Asks the user to confirm a tool call whose arguments look like a repetition loop rather than data, through Hermes' approval gate (denied or unanswered means refused); replaces flagged degenerate output with a short notice; warns when a session grows long.
GeneralFlags memory and skill_manage writes that state a user decision ("the user decided ...", "用户决定 ...") without the user's own words: each such sentence needs a quote that appears verbatim in text the user typed in this session, shares key words with the decision and does not drop a negation. A patch is checked by the text it produces. Observe mode (the default) records flagged writes and `hermes provenance-gate report` lists them; block mode refuses them and tells the model how to fix the write. Measured on an independent blind set of 133 lines: precision 1.000, recall 0.694.
General