@a9i5k4/dsh-ark9canvas
DSH 生图插件 Ark9canvas：Agent 生图(默认需用户批准) + 浮窗生图工作台(比例/透明背景/批量/提示词库/多渠道/生成记录)，OpenAI 兼容图片 API 文生图/图生图。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @a9i5k4/dsh-ark9canvas@0.4.1Ark9canvas 以 @a9i5k4/dsh-ark9canvas 发布，当前版本为 0.4.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
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.
| Approval gate by default | Every agent generation waits in the panel's Approvals tab — approve, deny, or let it time out; denial and timeout never bill |
| One workbench, two homes | Floating FAB + glass panel out of the box; auto-registers as a Better Sidebar tab when dsh-better-sidebar is installed; stacks above dsh-cua's FAB when both exist |
| Full size system | Aspect-ratio grid (12 presets + auto) computed with the quality-budget + 16px-alignment formula; manual W×H with 16-multiple snapping; gpt-image models auto-snap to the three native sizes |
| Batch up to 10 | Each image runs as an independent sub-task aggregated into one batch — partial success still returns what finished, with per-batch ok/fail counts |
| Transparent background | One toggle sends background:"transparent" (supported by gpt-image family) |
| Prompt library | Local favorites (☆) + custom JSON sources fetched through a host-side proxy — no CORS, no bundled third-party content |
| Multi-channel aggregation | Keep several OpenAI-compatible relays (baseURL + key + model each), switch the active one, fetch model lists per channel |
| Persistent generation logs | Every batch is recorded with params and outcomes; failed batches retry with one click; multi-select delete |
| Bilingual UI | One-tap Chinese/English switch (panel header + Settings), initialized from your browser language |
| AI-friendly by design | Tool results return saved file paths + dimensions — never base64 blobs — unless you explicitly ask for them; references accept dataURLs or previous output paths for iterative editing |
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.
[{title, prompt, tags?}]) proxied through the host to bypass browser CSP/CORSRatios 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.
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.
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.
~/.dsh/ark9-canvas-*.json; a server restart never orphans a pollPrerequisite: install DeepSeek Harness and start
dsh webat 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-ark9canvasworks the same. pnpm v11 blocks packages published <1 day ago: setminimumReleaseAge: 0in pnpm-workspace.yaml or pin an explicit version for same-day updates.
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).
cd ~/.dsh/profiles/web && pnpm up @a9i5k4/dsh-ark9canvas
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": ""
}
| Key | Meaning |
|---|---|
agentApproval | always (default) — agent generations need panel approval; never — unattended |
approvalTimeoutSec | 5–600 s; timeout cancels without billing |
channels / activeChannelId | Multi-channel aggregation; falls back to the top-level baseURL/apiKey/model when empty |
outputDir | Where images are saved; empty = ~/Pictures/ark9-canvas |
Tasks persist to ~/.dsh/ark9-canvas-tasks.json, generation logs to ~/.dsh/ark9-canvas-logs.json.
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 pagecordis.patch.yml — plugin registration rowdocs/ROADMAP.md — evolution roadmap: DSH-host synergies (AI prompt enhancement, memory-driven styles), cost dashboard, capability registry, mask editingsmoke-test.mjs — offline integration test (tools / routes / approval paths, no API calls)e2e-approval-test.mjs — real end-to-end generation test (bills!)This project is built human-machine collaboratively:
@a9i5k4/dsh-ark9canvas