TrueGrowth/openspec/changes/add-omni-intelligence-hub/design.md

Context

The current workbench shell owns the main navigation in WorkbenchShell.tsx, the app-center cards in app-definitions.tsx, and route rendering inside renderRoute(). Model selection is already implemented in Drawnix with runtime model discovery through useSelectableModels('text') and controlled model refs in ModelSelector.

The user-provided plan defines "万象智枢 / Omni Intelligence Hub" as the upstream decision and content-sourcing center. Its first release should make the feature discoverable and actionable without requiring the whole long-term platform to be complete.

Agent-Reach is an external project focused on agent access to external sources. It should be treated as a local capability provider rather than bundled frontend logic.

Goals

• Add a first-class workbench entry for Omni Intelligence Hub.
• Let users start research from a question, topic, product name, webpage URL, video URL, or GitHub repository.
• Reuse existing runtime-discovered text model selection instead of creating a separate model registry.
• Bridge to Agent-Reach for retrieval/source access and expose health/diagnostics in the UI.
• Keep the first implementation useful when Agent-Reach is unavailable by allowing draft setup, mode selection, model selection, and diagnostic guidance.

Non-Goals

• Do not implement every long-term submodule in full depth in the first release.
• Do not vendor or copy Agent-Reach source code into the web app.
• Do not replace existing image, video, audio, avatar, canvas, or asset workflows.
• Do not introduce a new model configuration surface separate from the existing provider/runtime model system.

Architecture

Workbench UI

• Add /intelligence as the top-level route.
• Add a sidebar item labeled 智枢 or 万象智枢 using a lucide icon such as Radar, SearchCheck, or BrainCircuit.
• Add an app-center definition with the product-plan description and route /intelligence.
• Implement an OmniIntelligencePage component in the workbench layer. The first screen should be the usable research workspace, not a marketing landing page.

Internal Navigation

Represent the planned submodules as tab/segmented navigation:

• 总览: 智枢首页
• 探索: 全域探索, 内容解析, 专题智研
• 洞察: 趋势雷达, 竞品战情
• 创作: 策源工坊
• 管理: 情报资产, 监控任务

The initial implementation may keep some tabs as structured empty states, but the main input, mode selection, model selector, Agent-Reach status, quick scenarios, and result preview must be functional.

Model Selection

• Use runtime-discovered text models via useSelectableModels('text').
• Reuse the existing controlled ModelSelector where practical.
• Store both modelId and modelRef in the page state.
• Include selected model metadata in Agent-Reach task submission so downstream summarization, verification, and content conversion use the chosen model.

Agent-Reach Bridge

Add a local bridge with these conceptual endpoints:

• GET /local-api/agent-reach/status: returns installed/runnable/auth/source capability state and doctor output when available.
• POST /local-api/agent-reach/research: submits a query, mode, model, source filters, depth, verification, and output options.
• POST /local-api/agent-reach/parse-link: submits a URL for webpage/video/repository parsing.

The bridge should execute Agent-Reach through a local runtime boundary and normalize responses into:

• summary
• keyFindings
• sources
• conflicts
• trendSignals
• contentIdeas
• scripts
• storyboards
• prompts
• rawDiagnostics

If Agent-Reach is missing, the status endpoint should return available: false with setup guidance. The UI should show a clear status panel and keep non-network planning controls usable.

Asset and Workflow Continuity

Omni Intelligence Hub results should be shaped so later changes can save them as intelligence assets, insert markdown reports into canvas, or launch image/video/avatar/script workflows. The first implementation can expose CTA placeholders for these handoffs while preserving the result data model.

Risks

• Agent-Reach command/API shape may change. Keep the bridge isolated and normalize the contract at /local-api/agent-reach/*.
• Authenticated community sources may require user login. Surface per-source capability and avoid treating missing auth as total failure.
• Research tasks can be long-running. If the first release uses synchronous calls, keep timeouts and progress states explicit; prefer adding task queue integration when the bridge proves stable.