ClawSeat

Generic solo TUI architecture/spec skill. Use this skill whenever the user asks an agent to act as a warden, professional code/architecture advisor, root-cause investigator, AI-friendly spec author, product test pilot, prompt translator, or lightweight coordinator for another TUI, SDK, memory seat, or coding agent. Also use it for concise task relay, Cartooner/product testing, issue investigation, and implementation-ready briefs. Do not use it for canonical project memory, planner fan-out, queue ownership, or thick seat protocol execution.

L3 project-memory hub for patrol, KB maintenance, queue/state tracking, faithful dispatch of operator/warden briefs, and E2E verification. Use when handling operator requests, context queries, decisions, deliveries, queue-drained receipts, or review/latest integration. Do not use for code, review, product-intent rewriting, direct specialist work, or seat lifecycle.

Federated memory oracle: synthesizes facts from ~/.agents/memory KB and preserves operator/warden brief intent during project-memory handoffs. Use when remembered facts, project-history synthesis, cross-seat evidence, queue-state context, or closeout recall are needed. Do not use for implementation, direct specialist work, or guessed facts.

Workflow author and dispatch orchestrator for ClawSeat tasks that need routing across specialist seats. Use when memory provides a brief, when workflow.md must be authored, when assigning owners, fan-out, liveness checks, SWALLOW fallback, or delivery consumption is required. Also use when coordinating review and next-step notifications. Covers workflow decomposition, schema validation, dispatch receipts, and planner summaries. Do NOT use for code implementation, independent review verdicts, visual QA, or project memory authority.

Multi-seat harness: bootstrap roster, start seats, dispatch tasks, complete handoffs with durable ACKs, provision frontstage heartbeat, render CLI operator console for a project profile.

Installer playbook for bootstrapping ClawSeat projects from the repository install flow. Use when the user asks to install ClawSeat, initialize a project, understand install.sh flags, recover an install, or follow docs/INSTALL.md setup steps. Also use when validating installer prerequisites or CLI-first onboarding. Covers install command discovery, environment assumptions, and resume guidance. Do NOT use for post-install task dispatch, normal /cs re-entry, code implementation, or modifying machine profiles without the install procedure.

Memory-driven project-group designer for ClawSeat v3 multi-team projects. Use when the first project memory needs to analyze a repo, recommend a module/layer-based team topology, write reviewable config proposals, or design an autonomous quality-docs testing group. For applying approved roster changes after initial design, pair with clawseat-roster-admin.

Implementation seat for ClawSeat workflow steps assigned by planner. Use when planner dispatches a brief, when you receive a [DISPATCH:] notification, or when fixing bugs, adding features, writing tests, refactoring code, or editing scripts and docs. Also use when implementing schema changes, templates, installers, or configuration from an in-flight workflow. Covers artifact authoring, local validation, and DELIVERY.md handoff with Docs Consulted (`N/A — <reason>` unless external SDK/API/CLI docs used). Do NOT use for code review, visual QA, scheduled patrol sweeps, operator intake, or memory authority decisions.

Controlled roster-change gate for ClawSeat project-memory. Use when an operator asks memory to add one builder/reviewer/patrol seat to an existing team, or to add one new subteam to an existing v3 multi-team project. Covers proposal, operator approval, allowed execution primitives, workspace refresh, launch/rehearsal, and evidence recording. Do NOT use for unsolicited scaling, direct TOML edits, deleting seats, changing secrets, or bypassing operator approval.

Creative and visual-quality seat for ClawSeat user-facing artifacts, prompts, multimedia assets, and experience review. Use when planner assigns visual design, copy, image prompts, multimodal analysis, UI/UX/a11y review, or creative artifact production. Also use when an output needs taste judgment beyond code correctness. Covers asset creation, design critique, content polish, and artifacts/ delivery. Do NOT use for backend implementation, logic-only code review, patrol sweeps, seat lifecycle, or secrets handling without privacy review.

Scheduled drift-inspection seat for ClawSeat code, docs, configuration, and evidence reports. Use when a cron or planner asks for patrol, when checking stale contracts, missing artifacts, schema drift, or operational health over time. Also use when emitting [PATROL-NOTIFY] findings. Covers report-only scans, drift evidence, KB findings, and patrol notifications. Do NOT use for feature verification, code fixes, active dispatch-chain ownership, user intake, or replacing reviewer verdicts.

Independent verification seat for ClawSeat diffs, tests, demos, delivery evidence, and browser-based UI/QA testing. Also supports visual consistency review of layout, spacing, color, and component alignment. Use when planner requests a review, when a builder delivery needs validation, when regression risk must be checked, or when a canonical Verdict is required. Also use when confirming acceptance criteria without changing artifacts. Covers diff review, visual consistency review, targeted test execution, demo verification, and canonical verdict reporting. Replaces designer seat in engineering template. Do NOT use for writing implementation patches, planning workflow ownership, visual design creation, scheduled patrols, or user intake.

Workflow design and execution skill. DESIGN mode: collaborative dialogue to assemble a workflow from capability-catalog tools. EXECUTE mode: runs an existing workflow file step by step, dispatching seats and handling gates. Does not implement an execution engine — dispatch logic is the planner's responsibility.

External peer protocol for non-canonical tmux workers. Use when a task is handled by a peer outside the project seat roster and needs durable peer deliveries, heartbeat tracking, or MiniMax readiness checks. Do not use for canonical seat dispatch, complete_handoff.py, or seat lifecycle changes.

Planner 专属 skill。接收 koder 传来的结构化用户需求 brief，将用户的真实操作步骤 映射到系统原子能力（seats / tools / skills），设计出可执行的工作流规格（workflow spec）， 并将其沉淀为可复用模板。不做用户访谈（访谈由 koder 的 clawseat-intake 完成）， 不负责工具参数填充（参数由 designer 席位负责），只负责工作流「骨架设计」和「dispatch」。

Privacy gate for ClawSeat messages, commits, and artifacts that may expose secrets or sensitive operator data. Use when preparing broadcasts, Feishu messages, DELIVERY summaries, commits, logs, or any content containing tokens, API keys, credentials, private paths, or customer data. Also use when a seat suspects accidental leakage. Covers secret detection, redaction guidance, and escalation for unsafe content. Do NOT use for general code review, product decisions, or approving content that has not been inspected.

ClawSeat memory seat's intake clarification skill for tmux operator sessions. Use when an operator request is ambiguous, has multiple interpretations, spans multiple layers, or carries high cost or irreversible consequences. Ask one Socratic question per turn with 2-4 concrete options before executing. Also use when creative type, engineering scope, or required resources are unclear. Do NOT use for clear direct commands, quick diagnostics, or when the operator says "just do it".

Reporting helper for recording ClawSeat dispatch and completion events into durable project status surfaces. Use when logging a new dispatch, a specialist delivery, a consumed ACK, or a STATUS.md registry update. Also use when a chain needs consistent chat tail blocks or auditable progress history. Covers event normalization, status registry maintenance, and delivery traceability. Do NOT use for deciding task ownership, writing implementation code, reviewing diffs, or replacing planner/memory authority.

OpenClaw Koder bridge for translating ClawSeat decision payloads and routing Feishu replies through the approved privacy boundary. Use when sending operator-facing decision summaries, processing Feishu responses, or syncing OpenClaw agent messages with ClawSeat memory. Also use when a handoff needs Koder-compatible wording or reply parsing. Covers payload translation, message routing, and privacy-aware response handling. Do NOT use for local code execution, planner dispatch, generic chat, or bypassing the privacy gate.

Memory-only planner report mode: sender routing, AUTO reports, and goal-drift recall.

Decision escalation gate for ClawSeat situations blocked by automation or missing operator authority. Use when a workflow needs a user decision, when privacy or Feishu routing requires explicit approval, or when the correct next action has multiple viable choices. Also use when preparing three-option decision payloads for memory/operator review. Covers option framing, blocking-context capture, and safe escalation. Do NOT use for routine dispatch, code implementation, autonomous planning, or direct Feishu delivery without memory authority.

Product-level ClawSeat entrypoint for install discovery and runtime handoff guidance across OpenClaw, Feishu, Claude Code, and Codex contexts. Use when a user says ClawSeat, wants to start the product, asks how to enter the system, or needs the correct install versus re-entry path. Also use when routing first-contact product intent. Covers entrypoint selection and handoff to install or /cs. Do NOT use for specialist workflow execution, direct code changes, detailed installer flag authority, or project-specific memory decisions.

Local ClawSeat /cs re-entry helper for operators who already have valid install state. Use when the user invokes /cs, asks to resume an existing ClawSeat runtime, reopen a project, or reconnect to the v0.7 resume contract after setup. Also use when distinguishing re-entry from first install. Covers local shortcut semantics and docs/INSTALL.md resume guidance. Do NOT use for fresh bootstrapping, product-level intake, specialist dispatch, implementation work, or creating install state from scratch.

tmux session control for ClawSeat seats. Covers session naming, essential commands (send-keys, capture-pane), ClawSeat script wrappers (send-and-verify, check-engineer-status, wait-for-text, detect-prompt-state), common pitfalls (paste overflow, Enter swallowed, focus mismatch, TMUX env), and self-diagnosis. Loaded by koder and planner for seat communication.

Thin Claude Code bundle for loading ClawSeat through the core gstack harness skill plus a project adapter skill, while delegating runtime/session behavior to the tmux-cli harness adapter.

飞书即时通讯：收发消息和管理群聊。发送和回复消息、搜索聊天记录、管理群聊成员、上传下载图片和文件（支持大文件分片下载）、管理表情回复。当用户需要发消息、查看或搜索聊天记录、下载聊天中的文件、查看群成员时使用。