The Powerduck Blog
Product updates, engineering notes, and guides for designing, debugging, testing, mocking, and serving MCP tools from one OpenAPI spec.
Browse the guidesWe 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 required region setting was unset, so the agent added a hardcoded fallback instead of failing fast. The service booted cleanly and shipped against the wrong target; there was no error because the default made the missing input look optional.
Read articleThe agent added a null guard to the exact route reported, tested that case, and closed the ticket. Days later the copied sibling threw the same exception: the patch fixed the occurrence, not the pattern. A local guard treats the symptom location; the defect lives wherever the block was copied.
Read articleThe agent wrote interfaces on both ends, ran tsc, saw green, and marked it done. The UI rendered undefined because the producer nests its payload under a wrapper the declared type never mentioned. A green check proves the code agrees with itself, not that the outside world returned the shape.
Read articleThe agent wrapped the whole order job in a retry loop; on a payment timeout it restarted from the top, charged again, and sent a second receipt. Replaying finished steps turns a timeout into duplicate side effects. Retry the failed step, not the task, and make the call idempotent.
Read articleThe agent renamed displayName to fullName across types and components; fresh-fixture tests stay green, but old rows still carry the old key, so a read returns undefined. A rename is a refactor only when every sink keeps the same shape; persisted rows, caches, and queues remember the old field.
Read articleThe agent folds a specific failure into a generic envelope, and the message that named the cause disappears. A shorter error string is not simplification when it removes the field on-call needs to triage; keep the diagnostic in structured fields and test the failure path.
Read articleGiven 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 article