hive-mind: Command for Claude Code

Capture a decision from this conversation into the hive-mind decision graph.

Arguments (optional): $ARGUMENTS — the decision, stated plainly. If empty, infer the most recent settled decision from the conversation.

Steps:

1. Draft the record from the conversation. Do not invent content; if a field has no support in the conversation, leave it empty and say so.
   • statement — the ruling, one sentence, imperative. "Use X, not Y."
   • decision_key — a stable lowercase kebab-case topic that survives wording changes, such as python-package-manager.
   • rationale — why. Cite the actual constraint that forced it.
   • alternatives_rejected — every approach considered and dropped, each with its reason. This field is the point of the exercise: it is what stops the same approach being proposed again next month.
   • scope — repo (this repo), client (all of this client's repos), or practice (every project, engineering-craft rulings).
   • decided_at — when the ruling was settled. Use the conversation timestamp; never substitute the current time when recording an older ruling.
2. Show the draft and ask for confirmation or edits. Do not call the tool yet.
3. On confirmation, call record_decision with the confirmed fields, including decision_key and decided_at.
4. If this decision replaces an existing one, first call recall_decisions to find the old uuid, then use supersede_decision instead — never record a contradicting decision alongside the one it replaces.

Review mode

If $ARGUMENTS starts with review, switch to reviewing the candidate queue — decisions extracted automatically from earlier sessions, none of which are in the graph yet.

1. List pending candidates:

   uv run --project "${HIVE_MIND_HOME:?Set HIVE_MIND_HOME to the Hive Mind checkout}" python -c "
   from hive_mind.candidates import pending
   for c in pending():
       print(f'{c.id}  [{c.scope}/{c.repo or c.client}]  '
             f'key={c.decision_key} decided={c.decided_at or \"unknown\"}  '
             f'{c.statement}')
   "
   

2. Walk them ONE AT A TIME. For each, show decision_key, statement, rationale, alternatives_rejected, and decided_at, then ask: accept / edit / discard. If a legacy candidate has no decided_at, inspect its source session timestamp or ask the user; never use extracted_at as a silent substitute.

   • accept — call record_decision with the candidate's fields, then mark it resolved (below). Pass source='harvested', the candidate's session_id, decision_key, and decided_at. Never call record_decision before the user has said accept.
   • edit — apply their changes, confirm the revision, then accept.
   • discard — mark it discarded. Say nothing further; a rejected candidate is a normal outcome, not a failure.

3. Mark resolved after each:

   uv run --project "${HIVE_MIND_HOME:?Set HIVE_MIND_HOME to the Hive Mind checkout}" python -c "
   from hive_mind.candidates import resolve
   resolve('<candidate-id>', '<accepted|discarded>')
   "
   

4. Stop after five in one sitting and offer to continue. A long review turns into rubber-stamping, which defeats the confirmation gate.

Extraction is deliberately noisy and the queue is where that noise is absorbed. If most candidates are task summaries rather than decisions, say so — the fix is the extraction prompt in hive_mind/extract.py, not more careful reviewing.