插件目录 / Developer / dsh-gemini-multimodal
dsh-gemini-multimodal
已验证 · 实测可装 RealAlexandreAI
功能简介
Developer 类别的 DeepSeek Harness 插件。
可用 — 实测通过,早期项目
Developer 类别的 DeepSeek Harness 插件。 实测能干净安装、正常启动。早期项目,但功能可用。
「已验证」表示我们的自动化 CI 在干净 profile 里实际执行了 dsh plugin add 并启动成功——仅此而已。功能描述与版本兼容性均为作者声明。这不是安全审计,也不代表对第三方代码的背书。
README
dsh-gemini-multimodal
Gives DeepSeek Harness multimodal perception: the harness model (e.g. DeepSeek) is text-only, so this plugin delegates seeing/hearing to Gemini — image/audio/video/document understanding, transcription, and image generation.
Why
DeepSeek models are text-only. When a task involves a screenshot, a chart, a voice memo, a video clip, or a PDF, this plugin lets the agent actually look and listen — and keep the judgment in the main model.
Two providers (pick one)
| provider | what it needs | when to use |
|---|---|---|
gemini_api |
a Gemini API key from aistudio.google.com | default; direct REST, fast, one HTTP call |
antigravity_cli |
local agy CLI, signed in with your Google account |
no key to manage; uses your Antigravity/Google One quota |
Where to get the key (gemini_api): go to aistudio.google.com → sign in with your Google account → Get API key → Create API key. Keys start with AIza... or AQ. — both work. Pro/Google One subscribers get the same free-tier API quota (image generation is rate-limited on the free tier).
antigravity_cli setup: curl -fsSL https://antigravity.google/cli/install.sh | bash, then run agy once and sign in. No key needed.
Quick start
dsh plugin --profile web add dsh-gemini-multimodal
Configure in your profile/settings layer (optional — zero-config: with no config at all it defaults to antigravity_cli and works if agy is installed):
- id: gemini-multimodal
name: dsh-gemini-multimodal
config:
provider: gemini_api # optional; default antigravity_cli
api_key: <your gemini key> # only for gemini_api
# output_dir: /path/for/images # generated images land here (default OS temp)
Tools
| tool | what it does |
|---|---|
media_understand |
analyze an image / audio / video / URL (OCR, charts, UI, spoken words) |
media_transcribe |
verbatim text of audio/video, timestamps where useful |
image_generate |
text prompt → PNG saved locally (free-tier image quota is low; 429s possible) |
read_document |
summarize or answer questions about PDF / Office / text files |
Security notes (antigravity_cli)
Headless agy -p cannot prompt for permissions, so the plugin passes --dangerously-skip-permissions by default. That grants the agent full tool access for that run. To tighten: set skip_permissions: false and add allow-rules in ~/.gemini/antigravity-cli/settings.json (e.g. read_file(...), deny shell). Prefer gemini_api when you need a narrower trust boundary.
Privacy
gemini_api: media is sent inline (base64) to Google's Gemini API — nothing stored by the plugin, key lives in your config only.antigravity_cli: media is read by the local agent and sent to Google per the agy flow.- Nothing is logged by this plugin.
Development
npm install
npm run typecheck
npm test # mime mapping / agy args / smoke (no network)
npm run build
License
MIT
Changelog
- 0.1.1 — bump for CI publish validation; zero-config default provider (antigravity_cli) with friendly setup guidance.
- 0.1.0 — initial release: dual providers (gemini_api / antigravity_cli), 4 multimodal tools.
安装
装一次目录插件,之后本站所有插件都能让 DeepSeek Harness 自动找、自动装:
dsh plugin add dshbase-catalog 然后对 agent 说「帮我装 dsh-gemini-multimodal」,它会在目录里找到并自动安装。文档:dshbase-catalog · 已验证场景包。
该插件是 GitHub 源码(未发 npm)——直接从仓库装:
Web profile:
dsh plugin --profile web add github:RealAlexandreAI/dsh-gemini-multimodal Headless(CLI)profile:
dsh plugin --profile headless add github:RealAlexandreAI/dsh-gemini-multimodal 实测报告
验证通过:从 GitHub 源码完成 L1 安装 + L2 加载 + L3 运行(dsh 0.1.0-rc.6)。