Running every test after every change may seem prudent after a regression, but at large scale it can make feedback too slow. Progressive validation stages checks by cost and relevance, escalating only as needed. section 10.3.4
Use three tiers:
- Run local linters and static checks first to catch trivial failures quickly and keep them out of shared CI queues. section 10.3.4
- Select CI tests based on changed files and their dependencies: a checkout change runs checkout tests and direct consumers’ tests, while a leaf utility may need only 12 tests. section 10.3.4
- If selected tests still fail, send the failure to the agent; cap repair attempts at two, then flag a human rather than spending more compute. section 10.3.4
The sequence depends on trustworthy tests: fix flaky or misleading tests before optimizing feedback speed. A fast, unreliable signal can send agents chasing false failures or create false confidence. section 10.9
