| name | repo-intent |
| description | Repo-wide mission, guardrails, and next slices for the image_creator MCP server. |
Repo intent
Mission
Build a thin MCP server that turns image-generation APIs into one stable local artifact contract: prompt in, file path out, with profile-first model choice.
Global constraints
- prefer one MCP tool with a small explicit API over many clever helper tools
- keep provider-specific request and response logic inside adapter modules
- the agent-facing success payload must stay provider-agnostic
- the agent should prefer curated profiles over raw model ids
- the file path on disk is the primary artifact; inline base64 is only an internal transport detail
- avoid orchestration theater, fake docs, and broad speculative scaffolding
Current repo state
- bootstrap docs and skill-native memory are installed
- OpenRouter generate and edit paths are implemented and live-proven in this environment
- default image model target is now
Gemini 3.1 Flash Image Preview
- curated profiles now define when to use
3.1 Flash Image Preview vs 3 Pro Image Preview
- transparent-background generation is now a GPT-image path and should not be attempted on Gemini-family routes
- transparent-background work is further split into balanced (
transparent_bg), fast (transparent_bg_fast), and higher-fidelity (cutout) paths
- deterministic background removal for an existing image is now a separate local tool instead of being forced through
edit_image
- Gemini adapter is implemented, but still lacks live proof here because
GEMINI_API_KEY is not configured
- repo truth must stay in sync with the actual code and verification evidence
Next slices
- keep curated profiles and docs aligned with live provider reality
- close Gemini live proof once
GEMINI_API_KEY is available
- only then decide whether
OpenAIProvider adds enough value to justify another adapter
Update rule
Update this skill when repo-wide intent, provider policy, or the top-level tool contract changes.