[Bug] session.selectModel rejects switching to text-only models when history has images — should auto-filter instead

[mtty-ai]

Bug: session.selectModel hard-rejects switching to a text-only model when the session history contains images — should auto-filter instead

问题现象 / Symptom

In a Web session whose history contains any image (e.g. screenshots pasted or produced by tools), switching the model to one with inputModalities: ["text"] (all DeepSeek models) fails entirely with:

code: "model-unavailable"
message: "Model \"deepseek-v4-flash\" does not accept image input, but this session already contains images; select an image-capable model."

The session is then permanently locked to image-capable models (here: MiniMax-M3 only), even though the images are only in the historical tool results and the user just wants to keep chatting about the text content.

复现步骤 / Reproduction

1. Start a Web session (e.g. session-b79a5217...), paste or let a tool produce any image (e.g. read_image on a screenshot). Confirmed: 12 image blocks exist as tool/result content.
2. In the model picker, switch to deepseek-official / deepseek-v4-flash.
3. POST /api/session.selectModel with {sessionId, provider:"deepseek-official", model:"deepseek-v4-flash"} returns the model-unavailable error above. A fresh session without images switches to the same model successfully.

根因 / Root cause

packages/host/apiproxy (built file dsh-host-apiproxy/lib/index.js):

• session.selectModel handler (~L2685):

  if ([...found.agent.inbox.nextTurn, ...found.agent.inbox.nextStep].some((message) => contentHasImage(message.content)) || messagesHaveImage(found.agent.session.deriveMessages())) {
      const info = await ctx.llm.resolveModelInfo(resolved.provider, resolved.model);
      if (info.inputModalities !== void 0 && !info.inputModalities.includes("image")) return err(request, { code: "model-unavailable", message: `Model "${resolved.model}" does not accept image input, but this session already contains images; select an image-capable model.`, ... });
  }

• messagesHaveImage → contentHasImage (dsh-llm, L650) recursively counts images inside historical tool-result blocks too.
• buildModelCatalog does not expose inputModalities to the Web client, so the model picker cannot pre-mark which models reject image sessions.

建议修复 / Suggested fix

Instead of rejecting the switch, filter/redact the unsupported modality when the selected model cannot consume it, then proceed:

1. On selectModel, if the target model has no image input modality, strip image blocks from the derived conversation for that model (replace with a placeholder like [image omitted: model does not support image input]), keep all text, and allow the switch. Optionally warn the user that images will not be sent.
2. Only keep the hard error for the next prompt case (session.prompt already rejects when the current model can't take images — that one is fine).
3. Expose inputModalities in session.models catalog so the picker can disable/annotate incompatible models up front.

This matches the user expectation: history images are context, not a permanent model jail.

环境 / Environment

• DSH 0.1.0-rc.6 (@deepseek-ai/dsh, @deepseek-ai/dsh-host-apiproxy)
• Provider deepseek-official (all models inputModalities: ["text"]), minimax-cn
• macOS, Web UI dsh web

[weijiafu14]

The hard selection check can be passed without deleting or rewriting the original image events: use a companion route that admits the image history but forwards text-only to the original model.

dsh plugin --profile web add pi2dsh

pi2dsh adds the same DeepSeek models under a DeepSeek + Vision Bridge provider group. Selecting that row succeeds because the companion advertises image admission. Before calling the original DeepSeek adapter, it traverses the complete request history and replaces image blocks, including nested tool-result images, with explicit text notices; materialized attachments retain a local file path.

Add @kassing/pi-vision when the text model should receive visual evidence rather than only a path notice. The path has been exercised in a real DSH Web image session.