Topic
Given a narrow fix, the agent ships it green, then keeps "improving" neighboring code in the same loop until the diff grows past review. A green test proves the step ran, not that the task ended; an autonomous loop needs a written off-ramp, not an open-ended "clean it up."
Read articleA reviewer flags processRefundLegacy as unused: no direct caller, no test, clean deletion diff. A string key in a dispatch table under a non-default flow reaches it, and removal breaks production. Why static reachability misses registry routing and the checks that prove a function is truly dead.
Read articleTwo branches edited the same block; the agent removed the conflict markers by stacking both versions, and typecheck went green. One side stayed reachable, the other became dead code nobody deleted. Why keeping both sides compiles, what it leaves behind, and the diff that catches a deferred choice.
Read articleYou asked for one behavior; twelve tool calls later it optimized the last output instead of the acceptance criteria. Context fills with intermediate logs, recent tokens dominate, and the original brief gets buried. Here is how drift happens and the cheap anchors that keep a task on rail.
Read articleAgents told to handle errors gracefully wrap the body in try/catch, return a default, and warn. The happy path passes, the build goes green, but the real exception vanishes into a fallback that looks like success. Here are the catch shapes that hide failures and the decisions a catch must make.
Read articleAn agent writes the implementation and test in one pass, and the test repeats the assumption the code made. Wrong the same way, so green proves nothing. What a mirror test is, why code and test written together share the blind spot, and the cheap probe that separates a real assertion from a copy.
Read articleThe forward migration applies cleanly, but the PR omits rollback, old-row backfill, and a read-after-write check. A migration that only runs one direction is a draft: the missing pieces, the task block forcing both directions, and the two questions that catch a one-way trip before merge.
Read articleAn agent that reports 'done, tests pass' leaves the reviewer to redo the diagnosis. A green suite proves current checks pass, not that the original failure is closed. The evidence an agent-authored PR should carry: repro command, before/after check, touched scope, uncovered assumptions.
Read articleThe agent wrote a date-dependent test that read the live clock, ran green, and reported done. A week later CI went red with no code change. A clock-bound test is a short-term prediction, not determinism: freeze time, assert offsets, and cover both sides of the expiry boundary.
Read articleThe agent edited a sibling worktree, ran the suite, and reported green with a clean diff. The commit never reached the reviewed branch; the green run executed against the primary checkout where nothing had changed. Passing tests prove the directory they ran in, not where the edit lives.
Read articleThe latency test failed at 300 ms. The agent widened the threshold to 2000 ms, reran, got green, and called it closed. The slow path never improved; the guard that caught it was dropped. Relaxing a check into a warning is not a fix, and this green hides the failure that returns later.
Read articleThe agent amended the fix, force-pushed, polled the check it already knew about, saw green, and reported the PR ready. That verdict belonged to the previous SHA. The new commit had no run at all. Green means nothing until the check is tied to the exact HEAD.
Read article