dsh-plugin-bilibili

中文 | English · Changelog · Releases

A Bilibili retrieval plugin for DeepSeek Harness. After install the agent gains five tools:

• bilibili_search — find videos by keyword: title, uploader, play count, duration, publish date.
• bilibili_video — full metadata for one video: counts, partition, multi-part pages, description.
• bilibili_subtitles — the subtitle transcript of one video, merged into plain text.
• bilibili_playurl — the direct mp4 play URL (and granted quality/size) for downloads or frame extraction.
• bilibili_frames — real video frames (preview sprite grid, cover fallback, or ffmpeg extraction) as image blocks, so image-capable models can watch the actual footage instead of relying on subtitles alone.

Anonymous by default: search bootstraps an anonymous cookie automatically and metadata always works. Set a SESSDATA cookie to unlock login-gated subtitle tracks (where most AI subtitles live), higher-quality play URLs, and the frames preview path. Effectively read-only: the plugin reads metadata, subtitles and frame images only — it never stores or re-uploads videos.

Unofficial community plugin — not affiliated with DeepSeek or Bilibili.

Harness compatibility: 0.2.3 is verified on DeepSeek Harness 0.1.7-rc.2 and 0.1.6-alpha.2 (Typert API Gateway) from one build. The 0.1.0–0.1.5 line works as well — nothing in the plugin uses the 0.1.6+ codec API. Version map in Compatibility.

Install

dsh plugin --profile web add dsh-plugin-bilibili

# or directly from Git:
dsh plugin --profile web add git+https://github.com/moxingovo/dsh-bilibili

Restart dsh web. New conversations gain bilibili_search, bilibili_video, bilibili_subtitles, bilibili_playurl and bilibili_frames.

Mount it in an agent preset instead when only that preset should see the tools:

- id: dsh-plugin-bilibili
  name: 'dsh-plugin-bilibili'

Optional SESSDATA

Log into bilibili.com, open DevTools → Application → Cookies → the bilibili.com entry, and copy the bare SESSDATA value — not the whole cookie header. Put it in the environment or in your $DSH_HOME/.env (the launcher injects every KEY=VALUE line of that file into the server process):

BILIBILI_SESSDATA=<your-bare-token>

Without it only publicly visible subtitle tracks are returned; videos whose tracks require login fail with the structured code BILIBILI_LOGIN_REQUIRED. A bare token is fine — the plugin prefixes SESSDATA= itself.

Configuration

Override any field in profiles/web/cordis.patch.yml — later layers win per row:

- id: bilibili
  name: 'dsh-plugin-bilibili'
  config:
    subtitleLanguage: ai-zh

This package is self-contained: it registers the tools and carries its own Bilibili REST provider. It does not publish a ctx.bilibili service, so no other plugin can consume it by service key (an earlier in-tree line did — see Compatibility).

Tools and error codes

Tools fail with structured errors carrying these codes:

bilibili_frames is registered on the plugin's own fiber like the other four tools, so it is visible everywhere the plugin is mounted (including a host-domain bundles mount). It needs two things at execution time: an attachments service in the calling scope, and a model route that declares image input. Either one missing fails that single call with a structured error (no attachment service is mounted in this scope / does not declare image input) instead of removing the tool from the registry — always visible, failing closed. See the CHANGELOG "fixes" section for why.

Skills

Two companion skills ship in skills/:

• plugin-tool-bilibili — tool usage: arguments, return shapes, error routing, fallbacks.
• plugin-web-bilibili — the host-side provider: config keys, the SESSDATA credential, WBI / risk-control behaviour, and the full error-code table.

DeepSeek Harness does not auto-load skills that live inside a plugin package. Copy them into your user skill root, or point a preset's skill-filesystem at the installed package:

# user root — available to every preset that mounts skill-filesystem
cp -r node_modules/dsh-plugin-bilibili/skills/* ~/.dsh/skills/

# or, in an agent preset: keep the skills in the package and reference them
- id: skill-filesystem
  name: '@deepseek-ai/dsh-skill-filesystem'
  config:
    customSkillDirs:
      - '<DSH_HOME>/profiles/web/node_modules/dsh-plugin-bilibili/skills'

Either way the agent consults the matching skill before calling the tools.

Security

• The cookie is read from the environment (or an explicit literal config value) only; it never enters logs or tool output.
• Every request refuses redirects, so the cookie can never be forwarded to another origin.
• The cookie is sent only to the configured API host (api.bilibili.com by default); subtitle CDN downloads carry no cookie.
• No video or audio download (bilibili_frames' extract mode downloads the mp4 to a temp directory to run ffmpeg, then deletes it).

Development

Node 22 or newer:

npm ci
npm test          # 20 tests, fully offline (mocked HTTP)
npm run typecheck # runs against the published DeepSeek Harness packages
npm run build

The repo pins its dependency tree in package-lock.json, and CI runs exactly these four commands (.github/workflows/ci.yml). Release steps live in RELEASE.md.

Known issue

Early rc releases of the official DeepSeek Harness packages declare an unpublished peer dependency: dsh-agent 0.0.1-rc.1/rc.2 and dsh-session 0.0.1-rc.1/rc.2 list @deepseek-ai/dsh-type-meta, which is not on the npm registry. A fresh install whose resolver lands on those versions fails with a 404 for @deepseek-ai/dsh-type-meta (reproduced with pnpm 11 and the npmmirror mirror; npm resolves 0.0.1-rc.5 and succeeds). Workarounds: npm with the committed package-lock.json (npm ci), or dsh plugin add inside an already-installed harness workspace, whose lockfile pins resolvable versions. This is an upstream rc-stage publishing issue and disappears once upstream fixes the metadata.

Compatibility

Live-verified on 0.1.7-rc.2: bilibili_search, bilibili_video, bilibili_playurl. bilibili_subtitles returns the documented BILIBILI_SUBTITLES_UNAVAILABLE for videos without a track. Typecheck, build and all 20 tests pass against @deepseek-ai/*@0.1.7-rc.2.

License

MIT — see LICENSE.