Inkotake / dsh-rich-artifacts

dsh-rich-artifacts

中文 | English

A DeepSeek Harness (dsh) dual-face plugin that turns workspace files into chat artifacts: raster images render inline in the conversation, other files become download cards.

No more file:///... links or “the file is in output/” — when the model creates a user-facing file, it calls publish_artifact and the file is durably delivered through the chat UI.

Features

• publish_artifact tool
  • path (required): workspace file to publish
  • title (optional): display title
  • display: auto (default) · inline · download
• ArtifactStore
  • SHA-256 content-addressed blob storage
  • metadata stored separately under ~/.dsh/artifacts/v1/
• Durable session event artifact/published
  • only artifact metadata enters the session log
  • historical sessions keep showing artifacts after reload / fork
• HTTP endpoints
  • GET /api/artifacts/:id/content — inline-safe preview
  • GET /api/artifacts/:id/download — attachment download
• Web UI turn-tail gallery
  • PNG / JPEG / WebP / GIF render inline
  • everything else gets a file card with a download button
• Security
  • workspace containment via fs.resolve + fs.contains
  • final-component symlink rejection
  • magic-byte MIME sniffing (a .png that is actually HTML will not inline)
  • SVG / HTML are download-only in V0.1
  • file size, image size and image-pixel limits

Install

dsh plugin --profile web add github:Inkotake/dsh-rich-artifacts

For a locked, reproducible source install:

dsh plugin --profile web add github:Inkotake/dsh-rich-artifacts#<commit>

Then restart dsh web (or the profile you installed into).

Usage

The plugin registers a system-prompt section, so the model should call the tool on its own when it creates a deliverable. You can also ask for it directly:

Create a training-loss chart, export it to CSV, and publish both files as artifacts.

The agent then runs:

publish_artifact({ "path": "output/loss.png", "display": "auto" })
publish_artifact({ "path": "output/loss.csv" })

and the turn tail shows the image inline plus a CSV download card.

Configuration

In the profile cordis.patch.yml:

- id: dsh-rich-artifacts
  name: dsh-rich-artifacts
  config:
    root: ~/.dsh/artifacts
    maxFileBytes: 104857600
    maxTurnBytes: 524288000
    maxImageBytes: 20971520
    maxImagePixels: 40000000
    inline:
      png: true
      jpeg: true
      webp: true
      gif: true
      svg: false

Field	Default	Meaning
root	~/.dsh/artifacts	Artifact storage root
maxFileBytes	104857600	Per-file size cap
maxTurnBytes	524288000	Reserved per-turn cap (not yet enforced in V0.1)
maxImageBytes	20971520	Max inline image size
maxImagePixels	40000000	Max inline image pixels (best-effort parse)
inline.*	png/jpeg/webp/gif: true, svg: false	Which image types may render inline

Build

pnpm install
pnpm run build

lib/index.js is the ESM host plugin, lib/client.js is the browser bundle (window.__ModuleLoader__.load(...)). Both are committed so source installs work without a build step.

Test

node tests/artifact-store.test.mjs

Architecture

DeepSeek Agent
     │  creates file
     ▼
Workspace File
     │  publish_artifact
     ▼
ArtifactStore ──────────────► artifact/published session event
     │                                  │
     │ (blob + metadata)                │ (artifact_id + metadata)
     ▼                                  ▼
HTTP /api/artifacts/:id/*        Session log (replayable)
     │                                  │
     └──────────────┬───────────────────┘
                    ▼
              Web UI turn-tail
              ┌─────────┴──────────┐
              ▼                    ▼
        ImageArtifact         FileArtifact

Notes

• artifact/published is a log-only event. Because the current harness has no registration surface for out-of-repo session event types, the host adds it to KNOWN_SESSION_EVENT_TYPES at startup so persisted sessions remain loadable while the plugin is installed.
• MIME type is decided by magic bytes first; extensions are only a secondary hint.
• SVG and HTML are intentionally not inlined in V0.1.

License

MIT