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
Agent feedback (off by default). The plugin option feedback (or KEYSTONE_FEEDBACK=on; KEYSTONE_FEEDBACK=off
switches it off for good, the benchmarks set it) makes a SessionStart hook put a short instruction into the context.
The chain skills — start, intent, plan, check, implement, verify, fix, review — then end with a self-report of at most
10 lines about the skill text (unclear / deviated / guessed / extra, each skill § section), in the language of the
conversation; the chat gets one line. Records are appended to ~/.claude/keystone-feedback/<project>/feedback.jsonl
(all worktrees of a repo share one folder; KEYSTONE_FEEDBACK_DIR moves it) — only on this machine, never sent
anywhere, by rule no project code, file contents or secrets (known secret patterns are masked — best effort), no automatic cleanup.
/keystone:feedback.summary [--since X.Y.Z] folds the records of all projects into repeated problems with a
count; judge <plan path> / judge runs the new read-only keystone:feedback-judge agent on the signals of the work
(tool errors, repeats, interrupts, verify rounds, user corrections — counted by feedback.mjs, no model): it confirms
or doubts each claim of the reports and adds what was left unsaid, with cause and class; auto <plan path> is called by
verify on M/L — once per task, only when the signals call for it. Feedback never blocks work, is not a gate and does
not touch the verdict.