Pi, now part of Earendil, is being built with its own agent‑driven tool. The project uses the issue tracker as a prompt source, turning each ticket into a diagnostic session. This dogfooding exposes how agent‑generated text reshapes issue quality.

Issues often start human, then a clanker rewrites them with confident yet wrong hypotheses. A slash command, /is, forces Pi to ignore the initial analysis and derive its own diagnosis from code. Yet the clanker’s expansion turns narrow bugs into sprawling, inaccurate narratives.

The volume of low‑quality, LLM‑assisted tickets spikes the tracker’s maintenance load. Over 3,000 external issues in 90 days faced auto‑closure; only 17% reopened. Most pull requests never merge, reflecting a churn that burdens maintainers.

At its core, Pi’s session log demands strict invariants. Allowing clankers to tolerate malformed data inflates complexity. The goal remains to prevent bad session data from entering the system, keeping maintenance predictable and code reliable.