Cloudflare's 'Flue' System Automates Open-Source Issue Triage with AI Agents
Cloudflare details its success in drastically reducing GitHub issues for the Astro project by implementing an AI-powered automated triage pipeline called 'Flue'.
The increasing volume of issues and pull requests in open-source projects is leading to maintainer burnout, a problem exacerbated by the ease with which AI can generate such reports. Cloudflare has addressed this challenge head-on by developing and deploying an automated triage pipeline, dubbed 'Flue,' for the Astro open-source project. This innovative system aims to tackle the overwhelming influx of GitHub issues by automating the reproduction, diagnosis, and even the initial steps toward fixing reported bugs.
At the core of Flue is a sophisticated pipeline that leverages isolated AI subagents running within GitHub Actions. This approach mirrors the manual process a developer would undertake: first, reproducing the reported issue in a controlled sandbox environment. Following successful reproduction, the system moves to diagnose the root cause of the bug, often by instrumenting the codebase and adding logging to pinpoint the exact source of the problem.
Once a bug is diagnosed, Flue's agents attempt to verify if the behavior is indeed an unintended bug or expected functionality by reviewing test suites and documentation. If confirmed as a bug, the system then proceeds to generate a fix. This involves converting the reproduction steps into failing unit tests, identifying a suitable solution based on the project's architecture, and deploying a preliminary fix.
To mitigate potential AI biases, each phase of this process is handled by a distinct, isolated subagent. These agents communicate sequentially, compiling their findings into a report.md file, ensuring a structured and auditable workflow. This modular design allows for specialized agents to focus on specific tasks, enhancing efficiency and accuracy.
The pipeline is driven by GitHub issue labels, acting as a state machine. A new submission begins with the 'triage needed' label, progressing through stages like 'fix verified' upon user confirmation. The system maintains minimal internal state, relying on the issue's comment history to determine the current status and next steps.
Upon identifying a fix, Flue automatically spins up a preview release using pkg.pr.new, posting the summary, full logs, and installation instructions back to the issue. If the original reporter confirms the patch resolves their issue, the automation then opens a pull request directly linked to the original issue, streamlining the contribution process.
What began as a specific skill for issue triage has evolved into Flue, an open, platform-agnostic framework. This generalization allows the agent automation principles—reacting to events, running isolated subagents, and separating reasoning from action—to be applied across various platforms beyond GitHub, including Slack or webhook-driven systems.
Since its implementation, Flue has dramatically reduced the number of open issues for Astro from over 200 to approximately 30, with the goal of reaching zero open issues for the first time in the project's five-year history. Cloudflare emphasizes that this was achieved through diligent automation, not by simply closing old tickets or ignoring reports, fostering a more efficient and responsive open-source development environment.