---

Compatibility

Surface	Status
Harness	DeepSeek Harness 0.1.0-rc.6 (compat declared for 0.1.0-rc.6)
Node	^22.19.0 || >=24.0.0
Engines	Any OpenAI-compatible images endpoint; presets for OpenAI Images (gpt-image-1) and Zhipu CogView (cogview-3-flash)
Surfaces	Host image_generate tool + web result card + Plugins settings tab

What you get

dsh-draw gives the harness one unified image_generate tool with standard parameters (prompt/size/count/quality/style/engine) that are translated per engine:

• Multi-engine routing — a config-driven chain (OpenAI Images, Zhipu CogView, or any OpenAI-compatible endpoint) walked top-down with health-aware fallback: consecutive failures push an engine into cooldown, and the next healthy engine serves the call.
• Durable results — generated images are saved as workspace attachments (content-addressed, under the harness's attachment policy) and returned as canonical file references.
• Quota accounting — per-session caps on generation calls and image bytes, folded from the durable session log and enforced before engine spend and before storage.
• Credentials as references — engine API keys are environment-variable names resolved per call through the official ctx.credentials seam; literal keys are never stored in configuration and never logged.
• Web surfaces — an in-conversation result card (images, engine, quota, one-click regenerate) and a Plugins settings tab (engine chain, credential status, probes, quota limits).

model                           harness
  │ image_generate {prompt, ...} ──▶ validate ──▶ quota check ──▶ router
  │                                  openai ──(fail)──▶ cogview ──▶ images
  │ ◀── canonical JSON + image blocks (durable attachment refs)
  │                       └── draw/generated session event (quota + audit)

Quick start

# 1. install the bundle into your profile
dsh plugin --profile web add "github:PerryLink/dsh-draw#main"

# or from npm (published releases)
dsh plugin --profile web add dsh-draw

# 2. provide the engine keys as credential references (environment variables)
#    OPENAI_API_KEY and/or ZHIPU_API_KEY — never in the profile patch

# 3. restart and verify the row
dsh --profile web --dump-config | grep -A2 'id: dsh-draw'

Then ask the agent to draw:

> Draw a 1536x1024 landscape of a lighthouse at dusk, vivid style.

Install & uninstall

• git channel (latest main): dsh plugin --profile web add "github:PerryLink/dsh-draw#main" — the prepare script builds with production dependencies only.
• npm channel (published releases): dsh plugin --profile web add dsh-draw.
• tarball channel: pnpm pack in this repo, then dsh plugin --profile web add ./dsh-draw-<version>.tgz.
• uninstall: dsh plugin --profile web remove dsh-draw (or remove the row from the profile patch).

If pnpm reports ERR_PNPM_IGNORED_BUILDS for this package (esbuild's harmless platform-binary validation), add allowBuilds: { esbuild: true } to your pnpm-workspace.yaml — the dsh CLI prints the exact snippet.

Configuration

All tunables are Schemastery Config fields (changeable from cordis.yml). An id-targeted override replaces the whole row — restate every key you need. cordis.patch.yml documents each key inline.

Key	Default	Meaning
engines	OpenAI + CogView presets	Ordered engine chain, walked top-down with fallback; each entry: id, baseUrl (no credentials), model, apiKeyRef (env-var name), enabled, sizeMap, qualitySupported, styleSupported, responseFormat (b64_json/url), imageMediaType
defaultEngine	openai	Engine id the router prefers; must name a configured engine
requestTimeoutMs	120000	Per-generation HTTP timeout (1000..600000)
maxImagesPerCall	4	Cap on images one call may produce (1..10)
maxPromptLength	4000	Prompt length cap in characters (1..32000)
maxGenerationsPerSession	200	Per-session generation-call cap (1..100000)
maxBytesPerSession	209715200	Per-session image-byte cap (1048576..4294967296)
failureThreshold	2	Consecutive failures before an engine enters cooldown (1..10)
cooldownMs	60000	Engine cooldown after the threshold trips (1000..3600000)

Example override in your profile patch:

- insert:
    - id: dsh-draw
      name: dsh-draw
      config:
        defaultEngine: cogview
        maxImagesPerCall: 2

Tools & surfaces

Surface	Notes
image_generate	Standard parameters; returns canonical JSON (engine/model/size, image references, quota, fallback flag, attempts) plus image content blocks
Result card (tool.call.toolview, key image_generate)	Images, engine/quota line, one-click regenerate (full drawer path: quota + routing + audit)
Settings tab (Plugins → Image generation)	Engine chain, credential status, set/remove API keys (credential references), connectivity probes, quota limits

Permissions & data

• Permissions: the plugin makes outbound HTTPS calls to the configured engine endpoints only; every other surface is read-only. The settings tab's only writes are credential set/remove calls on the official ctx.credentials seam.
• Data: generated images are saved through the official attachment store under the harness's own attachment policy. Quota usage is folded from the draw/generated session events — nothing else is stored.
• Session log: the draw/generated event records engine, model, standardized request, byte totals, and attachment ids — the audit facts, never the API keys.

Security boundaries

• Credential references, never literals. apiKeyRef names an environment variable; a baseUrl embedding credentials fails the load loudly.
• Sanitized display. URLs, probe notes, and error text are redacted (userinfo passwords, credential query values, bearer tokens, JWTs) before any display or log.
• Quota before spend. Generation and byte caps are checked before engine calls and before attachment storage; exhausted sessions fail fast without spending engine credits.
• Fail loud, fall back deliberately. Malformed responses surface as structured errors; a failing engine is skipped after its cooldown threshold, and a fully exhausted chain returns the complete attempt record instead of pretending success.

Known limitations

• Image models only. No video, audio, or edit endpoints; no vision understanding.
• Engine compatibility. Engines must speak the OpenAI POST /images/generations shape (base64 or URL delivery); provider-specific extras are out of scope.
• Cost awareness is structural. The plugin counts calls and bytes but does not know engine pricing — pair with dsh-budget for cost governance.

Development

pnpm install        # node ^22.19 || >=24
pnpm run typecheck  # tsc: src + tests against the local harness checkout
pnpm run typecheck:ci  # tsc against the published 0.1.0-rc.6 faces (no paths)
pnpm test           # vitest: 77 tests, 11 suites (scripted transport, real Context/Session/ToolRuntime)
pnpm run build      # tsc declarations + tsdown bundles (lib/)
pnpm run verify:self-contained  # dependency specs resolve from the registry
pnpm run verify:artifacts       # host ESM face + typert manifest + browser bundle + config files
pnpm pack           # the published tarball

Topics

dsh, dsh-plugin, deepseek-harness, deepseek, cordis, image-generation, openai-images, cogview, zhipu, text-to-image

Contributors

• @PerryLink — creator and maintainer: engine router, drawer, quota accounting, Typert wire vocabulary, browser half, and the five-language docs.

License

Apache License 2.0 © 2026 dsh-draw contributors