Taming a 40-Minute Lean CI: Three Rounds, Three Wrong Suspects

The pattern across all three rounds is the real finding here, every "obviously correct" hypothesis (split the big file, audit the cache key) was wrong in the same specific way: intuition pointed at the artifact's size, when the actual cost was in how many times an expensive operation got repeated across it. O(declarations × graph) collapsing to O(graph) via a shared visited set, and definitional unfolding cost hiding behind line count in Round 3, are structurally the same mistake wearing different clothes.

The cache bug is the one I'd flag as most instructive for anyone debugging CI generally: "read the timestamps before you audit the keys" is such a cheap check that gets skipped because key mismatches and eviction are the more sophisticated-sounding suspects. A composite action's internal step ordering being orthogonal to your own workflow's step ordering is exactly the kind of assumption that's invisible until you actually look at two adjacent log lines.

The PRD Loop section is the part I keep coming back to, though. "Make stopping honestly cheaper than pushing through dishonestly" reframes the whole agent-reliability problem correctly , it's not about trusting the agent to be honest, it's about removing every cheap path to a false "done" so honesty becomes the only remaining option. The miss report is the proof that worked: an agent writing "targets missed, this PR does not claim completion" and continuing on its own is only remarkable because the corner-cutting equilibrium was structurally blocked, not because the agent chose virtue unprompted.

Question on the completion audit: when it re-extracts conditions from the PRD independently and checks them against CI logs, does it ever have false negatives, flagging something as unmet when it actually was satisfied, just recorded in a place the auditor didn't check? Curious how much the auditor's own reliability had to be earned through the same kind of iteration as the main loop.