consensus

Consensus

The agent CLI selects the reviewer automatically. Never pass a model or agent name as a positional argument; <workdir> must immediately follow agent.

1. Use /request-adversarial-review to start the review and retain the spawned reviewer's session ID and review-file path. Read the review artifact.
2. Fix each legitimate material finding. Challenge an incorrect or overstated finding with concrete evidence in the same reviewer session.
3. After fixes, do not summarize earlier findings, fixes, resolved areas, or expected remaining problems. Do not add positive or negative subject-matter limits, expected conclusions, or desired verdicts. Resume the same reviewer with a minimal prompt.

Bad:

agent <workdir> --session <id> "/adversarial-review Re-check the fixes for
X and Y. Do not broaden into Z. Write to <next-review-file>."

Good:

agent <workdir> --session <id> "/adversarial-review Re-review the entire current
work against the same authoritative sources. Inspect the current implementation
and proof. Write the complete review to <next-review-file>. Do not edit any
other files."

1. Use a new ephemeral/reviews/...-round-NN.md path for every round; never overwrite an earlier review artifact.
2. Re-review the entire current target; previous findings are not the review scope.
3. Read each artifact and stop when its outcome is only nitpicks remain or no findings.

If the reviewer returns invalid review request, discard that session and restart with /request-adversarial-review; the rejected round cannot count as consensus.

If evidence does not resolve a disputed finding after three exchanges, stop and ask HITL to adjudicate. Consensus means either the independent reviewer has no material findings or HITL has resolved the remaining dissent.