Complete Milestone 2: OpenAI analysis, image QA, and scene editing
- OpenAI adapter (gpt-5.5 structured outputs for analysis/QA,
gpt-image-2 images.edit for scene editing) with lazy SDK import,
env-only credentials, and an injectable client for offline tests
- Capability-split provider protocols; mock provider untouched
- Shared retry service (transient-only classification incl. SDK
timeout/connection error names, bounded backoff) and cost service
(verified price table, per-run CostTracker, $10 default ceiling)
- Provider-dropdown nodes: AFS_AnalyzeProduct,
AFS_EvaluateImageCandidates, AFS_SceneEdit (12 nodes total)
- Backward-compatible cost fields on contracts; SceneEditRecord model
- Live analysis UI workflow; validator now flags unknown custom nodes
- Gated paid smoke tests (AIFS_ENABLE_PAID_TESTS=1): real gpt-5.5
analysis $0.0084/call, gpt-image-2 edit $0.2191/image
- Verified on ComfyUI 0.28.2 / Python 3.13.12 / torch 2.10.0 (MPS);
52 free tests pass, all workflows validate
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
