Challenge decision (Hermes edition)

Two modes. Pick by whether the decision has already been made.

Mode A — red-team (before)

1. sb_decision_context(subject=<the decision in one or two sentences>).
2. Argue from what comes back, not from principles. Every objection cites a note path. If similar_decisions is empty, say the red-team is weak for lack of prior art — do not compensate by inventing generic risks.
3. Give at most five objections, strongest first. For each: the claim it attacks, the evidence, and what would have to be true for the objection to be wrong.
4. Always cover these three, because they are the ones people skip:
   • What was reversed before — from reversed_precedents, and what is different this time.
   • Who is not in the room — from stakeholders, whoever the plan affects but never consulted.
   • How you would know you were wrong — a falsifiable reversal condition with a date to check it.
5. End with a verdict in one line: proceed, proceed with changes (name them), or hold and find out X first.
6. Write nothing. If the user commits afterwards, offer to record it with sb_create_note(type="decision", ...) including the reversal conditions from step 4.

Mode B — postmortem (after a reversal)

1. sb_decision_postmortem(path="05-decisions/...").
2. Name the lesson in one sentence, in the user's own vocabulary. "We underestimated the migration" is a lesson; "communication could be improved" is not.
3. For each entry in notes_resting_on_the_same_hypothesis, sb_read it and propose — do not apply — a concrete update. Show the list, ask for a go, then apply the ones the user approves.
4. Record the reversal on the decision note with sb_append_timeline(path, "Reversed — <lesson>", [...]). 05-decisions/ is append-only: never edit the rationale, never rewrite history.
5. File the lesson: sb_create_note(type="knowledge", fields={domain, confidence}, ...) linking back to the decision, then sb_curate(path).
6. sb_commit("postmortem: <decision>"). Report what was learned and which notes were updated.