dsh-image-gen

Generate images in DeepSeek Harness with OpenAI gpt-image-2, using a signed-in Codex subscription by default or an API key when explicitly configured.

中文说明

These illustrations mirror the shipped developing and completed card states. API-key mode can replace the light field with real streamed drafts; Codex subscription mode animates until its non-streaming response arrives. The completed state remains available as a durable DSH attachment with preview and download controls.

Capability map

Highlights

• Registers the Codex-compatible model tool name image_gen.
• Reuses the refreshable OAuth login owned by dsh-codex-connect; no OPENAI_API_KEY is required for Codex subscription mode.
• Streams up to three real provider partial images when the API-key Images endpoint is selected; Codex subscription mode keeps the developing animation active until its non-streaming response arrives.
• Cross-fades each partial over one animated developing plate, then sharpens into the final image.
• Stores the final image in DSH's immutable attachment store and supports replay, lightbox preview, and download.
• Keeps model-facing tool output text-only, so image generation does not make a text-only model route reject conversation history.
• Supports native tool calls and nested Code Mode calls with the same durable card.
• Resolves credentials per operation, rejects redirects, limits response sizes, validates image bytes through DSH, and retries only transient failures.
• Includes Chinese and English UI copy plus prefers-reduced-motion support.

Codex parity and improvements

OpenAI Codex's built-in image_gen tool hardcodes gpt-image-2, calls the ChatGPT Codex Images endpoint with subscription OAuth, records one working state, and saves the completed image under Codex's generated-images directory. Its public backend currently requests a non-streaming JSON response and its UI does not expose a distinctive diffusion animation.

dsh-image-gen keeps the compatible image_gen name and model while improving the visible process:

Primary references:

• Codex image-generation tool
• OpenAI image-generation skill
• GPT Image 2 model
• Image generation guide

Compatibility

Verified against:

• DeepSeek Harness 0.1.0-rc.6
• dsh-codex-connect 0.1.0-alpha.4.4
• Node.js 24.15.0 (package support: ^22.19.0 or >=24.0.0)
• DSH Web profile on Windows 11
• Real Codex subscription generation, durable replay, Blob preview, and download controls

Verification date: 2026-08-15.

Install

Review third-party source before installation and pin release tags or commits. For the default keyless subscription path, install Codex Connect, sign in once, then install this plugin:

dsh plugin --profile web add dsh-codex-connect
dsh openai-codex login
dsh plugin --profile web add github:LeemanCheung/dsh-image-gen#v0.2.0

For local development:

git clone https://github.com/LeemanCheung/dsh-image-gen.git
cd dsh-image-gen
npm install
npm run check
dsh plugin --profile web add .

The repository commits lib/index.js and lib/client.js, so a pinned Git install does not need to run a dependency build script. Restart the DSH Host after installation and refresh the Web page.

Authentication

authMode: auto is the default. It first asks the installed dsh-codex-connect package for its DSH-owned, refreshable ChatGPT OAuth credential and sends it only to the fixed first-party endpoint https://chatgpt.com/backend-api/codex/images/generations. If Codex credential resolution, compatibility checking, or refresh fails for any non-cancelled reason, it then tries the DSH credential reference named by apiKeyEnv (default OPENAI_API_KEY); if neither works, it returns the combined failure. codex-subscription never falls back.

Use authMode: codex-subscription to forbid API-key fallback, or authMode: api-key to use only the configured Images API account. Never put an OAuth token or API key in cordis.patch.yml, chat messages, Git, or screenshots. The plugin resolves authentication for every generation and does not retain it after the request.

Use

Ask naturally in a DSH conversation, for example:

Generate a cinematic 16:9 product photograph of a translucent mechanical keyboard on a dark glass desk, violet rim light, no text.

The model calls image_gen. While it runs, the card shows the developing animation. API-key mode replaces the light field with each real streamed partial; Codex subscription mode reveals the final image when its JSON response arrives. DSH's existing interrupt control cancels the request. The settled card supports preview and download.

Tool options:

• prompt: detailed generation instructions, 1–32,000 characters and at most 64,000 UTF-8 bytes.
• size: auto or arbitrary WIDTHxHEIGHT accepted by GPT Image 2: each edge divisible by 16, no edge above 3840, aspect ratio 1:3–3:1, and 655,360–8,294,400 total pixels.
• quality: auto, low, medium, or high.
• output_format: png, jpeg, or webp in API-key mode. Codex subscription mode currently returns PNG.
• output_compression: 0–100 for API-key JPEG/WebP only.
• background: auto or opaque. GPT Image 2 does not support transparent output.

Configure

The bundle inserts the image-gen row with safe defaults. Override it in the selected profile's cordis.patch.yml:

- id: image-gen
  name: dsh-image-gen
  config:
    authMode: auto # auto | codex-subscription | api-key
    apiKeyEnv: OPENAI_API_KEY
    baseUrl: https://api.openai.com/v1
    model: gpt-image-2
    defaultSize: auto
    defaultQuality: auto
    defaultOutputFormat: png
    defaultOutputCompression: 90
    defaultBackground: auto
    moderation: auto
    partialImages: 3
    requestTimeoutMs: 120000
    maxRetries: 2
    retryBaseMs: 1000
    maxConcurrent: 2

baseUrl, model, moderation, partialImages, output compression, and API pricing apply only to API-key mode. Codex subscription mode fixes the model to gpt-image-2, uses the fixed first-party Codex endpoint, returns PNG, and never sends OAuth to baseUrl. Configuration fails at load for an invalid provider URL or default image size. Plain HTTP is accepted only for loopback development endpoints. Every credential-bearing request uses redirect: "error".

API-key contract and operation bounds

Create a DSH credential named by apiKeyEnv (default OPENAI_API_KEY) or export that environment variable before starting the DSH Host, then set authMode: api-key. Never put the secret in the profile patch.

A custom baseUrl must expose <baseUrl>/images/generations and support either an OpenAI-compatible data[0].b64_json response or SSE image_generation.partial_image / image_generation.completed events carrying b64_json. It must be HTTPS outside loopback.

Only transient provider failures (429, 5xx, retryable protocol/response errors, and network failures) are retried. Provider moderation and user-input errors fail immediately.

Cost note

Codex subscription calls consume the image-generation allowance associated with the signed-in ChatGPT plan. API-key calls are billed by the selected quality and size; each requested partial costs additional image-output tokens according to the provider guide. partialImages does not apply to subscription mode.

Data, network, and permissions

• Network: subscription mode sends the prompt and supported options only to https://chatgpt.com/backend-api/codex/images/generations; API-key mode sends them to baseUrl.
• Credentials: subscription mode asks dsh-codex-connect for its DSH-owned OAuth credential; API-key mode resolves the configured DSH credential reference. Neither secret is stored in plugin state, logs, metadata, or session history.
• Storage: stores only completed images through the DSH attachment service. Partial frames stay in bounded Host memory while the call is active and are then discarded.
• Browser access: uses a loopback-only private RPC. A final image is returned only after the Host finds the exact attachment reference in the requested session and call record.
• Workspace files: does not read or write the session workspace.
• User data: prompts and tool arguments follow DSH's normal session logging. OpenAI receives the prompt under the terms governing the selected ChatGPT subscription or API account.

Troubleshooting

OpenAI Codex is signed out

Install dsh-codex-connect, sign in from its DSH settings page or run dsh openai-codex login, then retry. Never paste the OAuth token into chat.

No credential is configured for OPENAI_API_KEY

This appears in explicit API-key mode, or after automatic Codex fallback. Configure the credential for the DSH Host process or switch to a signed-in Codex subscription. Never send the key in chat.

Image generation was blocked

Revise the prompt. The plugin does not retry provider moderation or user-input errors.

Preview unavailable after success

Refresh the page. If the card still cannot load, inspect Host logs and verify that the profile still mounts dsh-image-gen and its attachment store is available.

dsh plugin --profile web why dsh-image-gen
dsh --profile web --dump-config

Requests time out

Increase requestTimeoutMs within its 10–300 second range or select a lower quality. API-key mode can also use JPEG. DSH interruption still aborts the upstream request.

Roll back or remove

Before changing a production profile, take a DSH config snapshot when the undo plugin is installed. To uninstall:

dsh plugin --profile web remove dsh-image-gen

Restart the DSH Host and refresh the page. Existing image records remain in session history; their custom card requires the plugin to be installed.

Development

npm install
npm run typecheck
npm test
npm run build
npm pack --dry-run

The keyless suite uses deterministic mocked SSE/JSON responses and a local redirect server. It covers both authentication modes without reading real secrets. Real-provider checks are manual because they consume a Codex subscription allowance or bill an API account. The 0.2.0 release was manually verified with one signed-in Codex subscription generation and cold-session browser replay.

The build emits:

• lib/index.js: Host Cordis plugin.
• lib/client.js: browser module-loader bundle.
• lib/client.js.map: browser source map.

Known limitations

• Version 0.2 generates new images. Codex-style reference-image editing is not yet exposed because a safe DSH attachment selector and explicit external-upload consent are needed.
• Final previews are intentionally loopback-only. Remote Web clients receive a clear unavailable state rather than image bytes.
• Current DSH credential resolution and attachment saving do not accept cancellation signals. The plugin checks cancellation before and after those stages and waits for them during teardown, but cannot interrupt a provider implementation that stalls inside either service.
• OpenAI may evolve arbitrary-size limits or event fields. The plugin fails closed on incompatible responses instead of guessing.

Security

See SECURITY.md for private reporting. Do not include OAuth tokens, API keys, private prompts, or generated private images in a public issue.

License

MIT