Sight. Nodes can now see images - the art-director pattern - and the brand-locked asset suite runs end to end: confirm or create a logo, generate eight exact-spec assets in parallel with the logo as a reference image, and score brand consistency with a vision judge (measured 8/10 overall in ~19 seconds for $0.38). Also fixes a provider defect shipped in 0.5.0. No breaking API changes.
Fixed:
- OpenAIProvider was broken against current OpenAI models (legacy max_tokens parameter rejected by GPT-5.x); now sends max_completion_tokens
Added:
- Vision input: ChatMessage.attachments mapped to each provider's native multimodal format; Node(attach_dep_artifacts=True) feeds a node its dependencies' generated images as pixels (examples/11_vision_judge.py - first live run caught a spelling error in a generated ad)
- Brand-locked asset suite benchmark: confirm-or-create logo, reference-image propagation across concurrent swarms, exact-spec finishing, vision-judged consistency scoring
- Repo Doctor MVP (skills/repo-doctor): offline-first release-readiness auditor built on smythe
Full details in CHANGELOG.md.