English | 简体中文

dsh-modellix

dsh-modellix brings Modellix media generation, LLM models, and Web research into DeepSeek Harness. One Modellix API Key enables a chat-first workflow: ask naturally, keep the context in the conversation, and inspect completed work without leaving the session.

Harness and this plugin currently use prerelease interfaces. Check the peer dependencies and CHANGELOG before upgrading Harness.

What changed in 0.2.1

• The Web switch now unregisters only modellix_web_search and modellix_web_fetch. Harness-native Web tools and providers remain available, so turning Modellix Web off falls back to the active profile's defaults.
• Updated the authoring and runtime integration for Harness 0.1.2-alpha.4, while retaining 0.1.1-rc.2 compatibility.

What changed in 0.2.0

• Removed the standalone Design tab. Chat is now the primary media workflow.
• Added six explicit Agent tools for media catalog, Schema, parameter preparation, uploads, generation, and result lookup.
• Added explicit Modellix Web Search and Fetch tools so the Agent can use them automatically when current or source-backed information is needed.
• Added Modellix Design as a right-side session panel. It is 360 px wide on desktop, full-width on narrow screens, and does not squeeze its internal content during opening or closing.
• Added live chat result cards. A submitted card updates in place when the background task finishes; the one-shot result lookup does not create a duplicate card.
• Added session-scoped result history, collapsible result lists and cards, image enlargement, native video/audio players, Add URL to chat, and Download.
• Kept the exact-parameter editor internally, but its entry is intentionally hidden in this release while the chat-first experience is finalized.
• Removed routine payment prompts from the UI. A configured Modellix Key already establishes the expected usage model; consumption and details remain available in Modellix.

Capabilities

Requirements

• DeepSeek Harness 0.1.2-alpha.4 (latest supported); 0.1.1-rc.2 remains supported
• Published-package runtime: Node.js ^22.19.0 || >=24.0.0
• Source development and release verification: Node.js 24.18.1, pnpm 11.24.0
• A valid Modellix API Key

The plugin contains its own Harness integration. It does not install or invoke modellix-cli at runtime.

Installation

Install the package in the target Web profile, inspect the merged configuration, and start or restart that profile:

dsh plugin --profile web add dsh-modellix
dsh --profile web --dump-config
dsh --profile web

--dump-config should contain the dsh-modellix Bundle layer and a plugin row with id modellix. Replace web if you use another profile.

To install a trusted local build:

pnpm install --frozen-lockfile
pnpm run verify:release:static
pnpm pack
dsh plugin --profile web add ./dsh-modellix-0.2.1.tgz

See local usage for the complete Windows-first workflow.

Configure the API Key

On first use, enter the Key in Connect Modellix, keep the required services enabled, and select Save and enable. The saved Credential is write-only: the Client receives only configuration status and source, never the stored Key.

Alternatively, provide MODELLIX_API_KEY to the Harness launch environment. Environment Credentials are read-only in the UI and require a Harness restart after replacement.

Do not put a real Key in a repository, command argument, URL, browser storage, log, screenshot, HAR, recording, or test snapshot.

Chat-first media workflow

Describe the outcome, not the tool sequence. For example:

Create a polished 16:9 architectural hero image of a glass botanical research pavilion floating above a dawn cloud sea, with restrained lapis-blue and warm-gold tones, realistic premium materials, no people, no text, and no watermark.

The Agent can then:

1. Search the live media catalog if a compatible model is not already known.
2. Read that model's live API Schema and use only published fields and values.
3. Reuse the latest relevant result URL for edit, image-to-video, or video-to-video requests instead of silently switching back to text-to-media.
4. Upload a local or conversation file when the selected Schema requires a public media URL.
5. Submit the generation exactly once. An unknown submission outcome is never replayed automatically.
6. Check once in the Agent turn. A background watcher then updates the existing card to its terminal state without another Agent tool call.

While a task is nonterminal, immutable assistant prose only confirms acceptance and points to the live result card and Modellix Design. It does not leave behind a stale “running” sentence.

Result behavior

• Running and failed: show a concise header/status only; Preview and JSON are not offered without a successful result.
• Succeeded: show Preview and JSON tabs. Images enlarge in a focus-managed dialog; video and audio use native players.
• One task, one card: a generate card takes ownership of its task; the corresponding get_result call does not render a second copy.
• Session isolation: the right panel shows only tasks owned by the current Harness conversation. Legacy records without a session owner are not injected into new sessions.
• Actions: Add URL to chat appends the selected resource URL to the composer; Download opens the upstream resource safely.
• Expiry: the displayed expiry follows the upstream result. If none is provided, the plugin applies a seven-day local display limit; it does not extend the upstream URL or keep a permanent media copy.

Modellix Design panel

The Modellix Design button sits at the far right of the conversation header, beside Session log. It opens a split-panel experience on large screens and a full-width overlay on narrow screens.

• The entire Results list is expanded by default and can be collapsed.
• Every result card is expanded by default and can be collapsed by selecting its header.
• The close button returns focus to the launcher.
• The advanced exact-parameter editor remains implemented but its entry is hidden in 0.2.1; ordinary users work through chat.
• At 560 px and below, the panel uses the available viewport width. At 360 px and below, the launcher becomes a reachable compact control.

LLM models

When LLM is enabled, the plugin reads the live Modellix catalog and adds those models to the Harness model selector. It does not fabricate fallback entries when the catalog is unavailable. Refresh status and model count are available in Modellix settings.

Automatic Web Search and Fetch

Users do not need to name a tool. For current, changing, external, or source-verification questions, the Agent is instructed to use modellix_web_search. When the user provides a public URL or a search result needs full-page reading, it uses modellix_web_fetch.

While Web is enabled, the plugin registers only the two explicit Modellix tools and asks the Agent to prefer them for matching work. Disabling Web unregisters those tools and removes the Modellix Web routing context; Harness-native web_search / web_fetch tools and their configured providers remain untouched and can be selected automatically. Failed or unknown Modellix Web requests are not repeated automatically.

Settings and recovery

The Modellix settings section shows:

• Credential configured/verification status and source;
• replacement and removal actions for a writable local Credential;
• independent Design, LLM, and Web switches; each switch changes only its own Modellix capability;
• live LLM catalog health, count, refresh time, and manual refresh.

Only HTTP 401 marks a Credential invalid. HTTP 402, 429, network failures, and 5xx retain their own recovery states. Concurrent 401 responses are coalesced into one Credential dialog.

Accessibility and responsive behavior

• Dialogs manage initial focus, Tab/Shift+Tab wrapping, background inertness, Escape behavior where allowed, and focus restoration.
• Result tabs support standard keyboard navigation; result/status changes use polite live regions.
• State is conveyed with text in addition to color.
• Layouts were checked at 320, 560, 768, and 1440 CSS px, at 200% text scaling, in light and dark themes, forced colors, coarse pointer, and reduced motion.
• Narrow layouts preserve every result action without horizontal page overflow.

Documentation

• Complete English user guide
• English release and acceptance checklist
• 完整中文用户指南
• 中文发布与验收清单
• Local source usage

The repository contains six English and six Chinese 1920×1080 screenshots captured from separate real-language sessions. They cover settings, chat image generation, the result drawer, media players, live LLM models, and automatic Search/Fetch. No capture contains a Key, request header, Network/HAR data, Credential file, or browser storage.

Development and release verification

pnpm install --frozen-lockfile
pnpm run check
pnpm run verify:pack
pnpm run verify:fresh-install
pnpm run verify:node22-install
pnpm run verify:release:static

pnpm run verify:release additionally requires fresh Secret-free browser and real API/Agent evidence bound to the exact package version and 40-character Git commit. The complete procedure is documented in RELEASE_CHECKLIST.md.

Current limitations

• The advanced exact-parameter editor entry is intentionally hidden in this release.
• Upstream generation cancellation is not exposed.
• Result history stores task metadata and upstream URLs, not permanent media copies.
• A complex unsupported Schema blocks submission instead of guessing or dropping constraints.
• The Modellix LLM catalog has no fabricated offline fallback.

Uninstallation

Remove a locally stored Credential from Modellix settings first, or revoke an environment Credential in its external secret manager. Then remove the plugin and restart the profile:

dsh plugin --profile web remove dsh-modellix
dsh --profile web --dump-config
dsh --profile web

Uninstallation does not remove upstream Modellix tasks, external environment variables, or all Harness profile data.

References

• Modellix getting started
• Modellix LLM overview
• Modellix model catalog
• DeepSeek Harness

License

See LICENSE.