5chan/docs/agent-runs/pretext-feed-sizing/progress.md

Progress Log

Append one entry per session.

2026-03-30 15:07

• Item: F001
• Summary: Built a production-path benchmark harness for real thread replies, fixed multiple estimator mismatches, and collected final desktop/mobile dom vs item-size measurements.
• Files: src/components/post-desktop/post-desktop.tsx, src/components/post-mobile/post-mobile.tsx, src/e2e/pretext-benchmark-harness.tsx, src/hooks/use-reply-height-estimates.ts, src/lib/utils/pretext-height-estimates.ts, src/views/post/post.tsx
• Verification: ./scripts/agent-init.sh --smoke, corepack yarn build, corepack yarn lint, corepack yarn type-check, corepack yarn doctor, corepack yarn test --run src/lib/utils/__tests__/pretext-height-estimates.test.ts src/components/__tests__/post-community-address-compat.test.tsx, Playwright benchmark runs on desktop and mobile for dom, estimates, and item-size
• Blockers: Reply virtualization now reduces DOM measurement reads substantially, but the user-visible performance gain is only modest on desktop and still not accurate enough to call mobile safe by default.
• Next: Start F002 by wiring a benchmarkable board-feed path and compare dom vs Pretext sizing on feed cards rather than thread replies.

2026-03-30 16:00

• Item: F002
• Summary: Added a board-feed benchmark surface, calibrated feed-card estimates from sampled DOM error, and wired the real board Virtuoso behind pretextFeed=item-size with hash-router query parsing support.
• Files: src/components/post-desktop/post-desktop.tsx, src/components/post-mobile/post-mobile.tsx, src/e2e/pretext-benchmark-harness.tsx, src/hooks/use-reply-height-estimates.ts, src/lib/utils/__tests__/pretext-height-estimates.test.ts, src/lib/utils/pretext-height-estimates.ts, src/views/board/board.tsx, src/views/post/post.tsx
• Verification: corepack yarn type-check, corepack yarn lint, corepack yarn build, corepack yarn doctor, corepack yarn test --run src/lib/utils/__tests__/pretext-height-estimates.test.ts src/components/__tests__/post-community-address-compat.test.tsx, Playwright board benchmark runs on desktop/mobile for dom, estimates, and item-size, real-route smoke test on #/mu?pretextFeed=item-size
• Blockers: The synthetic board benchmark is now clearly better than the reply path, but the live /mu route only loaded about 15 cards in this dev environment, so the production-route perf comparison is still too small and noisy to justify default-on rollout.
• Next: Continue F004 by profiling larger live board feeds with pretextFeed=item-size and decide whether board feed can ship default-on, mobile-first, or stay flag-gated while catalog/reply work continues.

2026-03-30 17:08

• Item: F001, F003, F004
• Summary: Finished catalog row sizing, removed catalog matched-filter render-time store writes, calibrated thread-reply estimates for the real Virtuoso path, extended the benchmark harness to cover catalog, and flipped board/catalog/reply virtualization defaults to item-size.
• Files: docs/agent-runs/pretext-feed-sizing/feature-list.json, src/components/catalog-row/__tests__/catalog-row.test.tsx, src/components/catalog-row/catalog-row.tsx, src/e2e/pretext-benchmark-harness.tsx, src/hooks/use-reply-height-estimates.ts, src/lib/utils/__tests__/pretext-height-estimates.test.ts, src/lib/utils/pretext-height-estimates.ts, src/views/board/board.tsx, src/views/catalog/__tests__/catalog.test.tsx, src/views/catalog/catalog.tsx
• Verification: corepack yarn type-check, corepack yarn test --run, corepack yarn build, corepack yarn lint (existing warnings only), corepack yarn doctor (baseline 86/100), corepack yarn knip, Playwright benchmark matrices for board/catalog/replies on desktop and mobile, live-route smoke checks on #/mu, #/mu/catalog, and #/mu/thread/QmWSMcz1TirxAcVJ9qKr8jygttVbXZ3bZ9mMpZyek5iu3n
• Blockers: Live-route perf deltas remain hard to observe in this dev environment because feeds are shallow and RPC noise is high, but the production-path benchmark harness now shows clear wins on all three surfaces and the real routes render correctly with the new defaults.
• Next: Keep the URL overrides (?pretextFeed=off, ?pretextCatalog=off, ?pretextReplies=off) available for rollback while gathering post-merge feedback on deeper live boards and threads.

2026-03-30 18:10

• Item: F002, F004
• Summary: Reproduced the user's /all board-feed regression against cached live data from a copied Helium Chromium profile, identified that reply-preview cards were massively overestimated in the strict board item-size path, and switched the board rollout to a hybrid strategy that only keeps cached Pretext heights on simple cards without preview replies.
• Files: src/components/__tests__/post-community-address-compat.test.tsx, src/components/post-desktop/post-desktop.tsx, src/components/post-mobile/post-mobile.tsx, src/lib/utils/pretext-height-estimates.ts
• Verification: corepack yarn test --run, corepack yarn build, corepack yarn lint (existing warnings only), corepack yarn type-check, corepack yarn doctor (baseline 86/100), Helium-profile Playwright probes on #/biz and #/all for master vs worktree plus desktop/mobile geometry checks on the worktree
• Blockers: Catalog remains the only clearly dramatic live win. Board feeds are now layout-safe and at least slightly smoother, but the hybrid fallback means the board path is a correctness-first improvement rather than a large DOM-read reduction.
• Next: Keep catalog and replies on the current Pretext default path, and continue iterating on a lower-error preview-reply estimator if we want a bigger board-feed win than the new hybrid mode provides.

2026-03-30 19:10

• Item: F001, F002, F004
• Summary: Removed the board-feed fallback again, separated preview-reply sizing from thread-reply sizing, preserved real line breaks in Pretext input, recalibrated desktop thread replies, compressed the itemSize lookup path, and reran deterministic desktop/mobile benchmark matrices. The board surface is now strict Pretext again with close geometry, desktop replies are materially better, and mobile replies remain the last blocker.
• Files: docs/agent-runs/pretext-feed-sizing/feature-list.json, src/components/__tests__/post-community-address-compat.test.tsx, src/components/post-desktop/post-desktop.tsx, src/components/post-mobile/post-mobile.tsx, src/hooks/use-reply-height-estimates.ts, src/lib/utils/pretext-height-estimates.ts, src/views/board/board.tsx
• Verification: corepack yarn build, corepack yarn lint (existing warnings only), corepack yarn type-check, corepack yarn doctor (85/100, one point below the earlier baseline because the repo still carries existing React Doctor findings), corepack yarn test --run src/lib/utils/__tests__/pretext-height-estimates.test.ts src/components/__tests__/post-community-address-compat.test.tsx src/views/board/__tests__/board.test.tsx, deterministic Playwright benchmark runs on desktop and mobile for board/catalog/replies dom vs item-size
• Blockers: Mobile thread replies still have a bad aggregate scroll-height estimate in the benchmark harness even after the latest calibration passes. Desktop board sizing is now accurate enough, but its deterministic benchmark is measurement-cheaper rather than obviously faster, so the strongest “clear win” story remains catalog, mobile board, and desktop replies.
• Next: Either keep iterating specifically on the mobile thread reply estimator or stop forcing item-size on that one surface while shipping the surfaces that are already clearly better.

2026-03-31 15:45

• Item: F001, F004
• Summary: Re-profiled mobile thread replies, replaced the broken blanket -350px mobile thread calibration with a feature-based mobile thread model, and disabled content-visibility:auto on the virtualized reply roots only when the Pretext item-size path is active. The key follow-up was a row-by-row mobile audit across the full 1200-reply benchmark: the sampled item-size reply tree sums to 722411px estimated vs 726218px actual, so the Pretext reply sizing is now within about 0.5% of the rendered total even though the old DOM Virtuoso path still reports a much smaller overall scroll height.
• Files: docs/agent-runs/pretext-feed-sizing/feature-list.json, src/components/post-desktop/post-desktop.tsx, src/components/post-mobile/post-mobile.tsx, src/lib/utils/pretext-height-estimates.ts, src/views/post/post.module.css
• Verification: corepack yarn type-check, corepack yarn lint (existing warnings only), corepack yarn test --run src/lib/utils/__tests__/pretext-height-estimates.test.ts src/components/__tests__/post-community-address-compat.test.tsx src/views/board/__tests__/board.test.tsx, corepack yarn build, corepack yarn doctor (86/100, repo baseline), corepack yarn knip, Playwright production-path reply benchmarks on mobile and desktop plus row-level mobile geometry sampling
• Blockers: Live deep-thread route verification is still limited by RPC/cached-feed availability in this dev environment, so the strongest proof remains the deterministic production-path harness rather than a real cached thread with thousands of replies.
• Next: If the user wants to ship, prepare the branch for commit/PR with the rollback query flags kept in place (?pretextFeed=off, ?pretextCatalog=off, ?pretextReplies=off) and do one more live-route smoke pass when cached live data is available.

2026-04-02 16:48

• Item: F002, F004
• Summary: Tightened the desktop board-feed height model by matching the actual OP body render path more closely. The board estimator now adds the desktop blockquote padding, models the board-view truncated-comment notice, exposes an explicit multiboard board-label height hook for /all-style feeds, and applies a small desktop preview-reply calibration instead of treating preview replies like full thread rows. That moved the board benchmark from “measurement-cheaper but visibly off” to near-parity runtime with much lower DOM reads.
• Files: docs/agent-runs/pretext-feed-sizing/feature-list.json, src/components/post-desktop/post-desktop.tsx, src/lib/utils/__tests__/pretext-height-estimates.test.ts, src/lib/utils/pretext-height-estimates.ts
• Verification: corepack yarn type-check, corepack yarn test --run src/lib/utils/__tests__/pretext-height-estimates.test.ts, corepack yarn build, corepack yarn lint (existing warnings only), corepack yarn doctor (86/100, repo baseline), corepack yarn test --run, Playwright desktop board benchmark at http://127.0.0.1:1356/?e2e=pretext-benchmark&surface=board&variant=production&count=800&seed=42&mode=dom and mode=item-size, fresh-route Playwright smoke on #/all and #/biz
• Blockers: Fresh browser route smoke still cannot validate deep live /all content because RPC/cached-feed availability is inconsistent in this environment, and ./scripts/agent-init.sh --smoke is not yet worktree-aware because it waits for the canonical https://5chan.localhost URL instead of the branch-scoped Portless route.
• Next: Re-test cached live /all and /biz against the current branch to confirm the board-feed geometry feels better in real browsing, then decide whether the board path is finally strong enough to merge along with catalog/replies.

2026-04-02 17:23

• Item: F002, F004
• Summary: Reprofiled live /all with the private Node RPC, found that desktop board cards were still underestimating almost exactly 14px per preview reply, removed that stale feed-level subtraction, and added a regression test so desktop feed cards now add the preview reply estimates directly. The live result changed materially: a 99-card /all sample went from 39.7px mean absolute error before the fix to 7.1px after the fix, with the controlled board benchmark regaining a cleaner item-size win.
• Files: docs/agent-runs/pretext-feed-sizing/feature-list.json, src/lib/utils/__tests__/pretext-height-estimates.test.ts, src/lib/utils/pretext-height-estimates.ts
• Verification: corepack yarn type-check, corepack yarn build, corepack yarn lint (existing warnings only), corepack yarn doctor (86/100, repo baseline), corepack yarn test --run src/lib/utils/__tests__/pretext-height-estimates.test.ts, corepack yarn test --run, live Playwright /all sampling on https://codex-feature-pretext-feed-sizing.5chan.localhost/#/all, live same-branch A/B on #/all vs #/all?pretextFeed=off, controlled Playwright board benchmark on ?e2e=pretext-benchmark&surface=board&variant=production&count=800&seed=42&mode=dom and mode=item-size
• Blockers: Live /all now looks much healthier, but the route-level perf delta versus pretextFeed=off is still modest/noisy rather than a dramatic blowout, and the master-origin comparison is not trustworthy yet because the master session is not loading the same feed depth as the worktree even after the settings pass.
• Next: Do another user-driven real-feed pass on /all and /biz with the new desktop board fix in place, then decide whether the remaining work should focus on squeezing a larger live-route win or on preparing the branch for merge with the current measurable improvement.

2026-04-02 17:27

• Item: F002, F004
• Summary: Confirmed that the remaining single-board feed glitching was not primarily a bad Pretext formula. Desktop board preview replies were still rendering under content-visibility:auto with the 120px intrinsic placeholder while the feed card itself was relying on Pretext sizing. I disabled deferred layout for board preview replies when the feed estimate path is active, added dev-only preview reply audit attributes on desktop, and reran live /biz sampling. That change dropped /biz from 52px mean absolute feed-card error to 6.7px.
• Files: docs/agent-runs/pretext-feed-sizing/feature-list.json, src/components/post-desktop/post-desktop.tsx, src/components/post-mobile/post-mobile.tsx
• Verification: corepack yarn type-check, corepack yarn build, corepack yarn lint (existing warnings only), corepack yarn doctor (86/100, repo baseline), corepack yarn test --run src/lib/utils/__tests__/pretext-height-estimates.test.ts, corepack yarn test --run, live Playwright /biz sampling before and after the deferred-layout fix on https://codex-feature-pretext-feed-sizing.5chan.localhost/#/biz
• Blockers: The branch is now much closer to “merge-ready on real feeds,” but the strongest user-facing proof is still the live geometry/jank sampling plus the deterministic harness rather than a clean master-origin A/B, because the master session is still not consuming the same feed depth reliably.
• Next: Do one more human-eye browsing pass on /all, /biz, and catalog with the current branch. If the route now feels consistently better than production, prepare the branch for merge with the rollback query flags retained.

2026-04-02 18:02

• Item: F002, F004
• Summary: Tracked the mobile /all stutter to two separate regressions. First, the Pretext worktree was missing master commit 99c0bbfac perf(board): reduce mobile reverse-scroll jank, so board.tsx was still snapshotting Virtuoso state on every scroll tick and comment-media.tsx was double-running the GIF first-frame hook per card. After porting that, live mobile /all audits still showed huge feed-card misses on preview-heavy media posts, so I reprofiled the live cards and found the mobile board preview calibration was stale: cards with five preview replies were underestimating by +371px / +367.9px. I replaced the old aggressive polynomial calibration with a bounded per-preview-count table, then skipped the full thread-reply estimate hook on board previews and memoized the expensive derived reply maps. That collapsed the live mobile /all mean absolute feed-card error from about 135px to about 18.8px and pulled the reverse-scroll probe much closer to master.
• Files: docs/agent-runs/pretext-feed-sizing/feature-list.json, src/components/comment-media/comment-media.tsx, src/components/post-desktop/post-desktop.tsx, src/components/post-mobile/post-mobile.tsx, src/hooks/use-reply-height-estimates.ts, src/lib/utils/__tests__/pretext-height-estimates.test.ts, src/lib/utils/pretext-height-estimates.ts, src/views/board/board.tsx
• Verification: corepack yarn type-check, corepack yarn test --run src/lib/utils/__tests__/pretext-height-estimates.test.ts src/components/__tests__/post-community-address-compat.test.tsx src/views/board/__tests__/board.test.tsx, corepack yarn build, corepack yarn lint (existing warnings only), corepack yarn doctor (86/100, repo baseline), live Playwright mobile /all geometry audit on https://codex-feature-pretext-feed-sizing.5chan.localhost/#/all, live mobile reverse-scroll probe on https://codex-feature-pretext-feed-sizing.5chan.localhost/#/all, and matched master control probe on https://5chan.localhost/#/all
• Blockers: Mobile /all is no longer catastrophically wrong, but the worktree still trails the current master control in the dev reverse-scroll probe. Latest matched sample: worktree seenCount=23, scrollHeight=18209, maxFrameMs=133.3, slowFrames16=11, slowFrames32=9; master seenCount=23, scrollHeight=17735, maxFrameMs=108.3, slowFrames16=8, slowFrames32=6.
• Next: Keep tuning the mobile board path until the reverse-scroll probe and human-eye feel are at least on par with master, likely by profiling rerender/commit churn on mobile /all now that the geometry bug is fixed.

2026-04-02 18:40

• Item: F002, F004
• Summary: Found the remaining mobile /all board regression in the board-level Virtuoso prop wiring rather than the Pretext formula. In off mode the worktree was still rendering <Virtuoso itemSize={undefined} />, which overrides React Virtuoso's internal default DOM measurer. That left mobile multiboard cards stuck on the defaultItemHeight fallback (420px) instead of reconciling to real DOM sizes, which explained the bad data-known-size values, wrong scroll height, and the remaining stutter. I changed board and catalog to only pass itemSize when the explicit item-size mode is active.
• Files: docs/agent-runs/pretext-feed-sizing/feature-list.json, src/views/board/board.tsx, src/views/catalog/catalog.tsx
• Verification: corepack yarn type-check, corepack yarn build, corepack yarn lint (existing warnings only), corepack yarn doctor (86/100, repo baseline), corepack yarn test --run src/views/board/__tests__/board.test.tsx src/lib/utils/__tests__/pretext-height-estimates.test.ts, fresh Playwright mobile /all load on https://codex-feature-pretext-feed-sizing.5chan.localhost/#/all, and persistent-session Playwright checks on https://codex-feature-pretext-feed-sizing.5chan.localhost/#/all
• Blockers: The biggest correctness bug is fixed, but the live mobile /all scroll probe in this noisy dev environment is still not a clean proof of parity versus master. The good signal is that board cards now reconcile correctly again: on the persistent RPC-backed session the first mobile /all card went from stale data-known-size=420 before the fix to data-known-size=413 after the fix, matching the real measured height.
• Next: Re-run human-eye A/B on mobile /all and /biz against master. If the feel issue is gone, the branch is much closer to merge-ready; if not, the remaining work is runtime churn rather than height reconciliation.

2026-04-02 18:46

• Item: F002, F004
• Summary: Re-ran matched live mobile A/B probes against master after the Virtuoso prop fix and confirmed the branch is no longer trailing on the board routes that mattered. On the persistent RPC-backed /all sessions, the worktree now renders the same visible card heights as master while posting slightly better reverse-scroll numbers in the probe (maxFrameMs 83.3 vs 99.1, slowFrames16 6 vs 7, slowFrames32 2 vs 3). On /biz, the branch keeps the Pretext path active (data-pretext-height nodes present) while matching the same visible card heights and total scroll height as master; probe timing is now in the same range instead of the clear regression the user saw earlier. I also added an inline comment in board/catalog so the itemSize={undefined} pitfall is documented in code.
• Files: docs/agent-runs/pretext-feed-sizing/feature-list.json, src/views/board/board.tsx, src/views/catalog/catalog.tsx
• Verification: corepack yarn type-check, corepack yarn build, corepack yarn lint (existing warnings only), corepack yarn doctor (86/100, repo baseline), corepack yarn test --run, matched Playwright mobile A/B on https://codex-feature-pretext-feed-sizing.5chan.localhost/#/all vs https://5chan.localhost/#/all, matched Playwright mobile A/B on https://codex-feature-pretext-feed-sizing.5chan.localhost/#/biz vs https://5chan.localhost/#/biz, and catalog smoke on both origins at #/all/catalog
• Blockers: None in the implementation itself. Remaining work before merge is review/commit/PR hygiene, plus the follow-up GitHub issue about future custom virtualization once this branch lands.
• Next: Prepare the branch for merge, keep FeedCacheContainer, and open the deferred “Pretext unlocks custom virtualization” issue only after the Pretext rollout is merged into master.