MR Chronicle

AI agent flow on GitLab Duo that uses Anthropic Claude to auto-generate changelogs, release notes, MR summaries, and doc suggestions from merge request diffs and project context.

Comment

Inspiration

Every engineering team has documentation debt. Developers open merge requests, write "fixed the thing" in the description, and merge without a changelog, release notes, or risk assessment. Three months later, nobody knows what changed or why — until a breaking change ships without migration steps or an incident hits and nobody can trace what went out.

I wanted to build something that eliminates this problem entirely — zero developer effort, fully automated.

What it does

MR Chronicle is a multi-agent AI pipeline that automatically reviews, documents, and creates actionable developer handoffs for every GitLab merge request.

When triggered, three specialized agents run in sequence:

  1. Diff Analyzer — reads every changed file, categorizes the change, detects breaking changes, and performs a risk assessment (security, performance, test coverage, dependencies)
  2. Context Gatherer — looks up linked issues, checks acceptance criteria compliance, reads existing changelog conventions, and gathers release context
  3. Chronicle Writer — posts a full report as an MR comment, commits a changelog update, and labels the MR

The report includes a Merge Readiness Score (green/yellow/red), and ends with a Developer Handoff — a copy-pasteable prompt block that any AI coding agent can execute to fix flagged issues immediately.

Each agent also works independently via GitLab Duo Chat.

How I built it

  • Platform: GitLab Duo Agent Platform (Flows + Custom Agents)
  • AI Model: Anthropic Claude via GitLab AI Gateway
  • Architecture: 3-agent pipeline with routed data flow between components
  • Trigger: @mention or assign-reviewer on any merge request
  • Actions: GitLab API tools — create_merge_request_note, create_commit, update_merge_request, get_issue, list_merge_request_diffs, get_repository_file

Why Anthropic Claude

  • Convention matching — Claude reads an existing changelog and reproduces the exact style (verb tense, link format, category headers). Not template filling — language understanding.
  • Risk reasoning — The Diff Analyzer understands that logging paymentData near a payment flow is a potential PII exposure, and explains why.
  • Multi-step planning — The Chronicle Writer reads context from two previous agents, generates 7 report sections, and executes 3 GitLab API actions in one turn.

Challenges I ran into

  • Multi-agent routing required the explicit from/as input format — simple string inputs silently broke data flow between agents
  • The platform is in beta, so debugging meant reading raw CI job logs to trace agent execution
  • Finding the right prompt length balance between specificity and reliability

What I learned

  • Documentation debt is universal but invisible until something breaks
  • The Duo Agent Platform's flow system is powerful for orchestrating multi-step AI workflows with real actions
  • The Developer Handoff turned out to be the most impactful feature — it closes the loop from review to fix

What's next

  • Auto-trigger on every MR via default reviewer (zero human involvement)
  • Merge blocking rules tied to readiness score
  • Cross-MR trend analysis for documentation quality over time

Built With

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

Updates