How Sentry uses a Claude skill to reproduce SDK bugs

Sentry maintains SDKs across 159 published packages spanning many languages and frameworks, and triaging bug reports on these open-source repos consumes significant engineering time. A major pain point is the tedious boilerplate needed just to reproduce an issue — matching language versions, framework setups, and SDK configurations before even confirming a bug is real. Rather than maintaining a shared repository of one-off reproduction apps, the team explored automating this setup work.

Their solution is a Claude skill that takes a GitHub issue URL, extracts language and version metadata, creates a branch, attempts a minimal reproduction using standard tooling like uv, npm, or bundle, runs it, and — if successful — opens a pull request with clear run instructions. If the reproduction proves too complex, the skill stops and explains why rather than struggling through it, which the author notes makes agents far more reliable than pushing them to complete every task.

Key lessons from building the skill include using CLIs for external system interactions, breaking work into clear steps, and adding explicit error-handling instructions. Full automation on incoming GitHub issues remains a future consideration, but the team is wary of adding to already-heavy bot noise, preferring to keep human review in the loop at the right points.