QA framework that's been watching my SaaS in production

[PecanStreetAI]

I open-sourced something today: morning-qa, the QA framework that's been watching my SaaS in production every morning since May.

It's a scheduled Claude Code agent that runs a roster of read-only checks each morning (dependency CVEs, cron-job heartbeat health, data freshness…)
and posts exactly one GitHub issue per day, with a severity model, an autonomy ladder it has to climb through recorded gates (it starts with zero write access), and a calibration ledger it is structurally barred from writing itself.

A few things I think make it worth a look:

• The agent's read-only tier is enforced mechanically by tool flags, not by prompt.
• Cost engineering with receipts: the docs keep the real measured numbers, including the month an unpinned model default silently ~30x'd the bill.
• The docs are pinned to the code by 116 tests, a doc edit that drifts from the mechanism fails CI.
• A case study of a real detect → diagnose → fix loop, scored honestly five days later, including the one prediction that only partially held.

It ships 3 worked-example checks plus templates for writing your own; the production instance runs 17 daily. MIT-licensed.

https://github.com/PecanStreetAI/morning-qa