dsh-mxpage
MxPage for DeepSeek Harness：基于从 ziguishian/MxPage 提取的、与主机无关的核心构建的电商产品图工作台（分析 → 规划 → Visual Prompt Agent → 生成）。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:kim1232aa/dsh-mxpage#8f825b5994c61f971a829c2562f16ca2c7a88925Mxpage 以 dsh-mxpage 发布，当前版本为 0.3.0。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。
English | 简体中文
DeepSeek Harness plugin that brings MxPage's
ecommerce product-image workbench into the official DSH web profile:
register photos → analyze selling points → plan hero/detail sections → Visual Prompt Agent → generate → edit → export
This is not a generic gallery, and it is not a thin prompt wrapper.
Community plugins such as dsh-imagegen register a one-shot generate_image.
dsh-mxpage does not. Every tool is named mxpage_*, and the model must walk
analyze → plan → VPA → generate so a product photo becomes a coherent
detail-page set rather than a pile of unrelated images.
Verified against DSH 0.1.2-rc.1 host packages. Requires the official
web profile (not sdk-minimal).
Not an official DeepSeek product. Not affiliated with DeepSeek AI. MxPage prompts/schemas remain MIT (灵矩绘境); see NOTICE.
v0.1 reimplemented MxPage's pipeline by hand and threw away the parts that make it a product. v0.2 inverts that: the upstream kernel is ported, the DSH adapter is thin, and the upstream UI is in scope.
src/core/ host-agnostic port of upstream lib/ (never imports @deepseek-ai/*)
└── ports/ Repository · ProviderResolver · Logger · StorageDriver · TaskRunner
src/host/ the five port implementations (JSON repository, fs storage, channels,
jobs-backed task runner)
src/shared/ route paths shared by host and browser
src/tools/ thin mxpage_* wrappers over core services
src/client/ the browser panel
src/core/** cannot import @deepseek-ai/*, schemastery, Next.js, Prisma or
React, and cannot touch process.cwd() / process.env — enforced by a test.
Long-running work goes through ctx.jobs (@deepseek-ai/dsh-jobs-local),
so the shell owns job identity, session scoping, lifecycle state, completion
notices and owner-disposal cancellation. src/host/task-runner.ts is the
fallback for hosts without a registry — the plugin still boots either way,
because jobs is read through the optional ctx.get('jobs') accessor.
Why the extraction was cheap, with evidence:
| Fact | Evidence |
|---|---|
Zero next/* imports inside upstream lib/ | the one exception, provider-runtime.ts, imported NextRequest solely to read two headers |
@prisma/client appears 9 times | two are type-only; four use only the Prisma namespace |
| The 1344-line OpenAI adapter had one hard coupling | import { inferCategory, logApiUsage } from "@/lib/monitor/api-usage" |
See src/core/README.md for the full seam list.
v0.1 required MXPAGE_IMAGE_API_KEY. v0.2 uses a channel list, so several
endpoints can be configured and rotated:
设置 → 插件 → MxPage → 渠道, or in cordis.patch.yml:
- insert:
- id: mxpage
name: dsh-mxpage
config:
channels:
- id: xai
label: xAI (Grok)
baseUrl: https://api.example.com/v1
apiKeyEnv: MXPAGE_XAI_KEY # preferred: keeps the secret out of the doc
models: [grok-imagine-image-2.0]
textModel: grok-4
imageModel: grok-imagine-image-2.0
rotateChannelOnQuotaExhausted: true
| Field | Meaning |
|---|---|
id / label | rotation key and display name |
baseUrl | OpenAI-compatible base URL; a missing /v1 is retried automatically |
apiKey / apiKeyEnv | literal secret, or (preferred) the name of an env var |
models | explicit image model ids; leave empty to discover via GET /models |
textModel / imageModel | preferred models for analyze+plan / generate on this channel |
disabled | skip this channel without deleting it |
Run mxpage_channels first whenever something fails — it reports which
channel is active, its catalog, and whether any image-capable model was found.
Model capability is inferred from the model name. Upstream deliberately skips real endpoint probing to avoid burning image quota, so "the name looks like an image model" does not prove the gateway can render images. A gateway that advertises a model it cannot serve is discovered by failing.
Generation and edit tools call an OpenAI-compatible Images API
(/images/generations, /images/edits). That API is paid. Each
mxpage_generate_section, mxpage_generate_page and mxpage_edit_section
consumes quota. The shipped skills instruct the model to confirm with you before
a whole-page job.
Secrets never appear in tool output, logs or session events — sk- and Bearer
tokens are redacted (src/util/redact.ts).
# from a checkout
dsh plugin add link:/absolute/path/to/mxpage
# or from the built tarball
dsh plugin add ./dsh-mxpage-0.3.0.tgz
dsh plugin add registers the bundle in the profile's
dsh.profile.bundles for you. Then configure a channel (above) and restart the
profile so the layer loads.
The skills ship inside the package. Copy them next to the DSH catalog so they are discoverable:
mkdir -p "${DSH_HOME:-$HOME/.dsh}/skills"
cp -R skills/mxpage-ecommerce-page skills/mxpage-xiaohongshu skills/mxpage-batch-sku \
"${DSH_HOME:-$HOME/.dsh}/skills/"
Attach a product photo and say:
根据这张商品图出一套淘宝详情页
Expected: the agent runs create → analyze → plan → generate and you get a hero
set plus detail sections, each as a new version (never overwriting), stored
under $DSH_HOME/mxpage/projects/<projectId>/ and returned as attachments.
The plugin ships a browser half (lib/client.js) that mounts a workbench into
the DSH centre column, with a sidebar toggle. This is the part upstream MxPage
actually lives in: a stateful workspace, not a prompt wrapper.
| Screen | What it does |
|---|---|
| 分析 | Project meta (name / platform / style, delete project), the product-asset grid (upload, reorder, set main image, delete), and the structured analysis editor with one-click analyze / save |
| 规划 | Output config (hero/detail counts, aspect, in-image language), analyze → plan, the project-level visual style guide, the section list with per-section generation and a whole-page job, plus whole-page translation into a target language |
| 编辑 | Per-section preview, inline editing of title/goal/copy/visualPrompt, generate / regenerate / repaint / enhance / translate, and the version list with activate |
| 导出 | One-click ZIP / JSON export, export notes, the model snapshot, and a gallery of everything currently exportable |
| 小红书 | The four-step carousel flow: plan → review each imagePrompt → generate → edit, with per-page download |
| 批量 SKU | Up to 20 product images per batch — one project per SKU, optional background analyze+plan with per-SKU failure isolation |
| 监控 | The API usage ledger: totals, token counts, top models / projects, quota-state classification, per-entry delete / clear, plus task history with retry for failed tasks |
| 渠道 | Channel diagnostics: the active channel, its model catalog, image/vision/text counts, per-channel key presence, connection test and model discovery with role recommendations, and the quota-rotation note |
Re-planning is guarded behind an explicit confirmation, because it deletes every section, version and generated image in the project.
Build-format note. The DSH web shell does not load client halves as ESM.
It hands each bundle a window.__ModuleLoader__.load({ id, factory }) façade and
a require that resolves the shell's live module table. tsdown therefore emits
the browser half as CJS into lib/client.raw.cjs, and
scripts/wrap-client.mjs wraps it into the envelope. test/client-bundle.test.ts
loads the built bundle through a simulated façade and asserts apply + inject
come back — that test exists because an ESM bundle would silently never apply.
The panel's data API lives at /api/dsh-mxpage/* (src/host/routes.ts),
registered on the host webServer and fenced to loopback requests.
mxpage_*)| Tool | Role |
|---|---|
mxpage_create_project | New project from 1–10 photos via attachment_ids (chat) and/or image_paths |
mxpage_add_asset | Append a photo; role: "main" swaps the primary reference |
mxpage_project_status | Read-only: analysis, sections, versions, running tasks |
mxpage_analyze_product | Vision analysis → category, materials, selling points, suggested plan |
mxpage_plan_page | Section plan + project-level visualStyleGuide. Re-planning deletes existing sections and images |
mxpage_generate_section | One frame; runs the VPA unless prompt_override is set |
mxpage_edit_section | repaint / enhance / translate; new version, never overwrites |
mxpage_generate_page | Whole page as a background job; mode: "missing" fills gaps only |
mxpage_job_status / mxpage_job_cancel | Job control; completed sections stay on disk |
mxpage_export_page | ZIP (00-头图/ + 01-详情页/ + export-manifest.json) or project JSON |
mxpage_xiaohongshu_plan | Xiaohongshu step 1 — has a fully local Chinese fallback plan |
mxpage_xiaohongshu_generate | Step 3 — one image per page, VPA-gated |
mxpage_xiaohongshu_edit | Step 4 — edit one page in place |
mxpage_translate_page | Whole-page translation as a background job — one translate edit per generated section |
mxpage_update_project | Rename a project or change its platform / style |
mxpage_delete_project | Delete a project and its workspace; requires confirm: true |
mxpage_set_main_asset | Swap the primary reference image by asset id or image path |
| Skill | When |
|---|---|
mxpage-ecommerce-page | Taobao / Tmall / JD / Shopee hero + detail pages |
mxpage-xiaohongshu | The four-step carousel flow |
mxpage-batch-sku | One project per SKU; never mix reference images |
Fixed (each documented at its call site):
requestRaw read
if (urls.length === 1 || options?.suppressUsageLog), conflating "skip usage
logging" with "skip the base-URL retry". The VPA is the only caller passing
suppressUsageLog: true, so it silently lost the /v1-vs-root fallback and
degraded to the template prompt on gateways needing a versioned base URL./api/files/[...path] route joined
rootDir() with an unvalidated relative path. The storage port rejects
escapes (normalizeRelPath) and the route is gone entirely.path.join output (backslashes on
Windows) while its URL builder converted back with split(path.sep).process.env.NODE_ENV !== "production", so production builds could not cancel.ProjectStatus.COMPLETED and
GenerationStatus.QUEUED were never written by any upstream service.Preserved deliberately (flagged, not silently changed):
shouldFallbackToNextImageModel
returns false for 429 / quota / 403 / 401, so an exhausted channel aborts
instead of trying the next candidate. Exposed as the
rotateChannelOnQuotaExhausted config flag.editSectionImage is not cancellable — upstream never registered an abort
controller on that path.archiver was replaced by a dependency-free ZIP writer, which also removes
upstream's process.cwd() temp file.npm install
npm run build # tsdown → lib/index.js (+ lib/client.js when src/client exists)
npm test # node --experimental-strip-types --test
npx tsc --noEmit # 0 errors expected
Tests worth knowing about:
test/bundle.test.ts — package manifest sanity and the host-agnostic
invariant on src/core/**test/core.test.ts — ZIP header/CRC format, repository semantics
(terminal-state stickiness, system-project hiding, stale recovery), storage
path containment, task cancellationtest/smoke.test.ts — loads the built lib/index.js, runs the real
apply() against a mock Cordis context, and asserts all 20 tools registerMIT. MxPage prompts, schemas and pipeline logic are MIT (灵矩绘境) — see NOTICE and LICENSE.
Topics: dsh-plugin · dsh · deepseek-harness · mxpage · ecommerce
mxpage_usage_stats | Usage-ledger summary: calls, tokens, quota events, top models / projects, recent errors |
mxpage_channels | Channel diagnostics |