Conversation
… constants Groundwork for nesquena#4343, and a dead-code fix. NOT a fix for the scroll oscillation itself — see the measurement below before assuming otherwise. _messageVirtualWindow's rowHeightFor() gives every UNMEASURED row a height so the virtual scroll geometry can be computed. It sourced that from the flat per-role constants in MESSAGE_VIRTUAL_DEFAULT_ROW_HEIGHTS: return roleForIdx ? _messageVirtualDefaultHeightForRole(roleForIdx(idx)) : defaultHeight; _updateMessageVirtualMeasurements already maintained a measurement-derived mean in _messageVirtualEstimatedRowHeight, and _currentMessageVirtualWindow already passed it down as `defaultHeight` — but roleForIdx is ALWAYS supplied, so that branch never ran and the calibrated number was dead code. Every row that never entered the render window kept the guess forever. On a 2000-message session only ~45 rows are ever measured, so 98% of the geometry came from the guess, and the guesses are wrong in both directions (measured median 214 against user:120 / assistant:160 / tool_call:400). The errors do not cancel: total scroll height came out 326,267px against a real 492,942px, 34% short. Track a running mean PER ROLE instead and use it for unmeasured rows (a single global mean would be worse than the constants, since a tool_call row and a user row differ enormously). A role falls back to its static seed until it has MESSAGE_VIRTUAL_ROLE_SAMPLE_MIN samples, so first paint is unchanged, and a session switch drops the samples with the rest of the height cache. Applied at all three unmeasured-row sites (rowHeightFor, _messageVirtualPrependedHeightDelta, _messageVirtualScrollTopForVisibleIdx) rather than just the one, so they cannot disagree about the same row's height. Effect, measured with real wheel events on that session: - scroll geometry: 326,267px -> 460,631px against a real 492,942px (34% short -> 6.6% short) - drift of the row the reader is looking at, after the measurement pass settles: 800px median / 848px max BEFORE, and 800px median / 848px max AFTER. Unchanged. Zero. So the geometry error is real and now largely fixed, but it is NOT what causes the nesquena#4343 oscillation. Trapping the scrollTop setter names the actual writer: _scrollAfterMessageRender -> _followMessagesAfterDomReplace -> scrollToBottom -> _setMessageScrollToBottom, firing on the measurement-driven re-render even though the caller passes preserveScroll:true and _messageUserUnpinned is already true; the delayed _settleMessageScrollToBottom writes look like the live culprit, re-scheduled by each re-render after _cancelBottomSettle clears them. That is a separate change and virtualize_transcript stays default-off until it lands. The existing virtualization tests all still pass with the static fallback (no samples recorded), so they would pass against a reverted fix and prove nothing on their own; test_unmeasured_rows_use_measured_role_mean_once_samples_exist pins the behaviour that actually changed and fails (120 != 300) when the calibration branch is removed. Five tests that eval these helpers in an isolated node scope needed the new state/companion declared. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018y9pr7ppZgDuCZCT6x96kY
… transcript refetch (nesquena#6826)
…x is provably identical; compression-anchor coverage (nesquena#6826 r3)
…U) for bounded tail read (nesquena#6826 r3 round-3)
… data_version guard (nesquena#6826 r4)
…al WAL interleaving test (nesquena#6826 r5)
Removes the local re-import of api.streaming._session_payload_with_full_messages inside test_in_tail_duplicate_guard_refuses_bounded_and_full_revision_accepted; it is already imported at module level (line 17) and unused in the function body, tripping F401 on CI's hosted lint. Mechanical maintainer fix per auto-fix policy. Co-authored-by: webtecnica <webtecnica@users.noreply.github.com>
Bounded sidecar-anchored tail read for regenerate (closes nesquena#6826). Thanks @webtecnica — six rounds of adversarial re-gate to convergence.
…ail read (nesquena#7204, @webtecnica) Adds the CHANGELOG entry for nesquena#7204 (closes nesquena#6826). Release-metadata only; the code change ships via the merge of nesquena#7204 (contributor webtecnica).
Release exp-v0.52.264: fast regenerate via bounded sidecar-anchored tail read (nesquena#7204, @webtecnica)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
While chasing a scroll-drift bug I found the virtual scroll geometry runs ~34% short of the real transcript height (326k px vs 493k px on a 2000-message session): only ~45 rows ever get measured, and every other row falls back to a flat per-role constant. The measured average the code already computes was passed in as
defaultHeightbut never actually reached, becauserowHeightFor()always took the constant instead.Now it keeps a measured mean per role and uses that for unmeasured rows, so the geometry is ~6.6% off instead of 34%. Roles fall back to the old constant until they have samples, so first paint is unchanged.
Honest caveat: this did not fix the drift I was chasing — that turned out to be elsewhere, and the drift measured identically before and after. So this is a correctness fix with no user-visible symptom attached. Nothing else changed; targeted suite green, and the new test fails if the calibration is reverted.
virtualize_transcriptonly. Feel free to merge or drop.