You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
implement now says a bug fix ships its regression test: one that fails on the unfixed code and passes on the fix. On ten held-out replays of real fix commits, the share of trials shipping such a test rose from 5 of 20 to 14 of 20 with correctness unchanged. Its self-review table also names a record the request says to skip, so a deliberate skip still leaves a trace.
Skills surface judgment calls and ask for unsaid constraints, answer deploy questions before requesting review, and self-review first. Reviews no longer request changes for red CI alone.
Correct stale facts and scope two rules after an audit against Claude Fable 5.2: the pinned-lease force push in git, demo-reel's verifier requirement, and several outdated paths, exit codes, and version stamps.
Add an eval harness for measuring whether a skill changes outcomes: container trials with the network off, a failure-class screen, and replays of real fix commits graded by the tests those commits shipped. The READMEs record each result with its limits, including the runs that showed no effect.