tmux-teams

Consult an Antigravity/Gemini advisor over ACP and get the answer back as a bmad-party-mode round-table, never as a single voice. Use when the user invokes $agy-advisor, wants an opinion from outside both the Claude and the OpenAI families, names the AGY seat, or adds --party <id> to seat a saved bmad-party-mode roster. Read-only: it advises, it never edits.

Consult a Claude-protocol advisor over ACP and get the answer back as a bmad-party-mode round-table, never as a single voice. Defaults to the top Claude model (fable at the session's effort); optionally takes a routing pair — $claude-advisor <bin> <model> — to reach k3, qwen3.8-max, deepseek or glm through a routed profile. Use when the user invokes $claude-advisor, asks for a second opinion, names a specific advisor seat, or adds --party <id> to seat a saved bmad-party-mode roster. Read-only by brief and by this skill issuing no write instruction; unlike the Codex lane there is no mode switch to enforce it.

Consult a Codex advisor over ACP and get the answer back as a bmad-party-mode round-table, never as a single voice. Takes an optional model — $codex-advisor [luna|sol|astra] — and locks reasoning effort to each model's maximum (ultra for astra and sol, max for luna; default sol[ultra]), which this adapter reports and the dispatch verifies. Use when the user invokes $codex-advisor, wants a second opinion from outside the Claude family, names a specific Codex seat, or adds --party <id> to seat a saved bmad-party-mode roster. Read-only: it advises, it never edits.

Use as the mandatory first-run setup for the delivery loop, right after installing this plugin, and whenever the declaration is missing, incomplete or rejected — an interactive wizard that asks until this repo has declared which Teams exist, how many workers each has, which model every seat runs on, and which workflows route over them, then writes and validates graph.json. Triggers: "setup loop", "set up the team graph", "ตั้งค่าลูป", "สร้างกราฟทีม", "graph.json", the graph page reporting the declaration could not be read, a fresh install that has never declared a loop.

Write or refresh HANDOFF.md at the project root — the state of play between sessions, composed for an AI agent to act on rather than for a person to skim. Use when the user invokes /handoff, says hand off, write the handoff, save state before compaction, or asks what the next session needs to know. ใช้เมื่อ Master สั่ง /handoff, ส่งงานต่อ, เขียน HANDOFF, หรือถามว่าเซสชันถัดไปต้องรู้อะไร. Runs through bmad-party-mode when installed, else party-mode, so a room argues about what belongs before anything is written. Do not use to write release notes, a changelog, a commit message, or documentation for humans.

Use as the mandatory setup whenever a review lane cannot start on this machine — when `acp_lanes` reports `setupRequired`, when a lane comes back `available: false`, or when `acp_lane_probe` refuses a lane with `executable_absent`, `executable_not_file` or `launcher_absent`. It reports what this machine is missing per lane, writes the per-machine override file, and re-checks that the lane became callable. Triggers: "lane setup", "ตั้งค่าเลน", "เลนใช้ไม่ได้", "executable_absent", a lane refused before it started, a fresh machine that has never run a review lane.

Subskill of party-mode for read-only advice, plan critique, architecture or tradeoff review, risk analysis, recommendations, and adversarial review without implementation. Use when the user explicitly invokes $party-advise, asks for Party Mode advice, asks 'คิดยังไง', 'ช่วยคิด', 'แนะนำ', 'review plan', 'advise', 'critique', 'tradeoff', or wants a recommendation before deciding whether to execute. Do not edit files; hand off to party-auto only when execution is requested.

Subskill of party-mode for executing moderately complex, high-risk, multi-file, or production-impacting work. Use when the user explicitly invokes $party-auto or asks to implement, fix, build, ship, execute, apply a plan, or do the work, and multi-agent planning, critique, file ownership, verification, and 3-model review reduce risk. Do not use for pure advice, analysis-only review, brainstorming, or trivial one-step edits; use party-advise instead.

Umbrella workflow for Party Mode with two subskills: party-auto for executing moderately complex/high-risk work with planning, multi-agent critique, grill-gate decision interview, file ownership, verification, and 3-model review; party-advise for read-only advice, plan critique, architecture/tradeoff review, risk analysis, or recommendations without implementation. Use when the user explicitly invokes $party-mode, $party-auto, $party-advise, asks for Party Mode, or asks to choose between executing and advising.

Act as PM over other agents — subagents or ACP lanes — from one delegation to a whole spec. Covers the brief contract, the STOP rule, what counts as evidence, and the autonomous run: tickets as a dependency graph, one research pass, parallel implementers each in its own worktree, merged as they land, reviewed against the spec, worktrees cleaned up. Use when dispatching work to subagents or ACP, writing a brief, judging a returned result, or implementing a whole spec with minimal supervision — 'implement the spec', 'work the tickets', 'ทำตาม spec', 'ลุยทั้ง spec'. ใช้เมื่อรับบทเป็น PM คุมงานเอเจนต์อื่น. Do not use for a single edit you can make yourself, and never to route work through the delivery loop — that is tmux-teams.

Audit or author requirements against a 12-point contract: 6 Grill Categories (Business/Functional, Validation, Exception, Security, Performance, Integration) for ANY requirement, plus 6 INVEST letters (Independent, Negotiable, Valuable, Estimable, Small, Testable) for story-shaped requirements. Use when writing stories/requirements in any planning workflow, reviewing one before dispatch or estimation, coaching an author, or when the user says audit/review this requirement or story, ตรวจ/ซัก requirement, does this story pass INVEST, or grill this story. Returns a per-dimension verdict with blockers and fixes — READY only when every applicable dimension passes.

วิเคราะห์ วางแผน หรือตัดสินใจอย่างเป็นระบบด้วย Sequential Thinking — คิดทีละขั้น ทบทวนได้ แตกสาขาได้ ก่อนให้คำตอบสุดท้าย ใช้ skill นี้เมื่อผู้ใช้ต้องการการคิดเชิงวิเคราะห์ที่มีโครงสร้าง เช่น วิเคราะห์, วางแผน, เปรียบเทียบทางเลือก, ชั่งน้ำหนักข้อดีข้อเสีย, ตัดสินใจเลือกระหว่างตัวเลือก, ออกแบบสถาปัตยกรรม, ประเมินความเสี่ยง, หรือ review โค้ด/ระบบ ทริกเกอร์เมื่อผู้ใช้สั่ง 'วิเคราะห์', 'วางแผน', 'เปรียบเทียบ', 'ช่วยคิด', 'ชั่งน้ำหนัก', 'ควรเลือกอะไร', 'พิจารณา', 'ประเมิน', 'analyze', 'plan', 'compare', 'evaluate', 'think through', 'break down', 'pros and cons', 'trade-offs', 'A หรือ B', หรือพิมพ์ '/sqthink' — รวมถึงคำถามเชิงตัดสินใจที่ต้องพิจารณาหลายมิติ แม้ไม่ได้ใช้คำเหล่านี้โดยตรง อย่าใช้ skill นี้กับงานง่ายๆ ที่ตอบได้ทันทีโดยไม่ต้องวิเคราะห์ เช่น parse JSON, แปลงหน่วย, หรือคำถาม factual ที่มีคำตอบเดียว

Test-adequacy evidence contract for agent-authored code: CRAP as risk/triage signal, mutation testing as evidence, policy decision separate from evidence state. Use when the user asks about CRAP score, mutation testing adequacy, test quality of changed code, or a test-quality policy/evaluation — not generic review.

Drive interactive CLI agents (codex, claude, claude-zai, opencode) inside tmux — from ONE live session to a whole PM-run team. Reliable prompt submission, working/idle/approval state detection, completion detection, output capture, and session resume. Use when delegating coding work to a live Codex or Claude TUI, babysitting a long run, handling an approval prompt, or acting as PM over several mixed agents. Codex-specific flags, markers and dialog behaviour are in references/codex-tmux.md; for a one-shot headless codex run prefer codex exec. Triggers: "สั่งงานผ่าน tmux", "ขับ codex", "ทีม codex/claude-zai", PM-via-tmux orchestration.