[TiantianFlow]

The lead should not take instructions in the goal for granted. The fleet researches and proposes, and the lead verifies and chooses. An audit of the skills and personas found four gaps, and this PR closes them.

Changes

• A method in the goal is a proposal. The lead persona and the lead skill now say that a method, cause, or fix that comes with the goal or in a later message is an input, not a decision. A seat checks it like any proposal. When the evidence contradicts it, the lead asks the human. It does not follow it unchecked or drop it quietly. The human's answer to that question is the decision.
• Verify, then accept. In the lead skill, "wait for that seat's board result and follow it" becomes: check the result against its acceptance and its evidence, then accept it or send it back.
• Who decides the plan. The plan is the order of the tasks and who takes each one. When the structure of the work is itself a choice, a seat proposes it. This matches the skill's existing rule that the lead does not decide "how the work is structured" alone.
• Operator relay. The operator relays the goal in the human's words instead of its own, and adds no method, cause, or plan.

docs/requirements.md section 8 (the lead and operator rows), the CHANGELOG and tests/test_skills.py are updated to match. The personas stay free of Foil mechanics, and the lead skill still never mentions the operator.

Verification

• uv run --frozen --extra dev pytest -m "not e2e_live": 574 passed. ruff check .: clean.
• Two rounds of independent review, privacy and publish-safety checks. Round 1 found that the lead could send the human's answer back to a seat for another check. That is fixed. Round 2 found nothing worth fixing.

---

Generated by Claude Code