--- name: jlceda-extension-test description: This skill should be used when testing or debugging a JLCEDA Pro extension (.eext) against a live JLCEDA project page. Covers the full workflow: (1) restarting Edge via mcp__extension-dev-mcp-tools__dev_plugin to reimport the .eext, (2) connecting to the user's already-logged-in Edge/Chrome via CDP, (3) navigating to a project tab and waiting for the sch canvas to render, (4) handling V3.2-specific dialogs ("另有一个标签页" / "导入扩展"), (5) mouse-drag selecting components on the schematic canvas, (6) clicking the extension's toolbar menu and verifying the resulting IFrame, (7) capturing screenshots + dialog text for comparison, (8) verifying settings-page toggles (e.g. saveToDisk) end-to-end, (9) running a 3-case saveToDisk regression (true / false / new-install), (10) polling the AI chat IFrame through reasoning-model timeouts (deepseek-v4-pro etc.). **MUST use mcp `dev_plugin` (NOT `import_plugin` — that one fails on V3.2 UI).** Use this skill whenever the user says "test the extension in Edge", "verify the menu triggers", "run the workflow end-to-end", "重开浏览器验证", "截图对比", "check the AI chat reply", "verify saveToDisk", or "打开 settings 测试开关". Do NOT use this skill for unit-testing the extension's source code with mocks — use the project's existing mock-test.js / ai-e2e-test.js for that. agent_created: true disable: false --- # JLCEDA Extension E2E Test Workflow Test a JLCEDA Pro extension (.eext) against the user's real, already-logged-in browser through CDP. This is the only reliable way to verify menu wiring, IFrame popups, and AI API calls — mock tests cannot cover the JLCEDA extension-host layer. ## When to use - The user has a built `.eext` (e.g. `build/dist/local-netlist-analyzer_v1.1.0.eext`). - The user wants to confirm the extension works in their real JLCEDA project (e.g. "测试", "在 Edge 里跑一下", "AI 回答对了没", "重开浏览器验证", "截图对比"). - The user wants to verify that clicking a specific toolbar menu item triggers the correct IFrame and that the IFrame's content matches expectation. ## When NOT to use - Pure unit / mock testing of the extension's TypeScript → use the project's `test/mock-test.js` and `test/ai-e2e-test.js`. - Verifying build output structure (`dist/index.js` IIFE shape) → use a `node` import check. - The user only wants to read source code or commit changes — do not spin up a browser. - Driving EDA APIs from a headless script (no browser UI) → use the WebSocket reverse bridge instead, see [references/ws-bridge.md](references/ws-bridge.md). ## Critical environment facts (read first) 1. **JLCEDA V3.0**: `eda.*` is injected into a hidden `blob:` frame. Walk `page.frames()` to find the frame with `window.eda && window.eda.sch_SelectControl`. **JLCEDA V3.2+** (V3.2.148+): `eda` lives in `window._EXTAPI_SCRIPT_SPACES_[extensionUuid].eda` (a sandbox object whose methods are on the prototype chain, not own properties). `page.frames()` walk does NOT work. Use `Object.getPrototypeOf(eda.sch_SelectControl)` to enumerate methods. See [references/v3.2-sandbox-api.md](references/v3.2-sandbox-api.md). 2. **The blob frame (V3.0) is on the same origin (`https://pro.lceda.cn`)** as the main page, so any code that touches blob iframes can also disturb sch iframes. NEVER remove/`.remove()` blob iframes that have not been opened by your extension — closing one can unload the sch API. 3. **The user's Edge runs with `--remote-debugging-port=9224` (default mcp profile) or `9269` (older mcp profile)**, already logged in. Use `chromium.connectOverCDP('http://localhost:9224')` or `9269`. Reuse, do not relaunch. 4. **mcp tools**: `mcp__extension-dev-mcp-tools__import_plugin` is **broken on V3.2** (it tries to find a `.tool-bottom-menu-more_SoDfO` locator that no longer exists in V3.2 UI, 30s timeout). **Use `mcp__extension-dev-mcp-tools__dev_plugin` instead** — it auto-starts Edge on a hashed port (9222-9321 range), imports the .eext, and listens for console errors. The user may need to re-scan QR. After `dev_plugin` returns, the Edge is on a new port — use `netstat -ano | findstr LISTENING | findstr :9` to discover it. 5. **Under `cll=debug` URL, re-importing the extension (高级 → 扩展管理器 → 导入) can break the sch API injection**. If `window.eda` (or `_EXTAPI_SCRIPT_SPACES_`) disappears after importing, ask the user to F5 the project page and wait ~15s. Reloading via `page.reload()` does NOT recover it; the user must press F5 in the browser. 6. **`playwright-core` location**: NOT in the project's `node_modules` by default. Two locations: - `C:/Users/13963/WorkBuddy/2026-07-12-00-12-10/extension-dev-mcp-tools/node_modules/playwright-core` (V3.0 era) - `C:/Users/13963/AppData/Roaming/npm/node_modules/@playwright/cli/node_modules/playwright-core` (V3.2 era, bundled with `@playwright/cli`) Use the absolute path in every script: `require('')`. 7. **V3.2 sch canvas rendering is slow** — after dblclick on a sch tab, wait for the modal "另有一个浏览器标签页" to appear, click 是, then wait ~10s for the second canvas (the main sch drawing surface, >1000px wide) to appear. Do not assume `page.locator('canvas').nth(1)` exists immediately. ## Workflow ### Step 0 — Decide: reuse logged-in Edge, or restart? - If the user's Edge (9224) is open with the right project already loaded AND the new `.eext` has not been imported yet, you can either: - (A) Import directly via Playwright clicking 高级 → 扩展管理器 → 导入 (see Step 2 V3.0 path below). - (B) Restart with `mcp dev_plugin` — auto-imports but forces QR re-scan. Preferred when you also want to test the import UX. - If the user explicitly says "重开浏览器验证" / "重启 Edge", go with (B). - Otherwise prefer (A) to avoid QR scan. ### Step 1 — Probe the browser ```js var { chromium } = require('C:/Users/13963/AppData/Roaming/npm/node_modules/@playwright/cli/node_modules/playwright-core'); var b = await chromium.connectOverCDP('http://localhost:9224'); // or 9269 // Reuse the first page; if none, open one. var ctx = b.contexts()[0]; var p = ctx.pages()[0]; if (!p || !p.url().includes('pro.lceda.cn')) { p = await ctx.newPage(); await p.goto('https://pro.lceda.cn/editor'); await p.waitForTimeout(5000); } console.log('Title:', await p.title()); ``` If no JLCEDA page is open, the user must navigate to a project URL ending in `#id=,tab=*` first. As a fallback, just `goto('https://pro.lceda.cn/editor')` and dblclick on the project in the left tree. ### Step 2 — Navigate to a sch tab (V3.2) ```js // dblclick the sch tab in the left tree const sch = await p.evaluate(() => { const all = document.querySelectorAll('li .tree-title_4dmpt, li'); for (const el of all) { const t = (el.textContent || '').trim(); if (/Schematic\d+_\d+/.test(t)) { const r = el.getBoundingClientRect(); if (r.width > 0) return { x: r.x + r.width / 2, y: r.y + r.height / 2, txt: t }; } } return null; }); if (sch) { await p.mouse.dblclick(sch.x, sch.y); await p.waitForTimeout(2000); // Handle the "另有一个浏览器标签页" modal — search for [class*="modal"] containing 是 for (let i = 0; i < 10; i++) { const yes = await p.evaluate(() => { const dlg = document.querySelector('.lc_modal_dialog, [class*="modal"]'); if (!dlg) return null; const all = dlg.querySelectorAll('button, .lc_btn, [class*="btn"]'); for (const el of all) { if ((el.textContent || '').trim() === '是') { const r = el.getBoundingClientRect(); if (r.width > 0) return { x: r.x + r.width / 2, y: r.y + r.height / 2 }; } } return null; }); if (yes) { await p.mouse.click(yes.x, yes.y); break; } await p.waitForTimeout(500); } await p.waitForTimeout(8000); } // Wait for the main sch canvas (index 1, >1000px wide) to render let canvasBox = null; for (let i = 0; i < 20; i++) { const cs = await p.locator('canvas').count(); if (cs >= 2) { const b = await p.locator('canvas').nth(1).boundingBox(); if (b && b.width > 800) { canvasBox = b; break; } } await p.waitForTimeout(1000); } ``` If `canvasBox` is still null after 20s, the sch worker is broken or the sch tab is not active. Dump `await p.screenshot(...)` and inspect. ### Step 2a — Find the EDA sch frame (V3.0 only — for V3.2 use the sandbox API path below) ```js // V3.0 only: blob frame walk var schFrame = null; for (var f of p.frames()) { var has = await f.evaluate(() => !!(window.eda && window.eda.sch_SelectControl)); if (has) { schFrame = f; break; } } if (!schFrame) { // wait up to 30s for EDA to mount (debug mode is slow) for (var t = 0; t < 30; t++) { await p.waitForTimeout(1000); for (var f of p.frames()) { var has = await f.evaluate(() => !!(window.eda && window.eda.sch_SelectControl)); if (has) { schFrame = f; break; } } if (schFrame) break; } } ``` If still not found after 30s, the EDA worker is broken — usually because the user previously closed a blob iframe. Ask them to F5. ### Step 2b — Access eda API in V3.2 sandbox (from main page) ```js // V3.2: eda lives in _EXTAPI_SCRIPT_SPACES_[extensionUuid] const spaceKey = await p.evaluate(() => { const spaces = window._EXTAPI_SCRIPT_SPACES_; if (!spaces) return null; // The uuid in extension.json — usually there's only one extension return Object.keys(spaces)[0]; }); const sandbox = await p.evaluate((k) => { const sp = window._EXTAPI_SCRIPT_SPACES_[k]; if (!sp) return null; return { hasEda: !!sp.eda, sch_SelectControl_type: typeof sp.eda?.sch_SelectControl, protoMethods: Object.getOwnPropertyNames(Object.getPrototypeOf(sp.eda.sch_SelectControl)), }; }, spaceKey); console.log('sandbox:', sandbox); ``` For full V3.2 API reference, see [references/v3.2-sandbox-api.md](references/v3.2-sandbox-api.md). ### Step 3 — Import the .eext (skip if already installed and visible in top menu) If the toolbar menu ("局部网表" or whatever the extension's name is) is **not visible**, the extension is not active. Two options: **Option A — Use mcp `dev_plugin` (PREFERRED for V3.2; `import_plugin` is broken)**: ``` mcp__extension-dev-mcp-tools__dev_plugin pluginPath= ``` This auto-starts Edge on a hashed port (9222-9321) and imports the .eext. After the tool returns, discover the new debug port: ```js // netstat -ano | findstr LISTENING | findstr :9 → pick the one with msedge ``` Then connect Playwright to that port and proceed. User may need to scan QR. **Option B — Reuse logged-in Edge via Playwright** (V3.0 only): ```js // 1. click 高级 var advanced = p.locator('span[data-test="Advanced"]'); try { await advanced.click({ timeout: 2000 }); } catch { await p.locator('.tool-bottom-menu-more_SoDfO').click(); await p.waitForTimeout(500); await p.locator('.tool-bottom-menu-more-container_NmJv7 span[data-test="Advanced"]') .evaluate(el => el.click()); } await p.waitForTimeout(300); // 2. open 扩展管理器 await p.getByText('扩展管理器', { exact: false }).click({ timeout: 10000 }); var modal = p.locator("[class*='lc_modal_dialog']").first(); await modal.waitFor({ state: 'visible', timeout: 10000 }); // 3. click 导入 → handle filechooser var [chooser] = await Promise.all([ p.waitForEvent('filechooser', { timeout: 10000 }), modal.locator('button', { hasText: '导入' }).click() ]); await chooser.setFiles(''); await p.waitForTimeout(2000); // 4. close modal await modal.locator("[class*='close']").first().click().catch(() => p.keyboard.press('Escape')); ``` **Caveat**: under `cll=debug`, after this import the sch API is lost. Tell the user to F5 the project page before continuing. ### Step 4 — Mouse-drag select a region ```js // use the canvas box from Step 2 (V3.2 path) or canvasInfo[0] (V3.0) var c = canvasBox; var sx = c.x + c.w * 0.2, sy = c.y + c.h * 0.2; var ex = c.x + c.w * 0.7, ey = c.y + c.h * 0.7; await p.mouse.move(sx, sy); await p.mouse.down(); for (var t = 0; t <= 20; t++) { await p.mouse.move(sx + (ex - sx) * t / 20, sy + (ey - sy) * t / 20); await p.waitForTimeout(20); } await p.mouse.up(); await p.waitForTimeout(800); // verify selection count via EDA API // V3.0: var sel = await schFrame.evaluate(async () => { var ids = await eda.sch_SelectControl.getAllSelectedPrimitives_PrimitiveId(); return { count: ids ? ids.length : 0 }; }); // V3.2: var sel = await p.evaluate(async (sk) => { const eda = window._EXTAPI_SCRIPT_SPACES_[sk].eda; const ids = await eda.sch_SelectControl.getAllSelectedPrimitives_PrimitiveId(); return { count: ids ? ids.length : 0 }; }, spaceKey); ``` ### Step 5 — Click the extension's toolbar menu Find the menu by its visible text. The menu's `` is usually inside a wrapper that has the click handler. Prefer the parent `.top-bar-menu_JIQpX` to disambiguate from sidebar tree items with the same text: ```js var menuPos = await p.evaluate(() => { // top menu only (not sidebar) const items = document.querySelectorAll('.top-bar-menu_JIQpX > *'); for (const el of items) { const t = (el.textContent || '').trim(); if (t === '') { const r = el.getBoundingClientRect(); if (r.width > 0) return { x: r.x + r.width/2, y: r.y + r.height/2, found: true }; } } return { found: false }; }); if (!menuPos.found) throw new Error('Top menu not found: '); await p.mouse.click(menuPos.x, menuPos.y); await p.waitForTimeout(800); // find the submenu item by exact text var subPos = await p.evaluate(() => { const all = document.querySelectorAll('*'); for (const el of all) { const t = (el.textContent || '').trim(); if (t === '' && el.children.length <= 2) { const r = el.getBoundingClientRect(); if (r.width > 0) return { x: r.x + r.width/2, y: r.y + r.height/2, found: true }; } } return { found: false }; }); if (subPos.found) { await p.mouse.click(subPos.x, subPos.y); await p.waitForTimeout(3500); } ``` If the submenu cannot be located, dump all visible items in the viewport to debug: ```js var items = await p.evaluate(() => { var list = []; function walk(root, depth) { if (depth > 10) return; var b = root.getBoundingClientRect(); var t = (root.textContent || '').trim(); if (b.width > 30 && b.width < 400 && b.height > 5 && b.y > 0 && b.y < 800 && t.length > 0 && t.length < 30 && root.children.length < 3) { list.push({ t, x: b.x, y: b.y, w: b.width, h: b.height }); } for (var c of (root.children || [])) walk(c, depth + 1); } walk(document.body, 0); return list; }); ``` ### Step 6 — Verify the resulting IFrame / dialog The extension typically opens an IFrame via `eda.sys_IFrame.openIFrame(...)` AND/OR a dialog via `eda.sys_Dialog.showInformationMessage(...)`. Check both: ```js // dialog var dialog = await p.evaluate(() => { const d = document.querySelector('.lc_modal_dialog_box_5wCN6, .lc_modal_dialog_zvEgQ'); if (d) return d.textContent.trim().slice(0, 500); return null; }); console.log('dialog:', dialog); // iframe var chatFrame = null; for (var t = 0; t < 30; t++) { await p.waitForTimeout(500); for (var f of p.frames()) { var title = await f.title(); if (title && title.indexOf('') >= 0) { chatFrame = f; break; } } if (chatFrame) break; } ``` Read the visible content: ```js var text = await chatFrame.evaluate(() => document.body.innerText); fs.writeFileSync('.txt', text, 'utf-8'); ``` For chat IFrames that also accept user input: ```js await chatFrame.fill('textarea', ''); await chatFrame.evaluate(() => { var b = Array.from(document.querySelectorAll('button')).find(x => (x.textContent || '').indexOf('发送') >= 0); if (b) b.click(); }); ``` ### Step 7 — Wait for async content (e.g. AI response) A single click-then-snapshot is not enough. Poll until the IFrame stops changing and exposes a final answer or error: ```js var lastLen = 0, stableRounds = 0; var start = Date.now(); while (Date.now() - start < 60000) { await p.waitForTimeout(1500); var info = await chatFrame.evaluate(() => ({ len: document.body.innerText.length, isSending: typeof isSending !== 'undefined' ? isSending : null, hasError: !!document.querySelector('.error-msg') })); if (info.len !== lastLen) { lastLen = info.len; stableRounds = 0; } else stableRounds++; if (info.isSending === false && stableRounds >= 4) break; } ``` The 4-round stable check absorbs the typing-indicator oscillation. 60s ceiling covers OpenAI-compatible reasoning tokens (deepseek-v4-pro takes 15-20s). ### Step 8 — Capture evidence Take a full-page screenshot for the human reviewer, plus the innerText dump: ```js await p.screenshot({ path: '.png' }); fs.writeFileSync('.txt', await chatFrame.evaluate(() => document.body.innerText), 'utf-8'); ``` For baseline-vs-after comparison, save two screenshots with different prefixes (e.g. `v134-...` and `v135-...`). ## Validation rules (what "pass" looks like) - Sch API discovered within 30s after F5 → EDA worker healthy. - `getAllSelectedPrimitives_PrimitiveId().length > 0` after mouse drag → drag works. - The clicked submenu opens a new IFrame whose `document.title` matches what `eda.sys_IFrame.openIFrame` was called with → menu wiring works. - Dialog text after click is NOT the generic "请先框选" / "请先在原理图中框选" — those indicate a bug (e.g. string enum mismatch in V3.2). - `document.body.innerText` of the IFrame contains a non-trivial analysis referencing the schematic's designators (e.g. `U1`, `U4`, `FPC1`) → AI API call succeeded. - A 60s timeout with no `assistant` message → API key / network problem; check `eda.sys_Storage.getExtensionUserConfig('__ai_config')` from inside the IFrame. ## Common pitfalls - **Closing the sch blob iframe** (V3.0) while doing unrelated IFrame cleanup → sch API disappears. Always filter by src content (`chat.html` / `result.html`) before calling `.remove()`. - **Reusing the Playwright `textarea.fill` twice in a row** → second send can fire while first fetch is still pending, producing `AbortError: signal is aborted without reason`. Wait for `isSending === false` between sends. - **Assuming `page.frames()[0]` is the sch frame** → it's the main editor page; sch is a sibling. - **Forgetting to call `await p.waitForTimeout(3000+)` after import** → the menu shows up immediately, but the sch frame is reloading in the background. - **Hardcoding 9224** when the user's debug port is different → probe `netstat -an | findstr LISTENING | findstr :9` first, or use `mcp dev_plugin` and discover the new port after. - **`p.reload()` does NOT recover EDA under `cll=debug`** → only a real F5 in the user's browser fixes it. - **V3.2: `Object.keys(eda.sch_SelectControl)` returns `[]`** — methods are on the prototype. Use `Object.getPrototypeOf` to enumerate. - **V3.2: `getState_OwnerComponentDesignator` does not exist** — use `eda.sch_PrimitiveComponent.getAllPinsByPrimitiveId(primitiveId)` to reverse-lookup pin→component. - **V3.2: `getState_PrimitiveType()` returns `'Component'` (string) not `'COMPONENT'` or `6`** — use string enum check. - **V3.2: sch API throws "获取所有器件失败"** when called from main page proxy while sch canvas is not active (e.g. user is on home page, or just opened a non-sch tab). Make sure the user dblclicked into the sch tab. - **`mcp import_plugin` 30s timeout** on V3.2 UI → always use `mcp dev_plugin` instead. - **Assuming an "AI 思考中" state means success** — also check the dialog text and IFrame body. A hung fetch can leave "AI 思考中" up forever. ## Diagnosing "AI chat does not reply" When the user opens the AI chat IFrame, types a question, clicks 发送, and nothing happens (no typing indicator, no reply, send button stays grey forever), the most common cause is **a global variable name collision inside the IFrame's own JS**. The IFrame is a `blob:` URL with its own `window`, but it still inherits many browser globals; redeclaring one as a `var` silently overwrites it. Symptom fingerprint (what to look for in `chatFrame.evaluate`): - `typeof chatMessages === 'undefined'` (or whatever the array variable should be named) — suspicious. - `typeof history === 'undefined'` **but** the IFrame shows "已加载网表" → the script crashed mid-init but UI still rendered the welcome message. - Click 发送 → `addMsg('user', text)` runs, then nothing — call to `history.push` (or similar) threw a TypeError; `try`/`catch`/`finally` for the fetch never reached. Diagnostic path (proven on local-netlist-analyzer v1.2.0): ```js // 1. confirm fetch endpoint is reachable from the IFrame var r = await chatFrame.evaluate(async () => { var r = await fetch('https:///v1/chat/completions', { method: 'POST', ... }); return { status: r.status, body: (await r.text()).length }; }); // → if status=200 and body>0, network/API is fine // 2. replace window.fetch and addMsg to see what doSend actually does await chatFrame.evaluate(() => { var origFetch = window.fetch; window.fetch = function(...a) { console.log('FETCH', a[0]); return origFetch.apply(this, a); }; var origAddMsg = window.addMsg; window.addMsg = function(role, c) { console.log('ADDMSG', role, (c||'').length); return origAddMsg(role, c); }; }); // 3. then click 发送 and read console output // → if you see "ADDMSG user" but NEVER "FETCH ..." → doSend crashed between // → if you see "FETCH ..." but no "ADDMSG assistant" → response parsing crashed ``` The fix is mechanical: rename the conflicting global to a non-reserved name (e.g. `history` → `chatMessages`, `location` → `myLocation`, `navigator` → `myNavigator`). Then rebuild and re-import via the MCP import_plugin tool. Safe names for in-iframe state: `msgs`, `chatMessages`, `conversation`, `messages`, `dialogState`, `pendingTool`. Avoid: `history`, `location`, `navigator`, `status`, `name`, `length`, `top`, `parent`, `self`, `frames`, `document`, `event`. ## V3.2-specific debugging recipes ### Reasoning-model timeout map (verified on local-netlist-analyzer v1.3.5/3.6) | Model | Real response time (16 器件 + 13 网络 prompt) | Recommended `setTimeout(controller.abort, ms)` | |-------|-----------------------------------------------|------------------------------------------------| | `gpt-4o-mini` / `gpt-3.5-turbo` | 2-4s | 60s (default) | | `deepseek-chat` (non-reasoning) | 3-8s | 60s | | `qwen2.5-72b-instruct` | 4-10s | 60s | | `deepseek-v4-pro` (reasoning) | 60-120s (76s 实测) | 180s | | `o1-preview` / `o1-mini` (reasoning) | 90-180s | 240s | | `claude-3.7-sonnet-thinking` (reasoning) | 60-150s | 180-240s | **Implementation tip**: store the per-model map in `cfg.timeoutMs` so users don't have to edit code: ```js const MODEL_TIMEOUT = { 'gpt-4o-mini': 60000, 'deepseek-chat': 60000, 'deepseek-v4-pro': 180000, 'o1-preview': 240000, }; var timeout = MODEL_TIMEOUT[cfg.model] || 180000; // default 3min for safety setTimeout(controller.abort, timeout); ``` The error message should mention the actual model and the time budget so users don't panic. ### "Save files" settings switch — verifying end-to-end (v1.3.8/3.9 pattern) When the extension adds a `__file_config` checkbox in `settings.html` that gates file writes from `analyzeSelection`, the skill needs a 3-case test: ```js // 0. Hook saveFile in the V3.2 sandbox, count invocations const r = await p.evaluate((sk) => { const eda = window._EXTAPI_SCRIPT_SPACES_[sk].eda; const orig = eda.sys_FileSystem.saveFile; window._saveCount = 0; window._lastSaved = null; eda.sys_FileSystem.saveFile = async function(blob, name) { window._saveCount++; window._lastSaved = name + ':' + blob.size; return orig.call(this, blob, name); }; return { hooked: true }; }, spaceKey); // CASE 1: saveToDisk = true (after user checks the box) await p.evaluate((sk) => { const eda = window._EXTAPI_SCRIPT_SPACES_[sk].eda; eda.sys_Storage.setExtensionUserConfig('__file_config', JSON.stringify({ saveToDisk: true })); }, spaceKey); // run analyzeSelection, then read _saveCount // expect: _saveCount === 2 (csv + json) // CASE 2: saveToDisk = false (default for new users) await p.evaluate((sk) => { const eda = window._EXTAPI_SCRIPT_SPACES_[sk].eda; eda.sys_Storage.setExtensionUserConfig('__file_config', JSON.stringify({ saveToDisk: false })); }, spaceKey); // run analyzeSelection, then read _saveCount // expect: _saveCount === 0 (no file IO) // CASE 3: storage empty / new user install (simulate first run) await p.evaluate((sk) => { const eda = window._EXTAPI_SCRIPT_SPACES_[sk].eda; eda.sys_Storage.setExtensionUserConfig('__file_config', null); }, spaceKey); // run analyzeSelection // expect: _saveCount === 0 (defaults to false, no surprise file writes) ``` **Common bug shape** (v1.3.8 初版): only the HTML `checked` attribute was set; loadCfg() never re-syncs from storage. Result: checkbox shows checked but `__file_config.saveToDisk` is `undefined`, and `loadFileConfig()` falls back to its default. Always sync storage → UI explicitly: ```js var fcfg = loadFileConfig(); // { saveToDisk: false } default cb.checked = (fcfg.saveToDisk === true); // explicit ===, not truthy ``` **Why a separate `__file_config` key** (not nested under `__ai_config`): resetting AI config should not reset file save preference, and JSON parse failures in one config should not poison the other. Future config groups (theme, shortcuts) follow the same pattern. ### "Object.keys(eda.sch_SelectControl) returns [] — is eda broken?" It is NOT broken. Methods are on the prototype. Confirm with: ```js const r = await p.evaluate((sk) => { const eda = window._EXTAPI_SCRIPT_SPACES_[sk].eda; const sc = eda.sch_SelectControl; return { type: typeof sc, ownKeys: Object.keys(sc), protoMethods: Object.getOwnPropertyNames(Object.getPrototypeOf(sc)), }; }, spaceKey); console.log(r); // → protoMethods should include 'getAllSelectedPrimitives' etc. ``` ### "AI 分析 always shows '请先框选' even after I selected 62 elements" This is the V3.2 string-enum bug. Your code probably does: ```js if (pt === 'COMPONENT' || pt === 6) { // ❌ V3.0 style ``` V3.2 returns `'Component'`. Fix: ```js if (pt === 'Component' || pt === 'COMPONENT' || pt === 6) { // ✓ accept all ``` ### "Pin reverse-lookup fails / owner is undefined" V3.2 removed `pin.getState_OwnerComponentDesignator()`. Use: ```js const allComps = await eda.sch_PrimitiveComponent.getAll(); for (const c of allComps) { const d = c.getState_Designator(); const cp = c.getState_PrimitiveId(); const pins = await eda.sch_PrimitiveComponent.getAllPinsByPrimitiveId(cp); for (const p of pins) { if (selectedPinPrimIds.has(p.getState_PrimitiveId())) { // pin belongs to component with designator `d` } } } ``` ### "Right-click menu is missing" You probably only have `extension.json.headerMenus` (V3.0 style). V3.2 also needs: ```json { "contextMenus": { "sch": [ { "id": "my-ctx", "title": "我的菜单", "menuItems": [ { "id": "analyze", "title": "分析", "registerFn": "analyzeSelection" } ] } ] } } ``` Or register at runtime via `eda.sys_RightClickMenu.changeMenu(menuId, items)`. ### "How do I confirm v1.X is the version currently loaded in Edge?" Take a screenshot of the top-menu and the resulting dialog. The dialog text format includes the summary count, e.g. "AI 分析: 16元件 13网络" for v1.3.5. Comparing dialog text across runs is the cheapest regression check. ### Font-rendering trap: ASCII `|` looks like 「中」 in JLCEDA dialogs (v1.3.7 fix) **Symptom**: Your dialog says "37选中中 54元件 40网络" — the separator between summary and detail looks like a Chinese character instead of a pipe. **Root cause**: JLCEDA dialogs render text in the page's monospace CJK fallback font. ASCII `|` (U+007C) has no dedicated glyph in many CJK fonts and is silently substituted with the visually similar 「中」. **Fix**: use a separator that has an explicit CJK glyph, in order of preference: - `·` (U+00B7 MIDDLE DOT) — universal, works everywhere - `｜` (U+FF5C FULLWIDTH VERTICAL BAR) — explicit CJK - ` | ` (space-pipe-space) — only if the font renders pipe correctly (rare) Verify by taking a screenshot and looking at the dialog pixel — text-based `innerText` won't catch the glyph substitution. ### `doSelectAll` fallback when the mouse drag selects zero (V3.0/V3.2) If the sch canvas geometry is awkward and the drag reliably misses the canvas, the script can drive selection via API directly: ```js // V3.0 await schFrame.evaluate(async () => { await eda.sch_SelectControl.doSelectAll(); return eda.sch_SelectControl.getAllSelectedPrimitives_PrimitiveId(); }); // V3.2 await p.evaluate(async (sk) => { const eda = window._EXTAPI_SCRIPT_SPACES_[sk].eda; await eda.sch_SelectControl.doSelectAll(); return eda.sch_SelectControl.getAllSelectedPrimitives_PrimitiveId(); }, spaceKey); ``` This is less authentic (no real "drag" gesture) but unblocks the rest of the E2E. Add a screenshot after to make the provenance clear. ### Three-tier `installV32Shim` decision (V3.0 / V3.2 sandbox / V3.2 main page) The extension's `activate()` must handle three independent code paths without breaking each other: | Environment | `window.eda` | `eda.sch_SelectControl.getAllSelectedPrimitives_PrimitiveId` | Action | |-------------|--------------|-------------------------------------------------------------|--------| | V3.0/3.1 (any) | populated, has sch_SelectControl | returns array | **No-op** — already compatible | | V3.2 sandbox (extension runs in worker) | populated, V3.0-compatible shim already in place | returns array | **No-op** — the sandbox proxy wraps V3.2 plain properties back to getters | | V3.2 main page (extension bundle accidentally lands there) | **missing**, only `_EXTAPI_ROOT_` exists | n/a | **Alias `eda = _EXTAPI_ROOT_`** + wrap 4 selection methods (`getAllSelectedPrimitives` / `getSelectedPrimitives` / `getPrimitivesByPrimitiveId` / `getPrimitiveByPrimitiveId`) to add legacy `getState_*` getters on each returned primitive | Common bug (v1.3.3): the v1.3.3 shim aliased `eda = _EXTAPI_ROOT_` in the V3.2 sandbox, which already has its own V3.0-compatible eda — the alias overwrote a working object and broke things. v1.3.4 fixed by tier-based detection: only shim when both `eda` is missing AND `_EXTAPI_ROOT_` is present (i.e. main page only). ```ts function installV32Shim() { if (g.eda && g.eda !== root && g.eda.sch_SelectControl) return; // V3.0/3.1 if (g.eda && g.eda.sch_SelectControl) { g.__edaShimInstalled = true; return; } // V3.2 sandbox if (!root) return; // neither g.eda = root; // V3.2 main page // wrap 4 primitive-returning methods } ``` ### esbuild + V3.2 extension bundle output shape (verified) The V3.2 extension bundle is shipped as an IIFE wrapped by EDA's bundle loader. Three gotchas when iterating on the build: - **`__toCommonJS(src_exports)`** wraps ESM `export`s into a CommonJS shim. The resulting `src_exports` is an object with `default` and named keys, NOT a real CJS module. Top-level `export function foo()` is fine, but if you write `module.exports = ...` it collides. - **All `eda` references inside the bundle run inside the worker**, not the main page — so `var eda = _EXTAPI_ROOT_` would refer to the worker's window, where eda is already the V3.0-compatible shim. Don't try to import `eda` from outside; the bundle captures it as a global. - **Build must be IIFE**, not ESM, because EDA loads the bundle via `