Repository navigation
perf(transcript): batch prompt geometry reads during row commits - #33960
teamleaderleo wants to merge 2 commits into
Conversation
Co-authored-by: Codex <199175422+chatgpt-codex-connector[bot]@users.noreply.github.com> Agent-Session: https://cmux.com/deeplink/agent?id=aid-d60a77e16a11b66e2ecf
…leanup Co-authored-by: Codex <199175422+chatgpt-codex-connector[bot]@users.noreply.github.com> Agent-Session: https://cmux.com/deeplink/agent?id=aid-d60a77e16a11b66e2ecf
|
cmux status (one comment, edited in place; each line is updated by the job that owns it) Review App of this head (app.zip, kept 1 day)
Gallery before/after for
States
Pick other states with a |
GitHub Actions
Inactive
|
Review (claude opus-5.5 medium): SHIP at dcf74d7 Visual verdict: this touches prompt rendering logic (when the fold measurement runs), but the intended appearance is unchanged. There are no images, and the first-paint geometry tests still need fleet results. This review covers only the seven-line delta; the rest is carried from the SHIP at bc8b4be. Nothing has run yet, so there are no test results or performance numbers, and I claim no speedup. What I checked in the delta
Follow-ups:
Validation and same-worker ABBA pending. The optional cancellation-specific assertion is already covered by the queue unit tests; the component assertion intentionally verifies observable no-stale-read behavior. |
|
Found 1 test failure on Blacksmith runners: Failure
|
![Fix with [code]smith](https://pr-comments-assets.blacksmith.sh/codesmith/fix-with-codesmith-light-3x.png)
|
Run 41658618283 failed in the measurement wrapper before it executed any checks or benchmarks: the broker's candidate snapshot did not contain historical baseline blobs, so The isolated comparison now reconstructs the baseline using the exact product-only reverse diff, verifies SHA-256 hashes of both source variants against their original Git objects, and restores the candidate before validation and after the comparison. Candidate tree and fixed harness hashes remain pinned. One corrected dispatch is in flight at the unchanged head |
|
Withdrawing this candidate: correctness and the measurement-order regression passed, but this run does not demonstrate a consistent overall scrolling speedup sufficient to justify the additional queue/microtask machinery. The adverse fling sample is retained, not dismissed as proven noise. No performance or zero-blank acceptance claim. Run41660063670 passed the full check, 30 unit/component tests, 3 browser tests (first-painted geometry and child-only correction included), and 3 streaming/fold tests. The same browser ordering test fails on the baseline as expected (first read index0 versus at least3 sibling writes). All source/harness/log hashes match; the driver restored the candidate. Head:
Headless diagnostic only. Native/nightly 120 Hz acceptance remains unproven. Render-start to commit is elapsed time, not React CPU time. Separate coverage passes avoid charging DOM sampling to the timing runs.
|
|
Withdrawn after same-worker ABBA run41660063670: mixed timing (fling p95 A124/148 → B165/116 → B106/112 → A112/130 ms), with one16ms blank frame per fast fling remaining. Correctness passed; no consistent overall speedup established to justify the added queue complexity. Full evidence is in the preceding report. The 120Hz scroll work continues with the separately scoped formatter candidate; this PR was not merged. |
Withdrawn after measurement
Withdrawing this candidate: correctness and the measurement-order regression passed, but this run does not demonstrate a consistent overall scrolling speedup sufficient to justify the additional queue/microtask machinery. The adverse fling sample is retained, not dismissed as proven noise. No performance or zero-blank acceptance claim.
Run41660063670 passed the full check, 30 unit/component tests, 3 browser tests (first-painted geometry and child-only correction included), and 3 streaming/fold tests. The same browser ordering test fails on the baseline as expected (first read index0 versus at least3 sibling writes). All source/harness/log hashes match; the driver restored the candidate.
Head:
dcf74d7c2ace4dbf5a58db755d730af6ad9de88a; baseline:7f6fe51cc3e65c5b18c93ca45def3fe4073530b4; worker:7b2ddfbac651b79fRun: https://github.com/manaflow-ai/cmux/actions/runs/41660063670
Accepted: True; restored: True; error: none
Headless diagnostic only. Native/nightly 120 Hz acceptance remains unproven. Render-start to commit is elapsed time, not React CPU time. Separate coverage passes avoid charging DOM sampling to the timing runs.
This candidate batched PromptText layout reads after sibling writes and before article-height reads, with a child-only microtask fallback. No product changes were merged.
Leo: “i feel like the app is slow on scrolling thru stuff... claude desktop is 120hz or something and ours is wack + it has sometimes not even rendered stuff!?”
Performance
The paired ABBA rows above retain all samples, including the adverse one. The 120 Hz nightly target remains unmet.
🤖 Generated with Codex
Agent session: agent session