Terrain: beauty/topo mode split, contour overlay, real grid LOD + mobile auto-LOD, pan restored#98
Open
AdaWorldAPI wants to merge 4 commits into
Open
Terrain: beauty/topo mode split, contour overlay, real grid LOD + mobile auto-LOD, pan restored#98AdaWorldAPI wants to merge 4 commits into
AdaWorldAPI wants to merge 4 commits into
Conversation
…nyon) Push the /garmin canyon toward OpenTopoMap quality — from data we already bake, no new download. ## Contour-line overlay (the big OTM-look step) Contours are Kind::Line features of GeoKind::Contour, so the overlay reuses the whole DRP1 pipeline: - garmin_bake: `--contour-level N` (default 3 ≈ 8k lines; level 4's 50k is too heavy for a browser overlay) emits `<stem>.contour.soa` via `drape::build_drape(&[Contour])` + `encode_drape` with a darker topo-brown palette (CONTOUR_LINE = [128,94,60]). - cockpit-server: `/api/garmin-contours/:location` + `resolve_garmin_contours` + test — a line-for-line mirror of the drape route (independent optional map). - GeoHelix.tsx: fetch contours (non-blocking) → `decodeDrape` → a thin semi-transparent LineSegments riding just under the road/river drape; a live `contours` toggle. Reuses the drape decoder + lazy-build-in-tick pattern. - body.manifest.json: `garmin_contours` entry for grand-canyon. ## Lake-vs-river fix The ×2 river widening (#97) also fattened the still lakes into chunky blobs. Probe found the discriminator is the Garmin type code, not shape: 0x4c is the flowing river-fill (the Colorado); 0x41/0x46/0x47 are still lakes. New `terrain::river_fill_grid()` (RIVER_FILL_TYPE = 0x4c, tested) stamps only the river; the bake dilates that mask ×2 and leaves lakes at their 1-cell stamp, so the Colorado stays bold while lakes shrink to OTM-style pinpoints. ## Verify - geo-hhtl lib: 25/25 tests (river_fill split + dilation). cockpit: tsc + build clean. Headless /garmin/grand-canyon: contour lines etch the relief like a topo map; both toggles work (features/contours); lakes are pinpoints. - cockpit-server could NOT be compiled locally: the rusty_v8 static-lib download host is blocked by org egress policy (403) and the cache is empty — CI will build it. The route change is a mechanical mirror of the deployed drape route. - Contours are a garmin_bake feature (the canyon); Iceland's raster DEM has no vector contours (deferred). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012jEwwaT5JZ5x8qWvcnaMYC
"The contours are a map layer, not the skin of the world." Separate the two
looks instead of drawing topo lines over the vivid surfel grade:
- BEAUTY (default): contours OFF by default. Still-water tanks/ponds on the
arid canyon are no longer bright blue lakes — the --arid bake retags them to
a new 10th palette slot (LAKE_TAG=9, LAKE_SUBTLE=[122,130,134]), a
barely-there grey fleck deliberately below the shader's blue-dominance `wet`
threshold so no vivid water treatment fires (1081 cells). Blue is reserved
for flowing water: the Colorado + perennial 0x4c tributary reaches. The
ver-8 wire carries its palette count (nK u8) and the client indexes
palette[kind] from the wire, so the extra slot is decode-safe.
- TOPO (the toggle, renamed `topo`): contour lines + a uTopo shader swap to
cartographic paper — pale beige-white relief ramp, green vegetation-KIND
tint, light carto-blue water, gentle hillshade. Classic OTM styling; the
beauty grade never shows under the line web. uTopo is tied to actual
contour-layer visibility in the tick, so the palette never swaps without
lines. Iceland (no contour sidecar) is untouched.
LOD honesty + stats HUD:
- Verified postLod() early-returns for every geo scene: the /api/body/lod
cascade culls by the ANATOMY body's block-bounds, so on /garmin/* the LOD
toggle was an inert placeholder. The button is now disabled there
("LOD n/a") with an explanatory tooltip; a per-scene .blocks cascade stays
future work.
- New stats HUD (bottom-right, ≤2 Hz, renderer.info ground truth): rendered
tris / line segments / draw calls / decoded verts / LOD status. Observed:
beauty = `tris 1.73M · lines 265k · calls 3`; topo on = `lines 745k ·
calls 4` — layer toggles provably do real work.
Rebaked canyon.v8grid{,.drape,.contour}.soa.gz (the drape/contour DRP1
palettes gained the 10th slot). geo 25/25 tests; tsc + vite build clean;
headless-verified both modes + HUD + disabled LOD button.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012jEwwaT5JZ5x8qWvcnaMYC
…ile auto-LOD, sidebar Four user-reported issues on the /garmin terrain scenes: 1. SIDEBAR: hidden on terrain scenes — a grid's "structures" are its row strips (1,024–3,584 identical unnamed entries), noise for a map. Anatomy / building scenes keep the searchable list. 2. PAN: right-drag (desktop) and two-finger drag (touch) now pan the orbit target along the camera basis; two-finger span change = pinch dolly; context menu suppressed. ROOT CAUSE of "two-finger did nothing" on mobile: the canvas had no `touch-action: none`, so the browser claimed multi-touch gestures before pointer events ever fired. pointercancel clears gesture state so browser-stolen touches can't wedge the pointer map. Single-pointer orbit math is unchanged (body scene unaffected). 3. REAL TERRAIN LOD: decodeGrid(buf, stride) sub-samples the ver-8 radix grid at decode time — grid cells are addresses, so stride 2 is exact selection (¼ verts/tris AND ¼ decode work), never resampling. The inflated wire is kept in a ref, so the LOD button re-decodes live without refetch. Grid scenes get a live toggle; mesh geo scenes stay honestly "LOD n/a"; the body's server cascade is untouched. 4. MOBILE AUTO-LOD: coarse-pointer + touch devices start with LOD on — a phone never pays the full-grid decode (the reported 35 s on the 16.5M-vert Iceland grid; ½-res is ~¼ the decode work and ¼ the GPU load). Headless-verified (log in the plan doc): sidebar gone; HUD proves the LOD toggle (tris 1.73M → 431k · verts 864k → 216k · "off · full grid" → "on · ½-res grid"); right-drag pan shifts the view (pixel diff 22.4); synthesized two-finger pointer events pan (diff 20.2); canvas touch-action none; mobile emulation (390×844, coarse+touch) auto-starts at ½-res with zero clicks. Known minor: toggling LOD re-decodes + remounts, which resets the camera. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012jEwwaT5JZ5x8qWvcnaMYC
… color model) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_012jEwwaT5JZ5x8qWvcnaMYC
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_83fc71be-b8e9-4c28-b403-58875cfc39ff) |
📝 WalkthroughWalkthroughAdds Garmin contour baking and serving, river/lake classification, client-side ver-8 grid LOD, TOPO rendering, contour overlays, touch gestures, and renderer statistics for terrain scenes. ChangesGarmin contour terrain flow
Estimated code review effort: 4 (Complex) | ~45 minutes Sequence Diagram(s)sequenceDiagram
participant TerrainScene
participant GarminContourAPI
participant GeoHelix
participant ThreeRenderer
TerrainScene->>GarminContourAPI: Request contour overlay
GarminContourAPI-->>GeoHelix: Return DRP1 gzip data
GeoHelix->>GeoHelix: Decode overlay and select grid stride
GeoHelix->>ThreeRenderer: Render terrain, contours, TOPO palette, and HUD
Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches📝 Generate docstrings
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (1)
geo/src/bin/garmin_bake.rs (1)
258-298: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick winContour sidecar duplicates the drape sidecar's build→encode→log shape.
The new contour block (277-298) repeats the same
build_drape→encode_drape→ statseprintln!pattern as the drape block just above (261-269), differing only in level/kinds/palette. Worth extracting a small helper (e.g.fn emit_overlay(name, dec, bbox, pos, w, h, level, kinds, palette) -> Vec<u8>) before a third overlay type repeats the copy-paste.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@geo/src/bin/garmin_bake.rs` around lines 258 - 298, The drape and contour sidecars duplicate their build, encode, statistics, and logging logic. Extract a reusable helper near the relevant code, such as emit_overlay, accepting the overlay name, decoder, bounds, terrain positions, dimensions, level, kinds, and palette, returning encoded bytes; move the shared build_drape, encode_drape, point counting, and eprintln behavior into it, then replace both blocks with helper calls while preserving their existing parameters and log labels.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@cockpit/src/GeoHelix.tsx`:
- Around line 151-165: Stride-based dimensions in decodeGrid do not ensure the
final full-grid row and column are sampled. Update the W/H calculation and
colF/rowF index generation so stride LOD always includes Wf - 1 and Hf - 1,
while retaining regular stride sampling and avoiding duplicate indices.
In `@geo/src/bin/garmin_bake.rs`:
- Around line 231-237: Replace the debug-only palette length check in the
palette construction logic with a runtime assertion that remains enabled in
release builds, preserving the invariant that LAKE_TAG is the next unused
palette slot before appending LAKE_SUBTLE. Keep the assertion tied to the
existing LAKE_TAG and palette symbols so any palette growth fails immediately
during baking instead of silently aliasing a real slot.
---
Nitpick comments:
In `@geo/src/bin/garmin_bake.rs`:
- Around line 258-298: The drape and contour sidecars duplicate their build,
encode, statistics, and logging logic. Extract a reusable helper near the
relevant code, such as emit_overlay, accepting the overlay name, decoder,
bounds, terrain positions, dimensions, level, kinds, and palette, returning
encoded bytes; move the shared build_drape, encode_drape, point counting, and
eprintln behavior into it, then replace both blocks with helper calls while
preserving their existing parameters and log labels.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro Plus
Run ID: 5562e997-404e-463e-a444-06ac3d871ddc
⛔ Files ignored due to path filters (3)
cockpit/public/canyon.v8grid.contour.soa.gzis excluded by!**/*.gzcockpit/public/canyon.v8grid.drape.soa.gzis excluded by!**/*.gzcockpit/public/canyon.v8grid.soa.gzis excluded by!**/*.gz
📒 Files selected for processing (6)
claude-notes/plans/2026-07-09-otm-contour-lines.mdcockpit/public/body.manifest.jsoncockpit/src/GeoHelix.tsxcrates/cockpit-server/src/main.rsgeo/src/bin/garmin_bake.rsgeo/src/garmin/terrain.rs
Comment on lines
+151
to
+165
| function decodeGrid(buf: ArrayBuffer, stride = 1): Decoded { | ||
| const dv = new DataView(buf); | ||
| const nC = dv.getUint32(6, true), W = dv.getUint32(10, true), H = dv.getUint32(14, true); | ||
| const nV = W * H, nT = (W - 1) * (H - 1) * 2; | ||
| const nC = dv.getUint32(6, true), Wf = dv.getUint32(10, true), Hf = dv.getUint32(14, true); | ||
| // Terrain LOD: sub-sample the radix grid by `stride` at DECODE time. Grid cells are | ||
| // ADDRESSES, so skipping is exact selection (never resampling) — stride 2 keeps every | ||
| // other height sample untouched and yields ¼ the verts/tris AND ¼ the decode work | ||
| // (the mobile fast path; the LOD toggle re-decodes live). Edge row/col clamp to the | ||
| // last full-grid sample so the tile rim survives. | ||
| const W = stride > 1 ? Math.ceil(Wf / stride) : Wf; | ||
| const H = stride > 1 ? Math.ceil(Hf / stride) : Hf; | ||
| const colF = new Uint32Array(W), rowF = new Uint32Array(H); | ||
| for (let c = 0; c < W; c++) colF[c] = Math.min(c * stride, Wf - 1); | ||
| for (let r = 0; r < H; r++) rowF[r] = Math.min(r * stride, Hf - 1); | ||
| const nVf = Wf * Hf; // wire counts (offsets/slices) | ||
| const nV = W * H, nT = (W - 1) * (H - 1) * 2; // display counts |
There was a problem hiding this comment.
🎯 Functional Correctness | 🟠 Major | ⚡ Quick win
Stride LOD drops the tile's rightmost column/bottommost row (rim not actually preserved).
The comment states the edge clamp keeps the tile rim, but Math.ceil(Wf/stride)/Math.ceil(Hf/stride) don't guarantee (W-1)*stride (or (H-1)*stride) reaches Wf-1/Hf-1. E.g. Wf=1024, stride=2 → W=512, last sampled column is index 1022, not 1023 — the true rim column is never read. This only triggers when stride > 1 (LOD active), i.e. the new mobile auto-LOD / manual LOD-toggle path this PR introduces.
🐛 Proposed fix — guarantee the last full-grid sample is always reachable
- const W = stride > 1 ? Math.ceil(Wf / stride) : Wf;
- const H = stride > 1 ? Math.ceil(Hf / stride) : Hf;
+ const W = stride > 1 ? Math.ceil((Wf - 1) / stride) + 1 : Wf;
+ const H = stride > 1 ? Math.ceil((Hf - 1) / stride) + 1 : Hf;📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
Suggested change
| function decodeGrid(buf: ArrayBuffer, stride = 1): Decoded { | |
| const dv = new DataView(buf); | |
| const nC = dv.getUint32(6, true), W = dv.getUint32(10, true), H = dv.getUint32(14, true); | |
| const nV = W * H, nT = (W - 1) * (H - 1) * 2; | |
| const nC = dv.getUint32(6, true), Wf = dv.getUint32(10, true), Hf = dv.getUint32(14, true); | |
| // Terrain LOD: sub-sample the radix grid by `stride` at DECODE time. Grid cells are | |
| // ADDRESSES, so skipping is exact selection (never resampling) — stride 2 keeps every | |
| // other height sample untouched and yields ¼ the verts/tris AND ¼ the decode work | |
| // (the mobile fast path; the LOD toggle re-decodes live). Edge row/col clamp to the | |
| // last full-grid sample so the tile rim survives. | |
| const W = stride > 1 ? Math.ceil(Wf / stride) : Wf; | |
| const H = stride > 1 ? Math.ceil(Hf / stride) : Hf; | |
| const colF = new Uint32Array(W), rowF = new Uint32Array(H); | |
| for (let c = 0; c < W; c++) colF[c] = Math.min(c * stride, Wf - 1); | |
| for (let r = 0; r < H; r++) rowF[r] = Math.min(r * stride, Hf - 1); | |
| const nVf = Wf * Hf; // wire counts (offsets/slices) | |
| const nV = W * H, nT = (W - 1) * (H - 1) * 2; // display counts | |
| function decodeGrid(buf: ArrayBuffer, stride = 1): Decoded { | |
| const dv = new DataView(buf); | |
| const nC = dv.getUint32(6, true), Wf = dv.getUint32(10, true), Hf = dv.getUint32(14, true); | |
| // Terrain LOD: sub-sample the radix grid by `stride` at DECODE time. Grid cells are | |
| // ADDRESSES, so skipping is exact selection (never resampling) — stride 2 keeps every | |
| // other height sample untouched and yields ¼ the verts/tris AND ¼ the decode work | |
| // (the mobile fast path; the LOD toggle re-decodes live). Edge row/col clamp to the | |
| // last full-grid sample so the tile rim survives. | |
| const W = stride > 1 ? Math.ceil((Wf - 1) / stride) + 1 : Wf; | |
| const H = stride > 1 ? Math.ceil((Hf - 1) / stride) + 1 : Hf; | |
| const colF = new Uint32Array(W), rowF = new Uint32Array(H); | |
| for (let c = 0; c < W; c++) colF[c] = Math.min(c * stride, Wf - 1); | |
| for (let r = 0; r < H; r++) rowF[r] = Math.min(r * stride, Hf - 1); | |
| const nVf = Wf * Hf; // wire counts (offsets/slices) | |
| const nV = W * H, nT = (W - 1) * (H - 1) * 2; // display counts |
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@cockpit/src/GeoHelix.tsx` around lines 151 - 165, Stride-based dimensions in
decodeGrid do not ensure the final full-grid row and column are sampled. Update
the W/H calculation and colF/rowF index generation so stride LOD always includes
Wf - 1 and Hf - 1, while retaining regular stride sampling and avoiding
duplicate indices.
Comment on lines
+231
to
+237
| // Slot 9 = LAKE_TAG. The ver-8 wire carries its palette count (nK u8) and the | ||
| // client indexes palette[kind] from the wire, so a 10th entry is decode-safe. | ||
| debug_assert_eq!(palette.len() as u8, LAKE_TAG); | ||
| palette.push(LAKE_SUBTLE); | ||
| eprintln!( | ||
| "arid palette: Stream drainage → rust-brown, Water river-blue, still lakes → subtle" | ||
| ); |
There was a problem hiding this comment.
🗄️ Data Integrity & Integration | 🟡 Minor | ⚡ Quick win
Palette invariant guarded by debug_assert_eq! is stripped in release builds.
If GeoKind::PALETTE ever grows beyond 9 entries, LAKE_TAG (hardcoded 9) would silently alias a real palette slot in a release bake — corrupting the shipped .soa/palette without any runtime signal, since debug_assert_eq! compiles to nothing in release.
🛡️ Proposed fix — use a runtime assertion that survives release builds
- debug_assert_eq!(palette.len() as u8, LAKE_TAG);
+ assert_eq!(
+ palette.len() as u8,
+ LAKE_TAG,
+ "GeoKind::PALETTE length drifted from LAKE_TAG — update the slot constant"
+ );📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
Suggested change
| // Slot 9 = LAKE_TAG. The ver-8 wire carries its palette count (nK u8) and the | |
| // client indexes palette[kind] from the wire, so a 10th entry is decode-safe. | |
| debug_assert_eq!(palette.len() as u8, LAKE_TAG); | |
| palette.push(LAKE_SUBTLE); | |
| eprintln!( | |
| "arid palette: Stream drainage → rust-brown, Water river-blue, still lakes → subtle" | |
| ); | |
| // Slot 9 = LAKE_TAG. The ver-8 wire carries its palette count (nK u8) and the | |
| // client indexes palette[kind] from the wire, so a 10th entry is decode-safe. | |
| assert_eq!( | |
| palette.len() as u8, | |
| LAKE_TAG, | |
| "GeoKind::PALETTE length drifted from LAKE_TAG — update the slot constant" | |
| ); | |
| palette.push(LAKE_SUBTLE); | |
| eprintln!( | |
| "arid palette: Stream drainage → rust-brown, Water river-blue, still lakes → subtle" | |
| ); |
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@geo/src/bin/garmin_bake.rs` around lines 231 - 237, Replace the debug-only
palette length check in the palette construction logic with a runtime assertion
that remains enabled in release builds, preserving the invariant that LAKE_TAG
is the next unused palette slot before appending LAKE_SUBTLE. Keep the assertion
tied to the existing LAKE_TAG and palette symbols so any palette growth fails
immediately during baking instead of silently aliasing a real slot.
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
2 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.
Builds on #97. Four commits: the OTM contour overlay, the beauty-vs-topo mode split, and a UX/architecture round (pan, real LOD, mobile, sidebar). All from data we already bake — no new downloads.
1. Contour-line overlay + lake-vs-river fix (
daac63d0)Kind::Linefeatures ofGeoKind::Contour, so the overlay reuses the DRP1 pipeline end-to-end:garmin_bake --contour-level N(default 3 ≈ 8k lines) emits a.contour.soasidecar; new/api/garmin-contours/:locationroute (mirror of the drape route, + test); client decodes with the samedecodeDrapeinto a thin semi-transparentLineSegments.0x4c= flowing river-fill = the Colorado;0x41/46/47= still lakes). New testedterrain::river_fill_grid()widens only the river ×2 — lakes stay pinpoints.2. Beauty vs topo mode split (
e7618a07)"The contours are a map layer, not the skin of the world."
--aridbake to a 10th palette slot (LAKE_SUBTLE) — a barely-there grey fleck below the shader's blue-dominance threshold, so no vivid water treatment fires (1,081 cells). Blue = flowing water only.topotoggle): contour lines + auToposhader swap to cartographic paper (beige-white relief, green vegetation tint, carto-blue water, gentle hillshade). The vivid grade never shows under the line web. Iceland untouched.postLod()early-returns on every geo scene (the/api/body/lodcascade culls by the anatomy body's block-bounds) — the LOD toggle was inert on terrain. New HUD (fromrenderer.infoground truth):tris · lines · calls · verts · LOD status.3. UX/architecture round (
474da00a)touch-action: none, so the browser claimed multi-touch gestures before pointer events fired.pointercancelclears gesture state. Single-pointer orbit math unchanged (body scene unaffected).decodeGrid(buf, stride)sub-samples the ver-8 radix grid at decode time — cells are addresses, so stride 2 is exact selection (¼ verts, ¼ tris, ¼ decode work), never resampling. The wire is kept in a ref; the LOD button re-decodes live without refetch.Verification (headless, pixel-diffed + HUD-scraped)
tris 1.73M → 431k · verts 864k → 216k · "off · full grid" → "on · ½-res grid".LOD on · ½-res grid, zero clicks.canvas.style.touchAction === 'none'.lines 265k → 745k · calls 3 → 4).tsc+ vite build clean.rusty_v8static-lib host is blocked by egress policy and the cache was empty) — the contour route is a line-for-line mirror of the deployed drape route; CI builds it.Known issue (accepted for now, follow-up filed in the plan doc)
Next round (post-merge, recorded in
claude-notes/plans/2026-07-09-otm-contour-lines.md)🤖 Generated with Claude Code
https://claude.ai/code/session_012jEwwaT5JZ5x8qWvcnaMYC
Generated by Claude Code
Summary by CodeRabbit
New Features
Documentation