Skip to content

armgit / cindy

Public Imported from GitHub
0

Consider it done. The open-source AI agent that works out of the box · 想到,就能做到。开源、开箱即用的 AI Agent。

fix(docs): keep inspect_pdf going when pdf.js cannot extract one page

Zhao Dengfengebb99c6426d2
diff --git a/apps/desktop/src/main/reviewer/__tests__/pdfInspect.test.ts b/apps/desktop/src/main/reviewer/__tests__/pdfInspect.test.ts
index a58ebbc821..23a3877b59 100644
--- a/apps/desktop/src/main/reviewer/__tests__/pdfInspect.test.ts
+++ b/apps/desktop/src/main/reviewer/__tests__/pdfInspect.test.ts
@@ -148,6 +148,24 @@ describe('inspectPdfPages', () => {
expect(Math.round(result.pages[0]!.height)).toBe(792);
});
+ it('单页文字提取失败只把该页标为未验证,其余页照常检查', async () => {
+ // pdf.js 5.7/6.4 在内容流里第一次显示文字前若尚无字体、且先设字号 0(Aspose 等工具写入的
+ // 隐形水印即此形态),getTextContent 会抛 "Cannot read properties of null (reading 'name')"。
+ // 文件本身完好(MuPDF 正常渲染),不能让这一页拖垮整份检查,也不能把它判成空白。
+ const pdf = buildPdf(['before', { raw: 'BT /F1 0 Tf (mark) Tj ET' }, 'after']);
+ const result = await inspectPdfPages(pdf, [], 10, 100);
+ expect(result.pages.map((p) => p.page)).toEqual([1, 2, 3]);
+ expect(result.pages[0]).toMatchObject({ textPreview: 'before', visibilityUnverified: false });
+ expect(result.pages[1]).toMatchObject({
+ textChars: 0,
+ blank: false,
+ visibilityUnverified: true,
+ });
+ expect(result.pages[1]!.textError).toContain("reading 'name'");
+ expect(result.pages[1]!.drawOps).toBeGreaterThan(0);
+ expect(result.pages[2]).toMatchObject({ textPreview: 'after', visibilityUnverified: false });
+ });
+
it('损坏的 PDF 抛错而不是返回一份"看起来正常"的空结果', async () => {
await expect(
inspectPdfPages(new Uint8Array(Buffer.from('not a pdf at all')), [], 10, 100),
diff --git a/apps/desktop/src/main/reviewer/__tests__/reviewPdfProcess.test.ts b/apps/desktop/src/main/reviewer/__tests__/reviewPdfProcess.test.ts
index 3b1a975e29..e2073a2d7b 100644
--- a/apps/desktop/src/main/reviewer/__tests__/reviewPdfProcess.test.ts
+++ b/apps/desktop/src/main/reviewer/__tests__/reviewPdfProcess.test.ts
@@ -91,8 +91,10 @@ describe('isolated PDF structure inspection', () => {
textPreview: '',
drawOps: 0,
imageOps: 0,
- blank: true,
- visibilityUnverified: false,
+ blank: false,
+ visibilityUnverified: true,
+ // 单页文字提取失败的原因要穿过进程边界交给 inspect_pdf,否则无法区分空文字与解析失败。
+ textError: "Cannot read properties of null (reading 'name')",
},
],
},
@@ -109,7 +111,11 @@ describe('isolated PDF structure inspection', () => {
pages: [2],
fork: () => child,
}),
- ).resolves.toMatchObject({ numPages: 3, pagesInspected: 1 });
+ ).resolves.toMatchObject({
+ numPages: 3,
+ pagesInspected: 1,
+ pages: [{ textError: "Cannot read properties of null (reading 'name')" }],
+ });
expect(sent).toMatchObject({ kind: 'inspect', pages: [2], maxPages: 5, previewChars: 100 });
expect(child.kill).toHaveBeenCalledTimes(1);
diff --git a/apps/desktop/src/main/reviewer/reviewPdfProcess.ts b/apps/desktop/src/main/reviewer/reviewPdfProcess.ts
index 11d278f04e..a7e0ca6918 100644
--- a/apps/desktop/src/main/reviewer/reviewPdfProcess.ts
+++ b/apps/desktop/src/main/reviewer/reviewPdfProcess.ts
@@ -218,6 +218,7 @@ function isReviewPdfPageInspection(value: unknown, maxPages: number): boolean {
nullableCount(page.imageOps) &&
typeof page.blank === 'boolean' &&
typeof page.visibilityUnverified === 'boolean' &&
+ (page.textError === undefined || typeof page.textError === 'string') &&
maxPages > 0
);
}
diff --git a/apps/desktop/src/main/reviewer/reviewPdfProcessProtocol.ts b/apps/desktop/src/main/reviewer/reviewPdfProcessProtocol.ts
index 80331fb7fd..6ff5f4b1f7 100644
--- a/apps/desktop/src/main/reviewer/reviewPdfProcessProtocol.ts
+++ b/apps/desktop/src/main/reviewer/reviewPdfProcessProtocol.ts
@@ -32,8 +32,10 @@ export interface ReviewPdfPageInspection {
imageOps: number | null;
/** 无文本且无任何绘图/图像算子。算子读不出来时恒为 false(不猜)。 */
blank: boolean;
- /** true = 算子读取不完整,结构检查未完成;false = 算子已成功读取。 */
+ /** true = 算子或文字读取不完整,结构检查未完成;false = 两者都已成功读取。 */
visibilityUnverified: boolean;
+ /** 该页文字提取失败时的解析器错误(此时 textChars=0 不代表没有文字)。 */
+ textError?: string;
}
export interface ReviewPdfInspectProcessResult {
diff --git a/apps/desktop/src/main/reviewer/reviewPdfUtilityProcess.ts b/apps/desktop/src/main/reviewer/reviewPdfUtilityProcess.ts
index 7656af8c75..debff29787 100644
--- a/apps/desktop/src/main/reviewer/reviewPdfUtilityProcess.ts
+++ b/apps/desktop/src/main/reviewer/reviewPdfUtilityProcess.ts
@@ -175,12 +175,13 @@ export function pageVisibilityUnverified(
textChars: number,
drawOps: number | null,
imageOps: number | null,
+ textFailed = false,
): boolean {
// This flag denotes an incomplete structural read, not every page that
// legitimately contains content. A successful operator parse must still
// allow the normal structural verdict to be `ok`.
void textChars;
- return drawOps === null || imageOps === null;
+ return textFailed || drawOps === null || imageOps === null;
}
export function selectPdfPageNumbers(
@@ -212,6 +213,12 @@ export function selectPdfPageNumbers(
* 算子表单独 try/catch:损坏页或超预算时降级成 null 而不是让整次检查失败 ——
* 半份结构信息也比"检查工具自己挂了"有用。null 时 blank 恒为 false,并标记
* visibilityUnverified,不把未完成的结构检查当成通过。
+ *
+ * 文字提取同样逐页 try/catch。pdf.js 对部分合法文件会在 getTextContent 里抛错
+ * (例如内容流开头先以字号 0 显示文字、此前未设字体时的
+ * "Cannot read properties of null (reading 'name')",常见于隐形水印),此时
+ * 文件本身可正常渲染。该页记为 textError + visibilityUnverified,blank 恒为 false,
+ * 其余页照常检查;整份文件打不开(InvalidPDFException 等)仍然整体抛错。
*/
export async function inspectPdfPages(
data: Uint8Array,
@@ -248,14 +255,20 @@ export async function inspectPdfPages(
const page = await document.getPage(pageNumber);
const viewport = page.getViewport({ scale: 1 });
- const textContent = await page.getTextContent();
- const parts: string[] = [];
- for (const item of textContent.items) {
- if (!item || typeof item !== 'object' || !('str' in item)) continue;
- if (typeof item.str !== 'string') continue;
- parts.push(item.str, item.hasEOL ? '\n' : ' ');
+ let text = '';
+ let textError: string | undefined;
+ try {
+ const textContent = await page.getTextContent();
+ const parts: string[] = [];
+ for (const item of textContent.items) {
+ if (!item || typeof item !== 'object' || !('str' in item)) continue;
+ if (typeof item.str !== 'string') continue;
+ parts.push(item.str, item.hasEOL ? '\n' : ' ');
+ }
+ text = parts.join('').replace(/\s+/g, ' ').trim();
+ } catch (error) {
+ textError = (error instanceof Error ? error.message : String(error)).slice(0, 300);
}
- const text = parts.join('').replace(/\s+/g, ' ').trim();
let drawOps: number | null = null;
let imageOps: number | null = null;
@@ -284,8 +297,14 @@ export async function inspectPdfPages(
textPreview: previewChars > 0 ? text.slice(0, previewChars) : '',
drawOps,
imageOps,
- blank: text.length === 0 && drawOps === 0 && imageOps === 0,
- visibilityUnverified: pageVisibilityUnverified(text.length, drawOps, imageOps),
+ blank: textError === undefined && text.length === 0 && drawOps === 0 && imageOps === 0,
+ visibilityUnverified: pageVisibilityUnverified(
+ text.length,
+ drawOps,
+ imageOps,
+ textError !== undefined,
+ ),
+ ...(textError !== undefined ? { textError } : {}),
});
}
return { numPages, pagesInspected: pages.length, pages };
diff --git a/packages/lizi-mcps/src/__tests__/cindyDocsMcpServer.test.ts b/packages/lizi-mcps/src/__tests__/cindyDocsMcpServer.test.ts
index ca641b9ad7..bbc95fdbf0 100644
--- a/packages/lizi-mcps/src/__tests__/cindyDocsMcpServer.test.ts
+++ b/packages/lizi-mcps/src/__tests__/cindyDocsMcpServer.test.ts
@@ -2710,6 +2710,29 @@ describe('inspect_pdf', () => {
expect(result.verdict).not.toBe('ok');
});
+ it('单页文字提取失败时给 warning 并指明页码,不当成空白或生成失败', async () => {
+ const client = await withPdf({
+ numPages: 2,
+ pagesInspected: 2,
+ pages: [
+ page(),
+ page({
+ page: 2,
+ textChars: 0,
+ textPreview: '',
+ blank: false,
+ visibilityUnverified: true,
+ textError: "Cannot read properties of null (reading 'name')",
+ }),
+ ],
+ });
+ const result = await callTool(client, 'inspect_pdf', { path: 'out.pdf' });
+ expect(result.verdict).toBe('warning');
+ expect(result.blankPages).toEqual([]);
+ expect(result.visibilityUnverifiedPages).toEqual([2]);
+ expect(result.warning).toContain('第 2 页是解析器读不出文字');
+ });
+
it('横向与非标准纸张都翻译成人能对照的说法', async () => {
const client = await withPdf({
numPages: 2,
@@ -2858,7 +2881,9 @@ describe('inspect_pdf', () => {
});
const failed = await callTool(broken, 'inspect_pdf', { path: 'out.pdf' });
expect(failed.errorCode).toBe('INSPECT_FAILED');
+ // 解析器缺陷也会走到这里:提示仍要求坏文件重做,但不能断言生成环节一定出错。
expect((failed.data as Record<string, string>).hint).toContain('重做');
+ expect((failed.data as Record<string, string>).hint).toContain('兼容性');
});
});
diff --git a/packages/lizi-mcps/src/cindy-docs/inspect_pdf.ts b/packages/lizi-mcps/src/cindy-docs/inspect_pdf.ts
index bc9f202f73..b807a5ad5e 100644
--- a/packages/lizi-mcps/src/cindy-docs/inspect_pdf.ts
+++ b/packages/lizi-mcps/src/cindy-docs/inspect_pdf.ts
@@ -71,7 +71,8 @@ const DESCRIPTION = [
'',
'【怎么判读】',
'- blankPages 非空 → 那几页是结构上确定为空的,大概率 CSS 把内容藏了或外部资源没加载上,重做;',
- '- visibilityUnverified=true → 某页算子解析未完成,未做位图级可见性确认,结构检查证据不完整;请重试或人工打开确认,不要把 verdict 当作视觉验收;',
+ '- visibilityUnverified=true → 某页算子或文字解析未完成,未做位图级可见性确认,结构检查证据不完整;请重试或人工打开确认,不要把 verdict 当作视觉验收;',
+ '- textError 非空 → 解析器读不出该页文字,textChars=0 不代表没有文字;常见于引用了原文件的页,用其他阅读器打开该页确认,不要仅凭此判定生成失败;',
'- numPages 远超预期 → 分页样式没生效(检查 page-break / break-inside);',
'- paper 不是你要的尺寸 → render_pdf 的 pageSize 传错了;',
'- textPreview 和你写的内容对不上 → 装错了内容或页序错乱。',
@@ -220,6 +221,9 @@ export function registerInspectPdfTool(
const visibilityUnverifiedPages = decorated
.filter((p) => p.visibilityUnverified)
.map((p) => p.page);
+ const textErrorPages = decorated
+ .filter((p) => p.textError !== undefined)
+ .map((p) => p.page);
const inspectedPageNumbers = new Set(decorated.map((page) => page.page));
let accumulatedThrough = Math.min(inspectedThrough, inspection.numPages);
while (inspectedPageNumbers.has(accumulatedThrough + 1)) accumulatedThrough += 1;
@@ -263,7 +267,7 @@ export function registerInspectPdfTool(
: accumulatedVisibilityUnverified
? {
verdict: 'warning',
- warning: `${visibilityUnverifiedPages.length > 0 ? `第 ${visibilityUnverifiedPages.join('、')} 页的结构算子解析未完成` : '此前批次存在结构算子解析未完成的页面'},未做位图级可见性确认,检查证据不完整。请重试并在必要时打开 PDF 确认。${coverageWarning ? ` ${coverageWarning}` : ''}`,
+ warning: `${visibilityUnverifiedPages.length > 0 ? `第 ${visibilityUnverifiedPages.join('、')} 页的结构算子解析未完成` : '此前批次存在结构算子解析未完成的页面'},未做位图级可见性确认,检查证据不完整。${textErrorPages.length > 0 ? `其中第 ${textErrorPages.join('、')} 页是解析器读不出文字(见 textError),可能是解析器兼容性问题而非文件损坏。` : ''}请重试并在必要时打开 PDF 确认。${coverageWarning ? ` ${coverageWarning}` : ''}`,
}
: partial
? { verdict: 'incomplete', warning: coverageWarning! }
@@ -292,7 +296,7 @@ export function registerInspectPdfTool(
timedOut ? 'INSPECT_TIMEOUT' : 'INSPECT_FAILED',
timedOut
? `解析超过 ${INSPECT_PDF_TIMEOUT_MS / 1000} 秒被中止,文件可能损坏或过于复杂。`
- : `读取 PDF 失败:${message}。如果这是刚生成的文件,说明生成环节就出了问题,请重做而不是交付。`,
+ : `读取 PDF 失败:${message}。可能是文件损坏,也可能是解析器兼容性问题;先用其他阅读器打开确认。打不开或内容缺失就重做,不要交付。`,
{ message },
);
}
diff --git a/packages/lizi-mcps/src/cindy-docs/types.ts b/packages/lizi-mcps/src/cindy-docs/types.ts
index 8238867e53..51de2aad87 100644
--- a/packages/lizi-mcps/src/cindy-docs/types.ts
+++ b/packages/lizi-mcps/src/cindy-docs/types.ts
@@ -94,6 +94,8 @@ export interface DocsPdfPageInspection {
blank: boolean;
/** 结构存在但未做位图级可见性验证。 */
visibilityUnverified: boolean;
+ /** 该页文字提取失败时的解析器错误(此时 textChars=0 不代表没有文字)。 */
+ textError?: string;
}
export interface DocsPdfInspection {