dsh-media-gen

English | 日本語

Generate images and videos in DeepSeek Harness (DSH), using the OpenRouter provider and API key already configured in DSH. Only OpenRouter is supported.

Features

• Ask the agent to write a model-specific prompt, or send your prompt unchanged through a generation card or command.
• Choose models per conversation, use input images, and reuse generated images for image or video generation.
• Save results in the working directory (generated-media/ by default). Video jobs run in the background and notify their originating session.
• English and Japanese UI; command responses are English.

Compatibility

Choose a release for your DSH version from GitHub Releases. Supported versions and changes are recorded in CHANGELOG.md.

Installation

Replace <tag> with the full release tag you selected:

dsh plugin --profile web add 'git+https://github.com/so16tm/dsh-media-gen.git#<tag>'

The repository includes built files, so no local build is needed. It is private: your git must be authenticated with a GitHub account that has access. Replace web if you use another profile.

To update, run the same add command with the new tag and restart DSH. Omitting the tag installs the latest main, including unreleased changes; always specify a tag when reinstalling a particular release.

To remove:

dsh plugin --profile web remove dsh-media-gen

Restart DSH after removal.

In DSH's Models settings, add an OpenRouter provider and save its API key.

Usage

From a chat request to a generated image, ready to reuse in another image or video. Presentation edited from a user-provided screenshot; the displayed cost is specific to this example.

Generation cards and chat

Open Generate… beside the chat input and choose image/video, model, prompt, and available generation options.

• Ask the agent lets the agent write or translate your prompt and ask about missing details.
• Generate as written sends your text unchanged and starts a billable request. Check the displayed estimate first.

You can also ask directly in chat. For different image variants, ask for one prompt per image; /image --count instead repeats the same prompt.

Generated images open in DSH's file preview. Use for image and Use for video open a card using the image as input (the first frame for video). Use in chat inserts its path and replaces the current draft.

Ask the agent to save model-specific prompt notes if needed. Notes persist across restarts and apply to all conversations; saving or clearing them always requires approval.

Commands

Place options before the prompt and quote paths containing spaces with ".

Command	Purpose
/image [--model id] [--aspect 1:1] [--resolution 1K] [--count n] [--seed n] [--image path]... [--yes] <prompt>	Generate images; repeat --image for multiple inputs
/video [--model id] [--duration seconds] [--aspect 16:9] [--resolution 720p] [--audio on|off] [--seed n] [--first path] [--last path] [--ref path]... [--yes] <prompt>	Submit a video job with optional first/last frames and reference images
/media-jobs [job-id]	List this session's video jobs or check one immediately
/media-model	Show selected models and defaults
/media-model image	List image model IDs
/media-model image <model-id> [provider]	Select a model for this conversation
/media-model image default	Clear the conversation selection
/media-approval on	Allow generation in this conversation without normal per-request confirmation
/media-approval off	Restore normal per-request confirmation

Use video instead of image for video model selection. Chat-input selectors do the same thing; selections persist across restarts. The agent prioritizes the conversation selection over its own model argument, then uses configured defaults when no model is specified. Choosing a different card model through Ask the agent updates an existing conversation selection.

Approval and costs

Generation costs money. With requireApproval enabled and no conversation permission, agent requests ask for approval with the actual prompt and estimated cost; direct commands show only an estimate unless you add --yes. The chat input's Media → Ask before generating menu also controls conversation permission; only the user can change it.

Agent tools always ask before sending images outside the working directory or making a second generation call in the same turn, even if normal confirmation is disabled. Direct commands follow requireApproval, --yes, and conversation permission instead; these extra agent checks do not apply to them.

Estimates may be unknown or lower than the actual bill, especially for input-image or token charges. Required agent approvals still ask when an estimate cannot be obtained. After a timeout or unusable response, the provider may already have charged: check before retrying. If part of a multi-image request fails, already saved images are returned with the failure.

Video jobs

Jobs resume tracking after a DSH restart when storage is available. Completed files are downloaded automatically; completion, failure, or expiry is reported to the session when it is open. Jobs started with /video do not wake the agent on completion; use /media-jobs or wait for the next agent turn.

After videoMaxWatchMs (3 hours by default), tracking stops. This does not cancel the provider job or guarantee a refund. Ask the agent to check the job with media_video_status and its job_id to resume tracking.

Configuration

Open Plugins → Media Generation → Components → Media Generation > for model defaults, chat-menu models, limits, approval, and notifications. All models… can select a model outside your chat-menu filter; the agent can still list all models.

Setting	Default	Purpose
defaultImageProvider / defaultImageModel	unset	Default image provider/model IDs
defaultVideoProvider / defaultVideoModel	unset	Default video provider/model IDs
chatImageModels / chatVideoModels	[]	Chat-menu model IDs; empty shows all
maxImagesPerCall	4	Images per call
maxVideoSeconds	10	Video duration limit; see Limitations
requireApproval	true	Normal per-generation confirmation
outputDir	generated-media	Relative output path within the working directory
notifyOnVideoCompletion	true	Notify the session when a video job ends

For polling, tracking, timeout, cache, and download-size settings, see the configuration schema.

Limitations

• Input images are sent to OpenRouter. Only PNG, JPEG, WebP, and GIF up to 20 MiB are accepted, and the model must support the requested input images or frames. API keys are not sent to the browser.
• Avoid common output-folder names such as Pictures. To display older outputs, the image route also accepts images in matching output folders without distinguishing working directories.
• Video previews support mp4, m4v, webm, and mov up to 32 MiB, independently of the download limit.
• When duration is omitted, maxVideoSeconds can limit the model's default only if its catalog lists supported durations.
• Without DSH storage, video tracking does not survive restarts.

Development

After cloning:

pnpm install --frozen-lockfile
pnpm check

pnpm check runs type checks, the build, tests, and CHANGELOG validation. To load the local checkout:

dsh plugin --profile web add <absolute-path-to-clone>

After editing src/, run pnpm build and commit the generated lib/ files with the source changes. Update both READMEs when usage changes; keep release and compatibility information in CHANGELOG.md.

License

MIT, as declared in package.json.