EastMoney Roadshow Digest (Transcript + Summary)｜东方财富路演纪要生成

EastMoney Roadshow Digest (Transcript + Summary)｜东方财富路演纪要生成. A reliability-first transcript & summary skill for public EastMoney roadshow replays, focused on...

Security Scan

The package is internally consistent with its stated purpose (fetch public EastMoney replay pages, extract subtitles/media, optionally run ASR, and optionally send transcript text to configured LLMs); nothing in the code attempts to access unrelated credentials or system paths, but it will send transcript text to external model providers if an API key is present so you should be deliberate about which keys are available in the host environment.

Purpose & Capability

Name/description match the code: providers/eastmoney.py validates and fetches public page/API data; providers/asr.py extracts and transcribes audio; main.py orchestrates, cleans text, writes outputs, and (optionally) calls LLMs. No unexpected cloud credentials or unrelated system services are requested.

Instruction Scope

SKILL.md explicitly limits scope to public roadshow URLs and forbids use of cookies/private credentials. The runtime honors that: requests target the public EastMoney detail API and media URLs. The only notable behavior is intentional and declared: when an LLM API key is present in the environment, cleaned transcript text (or intermediate text) will be POSTed to the configured LLM endpoint for enhancement. If no key is present the code falls back to rule-based processing and records downgrades in run_report.md.

Install Mechanism

No installer/remote download steps are present in the manifest. Python dependencies are standard (requests, faster-whisper, av) and ffmpeg is required at runtime; dependencies are declared in requirements.txt. No downloads from arbitrary URLs or extract/install steps are present.

Credentials

The package does not require environment secrets to operate. It reads optional LLM keys from environment to enable model enhancement (get_llm_config checks OPENAI_API_KEY, OPENROUTER_API_KEY, and MOONSHOT_API_KEY/KIMI). SKILL.md/README list a broader detection order (Anthropic, Google, xAI, etc.) but the code only implements a subset—this is an inconsistency (informational/sloppy) but not an access escalation. No unrelated secrets (AWS, GitHub tokens, etc.) are requested.

Persistence & Privilege

The skill does not request always:true and does not persistently modify other skills or system-wide settings. It writes its own outputs under outputs/ and inserts a local .vendor directory onto sys.path (typical for vendored deps); the entry guard requires a specific env marker but that is a contained enforcement, not privilege escalation.

Assessment

This package appears to do what it says: parse public EastMoney roadshow pages, prefer subtitles, fall back to media+ASR, clean text, and optionally call an external LLM if an API key is present. Before installing or running it: - Be aware that any transcript/cleaned text will be sent to an external model provider if you have OPENAI_API_KEY, OPENROUTER_API_KEY, or MOONSHOT_API_KEY/KIMI set in the environment. If you do not want transcript data leaving the host, do not set those keys. - SKILL.md / README claim detection for more LLM providers (Anthropic, Google, xAI) but the code only checks OpenAI, OpenRouter, and Moonshot/Kimi—expect this mismatch. If you rely on another provider, verify and/or modify get_llm_config before use. - The skill requires ffmpeg on PATH and faster-whisper + av to run ASR locally; running ASR may be CPU- or disk-intensive. - The package inserts a local .vendor path into sys.path; inspect that directory (if present) before running to ensure no unexpected vendored modules. - Follow the author's enforced entry: set EASTMONEY_ROADSHOW_ENTRY=python3-main when running main.py so the launcher does not exit. - For higher assurance, run the skill in an isolated environment (container/VM) and review the outputs/run_report.md after a run to confirm no unexpected network activity. If you want full parity with SKILL.md's claimed provider list, ask the author to align code and docs.

Purpose | 用途

Create a clean, reliability-first workflow for public EastMoney roadshow replay pages only.

仅处理公开可访问的东方财富路演回放页，不依赖本地 cookie、私有接口凭据或用户路径，强调执行可控与结果留痕。 本发布包不包含本地 vendor、缓存、输出物或调试残留，目标是可在新环境中按统一入口运行。

Input | 输入

• url: string
• Must match https://roadshow.eastmoney.com/luyan/<id>

Output | 输出

Write these files under outputs/:

• meta.json
• transcript.md
• clean_transcript.md
• summary.md
• brief.md
• run_report.md

Page parsing succeeds → always write meta.json. Any downstream failure or downgrade → still write run_report.md.

Workflow | 工作流

1. Validate URL
2. Parse page and fetch public metadata
3. Prefer subtitle candidates if available
4. Discover replay media and run ASR fallback if needed
5. Clean transcript
6. Generate summary
7. Generate executive brief
8. Always write run report

Role boundaries | 模块边界

• providers/eastmoney.py: validate URL, parse page, fetch public metadata, discover subtitle/media candidates only
• providers/asr.py: convert audio to text only
• main.py: orchestration, file writing, cleaning, summary generation, reporting

Reliability notes | 可靠性说明

• If subtitle is absent, attempt media+ASR fallback
• If ASR dependencies are missing, degrade gracefully and explain in run_report.md
• If page parsing succeeds, still write meta.json even when downstream steps fail
• LLM capability is optional environment capability, not a required dependency of the skill package
• When LLM capability exists in the runtime, use it automatically for cleaner transcript / summary / brief generation
• When environment-provided LLM capability is available, transcript text or intermediate cleaned text may be sent to an external model service for processing
• When LLM capability does not exist, fall back automatically to rule-based outputs and explain the path in run_report.md
• Do not bind the skill description to any specific LLM vendor or provider
• Keep outputs deterministic and conservative; do not invent facts
• Keep README.md, SKILL.md, and manifest.json aligned on scope, inputs, outputs, and dependency expectations

Practical guidance | 实操说明

• Use replay metadata from the public detail API
• Preserve source path in meta.json
• Prefer concise summaries grounded in extracted text
• If transcript quality is low, say so explicitly in run_report.md
• 唯一正确执行方式：EASTMONEY_ROADSHOW_ENTRY=python3-main python3 main.py --url <roadshow_url>
• 禁止使用 uv run、其他包装器、替代 Python 启动路径或任何未在 README / SKILL 明示的入口
• 调试、验证、重跑必须与正式执行使用同一标准入口；偏离入口的运行结果不得作为 skill 故障判断依据
• LLM 增强按环境中可用的 provider key 自动启用；若无 key，则自动降级到规则版输出

Entry enforcement | 入口约束

main.py contains a startup self-check. If the execution entry does not match the documented standard path, it must fail fast and exit with a clear error instead of continuing in a potentially inconsistent environment.

Dependencies | 依赖

• Python 3.9+
• requests required
• faster-whisper required for ASR
• av required as a runtime dependency for faster-whisper
• ffmpeg required on PATH for audio extraction / ASR fallback
• LLM capability is optional if the host runtime provides it; the skill remains usable without it

Model + keys | 模型与密钥

按环境变量自动检测可用 provider，优先顺序如下：

• OpenAI: OPENAI_API_KEY
• Anthropic: ANTHROPIC_API_KEY
• Google: GEMINI_API_KEY（别名：GOOGLE_GENERATIVE_AI_API_KEY, GOOGLE_API_KEY）
• xAI: XAI_API_KEY
• OpenRouter: OPENROUTER_API_KEY
• Moonshot/Kimi: MOONSHOT_API_KEY（别名：KIMI_API_KEY）

若检测到可用 key，clean transcript / summary / brief 可启用外部模型增强；若未检测到，则自动降级为规则版输出。

• 启用增强时，transcript 或中间清洗文本可能发送至外部模型服务处理。

Description | 描述

EastMoney Roadshow Digest (Transcript + Summary) is a reliability-first skill for turning public 东方财富路演回放页面 into structured transcript, clean transcript, summary, brief, and run report outputs.

本技能定位为：面向公开东方财富路演回放页的、以稳定性优先的纪要生成工具，可输出 transcript、clean transcript、summary、brief 与 run report。

Validation standard | 验证标准

• 调试、验证、重跑、正式执行全部使用同一入口：EASTMONEY_ROADSHOW_ENTRY=python3-main python3 main.py --url <roadshow_url>
• 对发布包的任何验证结论，必须基于该标准入口；偏离入口的结果无效

Non-goals | 非目标

• Do not handle private/live-only rooms
• Do not log in
• Do not require browser cookies
• Do not publish automatically