Context: kimi-code already ships an experimental feature to assign a specific model for sub-agent tasks (the [secondary]/[secondary_model] config section + SECONDARY_MODEL_ENV, experiment-gated). This is the model a user pins for sub-agent spawns.
Problem: Many coding models are text-only — they cannot inspect images. Web and mobile development regularly requires visual inspection (screenshots, UI states, layout renders, image diffs of a rendered page). Today there is no way for a user to pin a specific, vision-capable model for these visual tasks; they fall back to automatic/current-model selection, which is often a text-only model.
Proposal: add a visual model id slot mirroring the sub-agent model id — e.g. a [visual_model] config section with an env override, gated by an experiment flag, exactly parallel to [secondary_model]/resolveSecondaryModel. When set, visual/image-inspection tasks route to that configured model; when unset, behavior is unchanged (current default).
Expected acceptance: (1) user can configure a dedicated visual model id; (2) image/screenshot inspection tasks resolve to it when set; (3) unset = no behavior change; (4) covered by vitest tests.
Context: kimi-code already ships an experimental feature to assign a specific model for sub-agent tasks (the
[secondary]/[secondary_model]config section +SECONDARY_MODEL_ENV, experiment-gated). This is the model a user pins for sub-agent spawns.Problem: Many coding models are text-only — they cannot inspect images. Web and mobile development regularly requires visual inspection (screenshots, UI states, layout renders, image diffs of a rendered page). Today there is no way for a user to pin a specific, vision-capable model for these visual tasks; they fall back to automatic/current-model selection, which is often a text-only model.
Proposal: add a visual model id slot mirroring the sub-agent model id — e.g. a
[visual_model]config section with an env override, gated by an experiment flag, exactly parallel to[secondary_model]/resolveSecondaryModel. When set, visual/image-inspection tasks route to that configured model; when unset, behavior is unchanged (current default).Expected acceptance: (1) user can configure a dedicated visual model id; (2) image/screenshot inspection tasks resolve to it when set; (3) unset = no behavior change; (4) covered by vitest tests.