dsh-ark9canvas — Image Generation Workbench & Agent Tool for DeepSeek Harness

中文 · English · License BSD-3-Clause · pnpm add @a9i5k4/dsh-ark9canvas

> v0.4.0 UPDATE — Bilingual UI (Chinese/English, one-tap switch in the panel header and Settings) and a full stroke-SVG icon set replacing emoji, matching the DeepSeek Harness visual language. Evolution roadmap: docs/ROADMAP.md. — The image workbench is now feature-complete against the reference design: aspect-ratio grid with the same quality-budget + 16px-alignment formula, transparent background, batch generation up to 10 images (independent sub-tasks, partial success still returns what finished), a prompt library with custom JSON sources, multi-channel aggregation, persistent generation logs with retry, and config import/export.

An image-generation plugin for the DeepSeek Harness Web GUI: the agent paints on request via one tool, you paint on demand in a floating workbench — and every agent-initiated generation waits for your approval by default, so nothing bills without a human nod.

The problem it solves: image APIs bill per call, yet agent-initiated generation usually runs blind — a bad prompt retries itself, a loop burns your balance. This plugin puts a human gate in the loop: the tool blocks until you approve in the panel (or denies/timeouts with a clear message and zero cost), while the workbench itself stays one click away for your own un-gated use.

Highlights in 30 seconds

Feature tour

Approval gate — human in the loop, by default

When the agent calls ark9_generate_image, the request appears in the panel's Approvals tab with the prompt, parameters, and elapsed wait time. Approve → generation starts and bills; Deny → the tool returns a clear "user denied" message and the agent asks what to change instead of retrying; Timeout (configurable, 5–600 s) → cancelled, nothing billed. Set Settings → Ark9 生图 → 安全 to never if you want unattended auto-generation.

Workbench — five tabs

• 生成 Generate: prompt, reference images (upload or clipboard paste), model dropdown with per-channel fetch, aspect-ratio grid / manual W×H, quality, transparent toggle, 1–10 count
• 审批 Approvals: pending agent requests with one-click approve/deny
• 提示词 Prompts: search, click to apply, ☆ to favorite locally; custom JSON sources ([{title, prompt, tags?}]) proxied through the host to bypass browser CSP/CORS
• 记录 Logs: every generation with status pills (成功 / 部分成功 / 失败), retry, multi-select delete, click-to-preview
• 说明 About: quick reference

Size system — faithful to the reference formula

Ratios compute their pixel dimensions from a quality budget (low 1K² / medium 2K² / high 4K²) with 16-pixel alignment, exactly like the reference workbench. Because the gpt-image family only accepts three native sizes (1024×1024, 1536×1024, 1024×1536), gpt-image models automatically snap the computed size to the nearest native one; other models send the raw computed size. Manual W×H with a 16-multiple alignment toggle is always available.

Iterative editing — paths, not blobs

ark9_generate_image returns saved file paths with dimensions. Pass any previous output path back via images and the plugin reads the file and runs an /images/edits multipart call — multi-turn "make the robot red" works without ever stuffing base64 into the conversation. returnDataUrl: true opts into inline base64 when a client truly needs it.

Channels — aggregate your relays

Configure multiple OpenAI-compatible channels (name + baseURL + key + default model), mark one active, fetch each channel's model list from its own /models. The active channel serves both the agent tools and the workbench; single-channel setups from older versions migrate automatically.

Engineering core (restraint by design)

• Zero runtime dependencies beyond Node built-ins
• Batch aggregation: count N → N independent sub-tasks (n:1 each), merged into one batch view with ok/fail counts — one slow image never blocks the others
• Durable state: tasks and logs persist to ~/.dsh/ark9-canvas-*.json; a server restart never orphans a poll
• Loopback-only routes: every API route rejects non-localhost callers; file routes are name-sanitized against path traversal
• No third-party prompt content bundled: sources are user-provided URLs

Install (one command)

> Prerequisite: install DeepSeek Harness and start dsh web at least once.

Run in the profile directory (~/.dsh/profiles/web):

cd ~/.dsh/profiles/web
pnpm add @a9i5k4/dsh-ark9canvas

Then edit package.json in that directory and append to the dsh.profile.bundles array:

"@a9i5k4/dsh-ark9canvas"

Restart dsh web — the 🖼️ floating button appears (or a sidebar tab, with Better Sidebar installed). Open Settings → Ark9 生图 once to add a channel (baseURL with /v1, API key, model such as gpt-image-2).

> No pnpm? npm install @a9i5k4/dsh-ark9canvas works the same.
> pnpm v11 blocks packages published <1 day ago: set minimumReleaseAge: 0 in pnpm-workspace.yaml or pin an explicit version for same-day updates.

AI-era installation

Copy this to the AI assistant you're already using:

Install the npm package @a9i5k4/dsh-ark9canvas in the DeepSeek Harness web profile
directory ~/.dsh/profiles/web (pnpm add or npm install),
append "@a9i5k4/dsh-ark9canvas" to the dsh.profile.bundles array in package.json,
then restart dsh web. After that, open Settings → Ark9 生图 and add an
OpenAI-compatible image channel (baseURL with /v1, API key, model).

Updating

cd ~/.dsh/profiles/web && pnpm up @a9i5k4/dsh-ark9canvas

Configuration

Config file ~/.dsh/ark9-canvas.json (everything adjustable in the Settings GUI):

{
"baseURL": "https://your-relay.example/v1",
"apiKey": "sk-...",
"model": "gpt-image-2",
"quality": "high",
"size": "1536x1024",
"count": 1,
"agentApproval": "always",
"approvalTimeoutSec": 120,
"channels": [
{ "id": "c1", "name": "relay-a", "baseURL": "https://your-relay.example/v1", "apiKey": "sk-...", "model": "gpt-image-2" }
],
"activeChannelId": "c1",
"promptSources": [
{ "id": "ps1", "name": "my prompts", "url": "https://example.com/prompts.json" }
],
"outputDir": ""
}

Tasks persist to ~/.dsh/ark9-canvas-tasks.json, generation logs to ~/.dsh/ark9-canvas-logs.json.

Structure

• lib/index.js — Host half: two agent tools, eleven routes, OpenAI-compatible image proxy (async task protocol + batch aggregation), approval queue, persistent logs (zero runtime deps, Node built-ins only)
• lib/client.js — Browser half: floating FAB + glass workbench (shared vanilla-DOM implementation for floating panel and sidebar tab), settings page
• cordis.patch.yml — plugin registration row
• docs/ROADMAP.md — evolution roadmap: DSH-host synergies (AI prompt enhancement, memory-driven styles), cost dashboard, capability registry, mask editing
• smoke-test.mjs — offline integration test (tools / routes / approval paths, no API calls)
• e2e-approval-test.mjs — real end-to-end generation test (bills!)

Known limitations

• Video generation, mask/inpainting painting UI, Gemini-format calls, the infinite-canvas node editor, and WebDAV sync are out of scope (backend has no video model; config import/export stands in for sync).
• The prompt library ships without any third-party content — add your own sources.
• Panel-initiated (manual) generations are never approval-gated: pressing the button is the approval, and it bills.
• Plugin-set changes require a dsh restart.

Credits

This project is built human-machine collaboratively:

• Aik358 — project owner: product direction and engineering.
• ZCode (GLM, Z.ai) — autonomous engineering agent: plugin implementation, protocol reverse-engineering of the async-task/media-upload relay protocol, test suites.

Release

• GitHub: https://github.com/Aik358/dsh-ark9canvas
• npm: @a9i5k4/dsh-ark9canvas
• License: BSD-3-Clause · Independent implementation, contains no WorldCodes Canvas code or branding