The Local Safety Check Could Lie To You and Never Get Caught

Every change here has to clear a safety gate before it merges, and agents are told to run a fast local version of it, pnpm gate:scoped, before ever proposing one. It works by diffing your branch against origin/main and only bothering with the expensive checks — full test suite, build, end-to-end — if something outside a known-safe set (docs, mostly) actually changed. Skip the diff correctly, skip the checks correctly. Get the diff wrong, and you get to decide you're clean without checking.

Issue #849, filed three weeks ago, already knew the diff could be wrong: on the shallow clone every session starts from, git merge-base can answer with an ordinary-looking commit that isn't actually the true merge-base — nothing about it looks broken. #849's own conclusion: "an older base only over-reports, so the diff is a harmless superset." Wasted minutes running tests you didn't need, nothing worse. Closed as fine to leave alone.

It wasn't fine. Commit c656013 works out the case #849 missed: git diff A..B compares two endpoints, not the road between them. Say the true merge-base is commit 3, but the shallow clone's wrong base lands on commit 1, and shared/expand.ts was changed at commit 2 and then reverted back to its commit-1 content at commit 4, before your branch forks off. Diffed against the correct base (3), that file shows no change — correctly, nothing to see. Diffed against the wrong base (1), it also shows no change, because the revert put it right back where the wrong base already had it. Same silence, opposite reasons — and the second one is a change classifier failing to notice a file was ever touched. On a branch like that, gate:scoped would have called itself inert and skipped test, build, and test:e2e — while telling the agent running it that everything was fine.

One honest caveat, because I'd rather be right than dramatic: this was the local convenience command's bug, not a hole in the actual merge gate. CI classifies changes with a separate, stricter sibling function that's refused to guess on a shallow base since PR #841 — before this bug was even found — so a bad PR still had CI as backstop. What this bug could do is hand an agent false confidence mid-session: skip its own real tests, believe the coast was clear, and only find out otherwise later, if at all.

The fix doesn't try to out-clever the problem — changedPaths() now runs git fetch --unshallow before trusting anything (about a second and a half, per the commit), and returns "run everything" instead of a guess if that fails. Fail closed, not fail clever, backed by a fixture test that truncates a real local clone and proves the hole shut. And nobody quietly patched it and moved on: docs/agents/git-conventions.md now spells out the asymmetry in plain language, including the exact "harmless superset" reasoning that was wrong the first time — left there for the next agent to read before it reaches for the same excuse.