# 010 — Streaming Scope Hardening Status: D (diff-level implementation roadmap) ## Audit amendments (round 1, Sol reviewer Newton — GO-WITH-FIXES, 5 blockers, all folded) 1. **Latest-assistant-turn must be role-verified** (High): do not treat the last top-level match of the combined selector list as assistant. Selection rule: prefer the last node matching `[data-message-author-role="assistant"]`/`[data-turn="assistant"]`; use its nearest top-level conversation wrapper as the progress scope. A bare `article[data-testid^="conversation-turn"]` without an assistant-role descendant is NOT an assistant turn. 2. **Behavioral DOM tests, not string assertions** (High): the streaming-state tests must run `readChatGptStreamingState` against real jsdom/happy-dom fixtures — exact stop button hit, form-scoped fallback hit, out-of-form fallback miss, and each `:not()` exclusion (dictation/voice/read) as separate fixtures. 3. **No new broad completed-sidecar `includes('thought for')`** (High): the predicate must use the anchored duration grammar `/^thought for \d+[a-z]*( seconds?| minutes?)?( edit)?$/i` on normalized visible text (G7/G12 stay deferred; this plan must not introduce a NEW broad suppression). A growing live trace containing "Thought for 2s: Searching…" must remain live. 4. **Out-of-scope callers declared** (Medium): `chatgpt-deep-research.mjs:49-52`, `chatgpt-work-picker.mjs:690,950`, `chatgpt-multi-turn.mjs:54-56` keep their own stop predicates — WP2 scope is the general ChatGPT response path (`chatgpt.mjs` poller + observer + shared constants). Their migration is recorded as a deferred follow-up row in 040 (G1b). 5. **Activation table ↔ test mapping** (Medium): every activation row must name its executable test case; ARIA omitted-max and completed-sidecar counter-case get explicit fixtures. Format: DIFFLEVEL-ROADMAP-01 Work-phase: WP2 Upstream predicates verified with `git show 99b30cfa 0071c547 9f6703bf 93ccb79d` in `/tmp/oracle-chase-260724`. ## Objective Make ChatGPT streaming detection response-scoped and fail-safe against unrelated controls. Keep the exact `button[data-testid="stop-button"]` signal, but constrain the broad aria-label fallback to the composer form and reject dictation, voice, and read-aloud controls. Evaluate progress only in the latest assistant turn or in a metadata-verified thinking sidecar; determinate completed progress is idle. Geometry may narrow a sidecar candidate, but cannot establish that it is reasoning chrome. The implementation should move the browser-context predicate into `chatgpt-response-dom.mjs` so the authoritative poller and focused unit tests exercise one implementation. The observer continues to be an early-wake mechanism only, but inherits the same scoped stop selector constants. ## Gaps covered | Gap | Required closure | | --- | --- | | G1 | Composer-scope the aria-label Stop fallback and exclude dictation/voice/read controls everywhere the shared selectors are consumed. | | G2 | Search progress indicators only in the latest assistant turn, with a separate verified-sidecar path. | | G3 | Treat indeterminate progress or determinate `value < max` as live; treat `value >= max` as idle. | | G4 | Require thinking/reasoning/sidecar metadata before a right-side panel can veto completion. | ## File change map | Path | Action | Current path:line anchor | Functions / constants | | --- | --- | --- | --- | | `web-ai/chatgpt-response-dom.mjs` | MODIFY | `web-ai/chatgpt-response-dom.mjs:3-12`, `web-ai/chatgpt-response-dom.mjs:14-40` | `CHATGPT_STOP_SELECTORS`; add `readChatGptStreamingState()` and its browser-context helpers | | `web-ai/chatgpt.mjs` | MODIFY | `web-ai/chatgpt.mjs:55-60`, `web-ai/chatgpt.mjs:909-965` | response-DOM imports; `isStreaming()` | | `web-ai/chatgpt-response-observer.mjs` | MODIFY | `web-ai/chatgpt-response-observer.mjs:13-17`, `web-ai/chatgpt-response-observer.mjs:30-66`, `web-ai/chatgpt-response-observer.mjs:151-167` | shared `CHATGPT_STOP_SELECTORS` usages in `buildResponseObserverExpression()` and `readStreamingState()` | | `test/unit/web-ai-chatgpt-response-fragments.test.mjs` | MODIFY | `test/unit/web-ai-chatgpt-response-fragments.test.mjs:1-6`, `test/unit/web-ai-chatgpt-response-fragments.test.mjs:8-59`, `test/unit/web-ai-chatgpt-response-fragments.test.mjs:61-90` | imports; new `ChatGPT streaming state` cases and DOM fakes | | `test/unit/web-ai-chatgpt-response-observer.test.mjs` | MODIFY | `test/unit/web-ai-chatgpt-response-observer.test.mjs:1-7`, `test/unit/web-ai-chatgpt-response-observer.test.mjs:9-30` | observer expression selector-contract assertions | ## Proposed diffs ### 1. Shared stop selectors and response-scoped streaming predicate In `web-ai/chatgpt-response-dom.mjs`, replace the current broad stop selector and add a serializable browser-context predicate. The exact test-id remains document-wide; only the fallback is composer-scoped. Before: ```js export const CHATGPT_STOP_SELECTORS = [ 'button[data-testid="stop-button"]', 'button[aria-label*="Stop" i]', ]; ``` After: ```js export const CHATGPT_STOP_SELECTORS = [ 'button[data-testid="stop-button"]', 'form button[aria-label*="Stop" i]:not([aria-label*="dictat" i]):not([aria-label*="voice" i]):not([aria-label*="read" i])', ]; /** * Browser-context helper. Returns whether the current ChatGPT response has * positive live-generation evidence. * @param {{ assistantSelectors: string[], stopSelectors: string[] }} options * @returns {boolean} */ export function readChatGptStreamingState({ assistantSelectors, stopSelectors }) { const isVisible = (/** @type {Element} */ node) => { if (!(node instanceof HTMLElement)) return false; const rect = node.getBoundingClientRect(); return rect.width > 0 && rect.height > 0; }; const norm = (/** @type {unknown} */ value) => String(value || '').toLowerCase().replace(/\s+/g, ' ').trim(); const hasLiveProgress = (/** @type {ParentNode} */ scope) => { let nodes; try { nodes = scope.querySelectorAll('progress, [role="progressbar"]'); } catch { return false; } return Array.from(nodes).some(node => { if (!isVisible(node)) return false; if (node instanceof HTMLProgressElement) { if (!node.hasAttribute('value')) return true; return Number.isFinite(node.value) && Number.isFinite(node.max) ? node.value < node.max : true; } const rawNow = node.getAttribute('aria-valuenow'); if (rawNow == null) return true; const now = Number(rawNow); const rawMax = node.getAttribute('aria-valuemax'); const max = rawMax != null && Number.isFinite(Number(rawMax)) ? Number(rawMax) : 100; return Number.isFinite(now) ? now < max : true; }); }; for (const selector of stopSelectors) { let nodes; try { nodes = document.querySelectorAll(selector); } catch { continue; } if (Array.from(nodes).some(isVisible)) return true; } const activeAssistantSelectors = Array.isArray(assistantSelectors) && assistantSelectors.length ? assistantSelectors : [ '[data-message-author-role="assistant"]', '[data-turn="assistant"]', 'article[data-testid^="conversation-turn"]', ]; const assistantSelector = activeAssistantSelectors.join(', '); let turns; try { turns = Array.from(document.querySelectorAll(assistantSelector)); } catch { turns = []; } const topLevelTurns = turns.filter(node => !turns.some(other => other !== node && other.contains(node))); const latestTurn = topLevelTurns.at(-1); if (latestTurn && hasLiveProgress(latestTurn)) return true; let panels; try { panels = document.querySelectorAll( 'aside, [role="complementary"], [role="dialog"], [data-testid*="thinking" i], [data-testid*="reasoning" i], [class*="sidecar" i]', ); } catch { return false; } for (const panel of Array.from(panels)) { if (!isVisible(panel)) continue; const metadata = norm([ panel.getAttribute('aria-label'), panel.getAttribute('data-testid'), panel.getAttribute('class'), ].filter(Boolean).join(' ')); const verifiedThinkingPanel = metadata.includes('thinking') || metadata.includes('reasoning') || metadata.includes('sidecar'); if (!verifiedThinkingPanel) continue; const rect = panel.getBoundingClientRect(); const rightSide = rect.left >= window.innerWidth * 0.35 && rect.width >= 180 && rect.height >= 120; if (!rightSide) continue; if (hasLiveProgress(panel)) return true; const visibleText = norm(panel.textContent); if (visibleText.includes('thought for')) continue; if (visibleText.includes('thinking') || visibleText.includes('reasoning') || visibleText.includes('pro thinking')) return true; } return false; } ``` `hasAttribute('value')` is intentional for HTML ``: absent `value` means indeterminate, while the DOM `value` property alone would otherwise read as zero and lose that distinction. ARIA progress defaults `aria-valuemax` to 100, matching the upstream predicate. ### 2. Authoritative poller delegates to the shared predicate In `web-ai/chatgpt.mjs`, extend the current response-DOM import: Before: ```js CHATGPT_STOP_SELECTORS, readTopLevelAssistantTexts, readTopLevelAssistantTextsFromLocators, ``` After: ```js CHATGPT_STOP_SELECTORS, readChatGptStreamingState, readTopLevelAssistantTexts, readTopLevelAssistantTextsFromLocators, ``` Replace `isStreaming()` at current lines 909-965: Before: ```js async function isStreaming(page) { // document-wide stop/progress loops and geometry-led sidecar evaluation } ``` After: ```js async function isStreaming(page) { try { return Boolean(await page.evaluate( readChatGptStreamingState, { assistantSelectors: CHATGPT_ASSISTANT_SELECTORS, stopSelectors: CHATGPT_STOP_SELECTORS, }, )); } catch { return false; } } ``` This retains the current navigation-safe fail-open behavior of `isStreaming()` while deleting the duplicate page-global progress and sidecar implementation. ### 3. Observer usages consume the corrected selector contract No alternate stop selector may be introduced in `chatgpt-response-observer.mjs`. Its current imports and both usages remain structurally the same: ```js import { CHATGPT_ASSISTANT_SELECTORS, CHATGPT_STOP_SELECTORS, readTopLevelAssistantTexts, } from './chatgpt-response-dom.mjs'; ``` ```js const stopSelector = CHATGPT_STOP_SELECTORS.join(', '); ``` ```js for (const selector of CHATGPT_STOP_SELECTORS) { const first = page.locator?.(selector)?.first?.(); if (typeof first?.isVisible === 'function' && await first.isVisible().catch(() => false)) return true; } ``` The implementation edit here is a contract comment immediately above each usage, naming that the broad fallback is composer-scoped and exclusions must not be weakened. This makes the inherited G1 behavior explicit and prevents a later observer-only regression. ### 4. Focused tests Extend `web-ai-chatgpt-response-fragments.test.mjs` imports with `CHATGPT_STOP_SELECTORS` and `readChatGptStreamingState`. Add these cases using DOM fakes that implement `querySelectorAll`, `getAttribute`, `hasAttribute`, `contains`, and `getBoundingClientRect`: ```js describe('ChatGPT streaming state', () => { it('keeps the exact stop test id but ignores non-composer aria Stop controls', () => { expect(CHATGPT_STOP_SELECTORS[0]).toBe('button[data-testid="stop-button"]'); expect(CHATGPT_STOP_SELECTORS[1]).toMatch(/^form button/); expect(CHATGPT_STOP_SELECTORS[1]).toContain('dictat'); expect(CHATGPT_STOP_SELECTORS[1]).toContain('voice'); expect(CHATGPT_STOP_SELECTORS[1]).toContain('read'); }); it('ignores page-global progress outside the latest assistant turn', () => { const state = withStreamingDocument({ globalProgress: progressNode({ value: 10, max: 100 }) }, () => readChatGptStreamingState({ assistantSelectors: CHATGPT_ASSISTANT_SELECTORS, stopSelectors: CHATGPT_STOP_SELECTORS, })); expect(state).toBe(false); }); it('treats indeterminate and incomplete latest-turn progress as live', () => { expect(readStreamingFixture({ latestProgress: progressNode({ indeterminate: true }) })).toBe(true); expect(readStreamingFixture({ latestProgress: progressNode({ value: 40, max: 100 }) })).toBe(true); }); it('treats completed determinate latest-turn progress as idle', () => { expect(readStreamingFixture({ latestProgress: progressNode({ value: 100, max: 100 }) })).toBe(false); expect(readStreamingFixture({ latestProgress: ariaProgressNode({ now: 5, max: 5 }) })).toBe(false); }); it('requires sidecar metadata before right-side geometry can veto completion', () => { expect(readStreamingFixture({ panel: rightPanel({ text: 'Reasoning', metadata: '' }) })).toBe(false); expect(readStreamingFixture({ panel: rightPanel({ text: 'Reasoning', metadata: 'reasoning-sidecar' }) })).toBe(true); }); }); ``` Extend the observer expression tests: ```js it('embeds the composer-scoped Stop fallback exclusions', () => { const expr = buildResponseObserverExpression(); expect(expr).toContain('form button'); expect(expr).toContain('dictat'); expect(expr).toContain('voice'); expect(expr).toContain('read'); }); ``` ## Test plan The repository uses Vitest for these unit files (`package.json` scripts and existing `import { describe, expect, it } from 'vitest'`), not direct `node --test` for this surface. Run focused tests first: ```sh npx vitest run test/unit/web-ai-chatgpt-response-fragments.test.mjs test/unit/web-ai-chatgpt-response-observer.test.mjs ``` Then run the existing broader unit entry point: ```sh npm run test:unit ``` New focused test names: - `keeps the exact stop test id but ignores non-composer aria Stop controls` - `ignores page-global progress outside the latest assistant turn` - `treats indeterminate and incomplete latest-turn progress as live` - `treats completed determinate latest-turn progress as idle` - `requires sidecar metadata before right-side geometry can veto completion` - `embeds the composer-scoped Stop fallback exclusions` ## Accept criteria - G1: exact `button[data-testid="stop-button"]` remains accepted anywhere; the aria fallback begins with `form button` and contains all three exclusions (`dictat`, `voice`, `read`). Poller, observer expression, and observer recovery import the same constants. - G2: a visible page-shell progress bar outside the latest assistant turn does not make `isStreaming()` true; the same bar inside the latest assistant turn does. - G3: HTML and ARIA determinate bars at or beyond max are idle; incomplete and indeterminate bars are live. - G4: right-side dimensions alone never veto completion. Sidecar progress/text can veto only when aria-label, data-testid, or class metadata contains thinking/reasoning/sidecar. - No new page-global `progress` or `[role="progressbar"]` lookup exists in `chatgpt.mjs`. - Both focused test files and `npm run test:unit` pass. ### C-ACTIVATION-GROUNDING-01 | Conditional | Activation scenario | Observable proof that the branch fired | | --- | --- | --- | | Exact stop test-id | Mount a visible `button[data-testid="stop-button"]` outside a form. | `readChatGptStreamingState()` returns `true`. | | Scoped aria Stop fallback | Mount a visible `form button[aria-label="Stop generating"]`; repeat outside a form. | Inside form returns `true`; outside form returns `false`. | | Dictation/voice/read exclusion | Mount composer buttons labelled `Stop dictation`, `Stop voice`, and `Stop reading`. | Each fixture returns `false`; replacing label with `Stop generating` returns `true`. | | Latest-turn progress scope | Mount one shell progress bar and two assistant turns, with progress only in the newest turn on the second run. | First run returns `false`; second returns `true`. | | HTML determinate predicate | Mount visible ``, then change value to 99. | First returns `false`; second returns `true`. | | HTML indeterminate predicate | Mount visible `` with no `value` attribute. | Returns `true`. | | ARIA determinate predicate | Mount `[role="progressbar"][aria-valuenow="100"]` with omitted max, then now 99. | Default-max completed case returns `false`; 99 returns `true`. | | Verified sidecar metadata | Mount identical right-side panels with text `Reasoning`; only the second has `data-testid="reasoning-sidecar"`. | Unverified panel returns `false`; verified panel returns `true`. | | Completed sidecar summary exclusion | Mount a verified sidecar whose text contains `Thought for 12s` and has no live progress. | Returns `false`, proving retained-summary text did not veto. | ## Risks / rollback - DOM selector drift could hide a real composer Stop button. The exact test-id remains the primary signal; rollback is limited to reverting the shared fallback selector while retaining tests for non-generation controls. - Some ChatGPT variants may not wrap the latest assistant response in current selectors. That would omit response-scoped progress, but stop detection remains active and WP3 requires positive completion proof, so failure is a timeout/defer rather than premature finalization. Add a selector only with a captured DOM fixture. - Metadata verification may miss an unlabeled genuine sidecar. This is intentionally fail-neutral for streaming activity; WP3's scoped completion gate still fails closed. Roll back G4 independently by broadening only the metadata vocabulary, never by restoring geometry-only proof. - `page.evaluate()` failure returns `false`, matching current navigation handling. Do not make it throw through the poll loop.