Ronacher: AI-generated bug reports are poisoning open source issue trackers

Armin Ronacher reflects on dogfooding Pi — Earendil’s coding agent — to build Pi itself, and finds that the issue tracker has become the weakest link. Because issue text now feeds directly into agent prompts, a confidently-worded but wrong LLM-generated diagnosis isn’t merely annoying; it actively misleads the agent, which treats the report as evidence rather than rumor. A custom /is slash command instructing Pi to ignore the submitter’s analysis and re-derive root cause from code helps, but only partially, because the original human observation has already been laundered through an LLM that expanded scope and invented plausible-sounding hypotheses.

The second pattern Ronacher keeps hitting is over-engineering: told that a malformed session log crashes the reader, agents add tolerant parsers, fallbacks, migrations, and tests, rather than enforcing the invariant that bad session data should never be written in the first place. Maintainers end up repeatedly dragging the agent back from local defensive patches to global correctness, which is exhausting work.

The volume problem is concrete. Over 90 days, Pi auto-closed 2,504 of 3,145 external issues and PRs from non-approved contributors; only ~26% of issues and 8% of PRs were eventually merged or fixed. Ronacher pins blame less on GitHub’s tooling than on contributors who let agents like OpenClaw spray low-effort submissions at other people’s trackers. His ask: report what you actually observed — command, expectation, result, log — and stop outsourcing the diagnosis.