dsh-aigc-radar

English | 简体中文

AIGC Radar project search for DeepSeek Harness (dsh).

Stop rebuilding what already exists. While you plan and implement, the agent proactively checks the curated AIGC Radar library for mature, battle-tested projects that already solve your problem — before you write a line of code. Results render as native search cards in the dsh Web UI — not raw markdown — and survive session replay.

What you get

Tool	What it does
search_ai_projects	Searches the curated AIGC Radar library: GitHub projects above a 500-star floor, enriched with categories, bilingual (zh/en) tags and descriptions, and daily star-growth metrics
get_project_categories	Lists the category taxonomy (categories + subcategory counts) for filter discovery

Two routing layers make discovery automatic rather than opt-in:

• Explicit discovery — ask "找个能做 deep research 的开源框架" and the agent comes back with starred, categorized results, no "use the tool" needed
• Proactive reuse check — before the agent implements a major module or subsystem (auth, payments, workflow engines, search/indexing, protocol implementations, end-to-end RAG/Agent pipelines…), it runs one library check on its own initiative, so mature alternatives surface before anyone rebuilds them. Narrow work (bug fixes, renames, styling, CRUD) is deliberately excluded

Why a native plugin instead of the MCP server?

AIGC Radar also ships as an MCP server — and this plugin now rides that same MCP endpoint (POST /api/mcp), so every call counts against the same rate limits and quotas. What the plugin adds on top:

• Native web search cards — structured sources render as cards in the Web UI and are rebuilt faithfully on session replay (presentationMeta), which the MCP transport cannot express
• Typed canonical output — the result is one validated JSON value, so Code Mode can compose it programmatically (await tools.search_ai_projects({ q: 'mcp' })) with full type inference
• First-party prompt routing — the discovery-routing guidance lives in system-prompt assembly, not in MCP instructions that clients may truncate

Measured performance

The search tool is a single HTTPS call to the AIGC Radar public edge — the numbers below measure that full path, taken 2026-08-18 from a China home-broadband connection (GeoDNS → CN edge) with scripts/benchmark-search.sh (10 representative zh/en queries × 3 trials against https://aigcnews.cn/api/mcp):

Metric	Value
Search latency p50	355 ms
Search latency p95	810 ms
Curated projects served	18,426 — every one above the 500-star admission bar
Taxonomy	11 top-level categories, bilingual zh/en tags and descriptions

Index-quality benchmarks measured on the MCP interface — 3.5× faster than WebSearch end-to-end (7.4 s vs 25.8 s median) and 98.3% first-tool routing accuracy (59/60 trials) — live in the main repo's benchmark section; this plugin serves the same dataset over the same API, so those numbers carry over.

Install

Requires dsh (npx @deepseek-ai/dsh web).

Recommended — install the prebuilt package from npm:

…