Treat prompts and source media as user-owned inputs, keep MediaHub credentials/config/execution in one reusable middle block, and reserve outputs for Preview/Save nodes.
Constraint: Samples must show users exactly which block to copy into an existing workflow.
Rejected: Classifying API Key and execution nodes as inputs or outputs | exposes internal mechanics and weakens sample readability.
Confidence: high
Scope-risk: narrow
Directive: Keep sample responsibilities separated and all connections flowing left to right.
Tested: 124 pytest tests; JavaScript syntax; Python compilation; JSON validation; live ComfyUI accepted all eight prompt structures.
Not-tested: Manual visual import of all eight revised examples.
Separate user-facing layout from API fixtures, keep provider menus scalable, and lock media-family boundaries before more integrations are delegated.
Constraint: User workflows must remain readable as providers and media families grow.
Rejected: Mechanical row-major workflow layout | valid but obscures data flow and wastes space.
Confidence: high
Scope-risk: narrow
Directive: Add new providers under MediaHub/Provider and add new media families as independent vertical slices.
Tested: 123 pytest tests; JavaScript syntax; Python compilation; JSON validation; restarted ComfyUI object_info; all eight prompt smoke workflows.
Not-tested: Manual visual import of all eight updated user workflows.
Replace provider-switching task nodes with provider-specific config nodes and stable generic image/video execution contracts. Add Agnes Image 2.1 and Video 2.0 using behavior verified against the live API.
Constraint: Old workflows are intentionally unsupported; Agnes Image 2.0 remains documentation-only.
Rejected: Single task nodes with provider dropdowns | provider-specific union schemas create serialization risk and cross-provider coupling.
Confidence: high
Scope-risk: broad
Directive: New providers must register isolated operation adapters and reuse generic execution nodes without modifying existing provider packages.
Tested: 120 pytest tests; Python compileall; JavaScript syntax; 16 workflow JSON validations; restarted ComfyUI smoke; live Agnes Image 2.1 T2I/I2I; live Agnes Video 2.0 T2V completion and I2V Data URI submission.
Not-tested: User-driven visual inspection of every imported UI workflow and full completed Agnes I2V download.
T2I: MediaHub Text to Image with all documented MiniMax CN parameters.
I2I: MediaHub Image to Image with character reference support.
T2V: Three-node async pipeline:
- MediaHubT2VSubmit: validate params and submit task
- MediaHubT2VQueryStatus: poll until done (status indicator)
- MediaHubT2VRetrieveVideo: download and output VIDEO type
Key design decisions:
- T2V uses composable nodes instead of a single blocking node
- RetrieveVideo outputs standard VIDEO type (InputImpl.VideoFromFile)
connecting directly to ComfyUI built-in SaveVideo with preview
- Provider flows from Submit to downstream nodes via forceInput
- All model/resolution/duration cross-constraints are validated
- Added video development lessons to AGENTS.md
98 tests passed.
- Implemented MediaHubTextToImage and MediaHubAPIKey nodes for generating images using MiniMax CN API.
- Added frontend JavaScript for dynamic widget updates and API key visibility toggle.
- Created core helpers for image processing and API interactions in mediahub module.
- Developed tests for image tensor conversion and MiniMax CN API interactions.
- Included smoke tests for validating API workflow and node schema alignment.
- Added example workflow for testing and demonstration purposes.
Van/攻城狮鹿鸣君