Add zudui skill: chat-driven multi-agent team bootstrapper for Claude Code

[study8677]

Summary

Adds a Claude Code skill, zudui (组队 = "form a team"), that bootstraps a multi-agent AMP team via fully conversational chat — no tmux/iTerm boilerplate, no manual file editing.

The skill complements the existing setup.md flow: instead of asking each agent to self-register one by one, the operator says /zudui, answers a few prompts ("what roles do you want, what should each do?"), and the skill handles the rest:

1. Validates broker URL + API key (or creates a fresh key via login)
2. Registers each agent on the broker with a per-team prefix to avoid 409 collisions on the same account
3. Writes per-role AGENT.md for each agent
4. Drops a smart launcher (start-team.sh) that auto-detects tmux > iTerm2 > manual
5. Pre-accepts Claude Code's workspace-trust dialog and auto-dismisses the --dangerously-skip-permissions warning, so N agent panes boot to working state with zero human keystrokes

It pairs with two sibling skills (shangban 上班 = inbox watcher, xiaban 下班 = stop) to form a complete bootstrap → run → stop lifecycle for AMP teams.

What's in the PR

• .claude/skills/zudui/SKILL.md — full conversational protocol, idempotent rerun support
• .claude/skills/zudui/start-team.sh.template — smart launcher with pre_trust_dirs() and dismiss_bypass_dialogs() helpers
• .claude/skills/zudui/agents/openai.yaml — Codex CLI surfacing (so \$zudui works there too)
• README.md + README_CN.md — one new section between "Supported agent runtimes" / "注册 Agent" and "API overview" / "API 概览" pointing to the skill

Why two layers of dialog handling

Verified by testing on a 4-agent team:

• Workspace trust is a separate gate from tool permissions. --dangerously-skip-permissions does not cover it. The launcher pre-writes hasTrustDialogAccepted: true into ~/.claude.json for the mother dir + every agent subdir.
• Bypass-permissions warning is session-only (verified by JSON diff before/after manual accept) — re-fires every boot. The launcher polls each tmux pane via tmux capture-pane, sending keystroke 2 only on panes whose buffer contains "Yes, I accept", with per-pane dedup so a stray digit never lands in claude's input.

Test plan

• End-to-end on a 4-agent team (email_planner / email_coder / email_reviewer / email_runner):
  • Reset trust to false for 5 dirs → ./start-team.sh → 5 dirs flipped to true, 4 panes spawned, no dialogs visible, ⏵⏵ bypass permissions on shown in each pane, agents reached shangban polling state.
• Idempotent rerun preserves existing AGENT.mds; only fills gaps.
• Stop semantics: tmux kill-session cleanly stops all panes (CronList schema confirms crons are session-scoped, in-memory, so they die with the claude session).
• Cherry-picked cleanly onto current upstream main; README auto-merge resolved without conflicts.

🤖 Generated with Claude Code