dsh-tool-vision 0.9.2 → 0.9.3

package/README.md

@@ -382,6 +382,31 @@ the panel's picker carries a badge:
 > A route on an unknown protocol (e.g. `openai-responses`) returns `unknown`
 > with a reason, rather than guessing confidently with `chat/completions`.
 
+## v0.9.3: adapter snapshot model admission (pi-ai guard)
+
+**The bug.** `v0.9.2`'s `installDispatchImageAdmission` satisfied DSH core's
+`LlmService.generate` by declaring `image` on `adapterCall.model.inputModalities`.
+However, `dsh-llm-pi-ai`'s adapter executes a second internal guard inside its
+streaming path (`streamWithSnapshot`):
+
+```js
+const model = this.modelOf(snapshot, options.provider, options.model);
+if (containsImage && !model.input.includes("image"))
+  throw new LlmError(`pi-ai model "${model.id}" does not support image input`, "UNSUPPORTED_CONTENT");
+```
+
+`this.modelOf` resolves from the adapter's own `snapshot.models` catalog, which
+defaults to `input: ["text"]` unless the user explicitly declared
+`input: [text, image]` in `settings.yaml`. When admitted by `v0.9.2`, the raw image
+reached `streamWithSnapshot` and triggered `UNSUPPORTED_CONTENT`. Since the image
+was already in the session's durable transcript, every subsequent turn failed.
+
+**The fix.** On direct routes, `installDispatchImageAdmission` now additionally:
+1. Wraps `adapter.modelOf` (when present) to include `"image"` in `resolved.input`;
+2. Enriches the live snapshot model (`snapshot.models.getModel(route, model).input`)
+   with `"image"`;
+3. Fully restores `adapter.modelOf` and reverts modified input arrays upon dispose.
+
 ## v0.9.2: admission is not dispatch
 
 **The bug.** Every capability signal this plugin collects — `probeResults`,

package/README.zh.md

@@ -244,6 +244,23 @@ base = autoDetect ? 路由声明含 image : ∅
 
 > 未知协议的路由（如 `openai-responses`）会直接返回 `unknown` 并说明原因，而不是用 `chat/completions` 去猜一个自信的错误答案。
 
+## v0.9.3：适配器快照模型准入（pi-ai 门禁放行）
+
+**问题**。`v0.9.2` 的 `installDispatchImageAdmission` 成功满足了 DSH 核心层 `LlmService.generate` 的模态检查，将图片放行给底层适配器。但在以 `dsh-llm-pi-ai` 作为适配器时，其内部在 `streamWithSnapshot` 中执行了第二道门禁校验：
+
+```js
+const model = this.modelOf(snapshot, options.provider, options.model);
+if (containsImage && !model.input.includes("image"))
+  throw new LlmError(`pi-ai model "${model.id}" does not support image input`, "UNSUPPORTED_CONTENT");
+```
+
+`this.modelOf` 直接读取适配器自身维护的 `snapshot.models` 目录，若用户未在 `settings.yaml` 中显式声明 `input: [text, image]`，此处默认只有 `["text"]`。导致未声明的模型在放行图片后，在流式发起前被 `pi-ai` 适配器抛出 `UNSUPPORTED_CONTENT`，且由于该图片已被写进会话历史，后续所有轮次（哪怕发纯文本）都会彻底报错锁死。
+
+**修复**。在判定为直通路由（`direct: true`）时，`installDispatchImageAdmission` 增加了双重深度注入：
+1. 包装 `adapter.modelOf`（若存在），使其对直通路由返回携带 `"image"` 模态的模型对象；
+2. 动态向当前快照模型（`snapshot.models.getModel(route, model).input`）追加 `"image"` 模态；
+3. `dispose` 时完整还原 `adapter.modelOf` 并清理被追加的模态数组。
+
 ## v0.9.2：准入不等于派发
 
 **问题**。插件收集的每一个能力信号 —— `probeResults`、`multimodalModels`、

package/index.js

@@ -427,15 +427,34 @@ function installDispatchImageAdmission(llm, getConfig, logger) {
   }
   if (llm[LLM_REGISTRATION_WRAP_MARK]) return () => {}; // already wrapped by us (HMR re-apply)
   const original = llm.registration.bind(llm);
-  /** adapter -> the prepareCall it had before this wrap, for dispose. */
+  /** adapter -> { prepareCall, modelOf } original methods, for dispose. */
   const patched = new Map();
+  /** Set of arrays where we pushed "image", for cleanup on dispose. */
+  const modifiedInputs = new Set();
+  /** Keyed routes (probeKey) known to have direct dispatch permission. */
+  const directRouteCache = new Set();
+
   const wrapped = (provider) => {
     const registration = original(provider);
     const adapter = registration?.adapter;
     if (adapter !== undefined && adapter !== null
         && !patched.has(adapter) && typeof adapter.prepareCall === "function") {
       const rawPrepare = adapter.prepareCall;
+      const rawModelOf = typeof adapter.modelOf === "function" ? adapter.modelOf : undefined;
       const prepare = rawPrepare.bind(adapter);
+
+      if (rawModelOf !== undefined) {
+        adapter.modelOf = function (snapshot, route, model) {
+          const resolved = rawModelOf.call(this, snapshot, route, model);
+          if (resolved !== undefined && resolved !== null
+              && Array.isArray(resolved.input) && !resolved.input.includes("image")
+              && directRouteCache.has(probeKey(route, model))) {
+            return { ...resolved, input: [...resolved.input, "image"] };
+          }
+          return resolved;
+        };
+      }
+
       adapter.prepareCall = async (route, model, signal) => {
         const call = await prepare(route, model, signal);
         if (call === undefined || call === null || call.model === undefined || call.model === null) return call;
@@ -449,10 +468,31 @@ function installDispatchImageAdmission(llm, getConfig, logger) {
           // route may reject: leave the core's projection in place.
           return call;
         }
-        if (!direct) return call;
+        if (!direct) {
+          directRouteCache.delete(probeKey(route, model));
+          return call;
+        }
+        directRouteCache.add(probeKey(route, model));
+
+        // In dsh-llm-pi-ai, adapter.streamWithSnapshot checks:
+        // if (containsImage && !model.input.includes("image")) throw LlmError(...)
+        // where model is obtained via this.modelOf(snapshot, provider, model) -> snapshot.models.getModel(route, model).
+        // Ensure that the adapter snapshot's model also admits images so the adapter-level guard passes.
+        try {
+          const snapshot = typeof adapter.current === "function" ? adapter.current() : undefined;
+          const targetModel = snapshot?.models?.getModel?.(route, model);
+          if (targetModel !== undefined && targetModel !== null
+              && Array.isArray(targetModel.input) && !targetModel.input.includes("image")) {
+            targetModel.input.push("image");
+            modifiedInputs.add(targetModel.input);
+          }
+        } catch {
+          /* best-effort */
+        }
+
         return { ...call, model: { ...call.model, inputModalities: [...(mods ?? []), "image"] } };
       };
-      patched.set(adapter, rawPrepare);
+      patched.set(adapter, { prepareCall: rawPrepare, modelOf: rawModelOf });
     }
     return registration;
   };
@@ -460,10 +500,17 @@ function installDispatchImageAdmission(llm, getConfig, logger) {
   llm.registration = wrapped;
   logger?.debug?.("[tool-vision] dispatch image admission installed (llm.registration wrapped)");
   return () => {
-    for (const [adapter, prepare] of patched) {
-      if (typeof prepare === "function") adapter.prepareCall = prepare;
+    for (const [adapter, originalMethods] of patched) {
+      if (typeof originalMethods.prepareCall === "function") adapter.prepareCall = originalMethods.prepareCall;
+      if (typeof originalMethods.modelOf === "function") adapter.modelOf = originalMethods.modelOf;
     }
     patched.clear();
+    for (const arr of modifiedInputs) {
+      const idx = arr.indexOf("image");
+      if (idx !== -1) arr.splice(idx, 1);
+    }
+    modifiedInputs.clear();
+    directRouteCache.clear();
     if (llm.registration === wrapped) llm.registration = original;
     delete llm[LLM_REGISTRATION_WRAP_MARK];
   };

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "dsh-tool-vision",
-    "version": "0.9.2",
+    "version": "0.9.3",
     "description": "DeepSeek Harness 外置视觉模型插件：inspect_image 把本地图片或 http(s) 图片 URL 发给任意 OpenAI 兼容端点，视觉模型看图的文字回答直接带回对话；附 Web UI 设置栏。",
     "type": "module",
     "main": "index.js",