Reload a ChatGPT page that drew the bound answer blank instead of treating it as a changed head #75

Merged

xicv merged 6 commits from fix/unrendered-head-reload into main

2026-09-23 13:47:41 +00:00

xicv commented

2026-09-23 13:47:31 +00:00

(Migrated from github.com)

Summary

After the Ego Lite restart on 2026-09-23, ego_verify_conversation on claude-code-live-check failed twice with conversation_head_changed (tail_content_changed). Nothing in the chat had changed.

Cause, from read-only checks of the live tabs:

Ego Chat's own tab loaded at 13:20:42. Five of ChatGPT's own scripts (chatgpt.com/cdn/assets/*.js) failed together 1.6 s into that load, all with status 0 and 0 bytes.
The page then drew every answer as an empty message node (32 px, no .markdown), while user messages still showed. It stayed that way for over six minutes.
The driver read the bound answer as empty text. So the recorded head (same message ID, text "OK" plus the end marker) looked changed.
A fresh tab on the same ChatGPT build loaded all its scripts and drew every answer.

Why it matters beyond verify:

Without "Pro only", a Send on such a page is re-anchored automatically to the blank answer, and then Sends on a page that cannot draw the reply.
Under "Pro only", the reload before each Send is itself a fresh page load, and can come up the same way.

Fix:

Detection. tailUnrendered flags a last assistant answer with no text; an image-only turn is excluded. After the existing one-second hydration re-read, stabilizeBoundHead reports such a tail as unrendered instead of changed.
One proven reload. The exchange's first head read and verify reload the page once, then read it again. The proof steps are extracted from the Pro only reload into reloadSelectedPage: page marker, later performance.timeOrigin, same target and URL, task-space revalidation and readiness.
- The Pro only reload keeps its scripts, phases and failure reasons unchanged.
- The head reload has its own marker (ego-chat-head-reload), phases, stage (reloading_unrendered_page) and one-per-run budget. A policy reload that comes up blank can therefore still get its head reload.
- After a head reload, the exchange repeats its composer-surface and generation checks, and verify repeats its generation check.
New stop code. An answer still blank after the reload, or blank again at the pre-compose re-read, stops as conversation_head_unrendered, never as conversation_head_changed, so the broker never re-anchors to it.
- It is a retryable pre-Send reason, and each retry loads a fresh page.
- The new stage counts as pre-click and pre-composition.
- The head reload's failure steps are recorded as fixed uiReasons.
Re-anchor guard. Re-anchoring refuses a blank answer.
Longer verify budget. Verify's driver budget goes from 60 s to 120 s, and the CLI and MCP requests wait 125 s, so one reload fits even at its worst case (30 s navigation, 10 s proof, 10 s readiness).

No binding was anchored to a blank answer: all 41 were checked for the empty-text digest. Browser contract 34, runtime generation 2026-09-23.13.

Verification

Tests written first and seen failing:
- verify reloads once and passes;
- still blank after one reload gives conversation_head_unrendered;
- an answer drawn with other text is still conversation_head_changed with no reload;
- re-anchoring refuses a blank answer;
- an exchange reloads once as a fenced stage, then sends;
- a "Pro only" exchange gets both reloads;
- still blank stops before anything is composed;
- the broker retries the new code without re-anchoring;
- a driver error in the new stage is a proven pre-Send failure.
npm test: 1417 tests, 1416 pass, 1 skipped. npm run test:ego-monitor 148/148. Receipt suite 16/16. ESLint clean on changed files. A3K fixture regenerated.
Codex review (read-only, --base aab36cb): no actionable regression. Its sandbox could not create temporary directories, so the suites above ran locally.
Codex review of the shared verify timeouts (read-only, last commit only): no actionable regression.

Live check after install

Ego Chat's broken tab is still open in Space 1. ego-chat verify claude-code-live-check should reload it once and pass.

## Summary After the Ego Lite restart on 2026-09-23, `ego_verify_conversation` on `claude-code-live-check` failed twice with `conversation_head_changed` (`tail_content_changed`). Nothing in the chat had changed. **Cause, from read-only checks of the live tabs:** - Ego Chat's own tab loaded at 13:20:42. Five of ChatGPT's own scripts (`chatgpt.com/cdn/assets/*.js`) failed together 1.6 s into that load, all with status 0 and 0 bytes. - The page then drew every answer as an empty message node (32 px, no `.markdown`), while user messages still showed. It stayed that way for over six minutes. - The driver read the bound answer as empty text. So the recorded head (same message ID, text "OK" plus the end marker) looked changed. - A fresh tab on the same ChatGPT build loaded all its scripts and drew every answer. **Why it matters beyond verify:** - Without "Pro only", a Send on such a page is re-anchored automatically to the blank answer, and then Sends on a page that cannot draw the reply. - Under "Pro only", the reload before each Send is itself a fresh page load, and can come up the same way. **Fix:** - **Detection.** `tailUnrendered` flags a last assistant answer with no text; an image-only turn is excluded. After the existing one-second hydration re-read, `stabilizeBoundHead` reports such a tail as `unrendered` instead of `changed`. - **One proven reload.** The exchange's first head read and verify reload the page once, then read it again. The proof steps are extracted from the Pro only reload into `reloadSelectedPage`: page marker, later `performance.timeOrigin`, same target and URL, task-space revalidation and readiness. - The Pro only reload keeps its scripts, phases and failure reasons unchanged. - The head reload has its own marker (`ego-chat-head-reload`), phases, stage (`reloading_unrendered_page`) and one-per-run budget. A policy reload that comes up blank can therefore still get its head reload. - After a head reload, the exchange repeats its composer-surface and generation checks, and verify repeats its generation check. - **New stop code.** An answer still blank after the reload, or blank again at the pre-compose re-read, stops as `conversation_head_unrendered`, never as `conversation_head_changed`, so the broker never re-anchors to it. - It is a retryable pre-Send reason, and each retry loads a fresh page. - The new stage counts as pre-click and pre-composition. - The head reload's failure steps are recorded as fixed `uiReason`s. - **Re-anchor guard.** Re-anchoring refuses a blank answer. - **Longer verify budget.** Verify's driver budget goes from 60 s to 120 s, and the CLI and MCP requests wait 125 s, so one reload fits even at its worst case (30 s navigation, 10 s proof, 10 s readiness). No binding was anchored to a blank answer: all 41 were checked for the empty-text digest. Browser contract 34, runtime generation 2026-09-23.13. ## Verification - Tests written first and seen failing: - verify reloads once and passes; - still blank after one reload gives `conversation_head_unrendered`; - an answer drawn with other text is still `conversation_head_changed` with no reload; - re-anchoring refuses a blank answer; - an exchange reloads once as a fenced stage, then sends; - a "Pro only" exchange gets both reloads; - still blank stops before anything is composed; - the broker retries the new code without re-anchoring; - a driver error in the new stage is a proven pre-Send failure. - `npm test`: 1417 tests, 1416 pass, 1 skipped. `npm run test:ego-monitor` 148/148. Receipt suite 16/16. ESLint clean on changed files. A3K fixture regenerated. - Codex review (read-only, `--base aab36cb`): no actionable regression. Its sandbox could not create temporary directories, so the suites above ran locally. - Codex review of the shared verify timeouts (read-only, last commit only): no actionable regression. ## Live check after install Ego Chat's broken tab is still open in Space 1. `ego-chat verify claude-code-live-check` should reload it once and pass.