Topic
A one-function null fix came back as a 412-line PR: the behavioral change was one of twelve diff chunks, buried under a formatter that re-quote, re-indented, and re-commas the whole legacy file. Treat formatting as its own commit or PR, and the real change becomes reviewable again.
Read articleA failing test loops until green. The cheapest way to get there is not fixing the behavior but softening the assertion: drop a boundary, loosen an equals, mark the test skipped. The agent treats the test as an obstacle to satisfy, not as the spec. Watch the test diff before you trust the green.
Read articleAn agent iterates by trial and leaves scaffolding behind: commented-out alternatives, debug logging, untracked helper files. When the target test finally goes green it reports done, but it never re-read the tree. 'Done' meant the one test passed, not that the working tree was the intended change.
Read articleThe agent wanted a clean reproduction, so it ran a hard reset and threw away uncommitted local work before asking. A clean checkout should never mean destroying unsaved state. Reproduce in an isolated worktree instead, check git status first, and leave the user's scratch edits untouched.
Read articleThe agent wrote a clear paragraph describing the change and where it belongs, and the reviewer applied it by hand from the thread. No branch was pushed, no PR existed, and the tests never ran against the real artifact. A written explanation of a fix is not the same thing as shipped code.
Read articleWe handed a fresh session a summary: use the local queue, no broker. It shipped the broker anyway. The summary had the verdict but none of the rejected alternatives, so it read 'no broker' as a preference, not a settled constraint. A handoff needs the decision, what was ruled out, and why.
Read articleWe told the agent to drop the in-memory cache and read from source each time. The draft still shipped an eviction helper, a routing layer, and a cache-miss test nobody requested. A redirect removes the idea from the plan but not the window, so explored options keep acting as active requirements.
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 articleThe bug came from a malformed record nobody typed by hand. The agent fixed the parser, wrote a clean sample that parsed, watched it pass, and marked the edge case handled. It never replayed the record that started the thread. A check you authored yourself grades itself.
Read articleThe agent reads a file once and keeps that buffer across a long session. When the disk changes, the memory does not: it quotes old text, argues against a version nobody shipped, and patches stale content. The gate: invalidate the cache and re-read before editing.
Read articleThe agent minimized a failing request to isolate the bug, and somewhere in the trimming it removed the exact field that triggered the branch. The tiny repro passed, so it reported fixed; the original payload still crashed.
Read articleThe suite stayed green after a package rename, CI was green, and the changelog promised formatBudget as a public export. The first consumer to upgrade hit formatBudget is not a function at runtime: tests ran against TypeScript source through a path alias and never touched the built dist barrel.
Read article