fix(chatgpt): adapt to 2026-08 chatgpt.com UI and zero-size bridge viewports

[RusianHu]

Summary

Fixes a cluster of chatgpt adapter failures against the 2026-08 chatgpt.com web UI (root cause report: #2435). All findings were reproduced against a real logged-in account (zh-CN UI) and verified end-to-end after the fix; the adapter's unit tests pass 180/180.

Root causes found (verified live, see #2435 for details)

1. Zero-size viewport in bridge windows. The OpenCLI browser bridge runs tabs in a window where window.innerWidth/innerHeight === 0 while document.visibilityState === 'visible' (reproduced with --window foreground too). The new chatgpt.com UI sizes message/image containers from the viewport, so fluid-width nodes report rect.width === 0 while their text renders fine. Every isVisible that gates on rect.width > N then filters out exactly the nodes we need → ask TIMEOUT, read/detail EMPTY, image finds no generated images.
2. Sidebar anchors are gone. Conversation items now render as bare ul > li > button with no attributes and no <a href="/c/..."> anywhere in the document → history EMPTY_RESULT.
3. Model picker restructured. The trigger is now an unlabeled button showing a short effort label (observed extended → 高, min → 中, zh-CN), and the thinking-effort options live in a 选择模型 submenu that ignores synthetic pointer/keyboard events entirely (full pointerdown/mousedown/pointerup/mouseup/click sequence, trusted-key attempts — all no-ops), so the picker path cannot reach intelligence options.
4. Upload preview chips are aria-label-only ([Bug]: chatgpt image --image exits 1 with 'image upload preview did not appear' after a successful backend upload (waitForChatGPTUploadPreview misses aria-label-only preview chip) #2302) — filename matching against body.innerText misses them.

Changes (all in clis/chatgpt/utils.js)

• Viewport tolerance for layout-based visibility (15 isVisible sites): when document.documentElement.clientWidth <= 0, skip the rect gate; CSS display/visibility checks remain, so genuinely hidden aria-shadow copies are still excluded. On any normal viewport behavior is unchanged.
• Conversation discovery: getConversationList now prefers the backend API (/api/auth/session token → /backend-api/conversations) executed in the page context; the anchor extraction is kept as fallback.
• Model selection: added a last-resort effort patch via the settings API (/backend-api/settings/user_last_used_model_config) run in the page context (the CLI host may not be able to reach chatgpt.com directly; the browser session always can). It keeps the account's current model family from the oai-last-model-config cookie and patches only the effort (fast→min, balanced→standard, advanced→extended, very-high→xhigh — vocabulary confirmed by the API's own enum validation). The visible picker stays the primary path; the API fallback only fires when the picker cannot reach the option or the postcondition fails. Server-side eligibility rejects invalid combos with the config untouched.
• Model trigger discovery: fallback to the composer's only unlabeled short-text button; clicks dispatch the full pointer/mouse event sequence (plain .click() does not open the menu) with coordinates returned as before so native clicks still work on real viewports.
• Upload preview ([Bug]: chatgpt image --image exits 1 with 'image upload preview did not appear' after a successful backend upload (waitForChatGPTUploadPreview misses aria-label-only preview chip) #2302): match file names against aria-labels as well as visible text; media nodes in zero-size viewports are accepted via naturalWidth/videoWidth/background-image presence.
• Terminal no-image detection (policy refusals): when ChatGPT ends the turn with a text-only reply instead of an image (policy refusals finish in ~1 min), waitForChatGPTImages now fails fast with the refusal text in the error instead of burning the full --timeout on a state that will never produce images. Conservative gating: requires two consecutive quiet polls with substantive (≥16 char) assistant text; if message extraction is unavailable it falls back to the existing deadline behavior. Covered by two new unit tests (182/182 pass).
• Robustness: getCookies() calls race an 8s timeout (the full-jar call was observed to hang the daemon bridge); the post-PATCH reload is scheduled via setTimeout (reloading synchronously inside evaluate destroys the execution context and leaves the promise pending forever); the session fetch in buildChatGPTBackendHeaders catches network aborts instead of surfacing AbortError.

Verification

• Live: ask (new + continuation), history, detail, read, whoami, status, new, image (text-to-image), image --image (edit), model (fast → advanced round trip, confirmed server-side via the config cookie).
• npx vitest run clis/chatgpt → 180/180 pass.
• opencli validate chatgpt → PASS (14 commands).
• Refusal fast-fail covered by dedicated unit tests (text-only terminal state).

Notes for reviewers: the zero-size viewport is a property of our bridge environment (Windows, extension-connected daemon); the clientWidth <= 0 guards are no-ops on normal viewports. The history API path is strictly more robust than DOM anchors but the anchor fallback is retained in case the UI regresses or another locale renders anchors.

Update 2026-09-03: /c/WEB:<uuid> temporary route ids (new commit)

Re-testing ask against the live site on 2026-09-03 surfaced a further frontend change that also breaks upstream main, so it belongs in this PR:

Symptom. ask sends the message and the model replies (server-side conversation is created), but the CLI fails with ChatGPT did not create a conversation URL after sending the message.

Root cause (verified live). Brand-new conversations now route to /c/WEB:<uuid>:

• <uuid> is a client-side temporary id — it never appears in backend-api, and it is not the real conversation id. Navigating directly to /c/WEB:<uuid> bounces back to the home page.
• The id segment contains :, which the old parseChatGPTConversationId character class ([A-Za-z0-9_-]) rejects, so waitForConversationUrl times out after 30 s.
• Observed post-condition: once generation finishes, the frontend itself swaps the URL for the real /c/<server-id>.

Fix (second commit, clis/chatgpt/utils.js + clis/chatgpt/ask.js):

• parseChatGPTConversationId: the id character class now includes : so WEB: ids parse.
• ask: when the parsed id is a WEB: temporary id, wait for the response first, then re-read the URL — if the frontend has already swapped in the real id, use it; otherwise fall back to the newest entry of the backend-api conversation listing (new resolveWebConversationId helper, reuses the existing page-context fetch from getConversationList).
• The leave-conversation URL guard is skipped on that path: the WEB: → server-id swap is a legitimate navigation, and the old guard false-fires on it.

Verification (same battery as above, re-run after this commit):

• Live: ask (new conversation) → returns the real server-side id and correct response; ask --conversation <real-id> continuation works.
• npx vitest run --project adapter clis/chatgpt/ → 164/164 pass.
• npx tsc --noEmit → clean.
• npx tsx src/main.ts validate chatgpt → PASS (14 commands).

Fixes #2435