libr-agent

Analyze installed tools and existing assistant configurations to actively propose and create specialized LibrAgent assistants via agent__create. Use when the user wants to recruit, create, or build a domain expert or specialist agent, utilizing a proactive "Architect" workflow that maps available inventory tools to functional gaps. Not for workspace agents.md (agent-init), teamwork/org, or registering MCP servers.

Parse .trace.json files and correlate with backend events and tool calls. Use when debugging agent behavior, investigating tool call patterns, or validating concurrency limits.

Monitor and enforce the frontend bundle size budget. Use when checking bundle size, preventing regressions, or validating build output against budget constraints.

Audit React Context usage, detect prop drilling, verify compound component patterns. Use when reviewing frontend code for proper Context consumption, checking AgentChatContext/AgentSessionContext usage, or enforcing compound component patterns.

Audit builtin MCP tool implementations against the Tool Design Manifesto. Use when reviewing builtin tools for compliance, checking context_prompt usage, validating text-first MCPResult design, or auditing canonical naming and session isolation correctness.

Run and verify SeaORM database migrations against in-memory SQLite. Use when applying migrations, verifying schema changes, or testing migration correctness.

Scaffold integration tests in src-tauri/tests/ following the project's strict rules. Use when creating new Rust tests, ensuring CI compatibility, or preventing false sense of test coverage from unit tests in src/.

Validate that builtin servers and external MCP managers maintain proper session isolation. Use when auditing session isolation, checking for cross-session state leakage, or verifying per-session MCPServiceProxy instantiation.

Manage Windows-specific process isolation, stdio pipe decoding, and ANSI code page handling. Use when debugging Windows-specific MCP server issues, fixing stdio encoding, or handling Windows path discovery.

Guide for creating builtin MCP tools in LibrAgent. Use when creating a new builtin MCP server, adding tools to existing builtin servers, refactoring tool implementations to follow best practices, or auditing tool compliance with the Tool Design Manifesto v2.1.

Audit and critique builtin MCP tool implementations in LibrAgent. Use when auditing existing builtin MCP server implementations for compliance, reviewing pull requests that add or modify builtin tools, validating tool implementations against the Tool Design Manifesto v2.1, or identifying potential issues before they reach production.

Audit builtin MCP tool implementations in LibrAgent for schema accuracy, minimal complexity, and non-bloated hints. Use when reviewing builtin tools for compliance, validating tool schemas, checking for over-engineering, or ensuring next-action hints are concise and relevant.

Use when the user needs to diagnose and install runtime environment dependencies (Python, Node.js, uv) for running LibrAgent or MCP servers. Use on new machines, MCP "command not found" errors, or missing Python/Node runtime failures. LibrAgent 실행에 필요한 Python, Node.js 등의 환경을 자동으로 진단하고 설치하도록 안내하는 마법사입니다.

Guide for creating effective skills. This skill should be used when users want to create a new skill (or update an existing skill) that extends Claude's capabilities with specialized knowledge, workflows, or tool integrations.

Train custom LLM models locally using conversations exported from LibrAgent. Use when the user requests to fine-tune a model, train on past chats, export chat datasets, or configure custom agent models. Triggers on: "fine-tune model", "모델 학습시켜줘", "내 대화 데이터로 학습해줘", "export dataset", "대화 데이터셋 추출".

Implement a Creator-Reviewer quality gate loop. The Creator generates an artifact, and the Gatekeeper/Reviewer verifies it against quality criteria. If approved, it passes; if rejected, the Creator must perform rework and resubmit. Useful for strict code reviews, verification checks, or quality assurance. Triggers: "게이트키퍼", "리뷰 루프", "검증 게이트", "품질 검증", "gatekeeper", "reviewer loop", "quality gate".

Implement a central Coordinator (Hub) that manages, distributes, and merges tasks performed by specialized worker sessions (Spokes). Spokes communicate only with the Hub and not with each other. Useful for ad-hoc project management and complex workflow distribution. Triggers: "허브 스포크", "중앙 통제", "작업 배분", "hub and spoke", "coordinator pattern", "centralized workflow".

Analyzes conversation sessions to extract key insights, decisions, and technical details, then saves them to the knowledge base. Use this skill when asked to "summarize what I learned today," "record key points from recent sessions," or "sync current project context to long-term memory."

Restructure a living explicit org in LibrAgent without dissolving it. Use when an org already exists and you need to add/remove/merge roles in ROLES.md, update the org constitution (agents.md, MISSION.md), remove or retire org child sessions, reassign KANBAN owners, or record structural changes in coordination files. Not for initial org creation (use org), initial scaffolding (use teamwork), org dissolution, or archive.

Run explicit org-based teamwork in LibrAgent. Use when collaboration needs durable org identity, org-visible child sessions, org-root resume behavior, or clear parent/sibling org context while preserving normal parent workspace inheritance.

Implement a turn-based Driver-Navigator pair programming workflow. One agent (Driver) focuses on implementation and writes code, while the other (Navigator) reviews, validates, and suggests improvements in real time. Roles can swap dynamically based on tasks. Triggers: "페어 프로그래밍", "공동 작업", "드라이버 네비게이터", "pair programming", "driver navigator", "collaborative coding".

Execute sequential, stage-based assembly line tasks where the output of one stage becomes the input of the next stage. Useful for progressive refinement workflows like Research -> Draft -> Review -> Format. Not for parallel execution (use divide-conquer) or consensus (use consensus-delegation). Triggers: "파이프라인", "순차 처리", "단계별 처리", "pipeline", "sequential processing".

Use when the user needs to diagnose and install runtime environment dependencies (Python, Node.js, uv) for running LibrAgent or MCP servers. Use on new machines, MCP "command not found" errors, or missing Python/Node runtime failures. LibrAgent 실행에 필요한 Python, Node.js 등의 환경을 자동으로 진단하고 설치하도록 안내하는 마법사입니다.

Guide for creating and validating agent skills. Use when authoring or updating a SKILL.md, writing frontmatter, structuring scripts/references/assets, or checking a skill before deployment. Triggers on create skill, write skill, skill frontmatter, validate skill, fix skill YAML. For installing a finished skill, use skill-deployer instead.

Build and scaffold a multi-agent collaboration workspace with the right coordination model, shared operating files, and role-specific skills. Use when a user wants to create a team, crew, task force, asynchronous collaboration loop, specialist handoff workflow, or reusable shared workspace for multiple agents, then route execution to the right teamwork skill.

Extract and analyze LibrAgent debug logs with pattern matching and context. Use when debugging LibrAgent issues, analyzing agent workflows, investigating tool execution problems, or extracting specific log patterns like errors, warnings, planning operations, MCP tool calls, or browser automation traces. Supports extracting last N lines or searching for patterns with surrounding context.

Code quality validation for LibrAgent project. Use when performing code quality checks, refactoring validation, or ensuring code meets project standards for TypeScript/React frontend and Rust backend. Runs comprehensive validation including linting, formatting, type checking, build verification, and dead code detection.

Python CLI wrapper for X (Twitter) interaction using Twikit. Use when the user wants to: (1) post tweets (with text or media, including thread replies), (2) view home timeline, (3) view a user's tweets, (4) search tweets, (5) like/favorite a tweet, (6) retweet a tweet, or (7) delete a tweet. On first use, guide the user through credential setup (username, email, password, optional TOTP) or browser cookie setup. Store config in ~/.libragent/x_config.json and session cookies in ~/.libragent/x_cookies.json. Subsequent requests use the stored session without re-authentication. Triggers on requests like: "트윗 올려줘", "트위터 피드 보여줘", "post a tweet", "like tweet", "retweet", "delete tweet".

Python CLI wrapper for X (Twitter) interaction using Twikit. Use when the user wants to: (1) post tweets (with text or media, including thread replies), (2) view home timeline, (3) view a user's tweets, (4) search tweets, (5) like/favorite a tweet, (6) retweet a tweet, or (7) delete a tweet. On first use, guide the user through credential setup (username, email, password, optional TOTP) or browser cookie setup. Store config in ~/.libragent/x_config.json and session cookies in ~/.libragent/x_cookies.json. Subsequent requests use the stored session without re-authentication. Triggers on requests like: "트윗 올려줘", "트위터 피드 보여줘", "post a tweet", "like tweet", "retweet", "delete tweet".

A sample helper skill for Libr Assistant.

Python CLI wrapper for Telegram interaction using Telethon (MTProto protocol). Use when the user wants to: (1) send messages, (2) read messages from chats/channels/groups, (3) list chats, (4) download files, or (5) search messages. On first use, guide the user through authentication (API ID/Hash → phone → code → 2FA password). Store session and config in ~/.libragent/telegram_config.json. Subsequent requests use the stored session without re-authentication. Triggers on requests like: "텔레그램 메시지 보내줘", "텔레그램 확인", "send telegram", "텔레그램 채널 확인", "텔레그램 파일 다운로드".

Split a large task into independent subtasks, delegate them to child sessions for parallel execution, and merge the results. Use when a task can be broken into non-overlapping pieces that don't depend on each other's output. Not for consensus/review (use consensus-delegation) or benchmarks (use bench). Triggers: "분할 정복", "병렬 처리", "parallel execution", "split and process".

Resumable spec-driven guide for the Team Workspace / org lineage / scheduled task group refactoring in LibrAgent. Use when continuing, implementing, or reviewing this teamwork architecture so changes stay aligned with the agreed contract: workspace-scaffolded SSOT, teamwork as a meta-meta-skill, org as explicit org-created lineage teamwork, scheduled collaboration as separate task groups, and Settings-backed backend-enforced governance.

Run benchmark-style evaluations on AI agents or tools using parent-child session delegation. Use when the user wants to test an agent's performance on a set of problems (e.g. SWE-bench style, tool capability benchmark, coding challenge evaluation). The parent session orchestrates: spawns child sessions to solve individual problems, collects answers via checkSession, and generates a consolidated report with pass/fail scores and analysis. Triggers on: "bench 테스트 해줘", "SWE-bench 돌려줘", "agent benchmark", "tool evaluation".

Delegate work between LibrAgent AI agent sessions using sub-agent sessions. Use when an agent needs to spawn, brief, monitor, or troubleshoot a child session with `agent__startSession`, `agent__checkSession`, or `agent__messageToSession`, especially when deciding whether the child really needs parent workspace state, workspace instructions, or workspace-scoped skills.

Guide for designing and creating high-quality, reusable workflows using the playbook tool group. Use this skill when: (1) You want to capture a successful multi-step process for future reuse, (2) A user asks to 'create a playbook' or 'automate this task', (3) You need to design a structured automation workflow with clear inputs, outputs, and success criteria. Triggers: 'create playbook', 'automate this process', 'design workflow'.

Create and manage global scheduled tasks in LibrAgent (cron-based). Use when automation should outlive the current session, wake a specific assistant on a cron, or run recurring background work app-wide. For one-shot delays or session-bound recurrence inside the active conversation, use session-schedule instead.

Schedule one-shot delays or recurring runs bound to the current agent session in LibrAgent. Use delaySeconds for a single future injection, cronExpression for session-scoped recurrence, or when the user asks to continue, remind, or follow up later in this conversation. Does not require teamwork or task groups. For app-wide automation that outlives the session, use schedule instead.

Guide for creating high-quality tool servers (MCP servers) that enable LLMs to interact with external services through well-designed tools. Use when building tool servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).

Guide for creating high-quality tool servers (MCP servers) that enable LLMs to interact with external services through well-designed tools. Use when building tool servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).