feat:Enable DeepSeek and other text-only input models to perform image recognition and search functions.

[guigeng]

When using major models like DeepSeek, which only support text input, it’s possible to set up a fallback option or configure the model to handle visual inputs within the vision-related settings.

[lidge-jun]

Implemented on dev in 755fac5.

opencodex already ships the general mechanism this issue asks for — a vision sidecar: when a request routed to a text-only model carries images, the proxy describes each image with a vision-capable model (via the ChatGPT forward provider, visionSidecar.model, default gpt-5.2) and substitutes the description text in place of the image before the main call. The model catalog also advertises image input on behalf of those models so Codex allows attachments, and the web-search sidecar reuses the same flag.

What was missing was coverage for the direct DeepSeek provider (deepseek/*): only opencode-go's DeepSeek models were classified as text-only. Since every DeepSeek API model rejects image input upstream, all four registered models are now marked noVisionModels:

• deepseek-chat, deepseek-reasoner (until their upstream deprecation on 2026-07-24)
• deepseek-v4-pro, deepseek-v4-flash

This applies automatically to existing saved configs (registry classification is merged at request time), so after updating, attach an image to a deepseek/* conversation and the model will receive a detailed text description of it instead of the request failing.

Requirements: a ChatGPT-forwarded (OAuth) provider must be configured — the vision model runs through it. Without one, images are stripped fail-closed rather than breaking the upstream call. You can tune the describer via visionSidecar: { model, timeoutMs, enabled } in ~/.opencodex/config.json.

Verified with an end-to-end test that observes the sidecar firing (caption substituted into the upstream request, image bytes never forwarded to DeepSeek) plus a regression test that vision-capable models keep native image passthrough. Full suite: 2052 pass / 0 fail.

Ships in the next release.

[lidge-jun]

Closing — implemented on dev in 755fac5. Direct DeepSeek provider models are now classified as noVisionModels; the existing vision sidecar describes attached images on their behalf.