Idea: 纯文本模型贴图不再被拒——受理时转本地路径交给视觉技能 / accept pasted images with text-only models (fix ready on fork)

[deveuper]

DeepSeek Harness: pasted images are rejected before any vision skill can run

Standalone problem report for the warning "当前模型不支持图片输入 / This model does not support image input" shown when pasting an image into the web chat while the routed model is text-only, plus the source fix prepared on the fork deveuper/deepseek-harness (branch fix/text-only-model-image-admission).

这是一份把"纯文本模型贴图即弹警告"现象单独提出的问题报告，并附上在 fork 分支 fix/text-only-model-image-admission 上已完成的源码修复说明。

1. Problem / 现象

• What the user does / 用户操作: paste or drag an image into the DeepSeek Harness web chat input box, with a text-only main model selected (DeepSeek, Kimi, GLM text models, etc.).
• What happens now / 目前的效果: sending fails immediately. The UI banner shows 当前模型不支持图片，请切换支持图片的模型 (attachment-error, reason MODEL_DOES_NOT_SUPPORT_IMAGES). The image is refused at the door and never reaches the conversation, so no skill can ever read it.
• Root cause / 根因: the prompt admission code in packages/host/apiproxy/src/api-proxy.ts rejects any prompt containing image blocks when the routed model's inputModalities does not include image. This check runs before skills can act.
• Why this is a source bug, not a skill problem / 为什么是源码问题而非技能问题: the block sits in the harness admission layer, ahead of every skill. Any script-based vision skill — visionDS, vision-hub, or any other vision/OCR helper — is affected equally, because the image is rejected before the skill can receive it. The advertised workflow "keep the text-only main model, let another vision model look at the image" is therefore impossible out of the box.

2. The fix / 修复内容

Keep the main model text-only (unchanged), persist the uploaded images durably, and hand the model local file paths as text. The model then feeds those paths to a vision skill/script through its shell tools.

• packages/attachment/attachment/src/index.ts — add AttachmentStore.imagePath(ref) to the seam: resolve a stored image to its absolute local filesystem path.
• packages/attachment/attachment-local/src/index.ts / src/store.ts — implement imagePath and export attachmentObjectPath(root, ref).
• packages/host/apiproxy/src/api-proxy.ts — prompt admission: when the routed model declares inputModalities without image, persist the uploads exactly as before (durablePromptContent), then replace the image blocks with text blocks carrying each stored image's local path, its media type, and a hint to use the vision-ds skill. The durable user message stays text-only, so the text-only LLM adapter never sees an image block and the "model-visible ⟺ logged" invariant holds.
• Tests — new coverage in packages/host/apiproxy/tests/api-proxy-models.spec.ts (image → path-text conversion for a text-only model) and packages/attachment/attachment-local/tests/store.spec.ts (attachmentObjectPath); the new abstract member is implemented in the existing test/tooling attachment stores (packages/fs/tool-fs/tests/read-image.spec.ts, packages/host/apiproxy/tests/api-proxy-projections.spec.ts, packages/llm/llm-pi-ai/tests/adapter.spec.ts, packages/llm/llm-pi-ai/tests/provider-apis.e2e.ts, scripts/gen-tool-catalog.ts, scripts/test-invariants.ts).

3. Why remove the warning / 为什么去掉这条警告

• With the warning, the image is rejected at admission, so no vision skill can ever work with a text-only main model. Removing the rejection is the only way to enable the "main model stays, vision goes elsewhere" workflow.
• The change is additive and safe: the main model still receives only text; images are stored in the durable attachment store exactly as the vision-capable path does; models that DO declare image input keep the previous behavior (real image blocks) unchanged.
• The client-side copy mapping for MODEL_DOES_NOT_SUPPORT_IMAGES stays in place for any other surface that still emits it; only the ordinary session admission path stops emitting it. The separate subagent-continuation guard (SUBAGENT_IMAGE_UNSUPPORTED) is untouched by this change.

4. Verification / 验证

• tsc -b tsconfig.host.json — clean.
• Affected vitest suites — 104 passed (api-proxy-models, api-proxy-projections, attachment-local/store, tool-fs/read-image, llm-pi-ai/adapter).
• End-to-end on this machine — a pasted-style PNG is persisted under ~/.dsh/attachments/v1, the text-only model receives the local path, the vision-ds skill script reads it, and a real MiMo vision API call returned a correct description of the image (620 tokens). Windows built-in OCR fallback also verified.

5. Links / 相关链接

• Fix branch on the fork: https://github.com/deveuper/deepseek-harness/tree/fix/text-only-model-image-admission (commit 9f56c37e48)
• Ready-to-apply diff against official master: master...deveuper:fix/text-only-model-image-admission
• The skill that consumes the paths: https://github.com/deveuper/visionDS
• Official discussion (this report): Idea: 纯文本模型贴图不再被拒——受理时转本地路径交给视觉技能 / accept pasted images with text-only models (fix ready on fork) #427
• Official discussion (visionDS announcement): Show and tell: visionDS —— DeepSeek Harness 多提供商视觉/OCR 技能（可安装组合包）/ a multi-provider vision/OCR skill for DeepSeek Harness #384