2026.09.03 16:20 ~ 17:00 Meerkat

Ending "Isn't this incident caused by the same issue as that project six months ago? Why can't we prevent it?": cross-project incident prevention strategy implemented with Reusable Workflows x Claude

Development with AI 日本語

"Isn't this incident caused by the same issue as that project six months ago? Why can't we prevent it?" If you are an engineer managing multiple Android projects, you must have felt a sudden chill at these words. - User data disappeared after an OS update due to a missing Room migration. - Only the release build crashed due to a missing ProGuard / R8 keep rule. Both are classic examples where you'd want to say, "Didn't we run into this in another project six months ago as well?" Our team consists of about 20 members cross-functionally developing 15 projects, and team meetings or Slack were the only alerts when Android-common incidents caused by OS or SDK occurred. Furthermore, because engineers rotate between projects every few months, there was a structural risk of past incidents recurring in different projects. To resolve this issue, we centralized the Android common incident catalog into a single Markdown file. Triggered by updates to this file, we used GitHub Actions Reusable Workflows to automatically distribute it to 15 project repositories, and built a mechanism to have Copilot code review reference the incident information as instructions. With this, we intended to achieve a state where "updating the catalog once automatically deploys it to all repositories, enabling cross-project guardrails via AI code reviews." However, things did not go so smoothly. Even after several months, the number of additions to the Android incident catalog from members other than myself was zero. Faced with busy daily operations, we were confronted with the harsh reality that simply creating a system does not result in incident knowledge being added. Therefore, we decided to shift the main subject of growing the catalog from humans to AI. By integrating Claude into GitHub Actions, we analyze the diffs when merging PRs in each project and determine whether "it contains a common incident that other Android projects should reference." If yes, it structures the information in a form that is easy to reuse in other repositories and automatically creates a PR to the catalog. After review and merge, Reusable Workflows distributes it to 15 repositories, and Copilot references it during reviews. Even without humans actively moving, a system was completed where the Android common incident catalog continues to grow daily. Since its introduction, we have continuously collected and deployed incident information from 15 projects, maintaining 0 major common incidents in the mobile team. In this session, we will introduce the following so that you can introduce and operate it in your own projects starting tomorrow: - Mechanism and setup method for distributing Markdown across repositories using GitHub Actions "Reusable Workflows" - Copilot instructions sharing strategy to smooth connections between projects - Prompt design and exclusion tips for integrating Claude into GitHub Actions to judge "whether it is an Android common incident" from PR diffs - Format for symptoms, causes, and recurrence prevention measures from our operational Android common incident catalog, along with entry examples you can take home and use immediately While development speed increases dramatically with AI, the base number of incidents teams hit also continues to grow. You can take home the design and implementation of incident guardrails that run automatically once built. (Translated by the DroidKaigi Committee)

Intended audience

- Lead engineers / Project managers who manage multiple Android projects and are troubled by past Android issues recurring in other projects, such as crashes caused by Room migrations or lifecycles - Those who want to learn about GitHub Actions Reusable Workflows, operate them in earnest, or consider introducing them - Those who want to integrate generative AI such as Claude into CI/CD workflows and utilize it in Android projects - Those who want to utilize GitHub Copilot code review instructions to include company Android outage knowledge into the AI review context - Those in organizations that have experienced or are likely to experience the failure of "building a system only for it to become a formality"

Sessions in the same time slot

View all sessions