cue-skills (cue-omni-reader)

1. 安装并启动 DeepSeek Harness：npx @deepseek-ai/dsh web
2. 在终端执行上面的安装命令（CLI 会解析插件并核验来源）
3. 用 dsh plugins list 确认已安装，必要时重启 Harness 生效

插件以当前 dsh 进程的权限运行，安装时可能执行代码。请先通读仓库源码与许可证，确认无破坏性命令与越权访问；本站只做索引，不对第三方插件安全性作担保。

事实信息来自公开插件目录快照（2026-10-03），介绍文案由本站再加工。

以下为插件仓库 README 全文（原始内容，由公开目录抓取整理）。

# cue-skills

**[English](README.md) · [中文](README.zh-CN.md)**

Open-source agent skills published by [Cue](https://cuecue.cn) (sensedeal).

A **skill** is a portable instruction bundle that any AI agent ([Claude Code](https://docs.anthropic.com/en/docs/agents-and-tools/agent-skills), Codex CLI, Gemini CLI, …) can load to gain a new capability — without modifying the agent itself. This repo collects the skills Cue maintains for public use.

> **New here?** Read **[Why Cue](docs/landing.md)** · [中文](docs/landing.zh-CN.md) — source-cited research you can verify, and how to try it free.

## Install

```sh
npx skills add sensedeal/cue-skills                       # default: lists cue-buddy, cue-research, cue-omni-reader, cue-data-mcp
npx skills add sensedeal/cue-skills --skill cue-research   # add one skill
```

See [Using a skill](#using-a-skill) for manual install and per-skill notes. DSH users: jump to [DSH bundles](#dsh-users-install-the-cue-deep-research-bundles).

## DSH users: install the Cue Deep Research bundles

Running [DeepSeek Harness](https://github.com/deepseek-harness)? Add the three Cue bundles to a profile and get Cue as native `mcp__omni__*` (parse/read) and `mcp__cue_*` (financial/regulatory data) tools:

```sh
dsh plugin --profile web add @cueai/dsh-omni-reader            # Omni Reader parse/read tools
dsh plugin --profile web add @cueai/dsh-omni-reader-guard      # optional SSRF/consent guard
dsh plugin --profile web add @cueai/dsh-cue-data-mcp           # Cue public data MCP (15 domains / ~104 tools)
```

Restart dsh; the model then sees `mcp__omni__parse` / `…get_parse_status` / `…read_result` / `…read_outline` / `…save_result` / `…cancel_parse` / `…discard_result` and `mcp__cue___*` data tools. Set `CUE_API_KEY` (+ optional `OMNI_ALLOWED_ROOTS`) via `$DSH_HOME/.env`. Guide: [`dsh/`](./dsh) · [`usage.md`](./dsh/usage.md).

## DSH bundles

This repo also ships [DeepSeek Harness](https://github.com/deepseek-harness) **bundles** under [`dsh/`](./dsh) — thin composition packages that wire a Cue MCP server into a DSH profile so its tools surface natively (e.g. `mcp__omni__parse`). A skill is agent-loadable *instructions*; a bundle is a *composition package* (`package.json` + `cordis.patch.yml`) for the Harness. Install one with `dsh plugin --profile web add `; see [`dsh/README.md`](./dsh) and the [`cue-omni-reader` bundle](./dsh/cue-omni-reader).

## Skills in this repo

| Skill | Purpose | Status |
|---|---|---|
| [`cue-buddy/`](./cue-buddy) — **cue-buddy** | Lets business experts author, validate, test, tune, and pin-as-frequent [Cue](https://cuecue.cn) "buddy" research templates via natural conversation. No Python / API knowledge needed; the agent talks to the Cue production API on the user's behalf. | v0.3.6 |
| [`cue-research/`](./cue-research) — **cue-research** | Sibling to cue-buddy for *using* Cue from inside your AI agent: ask a question → skill matches ≤2 candidate buddies (or routes to free-form deep research via `/api/rewrite`) → you confirm credits → it runs in the background and retrieves the report via replay → on satisfaction you can distill a free-form run into a saved buddy via cue-buddy. Supports 仿写/mimic (imitate a reference URL or sample document's writing style). | v0.3.6 |
| [`cue-omni-reader/`](./cue-omni-reader) — **cue-omni-reader** | Helps external agents parse HTTP(S) URLs and authorized local documents, audio, or video through the official MCP tools, including safe Bridge bootstrap, recoverable operations, deliberate inline/artifact delivery, outline navigation, complete result reading, and cleanup. Thin instruction layer; no custom parser or protocol driver. | v0.5.0 |
| [`cue-data-mcp/`](./cue-data-mcp) — **cue-data-mcp** | Lets external agents query Cue's public MCP data services (regulatory, macro, disclosures, market data, legal full text, holdings, entity data, academic, IPO, ESOP, buyback, footnote details) by discovering live domains from the anonymous `/api/mcp-catalog` and connecting through each domain's routing DTO. Thin instruction layer; no hardcoded endpoints or tool names. | v0.3.6 |
| [`playbook/`](./playbook) — **playbook scene skills** | One agent-loadable `SKILL.md` per [Cue playbook](https://cuecue.cn/playbook) scene (equity research, credit diligence, wealth advisory, global macro, …). Load one and your agent runs that scenario's Cue deep research: it queries the **live** `/api/playbook` for the scene's *current* buddies, picks one, confirms credits, runs, and returns a source-cited report. Runtime-query-live — no baked `template_id`, so buddy add/edit/remove auto-reflects with no regeneration. Auto-generated by [`scripts/gen_scene_skills.py`](./scripts/gen_scene_skills.py). Works two ways: bundled here next to `cue-research`/`cue-buddy`, **or** published standalone to a third-party store — a standalone scene skill **self-bootstraps** (clones this repo, with a Gitee mirror fallback, to fetch the runner if not already present). | auto-generated |

More skills will be added here as Cue's surface grows.

## What is Cue / What is a buddy

[Cue](https://cuecue.cn) is a **Deep Research Agent + Intelligence Sentinel** platform for high-precision finance and business workflows. It picks tools from 300+ professional data sources (A-share / HK / US equity disclosures, fund AMAC registries, business registries, court records, regulatory feeds, sell-side research, capital flow data), cross-validates findings across sources, and produces structured, source-cited reports in minutes instead of hours.

A **"buddy" (搭子)** is a research playbook for a specific scenario — corporate-credit pre-diligence, public-record compliance snapshot, quarterly earnings review, private-fund manager DD, etc. — defined once and reused by supplying the subject. The `cue-buddy` skill in this repo is what business experts use to author these playbooks conversationally.

> **Scope boundary**: Cue's tool surface covers public data sources only (equity disclosures / business registries / court records / regulatory filings / capital flows / etc.). Scenarios requiring private data (real AML on bank-internal transactions, medical diagnosis, internal accounting) are **not appropriate** as Cue buddies — the supervisor cannot route them and falls back to generic web search. `cue-buddy +author` flow calls `+capabilities` to cross-check each declared evidence source against the actual catalog before persisting a template.

See [`cue-buddy/README.md`](./cue-buddy/README.md) for the full skill walkthrough.

## Using a skill

`cue-buddy` is self-contained: an entrypoint `SKILL.md`, supporting `references/`, and stdlib-only `scripts/`. `cue-research` is **self-contained too** — it vendors the shared Cue client (`cue_api` / `sse_report` / `paths`) in its own `scripts/`, so it runs standalone (no sibling `cue-buddy` required).

Preferred install is at the top of this README (`npx skills add`). The `skills` CLI lists every skill; `--skill` picks one.

**Manual install**: copy the skill folder into `~/.claude/skills/` (Claude Code) or reference it via `/use-skill `; other agents load `/SKILL.md` as a system instruction. Each skill is self-contained — `cue-buddy`/`cue-research` vendor their runtime, `cue-omni-reader`/`cue-data-mcp` are thin instruction layers.

**`cue-omni-reader`** ships no runtime parser scripts. Load its `SKILL.md` and connect an official Omni MCP surface. Existing URL tools need no local install; local files use the audited Bridge after minimum-root confirmation.

**`cue-data-mcp`** ships no runtime code. Load its `SKILL.md` and let the agent discover the live data domains from the anonymous `/api/mcp-catalog`, then connect through each domain's routing DTO. Self-contained: no sibling folders required.

The **`playbook/`** scene skills delegate the actual research run to `cue-research`'s runner. `cue-research` is self-contained (it vendors the shared Cue client), so install `playbook/` + `cue-research` as siblings under the same parent dir; `cue-buddy` is only needed if you also want to author or save buddies.

> **Credit grant to start**: every Cue account gets **500 credits on signup, plus 10 more each day**. Once you apply for an API key (sign in via the `cue` CLI) you can start — data queries and initial research draw from this grant, and each deep-research run still consumes credits, so the skills always confirm with you before spending.

## Contributing

Bug reports and skill suggestions: [open an issue](https://github.com/sensedeal/cue-skills/issues).

Pull requests welcome. Each skill has its own hard-rules / validator; see the skill's README for contribution guidelines.

## License

MIT — see [LICENSE](./LICENSE). Skill content is free to use, modify, and redistribute.