Refuse the Spike If the Assistant Nods First
A backend developer opened a blank scratch directory after a standup left one cache expiry rule unresolved. The teammate had claimed that every write should expire the whole keyspace, and the claim sounded operationally cheap. An assistant agreed in the first reply, before any failing example or constraint note had been read. That early nod became the spike subject, rather than a convenient…
A backend developer began coding without fully understanding a cache expiry rule that needed addressing. An assistant agreed without any hesitation, initiating the spike. The initial nod became the focus instead of starting to code right away. The goal of this narrow spike is to reject the rule if it's incorrect, even if a neat solution appears later.
If the assistant fails to reject the wrong rule, the spike is considered a failure. The spike runs for ninety minutes or ends earlier if no disagreement is found. Recent discussions have separated model improvement and software reliability, but this spike only checks if one assumption remains standing after interacting with a file the assistant cannot modify.
The assistant and server are both part of MonkeyCode, an open-source tool with free model access and a free server option. This spike does not evaluate any models, vendors, or industry news; it only verifies if a wrong assumption is still present. A free server is important as the spike should not interfere with the developer's daily work files.
The scratch directory can be removed after the spike, preventing any failed ideas from appearing in the main branch. The proposed setup creates a scratch directory, writes a cache module with a wrong comment, and uses an immutable fixture file to prevent any changes. A start note records the beginning time and refuses to consider any patch written after ninety minutes as valid.
The spike setup includes two Python files: one with the planted wrong comment and another with a test case that checks if the fixture contradicts the planted rule. The test case is made read-only, and the baseline command is expected to fail, indicating that the fixture is working as intended. If the baseline command succeeds, the spike should be immediately stopped, as the oracle is not performing any useful work.
The saved failure text serves as evidence and should be kept in the notes directory until the directory is deleted. The assistant is given a specific prompt to follow, and it can only edit the cache module, not any other files. The assistant must reject the planted rule or explain why the fixture is invalid, without opening a new hypothesis. The spike ends when ninety minutes have passed since the start time.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
