flow in .gitlab/duo/flows/phoenix_flow.yaml
  • Python - Core agent logic, GitLab API integration, log parsing, and fix generation
  • GitLab REST API - Pipeline logs, branch creation, commits, MR comments
  • Anthropic Claude - Powers fix generation via GitLab DAP (built-in, zero cost)
  • Green Scoring - Micro-pipeline only runs tests for the changed file, minimizing carbon footprint

Challenges we ran into

  • GitLab CI logs include timestamps and ANSI codes that required custom parsing to extract clean error messages
  • Shadow branch pipeline triggers required careful CI/CD rule configuration to support API-triggered pipelines
  • Wiring three agents in sequence while passing context between them required careful orchestration design

Accomplishments we're proud of

  • Built a fully working autonomous agent system in one day
  • Phoenix detects real pipeline failures, generates real fixes, and posts real comments to real MRs. end to end
  • Zero cost - runs entirely on GitLab DAP free tier
  • Green scoring reduces compute waste on every recovery

What we learned

  • GitLab Duo Agent Platform is genuinely powerful for multi-agent orchestration
  • Real agentic systems need robust error handling at every step, agents must "take action" not just "generate text"
  • The hardest part of agentic AI is not the AI, it's reliable tool execution

What's next for Project Phoenix

  • Add support for more error types (Docker, Kubernetes, Terraform)
  • Publish agents to GitLab AI Catalog for community use
  • Add @phoenix-bot mention trigger via DAP Agentic Chat
  • Integrate with GitLab Duo Chat for conversational pipeline debugging

Built With

  • anthropic-claude
  • gitlab-ci/cd
  • gitlab-duo-agent-platform
  • gitlab-rest-api
  • python
  • python-gitlab
  • yaml
Share this project:

Updates