dsh-meme

[!IMPORTANT] dsh-meme is a DeepSeek Harness plugin, not a standalone app.
Install DSH first, then add the plugin through the marketplace or CLI into a profile. It is a community open-source project, not an official DeepSeek product.

Features

• 😊 Click to send — open 😊 beside the composer, search or filter by pack and mood, then pick an image.
• 🎭 Proactive memes — the AI picks a meme that fits the conversation's mood without waiting to be asked; toggle “Proactive memes”, set a send chance (0–100%), or edit the prompt in settings.
• 📱 IM / mobile — through dsh-im, QQ / WeChat / Feishu and other IM chats send memes as native images with no text marker; on a phone or tablet, open DSH Web over your computer's LAN IP and turn on “Allow LAN access” (off by default). QQ tends to render large images as photos (stretched full-screen on tablets); turn on “QQ send size” in settings to send a 200–450 px copy while Web keeps the original.
• 🧠 Learn new images — upload an image and ask to save it; the AI identifies its category and caption. You can also drop image files into a pack folder and click “扫描新增 (Scan)” — the original bytes (including animated GIFs) are kept, and renamed/deleted files have their stale index rows cleaned up.
• 🗂️ Organize packs — ask the AI to “list my memes”, “find duplicates” or “recategorize this one” — the meme_list / meme_update tools handle it (deleting stays in the panel to avoid accidents).
• 📦 Manage and share — enable multiple packs, upload and edit images, import/export ZIPs, and install or update marketplace packs.
• 🚀 Ready to use — includes Big Fat Fish (24 images); the core has no third-party runtime dependencies and uses Node's built-in SQLite. QQ resizing is optional and uses sharp (already bundled with dsh-im; falls back to the original image when unavailable).

Promo video

Install

Requires Node.js ≥ 22.13.0.

Once installed, 😊 appears beside the composer in any conversation. The old package name dsh-expression is deprecated; its internal plugin ID is retained for compatibility.

Screenshots

Meme exchange: you send one, the AI replies with one

Picker: search, packs and mood filters

Manage packs

Open 设置 (Settings) → 表情包 (Memes):

• 图库 (Packs) — cards show image count and disk size; click a card to manage images. The category dropdown filters by real categories; “扫描新增 (Scan)” indexes images you copied into the folder and cleans up stale rows. Each switch controls whether the AI can use that pack; several can be enabled. Create, import/export ZIPs, submit, delete and update.
• 发现 (Discover) — preview marketplace packs and install them (Release ZIPs verify SHA-256).
• 设置 (Settings) — scan directory, proactive memes, prompt, send mode (automatic / text-only / image), send chance (toggle + 0/10/20/30/50/70/100%), QQ send size (toggle + 200/300/400/450 px long edge), layout (side-by-side / full-width row), LAN access, AGENTS.md.

Packs: installed packs with independent switches

Discover: marketplace previews and one-click install

Settings: send mode, send chance and pack directory

The default scan directory is ~/.dsh/meme-packs; settings live in ~/.dsh/dsh-expression.json and survive upgrades. ZIP imports accept standard compression and plain image folders.

Send mode only affects the AI: automatic follows the channel (Web marker, IM native image), text-only never sends images, image requires a real send every time. Send chance is off by default; when enabled, each turn rolls the selected chance to decide whether the AI proactively sends an image (automatic mode only), and a direct request for a meme always bypasses the roll. QQ send size is off by default; when enabled, QQ/IM sends a copy scaled to the selected long edge (300 px default), Web keeps the original, and GIF frames are preserved — the official QQ bot has no sticker message type, and smaller images are more likely to render as stickers instead of being stretched full-screen on tablets. Layout only affects rendering: side-by-side (default) or a full-width image row.

Troubleshooting

• QQ renders memes as stretched photos — the official QQ bot has no sticker message type, so large images are sent as photos. Turn on “QQ send size” and pick 300 px (200–450 works best); QQ/IM then sends a scaled copy while Web keeps the original. Scaling needs sharp, which dsh-im bundles; without it the original image is sent.
• Update failed / plugin disappeared — click Retry in the marketplace; if it still fails, quit DSH completely and update again. On older versions, or if it keeps failing, run pnpm add dsh-meme@latest in the profile directory with DSH closed. Packs and settings live in ~/.dsh and survive reinstalls. See the troubleshooting guide.
• Restrictive network / marketplace unavailable — mirrors are raced and the catalog is cached automatically; click Retry, then switch Catalog source to Mirrors. See the troubleshooting guide.
• Phone / LAN / reverse proxy — turn on Allow LAN access for LAN IPs; a reverse-proxy domain needs an explicit trusted host (DSH_MEME_TRUSTED_HOSTS or allowedHosts). See the troubleshooting guide.

Share your own pack

Create a pack, add images, click 导出 (Export) and send the ZIP; friends install it with 导入图库 (Import pack). To join the marketplace, click 投稿 (Submit) and attach the ZIP with image sources and licence details in the prefilled GitHub form. Use the plugin's exported ZIP — GitHub's “Download ZIP” is a source archive, not a pack.

Tools and extensions

send_meme tag=happy
send_meme query="猫 比心"
learn_meme
meme_list tag=work
meme_update path="/home/you/.dsh/meme-packs/work/a.jpg" caption="加班到趴桌"

Mood buckets: happy / angry / sad / shy / confused / daily. The model should copy a candidate's [表情: caption] verbatim; if nothing fits, retry with other words or reply with text.

Agent docs: AGENTS.md documents per-channel sending, path prefixes and the GIF-flattening caveat; pack maintenance lives in the skill .dsh/skills/dsh-meme-maintenance. The settings page can view/edit it.

Animated GIFs: learn_meme and upload recognition sample GIF frames into a filmstrip, so captions describe motion; the original keeps animating. WebP / APNG contribute metadata only and are recognized from the first frame.

IM delivery

Through dsh-im, send_meme candidates carry absolute paths and the model calls dsh_im_return_file; Web chats keep [表情: caption]. dsh-companion also provides QQ delivery.

Official QQ bot

For dsh-qqbot sessions, candidates carry absPath for qqbot_send_file's file_path. dsh-qqbot's path allowlist needs your pack directories in sendFile.extraRoots:

# the qqbot profile's cordis.patch.yml (or Settings → QQ bot)
- id: im-qqbot
  config:
    sendFile:
      extraRoots:
        - C:\Users\<you>\.dsh\meme-packs
        - C:\Users\<you>\.dsh\profiles\<profile>\node_modules\dsh-meme\memes

Replace C:\Users\<you> with ~ on Linux/macOS; include the bundled pack's path too.

Read-only meme service for other plugins

Other plugins in the same process can call ctx.get('dshMeme') (read-only):

tag is a mood bucket; tagStats gives the real category distribution on a miss. Every returned path is absolute and exists. For custom catalogs see the remote pack specification.

Plugin HTTP routes accept loopback hosts by default; LAN and reverse-proxy setup is in the troubleshooting guide.

Development

Plain ESM, no build step. Clone the repository and run the tests:

git clone https://github.com/yyh-001/dsh-meme.git
cd dsh-meme
npm test

Tests use Node's built-in runner (node --test test/*.test.mjs) and should be green before a release. With a link: install into a profile, restart DSH to pick up changes. Agent-facing conventions live in AGENTS.md.

Community & feedback

• Check the troubleshooting guide first, then open an issue.
• Chat: QQ group 1 (full) · QQ group 2 for dsh plugins
• Version history: npm.

Related projects

• DeepSeek Harness — the upstream agent platform
• dsh-im — QQ / WeChat / Feishu IM bridge that delivers memes as native images
• dsh-meme-packs — downloadable packs and a machine-readable catalog
• dsh-companion — companion-mode plugin with an optional QQ channel
• agent-expression — a local skill with the same pack idea, for OpenClaw / Hermes / Cursor

License

Code is licensed under MIT. Check each bundled, marketplace or third-party pack for image provenance and usage rights.