@youdie006/prodex 0.40.20 → 0.40.21

package/dist/chatgpt-browser.js

@@ -222,7 +222,15 @@ const CHATGPT_GENERATING_CONTROL_PATTERN = /\bstop\s+(?:generating|responding|re
 // are more reliable than the button label text, which the pattern above can
 // miss on an icon-only/relabeled control - a miss would let a mid-stream pause
 // be accepted as the final answer, silently truncating it.
-export const CHATGPT_STREAMING_SELECTOR = '[data-testid="stop-button"],[data-message-author-role="assistant"][aria-busy="true"],[data-message-author-role="assistant"] [aria-busy="true"]';
+//
+// Measured 2026-09-29: the stop control lost its test id and is now a composer
+// button labelled just "Stop". Sampled every 1.5 s through a streamed answer,
+// prodex reported generating=false on all 23 samples while that button was on
+// screen - so a pause mid-answer (a Pro reasoning gap, a network stall) could
+// be taken for the end and a truncated answer returned as complete. The label
+// is matched only inside the composer's form, where the send control lives,
+// so a "Stop" elsewhere on the page cannot answer for it.
+export const CHATGPT_STREAMING_SELECTOR = '[data-testid="stop-button"],form button[aria-label="Stop"],form button[aria-label="중지"],[data-message-author-role="assistant"][aria-busy="true"],[data-message-author-role="assistant"] [aria-busy="true"]';
 export function defaultChatGptProfileDir() {
     return path.join(os.homedir(), ".local", "share", "prodex", "chrome-chatgpt-pro");
 }
@@ -2127,15 +2135,36 @@ export function focusPowerSliderExpression() {
     return { ok: document.activeElement === slider };
   })()`;
 }
+/**
+ * The composer's model trigger - and nothing outside the composer.
+ *
+ * With no composer rendered yet the lookup used to fall back to the whole
+ * document and take the first menu button with text in it. On the current
+ * sidebar that is "Pinned" or "Explore": measured 2026-09-29, every send that
+ * started on a fresh chat - a new chat, an attachment, a tool, a temporary chat
+ * - clicked a sidebar menu, then failed with "model menu did not open". A
+ * project send passed only because the project page had already rendered its
+ * composer. Now the trigger is looked for inside the rendered composer's form
+ * only (the exact "Select ChatGPT model" control first, the older heuristic
+ * second), and an unrendered composer is reported as not found, which the
+ * caller's poll waits out.
+ */
 export function modelButtonRectExpression() {
     return `(() => {${CLICK_POINT_SNIPPET}
-    const c = document.querySelector('#prompt-textarea,[contenteditable="true"],textarea');
-    const scope = c ? (c.closest('form') || document) : document;
-    const b = [...scope.querySelectorAll('[aria-haspopup="menu"]')].find((el) => {
+    const composer = [...document.querySelectorAll('#prompt-textarea,[contenteditable="true"],textarea')].find((el) => {
+      const r = el.getBoundingClientRect();
+      return r.width > 0 && r.height > 0;
+    });
+    if (!composer) return { ok: false, reason: "model selector button not found: the composer has not rendered" };
+    const scope = composer.closest('form');
+    if (!scope) return { ok: false, reason: "model selector button not found: the composer has no form around it" };
+    const exact = [...scope.querySelectorAll('button[aria-label="Select ChatGPT model"][aria-haspopup="menu"],[data-codex-intelligence-trigger="true"][aria-haspopup="menu"]')];
+    const b = exact.length === 1 ? exact[0] : exact.length > 1 ? null : [...scope.querySelectorAll('[aria-haspopup="menu"]')].find((el) => {
       const t = (el.textContent || "").trim();
       const aria = el.getAttribute("aria-label") || "";
       return /\\S/.test(t) && !/파일|첨부|받아쓰기|음성|dictation|attach|file|voice|record|search|mic/i.test(t + aria);
     });
+    if (exact.length > 1) return { ok: false, reason: "model selector button not found: more than one model trigger in the composer" };
     if (!b) return { ok: false, reason: "model selector button not found" };
     // Return the label too: the caller compares it against the requested model
     // to skip opening the menu when it is already selected.
@@ -3002,14 +3031,42 @@ async function dispatchEnterKey(cdp) {
 // Creates a project in the user's real ChatGPT account: sidebar 새 프로젝트 →
 // type the name into the popover input → Enter → wait for navigation into the
 // new project. Backs out with Escape on any failure.
+/**
+ * Open the create-project dialog on the current sidebar by keyboard.
+ *
+ * Measured 2026-09-29: the sidebar's "Add new project" button sits at the
+ * sidebar's edge underneath the composer's overlay (z-20), so the point prodex
+ * clicked resolved to the overlay and nothing opened - `--project-new` failed
+ * with "new-project name input did not appear". The button carries a stable
+ * data-app-action-sidebar-project-create attribute; focusing it and pressing
+ * Enter opens the same "Create project" dialog without depending on what is
+ * drawn over it. Returns false on a sidebar without that control, where the
+ * older click path still applies.
+ */
+async function openNewProjectDialogByKeyboard(cdp) {
+    const focused = await cdp.evaluate(`(() => {
+    const button = document.querySelector('[data-app-action-sidebar-project-create]');
+    if (!button || button.disabled) return false;
+    button.focus();
+    return document.activeElement === button;
+  })()`);
+    if (!focused)
+        return false;
+    // A button is activated by the key's text, not by the keydown alone:
+    // without text the event reaches the page and nothing opens (measured).
+    await dispatchEnterKey(cdp);
+    return true;
+}
 async function createChatGptProject(cdp, name) {
     const hrefBefore = await cdp.evaluate("location.href");
-    const button = await cdp.evaluate(newProjectButtonRectExpression());
-    if (!button.ok || button.x === undefined || button.y === undefined) {
-        throw new Error(button.reason ?? "new-project button not found in the sidebar");
-    }
     try {
-        await verifiedClickAt(cdp, button.x, button.y, "새 프로젝트");
+        if (!(await openNewProjectDialogByKeyboard(cdp))) {
+            const button = await cdp.evaluate(newProjectButtonRectExpression());
+            if (!button.ok || button.x === undefined || button.y === undefined) {
+                throw new Error(button.reason ?? "new-project button not found in the sidebar");
+            }
+            await verifiedClickAt(cdp, button.x, button.y, "새 프로젝트");
+        }
         const inputVisible = await waitForExpressionTrue(cdp, newProjectNameInputVisibleExpression(), MENU_OPEN_TIMEOUT_MS);
         if (!inputVisible)
             throw new Error("new-project name input did not appear");
@@ -3921,6 +3978,8 @@ export async function sendChatGptPrompt(options) {
         : undefined;
     let consecutiveReadFailures = 0;
     let answerSettled = false;
+    let nonTextPolls = 0;
+    let nonTextSettled = false;
     const answerIsStable = createChatGptAnswerStabilityTracker();
     while (Date.now() - started < timeoutMs) {
         await sleep(1000);
@@ -3973,6 +4032,19 @@ export async function sendChatGptPrompt(options) {
         if (runtimeBlocker)
             throw new ChatGptBrowserBlockerError(submittedRequestBlocker(runtimeBlocker, requestId, pinnedThreadUrl));
         emitProgress("waiting", finalState.generating ? "generating" : "stabilizing");
+        // A turn that answered with an image leaves no assistant message, so the
+        // count below never moves and the send used to wait out its whole budget
+        // and report a timeout - measured again on 2026-09-29 after the page-only
+        // reader replaced the transcript. Two quiet polls in a row settle it.
+        if (finalState.nonTextAnswer && !finalState.generating && finalState.userMessageCount > beforeSubmit.userMessageCount) {
+            nonTextPolls += 1;
+            if (nonTextPolls >= 2) {
+                nonTextSettled = true;
+                break;
+            }
+            continue;
+        }
+        nonTextPolls = 0;
         if (!hasFreshChatGptAnswer(beforeSubmit.assistantMessageCount, finalState))
             continue;
         // A "fresh" answer must also be stable: ChatGPT can momentarily look done
@@ -3985,6 +4057,19 @@ export async function sendChatGptPrompt(options) {
         }
     }
     const completed = finalState;
+    if (nonTextSettled && completed) {
+        emitProgress("answered", "non-text");
+        return {
+            url: pinnedThreadUrl ?? completed.url,
+            title: completed.title,
+            answer: CHATGPT_NON_TEXT_ANSWER_NOTE,
+            modelHints: completed.modelHints,
+            ...(boundProjectId ? { boundProjectId } : {}),
+            requestId,
+            requestVerified: true,
+            warnings: withDialogNote([...sendWarnings]).filter((warning) => Boolean(warning))
+        };
+    }
     if (answerSettled && completed && hasFreshChatGptAnswer(beforeSubmit.assistantMessageCount, completed)) {
         emitProgress("answered");
         return {
@@ -5142,9 +5227,27 @@ export function defaultTimeoutForTools(tools, fallbackMs) {
     const wantsDeepResearch = tools.some((tool) => resolveComposerToolLabel(tool) === DEEP_RESEARCH_TOOL_LABEL);
     return wantsDeepResearch ? Math.max(fallbackMs, DEEP_RESEARCH_MIN_TIMEOUT_MS) : fallbackMs;
 }
+/**
+ * The composer's "+" control that opens its tools.
+ *
+ * Measured 2026-09-29: ChatGPT dropped data-testid="composer-plus-btn"; the
+ * control is now `Add files and more` with
+ * data-composer-navigation-target="add-context", and every `--tool` send
+ * failed with "composer tools button not found". Looked for inside the
+ * rendered composer's form, so a similarly labelled control elsewhere cannot
+ * answer for it; the old test id still wins where it exists.
+ */
 export function composerToolsButtonRectExpression() {
     return `(() => {${CLICK_POINT_SNIPPET}
-    const b = document.querySelector('[data-testid="composer-plus-btn"]');
+    const legacy = document.querySelector('[data-testid="composer-plus-btn"]');
+    if (legacy) return clickPoint(legacy);
+    const composer = [...document.querySelectorAll('#prompt-textarea,[contenteditable="true"],textarea')].find((el) => {
+      const r = el.getBoundingClientRect();
+      return r.width > 0 && r.height > 0;
+    });
+    const form = composer && composer.closest('form');
+    if (!form) return { ok: false, reason: "composer tools button not found: the composer has not rendered" };
+    const b = form.querySelector('button[data-composer-navigation-target="add-context"],button[aria-label="Add files and more"]');
     if (!b) return { ok: false, reason: "composer tools button not found" };
     return clickPoint(b);
   })()`;
@@ -5779,7 +5882,7 @@ export function answerExpression() {
         const tagged = node.closest('[data-message-model-slug]');
         if (tagged) modelSlug = tagged.getAttribute('data-message-model-slug') || undefined;
       }
-      return { role, text: textNode.innerText || "", modelSlug };
+      return { role, node, text: textNode.innerText || "", modelSlug };
     });
     const assistantMessages = messages.filter((message) => message.role === "assistant");
     const userMessages = messages.filter((message) => message.role === "user");
@@ -5816,6 +5919,22 @@ export function answerExpression() {
       assistantMessageCount: assistantMessages.length,
       userMessageCount: userMessages.length,
       lastUserText: userMessages.at(-1)?.text || "",
+      // An image result is rendered outside every message unit and leaves no
+      // assistant message at all (measured 2026-09-29), so it is reported here
+      // for the waiter to recognize: a generated image that comes after the
+      // last user turn, with no assistant turn after that user turn.
+      nonTextAnswer: (() => {
+        const lastUser = userMessages.at(-1);
+        if (!lastUser || !lastUser.node) return false;
+        const lastRole = messages.at(-1) ? messages.at(-1).role : null;
+        if (lastRole !== "user") return false;
+        return [...document.querySelectorAll('main img')].some((img) => {
+          const r = img.getBoundingClientRect();
+          if (r.width < 60 || r.height < 60) return false;
+          if (!/^generated image/i.test(img.getAttribute("alt") || "")) return false;
+          return Boolean(lastUser.node.compareDocumentPosition(img) & Node.DOCUMENT_POSITION_FOLLOWING);
+        });
+      })(),
       // ChatGPT tags each assistant message with the model that produced it -
       // the only ground truth for "did the Pro selection actually take".
       modelSlug: assistant ? assistant.modelSlug : undefined,

package/docs/releasing.md

@@ -16,6 +16,19 @@ Release target: `v0.40.13` on `main`. The GitHub Release records the resolved co
 
 The release adds a guarded, opt-in transition and more accurate blocker reporting. It does not claim that headless Pro consultation is verified on either deployment target or that authentication can be retained indefinitely.
 
+## 0.40.20 Release Record
+
+Release commit `b8f8004` on `main` (cut after PR #9, merge `b134eae`); tag `v0.40.20`. `publish.yml` run 36522237267: the first attempt failed on macOS 15 Intel, where `tests/cli.test.ts` "uses an explicit --cwd target for doctor checks" exceeded its 30 s timeout on the slowest runner (full suite 9.5 min there; the same code had passed that runner in PR #9's CI, and no earlier failed CI run showed this test). Nothing was published by that attempt. Re-running the failed job passed all six platforms, then npm publish with SLSA v1 provenance and the GitHub Release (2026-09-29 05:18 UTC).
+
+What it restores: project consults, and every consult on a machine with a pinned default project, after the ChatGPT redesign measured on 2026-09-29 (sidebar project rows, the hover-only "New chat in" control, the model picker, message markup). Verified live on the shared browser before release: a named-project send and a pinned-default Codex + Pro send both answered inside the requested project.
+
+Installation on 2026-09-29, global npm, each verified with `prodex --version` = 0.40.20:
+
+- WSL, M3, Mac mini: installed. Running services were not restarted: 3 stdio MCP servers on WSL and 4 on M3 keep older code until their clients reconnect; the Mac mini had none.
+- M4: not reachable (hostname does not resolve); not updated.
+
+Browser container, same day. The Claude Code MCP entry on WSL goes through `prodex-browser-browser-1`, which builds from `feat/headless-browser-compatibility`, not from npm. That branch had already adapted to the same redesign on its own (not yet merged to `main`), so the image was rebuilt from its CI-passing tip `b5eac7d` as-is - not merged with `main`, because the two parallel implementations conflict in 17 places in `src/chatgpt-browser.ts` and a hand-made hybrid would be unverified. New image `sha256:7b3e53b2fb7f`; it reports 0.40.18, the version in that branch's `package.json`. Verified: no consult in flight before the swap; container healthy with the profile volume kept; MCP handshake through `scripts/container-client.mjs`; `pro browser login --wait` found the session READY; a project consult inside the container answered in the requested project with the destination verified. The two implementations still have to be reconciled when that branch is merged.
+
 ## 0.40.19 Release Record
 
 Release commit `8650136` on `main` (cut after PR #8, merge `b5a223f`); tag `v0.40.19`. Published by `publish.yml` run 36510352704: all six native verification jobs passed (Ubuntu Node 20/22/24, macOS 15 arm64 and Intel, Windows Server 2025 / Node 22), then npm publish with SLSA v1 provenance, then the GitHub Release (2026-09-29 02:17 UTC). Before tagging, PR #8 passed the same six jobs and `npm run release:verify` passed locally on the release commit after a build (a fresh checkout without `dist` fails two package-bin cases, because `release:verify` runs tests before building).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@youdie006/prodex",
-  "version": "0.40.20",
+  "version": "0.40.21",
   "description": "Local receipt bus for coordinating Codex execution with ChatGPT Pro/Projects consultation.",
   "author": "youdie006",
   "license": "MIT",