Repository navigation
Code review redesign - #3383
Draft
backnotprop wants to merge 508 commits into
Draft
Code review redesign#3383backnotprop wants to merge 508 commits into
backnotprop wants to merge 508 commits into
Conversation
… feat/review-power-up
…ields get real labels The whole row used to dim, leaving the title at 3.5:1 and the hint at 2.5:1. The 10 px caption inside each key field read like a placeholder; each field now has a 13 px label above it. Claude-Session: https://claude.ai/code/session_01VJ6b1nJDj5YogGsP3Ymd9m
…abs hide whole into a counted menu, titles keep their extension, the view switch folds in a narrow pane Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
…slot The reserved pair is drawn as two disabled arrows at reduced opacity, so the header has no 60 px hole between Approve and the split button and still never shifts. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
…eight The Unicode ∆ becomes lucide Diff at 14 px, in the muted text colour like the row's label. The panel action rows (All changes, PR pages, Semantic diff, Call flow) are 24 px like the tree rows beside them, and 44 px in the phone shell. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
In the phone shell the Changes and All files trees use 44 px rows (the tree's itemHeight option, keyed so it redraws when the shell changes), and the viewed circle and git add targets are 44x44 through the ::part styling seam; the icons keep their size. Desktop rows stay 24 px. The Pierre patch is unchanged. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
The trail's badge column was reserved on every row whenever the tree was grouped, only for the group headings' viewed/total, so names like monthly-report.ts were cut with about 30 px free. Headings now get their own trail columns (::part(row-trail-heading)); file and folder rows reserve a badge column only when a file has a badge. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
'Claude · Sonnet 5.5 · …' cut the model. The summary line now gives way in the engine name and the options first; the model name never shrinks (it only truncates if it alone is wider than the row). The whole line stays in the title. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
… right after the last one shown; All changes never shrinks Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
… the gear; hover label stays below the file header Auto shows the strip when the content is taller than three screens, or taller than one screen when a file has more than 1,000 lines. The gear's switch becomes Minimap: Auto / Always / Off. The hover label follows the pointer and is clamped below the sticky file header. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
…dge pixel on its own side At DPR 2 a 100% image is smoothed across each edge (rgb 159,96,159 is a quarter of the neighbour, as bilinear filtering gives); a one pixel drift would read the neighbour whole. The test now reads the half of each edge pixel that lies away from the edge, and saves an edge crop when asked. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
…l edge It used to show whenever the pointer was anywhere over the panel, so with the Pierre tree it sat on the count column (−10). It stays inside its panel, beside the edge, and appears on hovering the edge or on keyboard focus. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
…rker The label lost its truncate class when the scope switch stopped truncating, so review-sidebar.spec.ts:209 waited for an element that no longer exists. It now measures [data-tree-scope-label]; the spec passes (4 of 4) against this branch's binary. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
…hichever switch shows Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
… Power Up; one label per key field Claude-Session: https://claude.ai/code/session_01VJ6b1nJDj5YogGsP3Ymd9m
DiffDisplayOptions.tsx: take the minimap's Auto / Always / Off row in place of its old switch, with the critic's divider spacing (my-2 -mx-2). ReviewDockRightActions.test.tsx: the switches are now Wrap, Line numbers, Backgrounds, Whole file, Hide whitespace, and semantic diffs' Show moved & renamed code (the minimap is a segmented row, not a switch). Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
App.tsx: take the layout's HeaderBranch for the repo header; keep the critic's gap-2 on the phone's branch line inside HeaderBranch. FileHeader.tsx: keep both the group/file-header class (the quiet open-in and hover reveals use group-hover/file-header) and the layout's @container/file-header. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
…agent's icon as the Inbox does The Inbox's own marks (ui inbox-thread HostMark and AuthorMark): the brand mark per agent, the neutral glyph for an unknown one, the reply glyph for your own messages, in the Inbox's 24 px message-header tile with no added height. ui's mention picker takes an optional mark per person. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
…fitted to their room, keeping the extension, with the ellipsis against it Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
FileTree.tsx imports: keep the critic window's focusFileHeaderWhenLost and the critic tree's useCompactTouchLayout. Claude-Session: https://claude.ai/code/session_01VJ6b1oES3jfkCyDmf8uKzv
# Conflicts: # packages/inbox/parts/MessageBodyView.tsx # packages/ui/HANDOFF.md # tests/dom-test-list.txt # tests/inbox-size-ceilings.json
…repo engine Delete packages/sem-diff and vendor/tree-sitter-kotlin. Shared, review-editor, the review app and the Pi extension (dev) depend on the package. The wasm list and grammar map come from @plannotator/sem-jiff/grammars, which adds Swift; the review build copies the wasm with the package's sem-jiff-copy-wasm, the binary embeds Swift, and Pi vendors the package's src files.
…he file name whole A rename entry takes its names from the engine's rename spans, one row per renamed name, keeping PHP's $ and Ruby's @ and : sigils. The file header in the Sem tab cuts folders first and fits the file name with the shared EllipsisName fitter.
Replaces the sem.4 tarball with sem.5 (sha256 15b785697eb0...), fork.patch from pierre-diffs (git diff 3eb8f393, version 1.7.0-beta.3-sem.5), and FORK.md's change (5). Light bands are unchanged at 15%; dark bands are 32% alpha with a 1 px 50% edge.
… left edge on both sides Split view, All changes: a file with a long line whose change is far right, drawn only once wheel steps bring it near, with the pointer over the old side and over the new side. Fails on integration/review-next (both sides land 13848 px sideways).
All changes ran revealFirstChange on every file's first render, and a file is first rendered when scrolling brings it near. So a vertical wheel step scrolled a file with a far-right change sideways, and the split view's side sync and the bottom bar carried it to the other side. The reveal now runs only for files drawn before the reader scrolls: the ones the view opens on.
…the dark edge Replaces the sem.4 tarball with sem.6 (sha256 635cb63d7557...), which carries sem.5's dark-theme bands (32% fill with a 1 px 50% edge; light unchanged at 15%) and change (6): a band whose ends span the same rows carries data-connector-level and draws without the edge, which outlined it as a rectangle. fork.patch is git diff 3eb8f393 on plannotator/diffs-fork at 006a785d.
Pin the four dependency lines to the published 0.2.0 and exempt the package from the minimum release age, like the other @plannotator packages.
…at/sem-jiff-package
… new renames test semDiffChanges.renames.test.ts (from integration) imports the package. The code review guide says where the engine lives, how it is wired and where its THIRD-PARTY-NOTICES file is.
The Sem tab's file header sits under tabular-nums, which in Inter widens the hyphen from 6.05 to 8.41 px; the canvas measure ignores it, so a cut name with hyphens drew past its box. Fitted names and the Sem path use normal numerics.
…x/sem-tab-rename-names
… view Next/previous file, a Guide outline jump, a Sem tab entry and a comment jump now mark their file to reveal its first change once its changed characters are drawn (within 2 s), whether the jump renders the file or lands on one already drawn. Wheel and other passive scrolling still never reveal. A tree click opens the file in its own tab, where DiffViewer already reveals it. The browser proof adds a tree click and a next-file jump after scrolling: the far-right file shows its change and the other files stay at the left edge.
# Conflicts: # bun.lock # package.json # vendor/pierre-diffs/FORK.md # vendor/pierre-diffs/fork.patch
This was referenced Nov 16, 2026
…tions, Collapse all files in the tree header, Pierre-like tree rows
# Conflicts: # apps/pi-extension/vendor.sh # packages/review-editor/components/DiffViewer.tsx # packages/review-editor/components/FileHeader.tsx # packages/ui/components/OpenInAppButton.tsx # tests/inbox-size-ceilings.json
…s, wrap the tools row when it does not fit Compact rows (under 220 px) hide only the numbers; Git add moves into the lead beside the viewed mark. The tools row wraps whenever the pills and tools do not fit, so Collapse all files never draws over All at 256 px with a long count. Merge the viewed pressed rules, fix the File kinds heading order and the touch wording in the guide, drop the spec header lines.
# Conflicts: # apps/hook/server/index.ts # apps/pi-extension/vendor.sh # tests/inbox-size-ceilings.json
This branch has not been 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.
The code review redesign as one branch. It merges every piece below (merge commits only, nothing rebased), plus main, so the redesign can be tested in its final state and land together. It lands as a merge, not a squash, so each piece keeps its history. Draft: the owner decides when it lands.
What it contains
@plannotator/sem-jiff0.2.0 (feat/sem-jiff-package); the in-repopackages/sem-diffis gone. It reads TypeScript and TSX, JavaScript, Python, Go, Rust, Java, C#, PHP, Ruby, Kotlin and Swift. The separate language PRs are closed.], Sem tab) still opens with its first change in view (fix/wheel-sideways). It comes with the Pierre semantic build at sem.6: a band on one row (an edit or rename in place) draws without the dark edge.semtool (Review: remove the old sem semantic-diff tool; the built-in engine feeds Sem alone #3394)Where the pieces met
Each place where two pieces met was resolved by hand on this branch. These are where bugs are most likely, so the list stays as a record. The full list, with each conflict, is in the integration notes (
integration.md, outside the repo). The main ones:AllFilesCodeView.tsx).OriginalFileView.tsx, whole-file on the old side) and in the panes' Current view, off on phones.FileHeader.tsxkeeps the layout's structure with the critic's count format, narrow-width menu, ligatures and overflow fixes. On phones it now keeps the critic's padding when the pane has tabs.extraTabs).packages/inbox/components/Thread.tsx(the Inbox window stack + live window): the review anchor line and Open the review kept; the reply-feedback imports went with main's move touseBodyText.apps/hook/server/index.ts: live review imports kept alongside main'sGIT_REPO_ENV.AllFilesCodeView.tsxkeeps Review: large diffs on par with Pierre #3391's whole-file prefetch and warm-up; semantic diffs start from the prefetched text once it fits the patch, and their item swap waits for scroll rest.FileTree.tsx, which the two-pane layout uses, with its space reserved.package.jsonis checked against main line by line each round.tests/dom-test-list.txt: both sides kept once each;FileTree.redraws.test.tsxdropped with the old tree.lineto "Open in" in the file header. Here the header's Open in button was moved (left of Viewed, with its extra menu items), so the line goes to that button and to the phone header's menu (CompactFileActions.tsx) instead of to main's old spot, which is dropped.DiffViewer.tsxkeeps main's line reader beside the minimap switch.apps/hook/server/index.ts: main'sloopback-no-proxyimport stays first; the live review import follows it.apps/pi-extension/vendor.sh: this branch's list (notourorsemantic-diff, plusreview-live,sem-diff-assets,sem-diff-opinion) plus main'sopen-in-locate,path-segmentandloopback-proxy.packages/guide-viewer/GuideFileCard.tsx: main's Collapse all fix (Guide viewer: Collapse all works after switching views #3467) and this branch's header heights merged cleanly; read by eye.tests/inbox-size-ceilings.json: the ceilings are the merged sizes (App.tsx6600,Thread.tsx6056,AttachmentPane.tsx3207). None is above main's.What was verified, and where
All runs were local, on the candidate's own build, with one browser turn per spec or probe. Speed is compared with Pierre's demo and the previous round, measured back to back in the same window.
Round 34j (5246aff), the candidate: 34i plus the review tree polish (f257789) and main at 341ab8d (06d1d29, then 341ab8d, with Open in: JetBrains IDEs, at the line #3372, Keep machine state out of config.json (#2852) #3458 and Guide viewer: Collapse all works after switching views #3467). Checks on the final head:
packages/server/inbox-open-in.test.ts:652expectsopen -a WebStorm <path>where main's launcher now givesopen -na WebStorm --args --line 12 <path>, and it fails the same way on a clean main (341ab8d). dom-2'susePaneOpens.test.tsx:166is a garbage-collection check that passed 7 of 7 three times alone. The first heavy-agents and heavy-mod runs failed before the build existed; with the build they are green.App.panelView.test.tsx: this branch's two grouping checks read the tree before it mounts (it mounts after the first code or 400 ms), so they passed in their lane but failed alone. The test now waits for the tree; it passes alone 3 of 3.Browser specs on 34j's build (5246aff), one browser turn per file:
:341skips itself here (no ripgrep)inbox-guides runs with the binary named
plannotator: its stop step checks the process name.Speed on 34j (5246aff), three runs at load 7 to 9.4, with 34i and Pierre's demo in the same turns, the order alternating:
Eight more pairs on 200 files, alternating, at load 7.7 to 10.3: 34j 88 to 94 ms (median 90; one 119 on the first launch of a turn), 34i 87 to 91 ms (median 89.5), so no change. Load layout shift 0.0055 on all three runs.
Round 34i (9b25b46), the candidate: 34f plus the sem-jiff package switch, the Sem tab name fix, the wheel fix with Pierre sem.6, then main at 3810fba (11 merges). Checks before the main merge (4dcab86):
scripts/install.test.tsalone three times: 327 of 327 eachAfter the main merge (9b25b46): typecheck green, guards 20 (Inbox structure on the merged result), heavy-mod and heavy-agents (13) green, unit-1 6661, unit-2 7719, unit-3 6692, dom-1 1787, dom-2 1901, dom-3 1736, dom-4 1899, all passing.
Browser specs on the merged build (9b25b46), one browser turn per file:
:341skips itself here (no ripgrep)Speed on 34i (4dcab86), three runs each, at load 3.5 to 4.7, with 34f and Pierre's demo in the same turns:
Eight more pairs on 200 files, 34i and 34f alternating: 34i 85 to 94 ms (median 87.5), 34f 86 to 92 ms (median 87), so no change. Load layout shift 0.0055 on all three runs.
Round 34f (06a89a1): round 34e with two changes. First, the vendored Pierre semantic build is now sem.5: connector bands read in dark themes at 32% alpha with a 1 px edge, and light themes are unchanged. Second, Sem tab renames read old → new in every language, and long paths keep the file name (fix/sem-tab-rename-names). Checks on it: pierre-drift matches upstream, typecheck green, review DOM 641/641, semDiff 69/69, review-semantic-diffs 4/4. The 34e shards stand for it, because neither change touches the Inbox.
Round 34e (0d4fb01): round 34d plus main at 233c27f (13 Inbox merges). Checks on it:
:341skips itself here: no ripgrep)Round 34d (f76bb0a): round 34c plus a fix for a race in the All changes view's line target: a click in the Sem tab could land on a file whose header never rendered (fix/sem-click-header). It also has a layout-proof dark-page check in
packed-page-load.spec.ts: after mount it counts light pixels instead of reading the centre pixel. Checks on it: typecheck green, review DOM 640/640 in the review-editor list, and Inbox shard 13 (wherereview-semantic-diffs.spec.ts:211had failed twice) run whole twice: 149 passed, 1 skipped, both times.Round 34c (84a8f5d): round 34b plus the Sem tab's name finder for all eleven languages (fix/sem-names-per-language): typecheck green, review DOM 1059/1059, unit and guard 884, semDiff tests 40/40, review-semantic-diffs 4/4.
Round 34b (d226398): round 34 plus the fix for the active tab name (fix/tree-name-layout cd0b074). Checks on it:
:341skips itself here: no ripgrep binary on this Mac)Round 34 (ed4c43b):
npm-check-updatesabsent from bun.lockEarlier rounds' full lanes: round 31: dom-1 1720, dom-2 1811, dom-3 1683, unit-1 6920, unit-3 6918, heavy-mod and heavy-agents green. Round 33: dom-1 1567, dom-2 2057, dom-3 1891.
Browser specs on round 34:
:217, the cut tab name, fixed in 34b)Speed on round 34, three runs each, interleaved with Pierre's demo in one turn:
Load layout shift: 0.0066 to 0.0072 on the sample and the demo fixture, 0.0035 on the worst fixture (bar under 0.1). Details:
perf-vs-pierre.md.Break-UI recheck on the five worst-case fixtures: 28 of 28 rows fixed (
break-ui/recheck-2.md), with row 9 reopened on round 34 and closed again on round 34b.Critic findings and what each fix answered:
critic/findings.md.Those notes are in the owner's review-redesign briefs folder, not in the repo.
Known open
inbox-reply-draft-files.spec.ts:80is main's own race. Main's binary fails it 2 of 3.packed-page-load.spec.ts:74was the spec's single-pixel read; the spec is fixed in 34d.review-explorer.spec.ts:217is fixed in 34b.review-semantic-diffs.spec.ts:211is fixed in 34d; shard 13 is green twice.inbox-keyboard-journeys.spec.ts:560is the known Mac-only failure.packages/server/inbox-open-in.test.ts:652(Open in WebStorm at a line, the macOS launch) fails on main at 341ab8d.packages/inbox/reviewPaneGuide.test.tsxfails its first one or two tests when run alone in a cold process on main as on this branch (its reader loads lazily and the test waits six ticks); in its lane it passes.packages/server/inbox-ended-reply.test.tsfails on main too (1 of 20 there, 3 of 20 here, run alone).inbox-pull-refresh.spec.ts:165failed once on GitHub's runner. It passed 12 of 12 locally against this branch's binary and against main's.Dependency changes
The lockfile is main's, plus what the redesign adds:
sem-jiff-copy-wasmcommand. A grammar loads only for files of its language. No tree-sitter npm packages and no vendored grammar remain in the repo.bunfig.tomladds@pierre/treesand@plannotator/sem-jifftominimumReleaseAgeExcludes, beside the other Pierre and Plannotator packages.One side effect reaches beyond code review:
packages/inbox/package.jsonpins 4.5.0 since the move to @pierre/diffs 1.7.Fork notes
patches/FORK.md: the patches on top of upstream packages.vendor/pierre-diffs/FORK.md: our vendored fork of@pierre/diffs(1.7) and what differs from upstream.Behaviour changes
/api/agents/capabilities. Git status is folded into the tree.semtool is gone, so the built-in engine feeds the Sem tab alone. Languages the engine does not read show the plain diff.PLANNOTATOR_SEM_PATHandPLANNOTATOR_SEM_VERSION,GET /api/semantic-diffin both servers, thesemanticDiffreply field, the Analysis "Semantic changes" switch, and the "sem · N" badge./api/diff,/api/diff/switch,/api/pr-switch,/api/pr-diff-scope, in both servers) carriesimageSizes. The server reads them within 250 ms and leaves them out if it takes longer./api/review-image-sizesstays as the fallback, and SVG sizes are now read too.pointerEventsOnScroll).Agent-facing wording changes
Tool descriptions, skills, CLI
--help, the mod's contract andINBOX_TOOL_COSTare unchanged. New or changed text an agent reads:/api/agents/capabilities{ id: "tour", name: "Code Tour" }TOUR_REVIEW_PROMPTand its schema deletedprompts.md(A review comment on \` lines - (<new|old> side, at ):+ the lines in a fence +Answer it with reply_to .)`(A code review of N comments. Answer each comment with reply_to its id, so the answer lands on its lines; answer the review as a whole with reply_to <lead id>.), then per commentComment <id> on <place>:+ fence + wordsN comments on the code.(On the review thread about \` line N ( side, at ); answer with reply_to .)`The person resolved it.when resolvedreview comment on <place>[, resolved]Review: <branch or PR title>I approved the review., orI approved the review, with notes:+ the exportThe live-thread wording was tested clean room in #3338: a fresh agent given only the delivered text answered each comment by its own id.