Repository navigation
Clean up bench: one-snapshot group queries, module split, stricter types - #129
Merged
Merged
Conversation
Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
bench/scripts/parity/run.sh serves a base git ref and the working tree side by side with `next start`, against one Postgres restored from a production dump, and compares them. compare-api.ts requests every group and chart route from both servers and compares status, content-type, cache-control and body, printing the route and first differing JSON path with both values. screenshots.spec.ts compares the landing page and five chart pages pixel by pixel at desktop and phone widths, and Playwright keeps the diff images. The database runs from a local Postgres 16 data directory, since the integration suite's container is not required. Group-cache fills run one at a time because they share a database-wide advisory lock, and each server starts with an empty Data Cache. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Every URL the API check requests comes from the base's own /api/groups, so a 404, a redirect or an empty body on either side now fails with "nothing was compared for this route", even when both servers agree on it. Slugs are percent-encoded, and the summary line tallies the statuses and body bytes that were compared. run.sh deletes the whole .next/cache of each side before it starts a server, so a Next.js version that keeps Data Cache payloads somewhere other than fetch-cache cannot serve payloads made by older code. The README and config comment now say what the screenshot comparison does: Playwright's comparator ignores pixels it classes as anti-aliasing, and any other differing pixel, or a layout shift of 1px or more, fails. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
GitHub's dependency graph does not read bun.lock, so dependency-review sees none of the site's transitive packages. scripts/dependency-snapshot.ts turns the lock into a snapshot for the Dependency Submission API, with one manifest per workspace. A package is runtime when a workspace's dependencies or optionalDependencies reach it, and development otherwise. Optional peers add no edge: Next.js lists Playwright as one, and following it would mark Playwright runtime. Every package in the current lock is in the snapshot, and every open high advisory in `bun audit` is development scope. The dependency-snapshot workflow runs it on pushes to main and on pull requests from this repository, with contents: write on that job only and no bun install. dependency-review now retries for up to 300 seconds while the snapshot arrives. fail-on-scopes stays at the action's default, runtime. CLAUDE.md layer 5 says so. The tests use node:test, so the root tsc needs no bun types. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
The submission script now waits 5 seconds and tries once more after a network error or a 5xx answer. Any other status, and a second failure, still exits 1. The wait can be set with SNAPSHOT_RETRY_DELAY_MS, which the tests use. The token is never logged. The workflow groups a push to main by commit, so a burst of pushes cannot cancel or replace a pending run and leave a commit without a snapshot. Pull requests keep one group per ref and cancel the older run. New tests run the script against a local mock of the API: 503 then 201, 503 twice, 422 with no retry, and a refused connection. Another test covers a workspace whose path (bench/web) differs from its package name (benchmarks-web), with a workspace-keyed copy of a package, a hoisted copy, and a decoy keyed by path. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
bench/web joins the root Bun workspace. One bun.lock replaces bench/web/pnpm-lock.yaml, and bench/web/pnpm-workspace.yaml goes: - The zod override (^4.6.7) moves to the root package.json. It also moves Lighthouse CI's chromium-bidi to zod 4, as decided, and clears the site's accepted zod advisory. The sharp override was already there. The postcss override is not carried over: Next.js 16.4.1 depends on postcss 8.5.23, which is patched, so until bench moves to Next.js 16 its next 15.5.28 brings postcss 8.4.31. CLAUDE.md lists those advisories under "bench advisories". - undici 7.30.1 is still inside the 14-day window, so bunfig.toml exempts undici until 2026-10-29 14:18 UTC, with the reason from the pnpm file. Every other version pnpm had locked is past the window. - The lockfile keeps pnpm's versions for every package only bench uses. The rest are the site's copies, which Bun dedupes onto. The site's resolution changes only in zod 3.25.76 -> 4.6.7. bunfig.toml sets linker = "isolated", Bun's default for a workspace. A hoisted tree hands bench's typescript-eslint the site's TypeScript 7, which it rejects. publicHoistPattern puts ESLint and Prettier packages in the root node_modules, where bench's ESLint config finds its plugins, as pnpm's default did. next.config.js traces from the repository root, where the packages now live. Every route traces the same packages at the same versions as the pnpm build on main. The bench-checks action, bench-ci.yml and bench-deploy.yml install with `bun install --frozen-lockfile` at the root and run bench's scripts with `bun run --cwd bench/web`. bench/web/vercel.json sets the same install command for `vercel build`, so the deploy does not depend on the project's settings. bench-deploy.yml also runs on changes to the root package.json, bun.lock and bunfig.toml. changed-paths counts the workspace's root files for both apps. CLAUDE.md and bench's docs use the Bun commands. trustedDependencies stays ["esbuild", "sharp"]. pnpm also let unrs-resolver build; its postinstall only checks for the native binding, which Bun installs as an optional dependency. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
bench moves to next 16.4.1 and react/react-dom ^19.3.1, the site's
versions, so both apps now share one copy. Next.js 16.4.1 depends on
postcss 8.5.23, which clears the four postcss advisories that next
15.5.28's postcss 8.4.31 carried; no postcss override is needed.
- POST /api/revalidate calls revalidateTag(BENCH_DATA_TAG, { expire: 0 }).
That is the documented equivalent of the old one-argument call; 'max'
would serve stale data once after an ingest. A new test pins it.
- Other bumps, each the newest version at least 14 days old: pg and
@types/pg 8.23.1, jsdom 30.1.2, vitest 5.0.4 (clears the two
@vitest/mocker advisories), eslint-config-next 16.4.1, @types/node
^24.19.2, @types/react(-dom) ^19.3.0, @vercel/functions ^3.10.1
(already locked). chart.js stays at 4.5.1.
- jsdom 30 uses undici 8. The newest 8.x past the window, 8.12.1, has
four high advisories fixed in 8.12.2, so the existing undici exemption
now covers 8.12.2 as well; the lock pins 8.12.2, and the exemption
list's removal time moves to 2026-10-29 15:06 UTC.
- The lock keeps @types/pg on the shared @types/node 24.19.2, as the
workspace move did.
- eslint-config-next 16 ships flat config, so eslint.config.mjs imports
it directly instead of through FlatCompat. Two new React Compiler rules
from eslint-plugin-react-hooks 7 are off, so the upgrade needs no code
change.
- next build (Turbopack, now the default) set tsconfig's jsx to
react-jsx and added .next/dev/types; committed so builds stay clean.
- Docs: bench/web/README.md, the architecture pages that describe
revalidateTag, the postcss.config.mjs and vitest.config.ts comments,
bunfig.toml, and CLAUDE.md (toolchain, audit counts 17 -> 11, the
bench advisories list).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Signed-off-by: Will Manning <[email protected]>
These two files were outside Prettier's glob and ran past 100 columns. Biome's check now covers them, so this is the one-time mechanical reformat, kept apart from the Biome switch. No logic changes. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
The root biome.json now covers bench. A bench/** override keeps its Prettier style (single quotes, trailing commas, 100 columns) and turns useExhaustiveDependencies on, in place of ESLint's react-hooks rule. It turns off five rules that fire on deliberate markup and CSS, and import sorting. bench/AGENTS.md gives the reasons, because biome.json cannot hold comments. Lint fixes keep behavior: a block around a switch case, an explicit type on one let, braces on void forEach callbacks, an optional chain, a template string, an import type. Two ESLint disable comments for exhaustive-deps become biome-ignore, and a third effect gets one. Stale ESLint disable comments are gone. Remove eslint, eslint-config-next, @eslint/eslintrc and prettier with their config files. Their plugins were the only reason for publicHoistPattern in bunfig.toml, so it goes too; a clean install, the suite and both builds pass without it. Fix the stale Next.js, TypeScript and ESLint statements in bunfig.toml and CLAUDE.md. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
bench's format and lint steps covered only bench/web, and ci.yml skips the root check:ci on bench-only changes, so bench/scripts went unchecked until the next site PR. One step, bun run biome ci bench, replaces both: it uses the workspace's pinned Biome with the root config and covers web, scripts and the parity harness. The job name does not change. Fix two notes in bench/AGENTS.md: organizeImports would rewrite 27 files (25 in bench/web, 2 in bench/scripts), and the site config also turns off noExplicitAny, so bench's old no-explicit-any check is gone. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Move bench/web to typescript ^7.0.2, as at the root, so both apps run the Go-based compiler. The tsconfig options the two apps shared now live in a root tsconfig.base.json (strict, noEmit, esModuleInterop, module, moduleResolution, resolveJsonModule, isolatedModules, skipLibCheck, allowJs, incremental, noUnusedLocals, noUnusedParameters, lib, jsx and the next plugin). Each tsconfig.json extends it and keeps its own target, paths and extra flags: the site its noUncheckedIndexedAccess and exactOptionalPropertyTypes, bench its noFallthroughCasesInSwitch and noUncheckedSideEffectImports. `tsc --showConfig` gives the same effective options for both apps as before. TypeScript 7 reports no errors in bench under the existing flags, so no source changes. bench-deploy.yml now also runs when tsconfig.base.json changes, since bench's build extends it. bunfig.toml and CLAUDE.md no longer say bench keeps its own TypeScript 5.9. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
The bench override in biome.json turned four a11y rules off for all of bench. New UI would then skip them unseen. Turn them back on and put a biome-ignore comment with its reason at each of the seven existing places. The markup does not change. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
…heck The parity check compared only the read API and light-theme screenshots. It now also requests the landing page and the five chart pages the screenshots use, and compares their status, content-type, cache-control and title. It probes an unknown chart slug (404) and a malformed slug (400 from the API, 404 from the page), and accepts only those expected statuses. Every other URL must still answer 200. The screenshots run in the dark theme too. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
bunfig.toml gave a reason for the isolated linker that no longer holds, now that both apps share TypeScript and ESLint is gone. State the reason that does hold. vitest.config.ts said Next.js owns tsconfig's jsx setting; it now comes from tsconfig.base.json. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
bench-checks lost its dependency cache when it moved from pnpm to Bun. Restore it with the same cache step as ci.yml. Run the parity check's own tests there too, by explicit path, so its planning and comparison logic is checked on every bench change. Correct a comment that said Biome checks bench's imports; import sorting is off for bench. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Dependency review passed without bun.lock's packages when the snapshot job failed, was skipped for a fork, or finished after the action's retry window. A guard now runs before the action. For a pull request from this repository, it waits up to 10 minutes for the snapshot job on the head commit and fails unless that job succeeded. For a fork, which gets no snapshot, it fails when the pull request changes a bun.lock. The job gains checks: read, and its checkout no longer keeps the token. The snapshot job now checks out the pull request's head commit, so the graph it files under that commit holds the head's own bun.lock. CLAUDE.md describes the guard, says that development-scope advisories (bench's included) no longer block merge and show in bun audit, and gives the real fallback for the zod override. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Give the publish times behind the statement that next 16.4.1 and react 19.3.1 were past the cooldown when bench moved to them, and say how the "13 days old" for the site differs. Rewrite the vite note, which still described pnpm. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
The database suites needed Docker, so they skipped on a Mac without it. Set BENCH_TEST_DATABASE_URL to a Postgres server and each suite creates a throwaway bench_test_* database there, applies the migrations, and drops the database when it ends. A Postgres advisory lock serializes the first applies, because the migrations create cluster-wide roles. Without the URL, the suites use testcontainers as before. When CI=true and neither Docker nor the URL is available, the suites now throw instead of skipping, so a broken runner cannot pass with them off. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Production signs in with a static BENCH_DB_PASSWORD for the read-only bench_read role, which has no rds_iam grant (migration 005). The IAM path in lib/db.ts never ran there. getPool() now requires BENCH_DB_PASSWORD through requireEnv, so a missing or empty value fails at first use with a clear message. passwordProvider, the region setting and the rds-signer import are gone, and the pool gets the password as a string. DbConfig stops being exported. Remove @aws-sdk/rds-signer. bun.lock loses only the signer and its transitive packages; nothing else is re-resolved. The README and the deploy runbook now list BENCH_DB_PASSWORD as required and drop BENCH_DB_REGION. BENCH_DB_REGION on the benchmarks-website Vercel project is now unused. Deleting it is a settings change for Will, so this commit leaves it. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
collectGroupCharts called chartPayload once per chart, two queries each: about 200 round trips for a 99-chart TPC-DS group. groupChartPayloads builds the same payloads from one statement for the last n commits and one for every chart's rows, a LATERAL join per chart. A chart's x-axis is the last min(n, total) commits of that shared window, so the JSON is byte-identical. chartPayload is unchanged and is the tests' oracle. On one connection the brief's SQL lost to the old 7-wide fan-out, so: - The commit count is an aggregate lateral plus a width_bucket binary search. A scalar subquery was flattened and ran once per output row. - A query chart's first row is a skip scan of idx_query_measurements_summary. MIN read ~27K heap rows per chart. - The compression families split charts by NULL variant, so each pin is sargable. IS NOT DISTINCT FROM planned a semi join per fact row. - Rows come in slices of 25 charts, one after another. One result for TPC-DS at ?n=all peaked at 1.6x the old memory. On the 2026-10-28 dump every group and window matches the old path. At n=100, warm: TPC-DS (NVMe) SF=1 1587 -> 321 ms, Clickbench 865 -> 143 ms, Compression 70 -> 120 ms. At ?n=all, TPC-DS takes 5.3 s against 4.9 s, with peak memory at most 1.34x the old. QUERY_CONCURRENCY now bounds only the summaries in collectGroups. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
groupChartPayloads ran the commit window and the first rows slice on two pooled connections, and later slices after them, each in its own READ COMMITTED snapshot. An ingest landing between them could give payloads the per-chart path never could: a commit counted in total_commits but missing from commits[], or a short chart losing its own first commit. Now the window and every rows slice run one after another on one client, in a REPEATABLE READ, READ ONLY transaction, rolled back on error, with the client always released. The clamp that patched one direction of the skew is gone. A group still runs two statements for up to 25 charts; a TPC-DS group (99 charts) runs five. The group fill now holds one more pool connection, after the summaries finish. New window-edge tests pin byte-equality with chartPayload for a random-access group, a vector-search group with a 0.1 threshold, and a compression-size group with a non-NULL variant block, with per-chart windows. The statement-count tests now read the pooled client. performance.md states the ?n=all memory as measured: under 1.5x, 1.2 to 1.8 GB against 1.3 GB. On the 2026-10-28 dump every group still matches the old path at n=100 and n=50. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Move lib/chart-store.ts into lib/chart-store/: queues, bundles, global-filter and group-store, with index.ts re-exporting each module so '@/lib/chart-store' keeps its export surface. The code moves verbatim. group-store reads the global filter through getGlobalFilterSnapshot() instead of the module variable. Add a pin test for the chart island before it is split: under StrictMode, one wheel event pans once, and unmount removes the listener. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Move the ChartController class and its helpers out of components/Chart.tsx into components/chart/: types, datasets, tooltip, config, card-dom, pan-zoom and controller. Chart.tsx keeps the 'use client' boundary, the architecture notes, ChartIslandProps and the Chart component. The code moves verbatim. The moved private methods (the slider, badge and window-chip renderers, wheel pan and the range strip) take the state, elements and abort signal as parameters, and the controller keeps a delegator with the same name for each. maybeConstruct gets the Chart.js config from lineChartConfig(). Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
PALETTE, colorFor, assignStableColors and labelForCommit in lib/chart-format.ts, and commitWindowUrlValue in lib/window.ts, have no importer outside their own tests. Delete them and their six tests. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Move noUncheckedIndexedAccess and exactOptionalPropertyTypes from the site's tsconfig.json into tsconfig.base.json, so bench inherits them. The site's effective config does not change. TypeScript 7 reported 162 errors in bench: 41 in app code, 121 in tests. The fixes change types only: - Optional properties that receive undefined get `| undefined`. - An index access gets `!` only where the index is in range by construction. commitDateLabel gets an explicit undefined check that returns '' as before. - Chart.js's types do not let a line chart or its datasets widen to the generic types under exactOptionalPropertyTypes, so three casts at that boundary bridge them. - Tests use `!`, or `?.` where Biome forbids `!` after `?.`. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
lib/chart-store.ts is now lib/chart-store/, and the Chart.js wiring left components/Chart.tsx for components/chart/. Fix the links and file references in AGENTS.md, performance.md (two links were broken) and design-decisions.md, and the chart-format.ts header. Two moved comments named the wrong place: config.ts now names pan-zoom.ts for the wheel listener, and card-dom.ts names renderDownsampleBadge. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
payloads[i]! in collectGroupCharts and values[vi]! in normalizeChartPayload removed null from the type as well as undefined. A chart with no rows is null, so the null check that follows looked dead to the compiler. Cast to the element type instead; this removes only undefined. Also state the parallel-length precondition of lttbIndices, pin the undefined-month and undefined-day paths of commitDateLabel with a test, and correct CLAUDE.md: the two strict flags now live in tsconfig.base.json. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
groupChartPayloads and the cache-fill lock hold a client from pool.connect(). pg emits a dropped connection as 'error' on the client, and the pool listens only while the client is idle, so a drop during a group build or a locked fill was an uncaught exception. pool.query, which the build used before, attaches such a listener for its query. Both now listen while they hold the client, and release it with the error so the pool discards it. The lock skips its unlock when the connection dropped, since the session took the lock with it, and a failed unlock now discards the client too, so its session cannot keep the lock. The success path does not change. Tests: a build whose backend is terminated rejects with no uncaught exception and leaves no client checked out; the lock client cases use the existing mocks. The statement checks now also assert that every statement ran on one client and that no listener is left behind. windowCommits' doc names the snapshot it runs in. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
The pool-size test wrapped pool.query, so it could not see the group build's connect()ed client, and its comment still described a fan-out of one query per chart. It now runs a fill under the advisory lock, as the Data Cache does, with nine query groups so the summaries fill all seven slots, and counts connection requests: at most 8, the lock and seven summaries. It counts requests, not checked-out clients, because a pool of 8 never checks out more than 8, so a wider fan-out would only queue. The PR 2 plan gets an "As built" note under Task 3: one REPEATABLE READ transaction on one client, sequential statements, slices of 25 charts, and the four SQL changes. The bench-checks comment no longer says a runner without Docker would skip the suite silently; the harness fails it in CI. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
🔦 Lighthouse audit
Run |
TruffleHog reports the test-harness parser's made-up connection URL (and its copy in the PR 2 plan) as an unverified Postgres result, which fails the scan. Mark both lines with trufflehog:ignore. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
main squash-merged the bench toolchain PR (#128). Its tree equals that PR's last commit, f92f33a, which this branch already contains, so the merge keeps this branch's tree. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
|
Bench preview for |
The design and plan documents for the toolchain move and the cleanup were working notes for those changes. The code, the architecture docs and CLAUDE.md now record what they decided. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
lwwmanning
added a commit
that referenced
this pull request
Oct 29, 2026
main squash-merged the bench cleanup PR (#129). Its tree equals that PR's last commit, 4365dee, which this branch already contains, so the merge keeps this branch's tree. Co-Authored-By: Claude Opus 5.5 <[email protected]> Signed-off-by: Will Manning <[email protected]>
This branch was successfully deployed
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
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.
Cleanup for bench with no visible change. #128 moved bench onto the site's toolchain; this PR builds on it.
A group's charts now come from one read-only snapshot in two statements (five for TPC-DS) instead of two per chart. A cold TPC-DS group drops from 1.6 s to 0.3 s on the production dump, and the JSON stays byte-identical. The unused IAM sign-in path and
@aws-sdk/rds-signerare gone, with eight transitive packages.Chart.tsxand the chart store are split into modules, and bench adopts the site's stricter TypeScript flags. The integration suite can run against a local Postgres, and CI now fails instead of skipping it. The working plans inbench/docs/plans/are removed; the code and the architecture docs record what they decided.Parity against main after #128: no difference in 881 responses and 28 screenshots.
🤖 Generated with Claude Code