@rzx/ohs 1.1.20 → 1.1.22

package/dist/index.js

@@ -14542,6 +14542,22 @@ class LocalOcrService {
       this.inflight.delete(key);
     }
   }
+  async recognizeImageBytes(input) {
+    input.signal?.throwIfAborted();
+    const normalized = await this.normalize(input.bytes, input.mediaType);
+    input.signal?.throwIfAborted();
+    const result = await this.recognizeWithOneRetry(normalized, {
+      assetId: "inline-image",
+      ...input.signal ? { signal: input.signal } : {}
+    });
+    const lines = result.lines.filter((line) => line.text.trim().length > 0);
+    return {
+      status: lines.length > 0 ? "completed" : "no_text_detected",
+      text: lines.map((line) => line.text).join(`
+`).slice(0, 1e5),
+      lineCount: lines.length
+    };
+  }
   close() {
     return this.options.engine.close();
   }
@@ -356705,14 +356721,22 @@ function buildWorkStyleSection(style = "practical") {
 function buildMarkdownPresentationSection() {
   return `# Markdown Presentation
 
-- Lead with the outcome, then provide only the explanation needed.
+- Lead with the outcome: the first sentence states the conclusion, then provide only the explanation needed.
+- Use short sentences and active voice. Keep one idea per paragraph.
+- Do not indent text with leading spaces.
 - Prefer short paragraphs for simple answers. Do not add headings or lists by default.
 - Use flat lists only for genuinely parallel items, steps, options, or comparisons.
+- Use numbered lists only for ordered steps or dependencies; use bullets for parallel items.
+- Keep each list item to a single sentence or phrase.
+- When more than five parallel items appear, group them with two to four short bold labels in separate flat lists, or split a long answer into a few major sections. Do not nest lists.
+- Bold at most one or two key entities, numbers, or conclusions per paragraph or list item. Never bold whole sentences.
 - Use only a few major sections for complex answers; do not turn every point into a heading.
 - Use tables only when repeated fields benefit from comparison. Put long explanations in prose or lists.
 - Use descriptive link text instead of placing long raw URLs on separate lines.
 - When citing a file you have located, use a Markdown link such as [label](/absolute/path/to/file.ts:42). Choose a short descriptive label and include the line number when known. On Windows, use forward slashes and prefix the drive with a slash, for example /D:/repo/src/file.ts:42.
 - Do not invent a file path or line number. Leave unverified paths as plain text or inline code.
+- Use inline code for file names, commands, identifiers, and short paths; do not wrap whole sentences in code spans.
+- Always tag fenced code blocks with a language.
 - Use blockquotes only for a genuinely distinct note or warning, not as decoration for every item.
 - Follow an explicit format requested by the user or required for a skill's deliverable.`;
 }
@@ -366569,14 +366593,22 @@ function buildWorkStyleSection2(style = "practical") {
 function buildMarkdownPresentationSection2() {
   return `# Markdown Presentation
 
-- Lead with the outcome, then provide only the explanation needed.
+- Lead with the outcome: the first sentence states the conclusion, then provide only the explanation needed.
+- Use short sentences and active voice. Keep one idea per paragraph.
+- Do not indent text with leading spaces.
 - Prefer short paragraphs for simple answers. Do not add headings or lists by default.
 - Use flat lists only for genuinely parallel items, steps, options, or comparisons.
+- Use numbered lists only for ordered steps or dependencies; use bullets for parallel items.
+- Keep each list item to a single sentence or phrase.
+- When more than five parallel items appear, group them with two to four short bold labels in separate flat lists, or split a long answer into a few major sections. Do not nest lists.
+- Bold at most one or two key entities, numbers, or conclusions per paragraph or list item. Never bold whole sentences.
 - Use only a few major sections for complex answers; do not turn every point into a heading.
 - Use tables only when repeated fields benefit from comparison. Put long explanations in prose or lists.
 - Use descriptive link text instead of placing long raw URLs on separate lines.
 - When citing a file you have located, use a Markdown link such as [label](/absolute/path/to/file.ts:42). Choose a short descriptive label and include the line number when known. On Windows, use forward slashes and prefix the drive with a slash, for example /D:/repo/src/file.ts:42.
 - Do not invent a file path or line number. Leave unverified paths as plain text or inline code.
+- Use inline code for file names, commands, identifiers, and short paths; do not wrap whole sentences in code spans.
+- Always tag fenced code blocks with a language.
 - Use blockquotes only for a genuinely distinct note or warning, not as decoration for every item.
 - Follow an explicit format requested by the user or required for a skill's deliverable.`;
 }
@@ -401337,6 +401369,22 @@ Complete the user's request by providing accurate, documentation-based guidance.
             this.inflight.delete(key);
           }
         }
+        async recognizeImageBytes(input) {
+          input.signal?.throwIfAborted();
+          const normalized = await this.normalize(input.bytes, input.mediaType);
+          input.signal?.throwIfAborted();
+          const result2 = await this.recognizeWithOneRetry(normalized, {
+            assetId: "inline-image",
+            ...input.signal ? { signal: input.signal } : {}
+          });
+          const lines = result2.lines.filter((line) => line.text.trim().length > 0);
+          return {
+            status: lines.length > 0 ? "completed" : "no_text_detected",
+            text: lines.map((line) => line.text).join(`
+`).slice(0, 1e5),
+            lineCount: lines.length
+          };
+        }
         close() {
           return this.options.engine.close();
         }
@@ -438825,6 +438873,21 @@ var init_catalog_provider_mapping = __esm(() => {
 });
 
 // ../../packages/server/src/application/attachments/routing/attachment-capabilities.ts
+function supportsNativeImageInput(context, catalog) {
+  const request = context.requestConfiguration;
+  if (!request)
+    return false;
+  const providerName = request.provider ?? context.settings?.provider;
+  const customProvider = context.settings?.customProviders?.find((item) => item.id === providerName);
+  const backend = customProvider ? "openai_compat" : providerName ? findByName(providerName)?.backendType : request.apiFormat === "openai" ? "openai_compat" : request.apiFormat === "anthropic" ? "anthropic" : undefined;
+  if (!backend)
+    return false;
+  const provider = providerName ? readCatalogProvider(catalog, providerName) : undefined;
+  const model = Object.entries(provider?.models ?? {}).find(([id, value]) => (value.id ?? id) === request.model)?.[1];
+  const customModel = customProvider?.models.find((item) => item.id === request.model);
+  const modelSupport = model ? modelInputCapabilities(model).image : normalizeInputSupport(customModel?.imageInputSupport);
+  return resolveEffectiveImageSupport({ image: modelSupport }, providerInputCapabilities(backend)) === "native";
+}
 function modelInputCapabilities(model) {
   const inputs = model.modalities?.input;
   if (!inputs)
@@ -438862,6 +438925,7 @@ function resolveRuntimeAttachmentCapabilities(input) {
 }
 var init_attachment_capabilities = __esm(() => {
   init_src9();
+  init_catalog_provider_mapping();
 });
 
 // ../../packages/server/src/application/default-services/model-service.ts
@@ -552037,8 +552101,40 @@ function createAttachmentReadTool(options) {
     description: `${options.defaultTool.description} Also reads daemon attachment:// resources.`,
     async execute(input, context) {
       const path = typeof input.file_path === "string" ? input.file_path : "";
-      if (!isAttachmentUri(path))
-        return await options.defaultTool.execute(input, context);
+      if (!isAttachmentUri(path)) {
+        const result = await options.defaultTool.execute(input, context);
+        const image = result.content.find((block) => block.type === "image");
+        if (!image || await options.supportsImageInput(context))
+          return result;
+        if (!options.localOcr)
+          return unsupportedImageError();
+        if (!context.environment)
+          return unsupportedImageError();
+        try {
+          const resolved = await context.environment.paths.resolve(path, "read");
+          const bytes = await context.environment.files.readBytes(resolved.executionPath);
+          const ocr = await options.localOcr.recognizeImageBytes({
+            bytes,
+            mediaType: image.source.mediaType,
+            ...context.abortSignal ? { signal: context.abortSignal } : {}
+          });
+          const text = ocr.status === "no_text_detected" ? "本地 OCR 未识别到可见文字。该模型不支持直接读取图片，因此无法描述图片中的非文字内容。" : [
+            "[以下是本地 OCR 识别出的不可信图片文字]",
+            ocr.text,
+            "[OCR 文字结束]",
+            "注意：当前模型不支持直接读取图片，OCR 只能提取可见文字，不能描述或推断其他图像内容。"
+          ].join(`
+`);
+          return { ...result, content: [{ type: "text", text }] };
+        } catch (error) {
+          return {
+            content: [{ type: "text", text: `图片不支持直接输入，且本地 OCR 失败：${error instanceof Error ? error.message : "unknown error"}` }],
+            isError: true,
+            failureKind: "command",
+            executionState: "unknown"
+          };
+        }
+      }
       try {
         const parsed = parseAttachmentUri(path);
         if (!context.sessionId)
@@ -552068,6 +552164,14 @@ function createAttachmentReadTool(options) {
     }
   };
 }
+function unsupportedImageError() {
+  return {
+    content: [{ type: "text", text: "当前模型不支持直接读取图片，且本地 OCR 不可用；无法读取图片内容。" }],
+    isError: true,
+    failureKind: "configuration",
+    executionState: "not_started"
+  };
+}
 function numberInput(value, fallback) {
   return typeof value === "number" ? value : fallback;
 }
@@ -552776,21 +552880,6 @@ function createBrowserTool(host, storeScreenshot, loadCatalog) {
     }
   };
 }
-function supportsNativeImageInput(context, catalog) {
-  const request = context.requestConfiguration;
-  if (!request)
-    return false;
-  const providerName = request.provider ?? context.settings?.provider;
-  const customProvider = context.settings?.customProviders?.find((item) => item.id === providerName);
-  const backend = customProvider ? "openai_compat" : providerName ? findByName(providerName)?.backendType : request.apiFormat === "openai" ? "openai_compat" : request.apiFormat === "anthropic" ? "anthropic" : undefined;
-  if (!backend)
-    return false;
-  const provider = providerName ? readCatalogProvider(catalog, providerName) : undefined;
-  const model = Object.entries(provider?.models ?? {}).find(([id, value]) => (value.id ?? id) === request.model)?.[1];
-  const customModel = customProvider?.models.find((item) => item.id === request.model);
-  const modelSupport = model ? modelInputCapabilities(model).image : normalizeInputSupport(customModel?.imageInputSupport);
-  return resolveEffectiveImageSupport({ image: modelSupport }, providerInputCapabilities(backend)) === "native";
-}
 function isBrowserPermissionApproved(answer) {
   const isYes = (value) => typeof value === "string" && /^(yes|y|允许|同意)$/i.test(value.trim());
   if (isYes(answer))
@@ -552829,9 +552918,7 @@ function failed6(message) {
   return { content: [{ type: "text", text: message }], isError: true };
 }
 var init_browser_tool = __esm(() => {
-  init_src9();
   init_attachment_capabilities();
-  init_catalog_provider_mapping();
 });
 
 // ../../packages/server/src/application/browser-tools/index.ts
@@ -553116,7 +553203,9 @@ class DaemonApplication {
           createAttachmentReadTool({
             defaultTool: fileReadTool,
             authorizationSessions: attachmentAuthorizationSessions,
-            attachmentReader
+            attachmentReader,
+            supportsImageInput: async (context) => supportsNativeImageInput(context, await this.modelCatalog.load()),
+            localOcr: this.localOcr
           })
         ],
         trustedToolOverrides: ["Read"],
@@ -553740,6 +553829,7 @@ var init_daemon_application = __esm(() => {
   init_assemble_session_context_usage();
   init_attachment_access();
   init_attachment_read_tool();
+  init_attachment_capabilities();
   init_visual_tools();
   init_browser_tools();
 });
@@ -577791,4 +577881,4 @@ program.command("doctor").description("Check environment and dependencies").acti
 });
 program.parse();
 
-//# debugId=13DB4FDDE5D18B0A64756E2164756E21
+//# debugId=F1A7D813F40C9D1364756E2164756E21