Vision / Prism — one image-grounded conversation #49

Planning representation replaced — 6 November 2026. The accepted Image-grounded conversation outcome is retained in #62. Reopen one bounded task when selected. This closure does not mean the capability is complete or cancelled. The original scope and evidence below remain preserved; current selected work starts at #52.

Attach one image, ask a question and receive a source-grounded response from one admitted local visual profile.

Starting point: Vision admission carries modality metadata; image custody, forwarding and executable visual calls are missing.

First working slice

• Admit one image into Artifacts — owned storage and quarantined results #23 custody with owner, digest, classification and retention; bound format, dimensions, decompression and metadata handling.
• Register one exact image-capable chat or visual-analysis profile and prove its Connector request/response dialect; an image input declaration does not imply image generation.
• Carry artifact references and required modality from a minimal Bridge attachment through Intent, Run and Graph to the typed call; retain transform lineage and attributed output.
• Verify cold activation, text-only capability refusal, malformed media, cancellation and deletion; never silently drop the image or switch to a remote provider.

Connections: Uses #23, #42, #5/#8 and #6/#7; extends #11 with one bounded attachment journey. Shared runtime acceptance belongs to #50, before #16/#17 policy expansion.

Later: Multiple images, OCR/document paths, video, generation and live perception, each with its own typed contract.

References: ADR 36 · ADR 21 · ADR 22.

Contribution entry: Contribute image validation/attachment or the exact visual connector dialect.
Open decision: Select one image-capable profile and input-size/transform policy.
Verification: Malformed/media fixtures; one named local visual profile for closure. Shared contribution and touched-docs rules: #52.