← 返回 Skills 市场

Meeting Summary

Name: Meeting Summary
Author: jack-yang-ai

作者 Jack-Yang-ai · GitHub ↗ · v2.3.0 · MIT-0

cross-platform ⚠ suspicious

292

总下载

当前安装

版本数

在 OpenClaw 中安装

/install meeting-summary

功能描述

会议录音 → 结构化纪要。支持 Step ASR 转写、pyannote 说话人分割、声纹实名匹配、LLM 纪要生成，以及长录音 chunk 化增量处理。

安全使用建议

What to check before installing/using this skill: - Verify external endpoints: the scripts call StepFun ASR/LLM endpoints (api.stepfun.com) and the setup guide pulls models from Hugging Face / GitHub. Decide whether you trust these services to receive your audio and derived data. - Confirm the model routing: SKILL.md instructs spawned subagents to use anthropic/claude-opus-4-6, but the code uses StepFun models. Ask the author which external LLM will actually receive prompts/audio; mismatches can cause unexpected data flows. - Protect API keys: the skill expects a StepFun API key (~/.stepfun_api_key or STEPFUN_API_KEY) and may read ~/.huggingface_token. Only provide keys you control and store them with tight permissions. Consider using ephemeral/test keys first. - Be aware of biometric data storage: enrolled voiceprints are persisted under ~/.openclaw/workspace/memory/voiceprints. If you will enroll users' voices, obtain consent and secure/remove these files appropriately. - Run in a sandbox first: execute the scripts in an isolated environment (throwaway VM or container) to confirm network calls and file writes match your expectations. - Inspect code paths for exfiltration: review transcribe.py and other scripts to verify which data is sent in requests (audio files, base64 payloads, any extra metadata like open_id). If you cannot audit, avoid providing real sensitive recordings. - If you proceed, prefer local-only operation: do not enroll voiceprints with real identities until you trust the storage and service endpoints; or remove STEP* keys to force offline-only fallback modes. If you want, I can: (a) list the exact network calls and files the scripts will access, (b) point to the lines where open_id or other identifiers would be passed to spawned agents, or (c) draft a safe minimal-run checklist to exercise the skill in a sandbox.

功能分析

Type: OpenClaw Skill Name: meeting-summary Version: 2.3.0 The meeting-summary skill is a well-structured audio processing pipeline that transcribes and summarizes recordings using StepFun APIs and local speaker diarization (pyannote/wespeaker). The scripts (meeting-summarize.py, transcribe.py, voiceprint-manager.py) perform legitimate orchestration tasks, including audio preprocessing with ffmpeg, API interaction with api.stepfun.com, and local caching of results in a dedicated workspace. While it accesses sensitive API keys (~/.stepfun_api_key) and Hugging Face tokens, this behavior is explicitly documented and necessary for the stated functionality. The SKILL.md instructions are focused on task execution and include safety checks to prevent unauthorized environment modifications.

能力评估

⚠ Purpose & Capability

The skill claims to perform ASR, diarization, voiceprint matching and LLM-based summarization — and the included scripts implement those flows (transcribe.py, pyannote wrapper, voiceprint manager). However the package metadata declares no required env vars/credentials while the SKILL.md and scripts require a StepFun API key and optionally a Hugging Face token / pyannote venv. Also SKILL.md instructs spawned subagents to use anthropic/claude-opus-4-6, but the scripts call StepFun ASR/LLM endpoints (step-*/api.stepfun.com). This mismatch between declared requirements, runtime instructions, and code is incoherent and unexpected.

⚠ Instruction Scope

SKILL.md explicitly instructs the agent to spawn a subagent and pass audio path and user open_id; it also requires reading local files (~/.stepfun_api_key, ~/.huggingface_token) and uses those values in network calls. The scripts perform outbound network calls to StepFun endpoints and optionally load models from Hugging Face. Passing user open_id and audio to a spawned subagent (and forcing a particular external model in the instruction) could cause user-identifying data and audio to be sent to third-party services not declared in the registry metadata. The instruction set forbids installing dependencies or modifying code, which prevents the agent from fixing environment mismatches automatically — again unusual.

ℹ Install Mechanism

There is no install spec (instruction-only at registry level) but the repo includes runnable scripts and a setup guide that instructs downloading models (ONNX from GitHub releases) and using Hugging Face to fetch pyannote models. Downloading models from GitHub/HuggingFace is expected for this purpose; no opaque shorteners or unknown personal servers are used in the provided guide. Overall install mechanism risk is moderate but conventional for ML tooling.

⚠ Credentials

The skill requires sensitive credentials (STEPFUN_API_KEY stored in ~/.stepfun_api_key or STEPFUN_API_KEY env var; optional HF_TOKEN and pyannote venv) and stores voiceprints locally at ~/.openclaw/workspace/memory/voiceprints. The registry metadata claims 'Required env vars: none' which is incorrect. Requesting and using an ASR/LLM API key is proportional for the described functionality, but the undeclared nature and the forced spawn-model mismatch (Anthropic vs StepFun) raise the risk that audio or identifiers could be routed to an unexpected third party. Also the skill persists biometric identifiers (voiceprints) locally—this is sensitive and should be explicitly acknowledged before install.

ℹ Persistence & Privilege

always:false (good). The skill writes and reads files under ~/.openclaw/workspace (cache, models, memory/voiceprints) — persistent local storage of outputs and enrolled voiceprints is intended by the skill. It does not request system-wide privileges or modify other skills. Still, the persistent storage of voiceprints and caches means sensitive biometric and audio-derived data will remain on disk unless the user removes it.

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install meeting-summary
安装完成后，直接呼叫该 Skill 的名称或使用 /meeting-summary 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v2.3.0

v2.3.0: 稳定版发布，SKILL.md 防呆化，强制说话人确认，环境修复禁令

v2.2.0

**Summary: Major update focused on execution workflow, safety, and output quality for meeting audio summarization.** - Switched routing to always use the Claude Opus-4-6 model as subagent, main session now spawns subagent on audio input. - Enforced strict step-by-step execution: environment check → main script execution → output inspection → user speaker mapping confirmation → summary delivery. - Speaker confirmation is now mandatory for all ≥2-speaker meetings, even with voiceprint matches—users must confirm identities before finalizing summary. - Enhanced summary output structure: delivers only structured content (not raw script markdown), always includes participants, overview, key points, decisions, actions, risks. - Improved environment and error handling: no installation actions allowed, clear stop/notify logic on unmet dependencies or errors, explicit downgrade path when pyannote is unavailable. - Updated and detailed cache usage and management rules to prevent data loss and ensure efficiency.

v2.1.1

补充 StepFun API 申请链接、Hugging Face 本地模型下载说明、FAQ 新增模型下载问题

v2.1.0

统一中文版文档，补充模型信息(step-asr/step-3.5-flash)、HF下载说明、FAQ、效果示例、适合谁/不适合谁

v2.0.0

Full pipeline: Step ASR + pyannote diarization + chunked processing + cache + voiceprint matching + LLM summary

元数据

Slug meeting-summary

版本 2.3.0

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 5

常见问题

Meeting Summary 是什么？

会议录音 → 结构化纪要。支持 Step ASR 转写、pyannote 说话人分割、声纹实名匹配、LLM 纪要生成，以及长录音 chunk 化增量处理。它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 292 次。

如何安装 Meeting Summary？

在 OpenClaw 或 Claude Code 对话框中运行命令「/install meeting-summary」即可一键安装，无需额外配置。

Meeting Summary 是免费的吗？

是的，Meeting Summary 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。

Meeting Summary 支持哪些平台？

Meeting Summary 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。

谁开发了 Meeting Summary？

由 Jack-Yang-ai（@jack-yang-ai）开发并维护，当前版本 v2.3.0。