Repository navigation
feat(deploy): hand the workflow to a coding agent as a Comfy API brief - #17966
Conversation
|
Navigate logical layers of code changes, visualize relationships, and explore their blast radius. Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Repository: Comfy-Org/ComfyUI_frontend/.coderabbit.yaml Review profile: ASSERTIVE Plan: Advanced Run ID: 📒 Files selected for processing (1)
Included review availability: This review used your included allowance. Your plan provides up to 8 included reviews per hour; 7 remain after this review. 📝 WalkthroughWalkthroughThe workflow deploy flow collects build inputs, generates and copies distribution-specific agent handoff instructions, and downloads the captured graph for Cloud. The deploy card shows workflow input counts and the copy action’s loading and success states. Clipboard copying returns a success value. ChangesWorkflow agent handoff
Priority: ➖ Normal Estimated code review effort: 3 (Moderate) | ~30 minutes Change: Feature Sequence Diagram(s)sequenceDiagram
participant DeployToComfyApiCard
participant useAgentHandoff
participant copyToClipboard
participant CloudDownload
DeployToComfyApiCard->>useAgentHandoff: copyBrief()
useAgentHandoff->>useAgentHandoff: Capture workflow and derive build inputs
useAgentHandoff->>copyToClipboard: Copy generated handoff brief
copyToClipboard-->>useAgentHandoff: Return copy result
useAgentHandoff->>CloudDownload: Download captured graph after successful Cloud copy
useAgentHandoff-->>DeployToComfyApiCard: Return copy result
Suggested reviewers: Merge Risk: 🔵 Low · up to This change adds a Build with your agent button that copies a brief for the open workflow. It has no known production defects. One test may be too weak to catch a missing upload description, so the owner should confirm that before merging. Security Architecture ReviewSecurity architecture risk: 🔵 Low · up to The handoff explicitly requests approval before uploads, pushes, releases, and retries, and separates release creation from deployment. No new vulnerability was established, but enforcement and recovery after the handoff remain unverified. Retained concerns Security review detailsSecurity Blast Radius
Security Findings and Attack Paths
Trust Boundaries and Controls
Resilience and Maintainability Implications
Hardening Proposals
🚥 Pre-merge checks | ✅ 6 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (6 passed)
Full details: End-To-End Regression Coverage For FixesExplanation The PR includes bug-fix signals in the commit subjects: ✨ Finishing Touches 💡 1🛠️ Fix failing CI checks 💡
🧪 Generate unit tests (beta)
Comment |
🎨 Storybook: ✅ Built — View Storybook🎭 Playwright: ✅ 2197 passed, 0 failed · 2 flaky📊 Browser Reports
📦 Bundle: 10.2 MB gzip 🔴 +827 kBDetailsSummary
Category Glance App Entry Points — 3.92 kB (baseline 4.46 kB) • 🟢 -550 BMain entry bundles and manifests
Status: 1 added / 1 removed Graph Workspace — 1.47 MB (baseline 2.15 MB) • 🟢 -683 kBGraph editor runtime, canvas, workflow orchestration
Status: 2 added / 2 removed / 1 unchanged Views & Navigation — 135 kB (baseline 154 kB) • 🟢 -18.7 kBTop-level views, pages, and routed surfaces
Status: 16 added / 17 removed / 2 unchanged Panels & Settings — 648 kB (baseline 671 kB) • 🟢 -23.1 kBConfiguration panels, inspectors, and settings screens
Status: 28 added / 30 removed User & Accounts — 53.7 kB (baseline 54.4 kB) • 🟢 -753 BAuthentication, profile, and account management bundles
Status: 11 added / 11 removed / 4 unchanged Editors & Dialogs — 147 kB (baseline 326 kB) • 🟢 -179 kBModals, dialogs, drawers, and in-app editors
Status: 9 added / 13 removed UI Components — 194 kB (baseline 201 kB) • 🟢 -7.24 kBReusable component library chunks
Status: 18 added / 24 removed Data & Services — 4.16 MB (baseline 3.07 MB) • 🔴 +1.09 MBStores, services, APIs, and repositories
Status: 13 added / 22 removed / 1 unchanged Utilities & Hooks — 587 kB (baseline 813 kB) • 🟢 -225 kBHelpers, composables, and utility bundles
Status: 31 added / 43 removed / 7 unchanged Vendor & Third-Party — 17.2 MB (baseline 14.2 MB) • 🔴 +3 MBExternal libraries and shared vendor chunks
Status: 7 added / 7 removed / 11 unchanged Other — 18.4 MB (baseline 18.7 MB) • 🟢 -335 kBBundles that do not match a named category
Status: 168 added / 206 removed / 141 unchanged ⚡ Performance Report
Show regressions
All metrics
Historical variance (last 15 runs)
Trend (last 15 commits on main)
Raw data{
"timestamp": "2026-10-08T20:54:00.190Z",
"gitSha": "ee535a4a3a0cffa4d53de661b2f2f12ba0780c2c",
"branch": "willie/deploy-to-comfy-api-agent-handoff",
"measurements": [
{
"name": "canvas-idle",
"durationMs": 2008.5809999999924,
"styleRecalcs": 10,
"styleRecalcDurationMs": 3.8209999999999997,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 400.86600000000004,
"heapDeltaBytes": 502332,
"heapUsedBytes": 68088620,
"domNodes": 0,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 8.266999999999996,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666115702479342,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-idle",
"durationMs": 2008.705999999961,
"styleRecalcs": 10,
"styleRecalcDurationMs": 3.88,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 399.927,
"heapDeltaBytes": 486724,
"heapUsedBytes": 69248208,
"domNodes": 0,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 7.512999999999999,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.665289256198346,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-mouse-sweep",
"durationMs": 1746.8880000000127,
"styleRecalcs": 74,
"styleRecalcDurationMs": 16.525,
"layouts": 12,
"layoutDurationMs": 3.724,
"taskDurationMs": 745.0859999999999,
"heapDeltaBytes": -2094464,
"heapUsedBytes": 66413816,
"domNodes": 0,
"jsHeapTotalBytes": 524288,
"scriptDurationMs": 154,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "canvas-mouse-sweep",
"durationMs": 1762.1239999999716,
"styleRecalcs": 73,
"styleRecalcDurationMs": 16.232999999999997,
"layouts": 12,
"layoutDurationMs": 4.389,
"taskDurationMs": 780.8980000000001,
"heapDeltaBytes": -2459244,
"heapUsedBytes": 65470096,
"domNodes": 0,
"jsHeapTotalBytes": 524288,
"scriptDurationMs": 159.789,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66698113207547,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-zoom-sweep",
"durationMs": 1702.1159999999895,
"styleRecalcs": 32,
"styleRecalcDurationMs": 9.494000000000002,
"layouts": 6,
"layoutDurationMs": 1.1550000000000002,
"taskDurationMs": 320.8299999999999,
"heapDeltaBytes": 3349508,
"heapUsedBytes": 71263520,
"domNodes": 53,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 10.587000000000007,
"eventListeners": 15,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.6656862745098,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-zoom-sweep",
"durationMs": 1708.7839999999233,
"styleRecalcs": 31,
"styleRecalcDurationMs": 9.120999999999999,
"layouts": 6,
"layoutDurationMs": 1.2309999999999999,
"taskDurationMs": 319.61300000000006,
"heapDeltaBytes": 3400956,
"heapUsedBytes": 71723336,
"domNodes": 53,
"jsHeapTotalBytes": 262144,
"scriptDurationMs": 10.541999999999996,
"eventListeners": 15,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666019417475724,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "dom-widget-clipping",
"durationMs": 643.7250000000176,
"styleRecalcs": 13,
"styleRecalcDurationMs": 3.1999999999999997,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 433.42,
"heapDeltaBytes": -8510292,
"heapUsedBytes": 59183788,
"domNodes": 0,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 159.526,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.665789473684214,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "dom-widget-clipping",
"durationMs": 687.0790000000397,
"styleRecalcs": 14,
"styleRecalcDurationMs": 3.5150000000000006,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 451.965,
"heapDeltaBytes": -8103852,
"heapUsedBytes": 60286740,
"domNodes": 0,
"jsHeapTotalBytes": 262144,
"scriptDurationMs": 165.107,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66585365853659,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "dom-widget-order-dense-n200-w200-i120",
"durationMs": 1140.515999999991,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 781.7180000000001,
"heapDeltaBytes": 15415220,
"heapUsedBytes": 84692376,
"domNodes": 0,
"jsHeapTotalBytes": 3047424,
"scriptDurationMs": 66.98799999999999,
"eventListeners": 0,
"totalBlockingTimeMs": 505,
"frameDurationMs": 33.335294117647074,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "dom-widget-order-dense-n200-w200-i120",
"durationMs": 1121.3579999999865,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 736.092,
"heapDeltaBytes": 15172404,
"heapUsedBytes": 84301184,
"domNodes": 0,
"jsHeapTotalBytes": 4096000,
"scriptDurationMs": 65.51799999999999,
"eventListeners": 0,
"totalBlockingTimeMs": 506,
"frameDurationMs": 33.839393939393936,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "dom-widget-order-sparse-n1000-w20-i120",
"durationMs": 847.1030000000042,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 460.94700000000046,
"heapDeltaBytes": 18204220,
"heapUsedBytes": 104450928,
"domNodes": 0,
"jsHeapTotalBytes": 266240,
"scriptDurationMs": 166.35399999999996,
"eventListeners": 0,
"totalBlockingTimeMs": 216,
"frameDurationMs": 25.0029411764706,
"p95FrameDurationMs": 16.80000000000109
},
{
"name": "dom-widget-order-sparse-n1000-w20-i120",
"durationMs": 864.0460000000303,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 461.1710000000002,
"heapDeltaBytes": 17850684,
"heapUsedBytes": 104115492,
"domNodes": 0,
"jsHeapTotalBytes": 266240,
"scriptDurationMs": 156.00800000000004,
"eventListeners": 0,
"totalBlockingTimeMs": 219,
"frameDurationMs": 24.76000000000001,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "extension-clean-combined-draw",
"durationMs": 8920.767999999953,
"styleRecalcs": 2,
"styleRecalcDurationMs": 0.4530000000000003,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 7967.155000000001,
"heapDeltaBytes": -5338688,
"heapUsedBytes": 72113292,
"domNodes": -171,
"jsHeapTotalBytes": -11407360,
"scriptDurationMs": 6766.447,
"eventListeners": -78,
"totalBlockingTimeMs": 661,
"frameDurationMs": 33.83712121212122,
"p95FrameDurationMs": 66.60000000000036
},
{
"name": "extension-clean-combined-draw",
"durationMs": 8934.38900000001,
"styleRecalcs": 2,
"styleRecalcDurationMs": 0.47999999999999954,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 8180.253000000001,
"heapDeltaBytes": 2574040,
"heapUsedBytes": 80131084,
"domNodes": -142,
"jsHeapTotalBytes": -11931648,
"scriptDurationMs": 6711.025,
"eventListeners": -78,
"totalBlockingTimeMs": 478,
"frameDurationMs": 32.483636363636364,
"p95FrameDurationMs": 50.100000000000364
},
{
"name": "extension-rgthree-prototype-draw",
"durationMs": 9011.544000000014,
"styleRecalcs": 2,
"styleRecalcDurationMs": 0.40900000000000136,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 8311.451000000001,
"heapDeltaBytes": 19287968,
"heapUsedBytes": 97131520,
"domNodes": -142,
"jsHeapTotalBytes": -8261632,
"scriptDurationMs": 6838.067,
"eventListeners": -78,
"totalBlockingTimeMs": 578,
"frameDurationMs": 33.27084870848709,
"p95FrameDurationMs": 50.100000000000364
},
{
"name": "extension-rgthree-prototype-draw",
"durationMs": 8608.189000000039,
"styleRecalcs": 1,
"styleRecalcDurationMs": 0.3359999999999995,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 7893.196,
"heapDeltaBytes": 2202592,
"heapUsedBytes": 80100244,
"domNodes": -142,
"jsHeapTotalBytes": -12193792,
"scriptDurationMs": 6475.809,
"eventListeners": -78,
"totalBlockingTimeMs": 426,
"frameDurationMs": 32.45169811320754,
"p95FrameDurationMs": 50.100000000000364
},
{
"name": "large-graph-idle",
"durationMs": 2014.7439999999506,
"styleRecalcs": 10,
"styleRecalcDurationMs": 5.371,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 478.60099999999994,
"heapDeltaBytes": -10202176,
"heapUsedBytes": 78361552,
"domNodes": 0,
"jsHeapTotalBytes": -225280,
"scriptDurationMs": 32.637,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.665289256198346,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "large-graph-idle",
"durationMs": 2022.8530000000546,
"styleRecalcs": 10,
"styleRecalcDurationMs": 4.086000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 479.36100000000005,
"heapDeltaBytes": -10486816,
"heapUsedBytes": 77532436,
"domNodes": 0,
"jsHeapTotalBytes": -225280,
"scriptDurationMs": 31.88200000000001,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666115702479342,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "large-graph-pan",
"durationMs": 2213.0960000000073,
"styleRecalcs": 70,
"styleRecalcDurationMs": 10.32,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1480.9889999999996,
"heapDeltaBytes": -18996520,
"heapUsedBytes": 70252592,
"domNodes": 0,
"jsHeapTotalBytes": 1085440,
"scriptDurationMs": 793.9020000000002,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "large-graph-pan",
"durationMs": 2208.3390000000236,
"styleRecalcs": 69,
"styleRecalcDurationMs": 10.392999999999999,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1505.7350000000001,
"heapDeltaBytes": -19417264,
"heapUsedBytes": 70142332,
"domNodes": 0,
"jsHeapTotalBytes": 1085440,
"scriptDurationMs": 809.879,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "large-graph-zoom",
"durationMs": 3538.088000000016,
"styleRecalcs": 73,
"styleRecalcDurationMs": 13.299,
"layouts": 60,
"layoutDurationMs": 13.644,
"taskDurationMs": 1585.3429999999998,
"heapDeltaBytes": 892336,
"heapUsedBytes": 69610548,
"domNodes": 24,
"jsHeapTotalBytes": 1310720,
"scriptDurationMs": 766.878,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66556603773585,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "large-graph-zoom",
"durationMs": 3534.4350000000304,
"styleRecalcs": 72,
"styleRecalcDurationMs": 13.366000000000001,
"layouts": 60,
"layoutDurationMs": 13.665000000000001,
"taskDurationMs": 1599.8210000000001,
"heapDeltaBytes": 1677480,
"heapUsedBytes": 69611596,
"domNodes": 24,
"jsHeapTotalBytes": 1310720,
"scriptDurationMs": 762.1669999999999,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.665566037735847,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "legacy-node-drag",
"durationMs": 3059.131999999977,
"styleRecalcs": 48,
"styleRecalcDurationMs": 7.398999999999999,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1831.8380000000002,
"heapDeltaBytes": 3905580,
"heapUsedBytes": 73492016,
"domNodes": 0,
"jsHeapTotalBytes": 1048576,
"scriptDurationMs": 964.904,
"eventListeners": 180,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.75769230769231,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "legacy-node-drag",
"durationMs": 3029.7649999999976,
"styleRecalcs": 45,
"styleRecalcDurationMs": 6.247000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1812.516,
"heapDeltaBytes": 8375976,
"heapUsedBytes": 76412696,
"domNodes": 0,
"jsHeapTotalBytes": 1536000,
"scriptDurationMs": 955.967,
"eventListeners": 180,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.75833333333334,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "minimap-idle",
"durationMs": 2015.340999999978,
"styleRecalcs": 8,
"styleRecalcDurationMs": 3.2829999999999995,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 474.2510000000001,
"heapDeltaBytes": -9730888,
"heapUsedBytes": 80138952,
"domNodes": 0,
"jsHeapTotalBytes": -225280,
"scriptDurationMs": 30.720999999999997,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666115702479335,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "minimap-idle",
"durationMs": 2012.788999999998,
"styleRecalcs": 8,
"styleRecalcDurationMs": 3.406000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 469.7819999999999,
"heapDeltaBytes": -10338212,
"heapUsedBytes": 79525972,
"domNodes": 0,
"jsHeapTotalBytes": 36864,
"scriptDurationMs": 32.70000000000001,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666115702479342,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-dom-widget-clipping",
"durationMs": 592.1020000000112,
"styleRecalcs": 48,
"styleRecalcDurationMs": 5.894,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 383.9920000000001,
"heapDeltaBytes": -11668604,
"heapUsedBytes": 57937476,
"domNodes": 0,
"jsHeapTotalBytes": 524288,
"scriptDurationMs": 138.148,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.665714285714266,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-dom-widget-clipping",
"durationMs": 591.6799999999967,
"styleRecalcs": 47,
"styleRecalcDurationMs": 5.746999999999999,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 385.12899999999996,
"heapDeltaBytes": -12635412,
"heapUsedBytes": 57381796,
"domNodes": 0,
"jsHeapTotalBytes": 524288,
"scriptDurationMs": 139.64000000000001,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66571428571429,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-idle",
"durationMs": 2008.4719999999834,
"styleRecalcs": 12,
"styleRecalcDurationMs": 4.748,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 399.251,
"heapDeltaBytes": 619656,
"heapUsedBytes": 69603268,
"domNodes": 0,
"jsHeapTotalBytes": 262144,
"scriptDurationMs": 7.289000000000004,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666115702479335,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-idle",
"durationMs": 2010.2139999999054,
"styleRecalcs": 12,
"styleRecalcDurationMs": 4.547999999999998,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 404.814,
"heapDeltaBytes": 939504,
"heapUsedBytes": 70834340,
"domNodes": 0,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 7.360999999999993,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.665289256198346,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-mouse-sweep",
"durationMs": 1697.5790000000188,
"styleRecalcs": 77,
"styleRecalcDurationMs": 17.406000000000002,
"layouts": 16,
"layoutDurationMs": 5.4990000000000006,
"taskDurationMs": 714.4119999999999,
"heapDeltaBytes": -6300404,
"heapUsedBytes": 63170320,
"domNodes": 0,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 104.965,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66568627450981,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-mouse-sweep",
"durationMs": 1702.0919999999933,
"styleRecalcs": 77,
"styleRecalcDurationMs": 16.455000000000002,
"layouts": 16,
"layoutDurationMs": 4.999999999999999,
"taskDurationMs": 700.623,
"heapDeltaBytes": -6303500,
"heapUsedBytes": 63334972,
"domNodes": 0,
"jsHeapTotalBytes": 262144,
"scriptDurationMs": 103.44800000000001,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.6656862745098,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-transition-enter",
"durationMs": 1384.992000000011,
"styleRecalcs": 17,
"styleRecalcDurationMs": 29.636999999999997,
"layouts": 5,
"layoutDurationMs": 14.562999999999999,
"taskDurationMs": 873.4190000000002,
"heapDeltaBytes": -610288,
"heapUsedBytes": 90513332,
"domNodes": 14227,
"jsHeapTotalBytes": 2097152,
"scriptDurationMs": 28.088,
"eventListeners": 2463,
"totalBlockingTimeMs": 127,
"frameDurationMs": 19.44305555555555,
"p95FrameDurationMs": 16.80000000000109
},
{
"name": "viewport-pan-sweep",
"durationMs": 8272.219000000006,
"styleRecalcs": 251,
"styleRecalcDurationMs": 34.034,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 5297.1269999999995,
"heapDeltaBytes": 7870976,
"heapUsedBytes": 95700996,
"domNodes": 0,
"jsHeapTotalBytes": 1015808,
"scriptDurationMs": 2656.116,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.665927419354837,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "viewport-pan-sweep",
"durationMs": 8237.256000000001,
"styleRecalcs": 250,
"styleRecalcDurationMs": 33.13000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 5238.412,
"heapDeltaBytes": 7840964,
"heapUsedBytes": 95763968,
"domNodes": 0,
"jsHeapTotalBytes": 1015808,
"scriptDurationMs": 2586.6580000000004,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66578947368421,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "vue-large-graph-idle",
"durationMs": 16936.13700000003,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 16394.413,
"heapDeltaBytes": -18830788,
"heapUsedBytes": 180118928,
"domNodes": -3045,
"jsHeapTotalBytes": 8323072,
"scriptDurationMs": 237.68800000000002,
"eventListeners": -14915,
"totalBlockingTimeMs": 0,
"frameDurationMs": 28.506060606060604,
"p95FrameDurationMs": 66.59999999999854
},
{
"name": "vue-large-graph-idle",
"durationMs": 17253.179999999928,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 16569.73,
"heapDeltaBytes": -47574656,
"heapUsedBytes": 178953240,
"domNodes": -3045,
"jsHeapTotalBytes": 7217152,
"scriptDurationMs": 244.462,
"eventListeners": -14915,
"totalBlockingTimeMs": 0,
"frameDurationMs": 29.385519591141396,
"p95FrameDurationMs": 66.59999999999854
},
{
"name": "vue-large-graph-pan",
"durationMs": 21433.741999999995,
"styleRecalcs": 175,
"styleRecalcDurationMs": 17.54800000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 20965.16,
"heapDeltaBytes": -52069772,
"heapUsedBytes": 174761160,
"domNodes": -3045,
"jsHeapTotalBytes": 6750208,
"scriptDurationMs": 840.764,
"eventListeners": -14915,
"totalBlockingTimeMs": 12,
"frameDurationMs": 39.61645101663586,
"p95FrameDurationMs": 83.39999999999782
},
{
"name": "vue-large-graph-pan",
"durationMs": 21099.03499999996,
"styleRecalcs": 171,
"styleRecalcDurationMs": 16.71699999999998,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 20631.607,
"heapDeltaBytes": -26534660,
"heapUsedBytes": 197334656,
"domNodes": -3045,
"jsHeapTotalBytes": -1011712,
"scriptDurationMs": 819.6279999999999,
"eventListeners": -14915,
"totalBlockingTimeMs": 16,
"frameDurationMs": 38.153887884267625,
"p95FrameDurationMs": 83.30000000000291
},
{
"name": "vue-node-resize-workload",
"durationMs": 6303.012000000024,
"styleRecalcs": 20,
"styleRecalcDurationMs": 77.57799999999997,
"layouts": 20,
"layoutDurationMs": 22.738000000000007,
"taskDurationMs": 6169.303999999999,
"heapDeltaBytes": 949760,
"heapUsedBytes": 217394216,
"domNodes": 0,
"jsHeapTotalBytes": 1310720,
"scriptDurationMs": 98.44600000000003,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 36.95882352941177,
"p95FrameDurationMs": 66.70000000000073
},
{
"name": "vue-node-resize-workload",
"durationMs": 6111.8020000000115,
"styleRecalcs": 20,
"styleRecalcDurationMs": 80.43,
"layouts": 20,
"layoutDurationMs": 23.058999999999997,
"taskDurationMs": 5981.054,
"heapDeltaBytes": 1216880,
"heapUsedBytes": 213870616,
"domNodes": 0,
"jsHeapTotalBytes": 786432,
"scriptDurationMs": 97.57400000000001,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 34.16983240223465,
"p95FrameDurationMs": 66.70000000000073
},
{
"name": "vue-zoom-out-idle",
"durationMs": 8504.062000000033,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 8169.713,
"heapDeltaBytes": -22164196,
"heapUsedBytes": 180497116,
"domNodes": -3045,
"jsHeapTotalBytes": 6049792,
"scriptDurationMs": 124.65699999999997,
"eventListeners": -14915,
"totalBlockingTimeMs": 0,
"frameDurationMs": 28.332333333333334,
"p95FrameDurationMs": 50.10000000000218
},
{
"name": "vue-zoom-out-idle",
"durationMs": 8503.018999999995,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 8169.735,
"heapDeltaBytes": -20624832,
"heapUsedBytes": 182018736,
"domNodes": -3045,
"jsHeapTotalBytes": 9195520,
"scriptDurationMs": 115.67499999999997,
"eventListeners": -14915,
"totalBlockingTimeMs": 0,
"frameDurationMs": 29.512500000000006,
"p95FrameDurationMs": 50.10000000000218
},
{
"name": "workflow-execution",
"durationMs": 122.50200000005407,
"styleRecalcs": 7,
"styleRecalcDurationMs": 10.641,
"layouts": 3,
"layoutDurationMs": 1.5610000000000002,
"taskDurationMs": 83.47900000000003,
"heapDeltaBytes": 3858952,
"heapUsedBytes": 69616324,
"domNodes": 115,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 10.649,
"eventListeners": 17,
"totalBlockingTimeMs": 0,
"frameDurationMs": 19.466666666666697,
"p95FrameDurationMs": 33.30000000000018
},
{
"name": "workflow-execution",
"durationMs": 119.91899999998168,
"styleRecalcs": 8,
"styleRecalcDurationMs": 8.129000000000001,
"layouts": 2,
"layoutDurationMs": 0.6179999999999999,
"taskDurationMs": 77.292,
"heapDeltaBytes": 3763520,
"heapUsedBytes": 70104192,
"domNodes": 95,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 8.343,
"eventListeners": 33,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.671428571428546,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "agent-composer-empty-conversation-typing",
"durationMs": 767.2719999999345,
"styleRecalcs": 78,
"styleRecalcDurationMs": 93.856,
"layouts": 76,
"layoutDurationMs": 22.468,
"taskDurationMs": 731.9219999999999,
"heapDeltaBytes": 16788984,
"heapUsedBytes": 81649068,
"domNodes": 76,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 297.873,
"eventListeners": 876,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.665217391304335,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "agent-composer-empty-conversation-typing",
"durationMs": 835.0719999999683,
"styleRecalcs": 78,
"styleRecalcDurationMs": 95.997,
"layouts": 76,
"layoutDurationMs": 23.345000000000002,
"taskDurationMs": 797.6020000000001,
"heapDeltaBytes": -14641560,
"heapUsedBytes": 72098956,
"domNodes": 76,
"jsHeapTotalBytes": 9961472,
"scriptDurationMs": 331.769,
"eventListeners": 876,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666000000000004,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "agent-composer-long-conversation-typing",
"durationMs": 1328.1190000000151,
"styleRecalcs": 78,
"styleRecalcDurationMs": 246.921,
"layouts": 76,
"layoutDurationMs": 26.504,
"taskDurationMs": 1301.6340000000002,
"heapDeltaBytes": 8446256,
"heapUsedBytes": 114040572,
"domNodes": 76,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 438.85400000000004,
"eventListeners": 876,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.875949367088616,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "agent-composer-long-conversation-typing",
"durationMs": 1373.9379999999528,
"styleRecalcs": 78,
"styleRecalcDurationMs": 257.533,
"layouts": 76,
"layoutDurationMs": 27.799999999999997,
"taskDurationMs": 1352.0939999999998,
"heapDeltaBytes": 8631984,
"heapUsedBytes": 114779056,
"domNodes": 76,
"jsHeapTotalBytes": 262144,
"scriptDurationMs": 453.416,
"eventListeners": 876,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.289999999999985,
"p95FrameDurationMs": 16.799999999999272
}
]
}🔬 E2E Coverage
Top 10 files by uncovered lines
🌾 Fallow: ✅ PassedFallow audits the diff for duplication, complexity and dead code. It gates on new findings only, so when the audit completes, a red is something this PR introduced or moved — never inherited backlog. A red can also mean the audit itself did not finish; the details below say which happened. No new findings in the changed files. Reproduce locally# the base the gate compared against
npx fallow audit --root . --quiet --format json \
--changed-since e2e6849eefd41e39dcf32804b330b1c64eb4648e
Full docs: ADR-DEVEX-LINT-0015 🧩 Custom Nodes Ecosystem Matrix: ✅ PassedExecutes the frontend JS of every registry pack that ships any, against this change, in both renderers. The eight
On a pin-bump PR, a red gate means the ecosystem moved — not that your diff broke something. That is what the bump PR is for. Outside a pin bump or a cold rebuild of the unpinned tail, a red means the diff. Note the corpus size and the executed count are different numbers, and the verdict is against the executed count: roughly 3,100 of ~5,100 packs ship no extension-shaped JS and are skipped. Reproduce locally# build one shard's fixture, then run it
python3 scripts/registry-census/build_matrix.py --shard 1/4
MATRIX_RENDERER=vue MATRIX_VUE=1 \
npx vitest run --config vitest.matrix.config.mts
Full docs: scripts/registry-census/README.md |
Codecov Report❌ Patch coverage is
@@ Coverage Diff @@
## main #17966 +/- ##
==========================================
+ Coverage 85.12% 85.80% +0.68%
==========================================
Files 3480 2452 -1028
Lines 167778 140335 -27443
Branches 43971 38517 -5454
==========================================
- Hits 142821 120418 -22403
+ Misses 22978 19539 -3439
+ Partials 1979 378 -1601
Flags with carried forward coverage won't be shown. Click here to find out more.
... and 1053 files with indirect coverage changes 🚀 New features to boost your workflow:
|
There was a problem hiding this comment.
🔍 Cursor Review — Consolidated panel
Triggered by @shrimbly.
Found 10 finding(s).
| Severity | Count |
|---|---|
| 🟠 High | 2 |
| 🟡 Medium | 6 |
| 🟢 Low | 2 |
Panel: 8/8 reviewers contributed findings.
4523fce to
b89d57e
Compare
There was a problem hiding this comment.
🔍 Cursor Review — Consolidated panel
Triggered by @shrimbly.
Round 2 — ledger: 10 prior finding(s) across 1 round(s) (0 never answered).
Found 10 finding(s).
| Severity | Count |
|---|---|
| 🟠 High | 1 |
| 🟡 Medium | 5 |
| 🟢 Low | 4 |
Panel: 7/8 reviewers contributed findings.
Reviewers that did not contribute: gpt-5.6-sol-max:edge-case (error)
DrJKL
left a comment
There was a problem hiding this comment.
Verdict: changes requested because the agent handoff bypasses its rollout gate and emits commands or filenames that fail for valid user paths and names.
Pre-merge blockers:
- [P1] The new agent action is exposed without the claimed
distributions_enabledgate. - [P1] A backtick in the workflow name makes the filename in the brief differ from the downloaded filename.
- [P1] The localhost and Desktop recipes do not quote path placeholders, so valid paths containing spaces produce broken commands.
Findings without an anchorable line
- [P3] The PR verification section is stale. It reports 24 passing deploy tests and a cancelled-export case, but the current focused run passed 34 tests and this head no longer has an export cancellation path.
Existing findings confirmed or contested
- Round 1 workflow-value injection: Confirmed fixed for added lines and fences.
singleLinefolds controls and unsafe listed values are omitted. The inline finding on the heading covers the remaining raw Markdown syntax. - Round 1 export cancellation and filename drift: Confirmed fixed for prompt cancellation and ordinary names by direct
downloadBlob. The inline backtick finding identifies a remaining filename mismatch. - Round 1 unhandled preparation rejection: Confirmed fixed. Snapshot and document construction now run inside the
copyBriefcatch, which reports and toasts. - Round 1 shell quoting and control characters: Confirmed fixed by
singleLineandquoteForShell, with focused tests. - Round 1 lost clipboard activation: Confirmed fixed. The clipboard write precedes the download.
- Round 1 unresolved Desktop install path: Confirmed fixed by the Desktop path guidance. The inline path-quoting finding is separate.
- Round 1 exponential subgraph flattening: Confirmed fixed. Collected definitions are traversed once and the reuse test covers the case.
- Round 1 model detection: Confirmed partly unresolved.
properties.modelsis now included, but every string widget still goes through a suffix test, so prompt text such asa photo of my cat.ptis still reported as a model. Asset-backed IDs also remain outside this scan. - Round 1 discarded pack version: Confirmed fixed.
properties.veris retained and rendered. - Round 1 dialog-open snapshot: Confirmed fixed. The active workflow is read when the user clicks.
- Round 2 shell use of workflow values: Confirmed fixed for classes, models, and packs. Values with shell metacharacters are omitted.
- Round 2 preparation outside the catch: Confirmed fixed. Both snapshot and document generation are inside the try block.
- Round 2 clipboard false branch without telemetry: Confirmed fixed by the
reportErrorcall on the false branch. - Round 2
comfy-corepack: Confirmed fixed by the core sentinel exclusion. - Round 2 Windows interpreter path: Confirmed fixed for platform discovery. The brief now names Unix, Windows venv, and portable paths and identifies its commands as POSIX.
- Round 2 doubled snapshot suffix: Confirmed fixed.
<newest-snapshot>no longer receives another.jsonsuffix. - Round 2 stale focused-field state: Confirmed implementation fix.
prepareForSave()runs beforeactiveStateis read. The inline test finding explains why the current test does not prove that ordering. - Round 2 unused subgraph definitions: Confirmed still present. Every collected definition contributes classes, packs, models, and card counts even when no root node reaches it.
- Round 2 build output in the caller directory: Confirmed fixed for Cloud and Desktop by creating and entering
comfy-build. - Round 2 copied state after failure: Confirmed fixed. The label is cleared before each attempt and a success-then-failure test covers it.
Checked and cleared
- I rejected the possible post-copy graph mutation finding.
prepareForSaveclones the canvas serialization, and later captures replaceactiveStaterather than mutating the captured object. This root-only action has no in-place graph mutator on its path. - Fallow's 15 dev-dependency notices are baseline issues. This PR changes no dependency declarations.
- Fallow's icon token, arbitrary max-width, and arbitrary max-height notices are unchanged from the base branch.
- I traced the widened clipboard return through outside consumers. Existing callers ignore the result, so the default contract remains compatible.
- I traced the widened
CopyHandlerthroughJobDetailsPopover. It still ignores the callback result, andPromise<unknown>accepts the clipboard composable's boolean without changing behavior.
Coverage
Standard depth. I ran correctness, maintainability, idiomaticity, tests, and documentation lanes over all 12 changed files. Standard specialist lenses included blast-radius and regression history, type-system discipline, deletion-first and comment review, test-suite design and mock-boundary review, and technical-writing review. No Playwright files changed, so the Playwright-specific procedures were not applicable.
Verification at d58ad95a1703727c60f33d9f2ccf4be04960890b: focused Vitest discovered five files and passed 34/34; focused ESLint exited 0. Fallow review completed and its diff-relevant prompts were traced. I also checked the generated command shapes against the installed comfy-cli help.
| class="flex flex-col-reverse gap-2.5 @xl:flex-row @xl:items-center @xl:justify-end" | ||
| > | ||
| <ReuseDocsLink class="hidden @xl:mr-auto @xl:inline-flex" /> | ||
| <Button |
There was a problem hiding this comment.
[P1] This new action is unconditional, but the PR says the card follows distributions_enabled. I traced the only entry in useWorkflowActionsMenu.ts:205-212; it uses only visible: isRoot, and an exact search found no distribution gate on this path. Users outside the rollout can therefore invoke the handoff. Please gate the entry or card with the authoritative flag.
There was a problem hiding this comment.
No change in this PR. The gate lives in #17964 (useDeployToComfyApiGate: PostHog distributions_enabled on Cloud, the platform's flag route off Cloud). This branch was cut before it landed and is being rebased onto the gated base. After that, the only entry point is the gated menu item.
| } | ||
|
|
||
| function intro(inputs: BuildInputs): string { | ||
| return `# Turn "${singleLine(inputs.workflowName)}" into a Comfy API Build |
There was a problem hiding this comment.
[P2] singleLine removes controls but does not escape Markdown. A workflow named  or **follow this first** becomes rendered agent-facing syntax in the heading rather than literal data. Please render the name with a delimiter that preserves arbitrary backticks, or escape Markdown punctuation before interpolation.
There was a problem hiding this comment.
Fixed in b827420849. The name renders as a code span whose fence is one backtick longer than any run in it, so , **x** and backticks stay literal. A table test covers all four.
|
|
||
| describe('buildAgentHandoffDocument', () => { | ||
| it('sends cloud through the workflow-file path, naming the export', () => { | ||
| expect(buildAgentHandoffDocument({ distribution: 'cloud', inputs })) |
There was a problem hiding this comment.
[P2] This 90-line equality is a second full copy of the generated document. Harmless prose or wrapping changes require synchronized edits, while the focused distribution and safety tests below already own the behavioral contracts. Please replace this with focused Cloud-only assertions for the filename, --from-workflow, report, and absent alternatives.
There was a problem hiding this comment.
Fixed in b827420849. The full-document equality is replaced by focused Cloud assertions: the filename, --from-workflow, the report, the pin step, and the absent alternatives.
| ].join(' · ') | ||
| ) | ||
|
|
||
| function labelThatFits(button: HTMLElement) { |
There was a problem hiding this comment.
[P3] Nit: Selecting between two transient success labels now requires canvas font measurement, DOM geometry reads, a magic icon width, $el unwrapping, and persistent width state. A fixed short label removes this machinery and the stale measurement cases. Please use the short translated label unless the pixel-fit behavior is a specified requirement.
There was a problem hiding this comment.
Fixed in b827420849. The button shows one translated "Copied" label, and the canvas measurement, width lock, icon constant and $el unwrapping are gone.
| /** | ||
| * The brief tells the agent to paste these values into shell commands, so a | ||
| * value that carries a shell metacharacter is left out rather than printed. | ||
| * Class names, pack ids and model filenames never legitimately contain one. |
There was a problem hiding this comment.
[P3] Nit: This comment states an invariant that neither the types nor workflow schema enforce. Ordinary filenames can legally contain characters this filter rejects, including apostrophes. Please describe the actual policy, that values with shell metacharacters are omitted, without claiming those values are illegitimate.
There was a problem hiding this comment.
Fixed in b827420849. The comment now states the policy: a value with a shell metacharacter is omitted and counted. It no longer calls such values illegitimate.
| await useAgentHandoff().copyBrief() | ||
|
|
||
| expect(prepareForSave).toHaveBeenCalled() | ||
| expect(prepareForSave.mock.invocationCallOrder[0]).toBeLessThan( |
There was a problem hiding this comment.
[P2] This only proves prepareForSave precedes the later clipboard call. Moving the activeState read above prepareForSave still leaves this assertion green, which is the stale-focused-field regression the test is meant to catch. Please make pre-prepare and post-prepare graph values differ, then assert the post-prepare value reaches the brief or blob.
There was a problem hiding this comment.
Fixed in b827420849. prepareForSave now replaces the stale graph with a fresh one, and the test asserts the fresh class reaches the brief and the blob, with no trace of the stale one.
| } | ||
| if (isCloud) { | ||
| downloadBlob( | ||
| inputs.workflowFileName, |
There was a problem hiding this comment.
[P1] The brief and download still disagree for a valid workflow name containing a backtick. inlineCode removes backticks from the displayed workflowFileName, while this call downloads the unchanged filename. For a name formed from release, a backtick, and v2.json, the agent looks for releasev2.json but the browser writes the original name. Normalize the actual filename and displayed filename once, and add this adversarial composition case to the composable test.
There was a problem hiding this comment.
Fixed in b827420849. handoffFileName computes the name once, dropping backticks and path separators. The brief and downloadBlob both use it. The composable test covers release, a backtick, and v2.json, and a mutation passing the raw name fails it.
| \`<install>\\python_embeded\\python.exe\` for the Windows portable build. | ||
|
|
||
| \`\`\`bash | ||
| comfy build init <install> --name "${name}" --python <python> |
There was a problem hiding this comment.
[P1] These path placeholders are unquoted. Substituting a supported path such as /Users/me/Comfy UI splits it into multiple shell arguments; the Desktop ls and snapshot commands have the same problem. Please quote every path placeholder in the POSIX examples and tell the agent to preserve those quotes when substituting.
There was a problem hiding this comment.
Fixed in b827420849. Every path placeholder is quoted: "<install>", "<python>", "<newest-snapshot>", "<path-to-file>", and the snapshots glob. The brief tells the agent to keep the quotes. Removing them fails the localhost and Desktop tests.
|
|
||
| - Set the ComfyUI version: \`comfy build update . --comfy-version <ref>\` | ||
| - Resolve every model the report lists: \`comfy build refs resolve '<filename>'\` | ||
| - Pin any pack that arrived without a \`gitRef\` to the version listed below, or |
There was a problem hiding this comment.
[P2] This required step gives the agent no command, build-spec file, or field to edit. The installed CLI help has no pack-pin option, so the agent must guess at the generated schema or skip the pin. Please name the exact file and field to change, provide the supported command, or direct the agent to an authoritative builder instruction.
There was a problem hiding this comment.
Fixed in b827420849. comfy-cli has no pin command, so the step names the fields: each customNodes entry in comfy-build.yaml needs registryVersion with the registry id in id, or a repository with a 40-character commit in gitRef. It points to comfy skills show comfy-build-authoring for the full rules.
| "readDocs": "Read the docs", | ||
| "deployOnPlatform": "Deploy on Platform" | ||
| "deployOnPlatform": "Deploy on Platform", | ||
| "deployWithAgent": "Deploy with your agent", |
There was a problem hiding this comment.
[P2] This CTA promises deployment, but the generated brief explicitly says a Build is not a deployment and instructs the agent not to deploy. The click itself only copies instructions and may download JSON. Please name the action it performs, such as Copy agent build brief, or extend the flow so it actually deploys.
There was a problem hiding this comment.
No change for now. The CTA copy is the designer's, and I have flagged this to them. The label will follow their call.
There was a problem hiding this comment.
Renamed in eb4270a04e: the button now reads "Build with your agent".
|
@DrJKL thanks for the thorough pass. Everything inline is answered on its thread; fixes are in On the blockers:
On the findings without a line:
|
DrJKL
left a comment
There was a problem hiding this comment.
Verdict: changes requested because the current brief cannot reach its promised green state and the stacked gate is still absent from the actual base.
Still open from my previous review:
- [P1] The current base at
4345f9a6still exposes the menu item withvisible: isRoot; it does not contain thedistributions_enabledgate. I understand the branch is awaiting a rebase onto #17964, but I have to review the actual base/head pair, which is currently conflicted. - [P2] The Deploy with your agent CTA still hands off a build-only brief that explicitly prohibits deployment. I am carrying this while the designer decides the final copy.
- [P3] The PR body is still stale in the REST response. It says the card is gated, the Cloud document is asserted in full, the button measures long/short copied labels, model detection is suffix-only because widget kinds are unknown, every workflow value appears in the brief, and the deploy suite has 24 tests including cancelled export. Those statements no longer match this head. The issue reply says the verification section was updated to 53 tests, but the current PR body still contains the old text.
The prior Markdown, full-document test, copied-label measurement, malformed-input parsing, distribution typing, clipboard-result assertions, fresh-graph capture, filename consistency, path quoting, and pack-pinning findings are fixed. I also independently confirmed the other reviewer's orphan-subgraph finding is fixed. I confirm their model false-positive finding remains partially open for the reason in the inline comment.
I ran pnpm vitest run src/platform/workflow/deploy src/composables/useCopyToClipboard.test.ts (53 passed), focused ESLint, pnpm typecheck, pnpm format:check, and Fallow. All passed.
| Build is a definition of everything needed to run the workflow on a serverless | ||
| API: the ComfyUI version, the custom node packs and the models. | ||
|
|
||
| A Build is not a deployment. Deploying it is a separate step. Stop at a green |
There was a problem hiding this comment.
[P1] I traced this terminal state against the current official comfy-build skill. comfy build push only stores an editable Build definition; comfy build release create ... --watch creates the artifact, and deployable: true on that release is the green result. Every branch below stops at push, while the closing section explicitly forbids the release cut, so an agent following the brief cannot satisfy this instruction. Please either stop at a validated/pushed Build definition and use that exact term, or include the release step needed to reach a green release.
There was a problem hiding this comment.
Fixed in 229d0ca51d. The brief ends in one shared section for every distribution: disclose and wait for a yes, list comfy build refs build-targets, then push and comfy build release create ... --watch. Green is complete with deployable: true, and the brief still stops before deploying. A per-distribution test pins it, and dropping the release line fails all three rows.
| function canLoadModels(nodeType: string): boolean | undefined { | ||
| const definitions = nodeDefStore.nodeDefsByName | ||
| if (!Object.hasOwn(definitions, nodeType)) return undefined | ||
| return Object.values(definitions[nodeType].inputs).some(isComboInputSpec) |
There was a problem hiding this comment.
[P2] This still does not establish that the node can load a model. COMBO is the generic dropdown input type, so a node with an unrelated dropdown and a free-text widget ending in .pt passes this predicate; unknown classes also retain the all-string suffix fallback. modelNames() then scans every serialized string rather than the specific combo widget. I confirmed the existing missing-model path classifies a concrete widget instead. Please map serialized values to the model-capable input, or limit unknown classes to properties.models, rather than treating any combo as node-wide model-loader proof.
There was a problem hiding this comment.
Fixed in 229d0ca51d. The COMBO predicate is gone. A string is a model only in the model input of a known loader, taken from the model-folder registry that the model browser and the missing-model scan use (getRegisteredNodeTypes()). A named widgets_values object is read at that key. Any other node, known or unknown, contributes only properties.models.
| } | ||
| } | ||
|
|
||
| /** The nodes of every subgraph definition the root graph reaches, each once. */ |
There was a problem hiding this comment.
[P3] Nit: this adds a second reachable-subgraph walker beside flattenWorkflowNodes() / buildSubgraphExecutionPaths() in workflowFlattening.ts. The derived classes, packs, and models are deduplicated immediately, so visiting reused definitions per execution path does not change the result. Please reuse the existing traversal after the tolerant outer parse instead of creating a second owner for cycle and reachability behavior.
There was a problem hiding this comment.
No change. flattenWorkflowNodes and buildSubgraphExecutionPaths recurse once per instantiation path, which grows exponentially for a definition reused at each level. That was the round-1 finding this walker fixed, and reads a subgraph the graph reuses at several levels once pins it. Reusing either would bring the blow-up back on dialog open.
| } | ||
|
|
||
| const collator = new Intl.Collator('en') | ||
| const sorted = (values: Set<string>) => |
There was a problem hiding this comment.
[P3] Nit: the engineering guidance says to use a function declaration when one is possible. This helper has no lexical requirement for an arrow expression; please make it a local function declaration.
There was a problem hiding this comment.
Fixed in 229d0ca51d. sorted is now a local function declaration.
| }) | ||
| } | ||
|
|
||
| describe('useAgentHandoff', () => { |
There was a problem hiding this comment.
[P2] The pure utility test injects a hand-written model-loader callback, but no composable test proves the real node-definition wiring. I replaced canLoadModels with an unconditional undefined and both this suite and buildInputs.test.ts still passed (21 tests). Please seed nodeDefsByName with a known text-only node and a known loader, then assert the copied brief omits prompt text ending in .pt and keeps the loader model.
There was a problem hiding this comment.
Fixed in 229d0ca51d. The composable test seeds real node definitions for CheckpointLoaderSimple and CLIPTextEncode and asserts the brief keeps the loader's model and drops a photo of my cat.pt. Replacing the wiring with () => undefined fails it.
| expect(order).toEqual(['copy', 'download']) | ||
| }) | ||
|
|
||
| it('downloads nothing off Cloud, where the agent reads the install', async () => { |
There was a problem hiding this comment.
[P2] This composition suite covers Cloud and localhost but not Desktop. I changed the composable to route Desktop through localhost and this suite plus agentHandoff.test.ts still passed (28 tests), despite Desktop users then receiving comfy which instead of snapshot instructions. Please table this off-Cloud case for localhost and Desktop and assert the distinguishing command as well as no download.
There was a problem hiding this comment.
Fixed in 229d0ca51d. The off-Cloud case is a table for localhost and Desktop that asserts each distribution's own recipe (comfy which or --from-snapshot), the absence of the other, and no download. Routing Desktop through localhost fails the Desktop row.
DrJKL
left a comment
There was a problem hiding this comment.
Verdict: changes requested because the coding brief still promises a green result while prohibiting the release step that creates one.
The merge fixes the gate: the current menu now uses isRoot && deployGate.enabled.value, the branch is mergeable, and the conflict resolution preserves both agent and platform actions. The PR body is also current now.
Still open:
- [P2] The Deploy with your agent CTA still hands off a build-only brief that prohibits deployment. The PR body says design is reviewing this label, so I am carrying it until that decision lands.
- The six inline findings from my immediately preceding review remain unchanged on this merged head. I have re-anchored them here because that review was submitted against
b8274208four seconds after this merge landed.
I ran the deploy, clipboard, and workflow-menu suites on the merged base/head pair: 8 files and 103 tests passed. Focused ESLint and the full typecheck also passed. I found no new defect in the merge resolution.
| Build is a definition of everything needed to run the workflow on a serverless | ||
| API: the ComfyUI version, the custom node packs and the models. | ||
|
|
||
| A Build is not a deployment. Deploying it is a separate step. Stop at a green |
There was a problem hiding this comment.
[P1] Still open on the merged head: comfy build push stores the editable Build definition, while the official comfy-build flow uses comfy build release create ... --watch and deployable: true for the green artifact. Every branch stops at push, and the closing section forbids the release cut. Please either say to stop at a validated/pushed Build definition or include the release step required to reach a green release.
There was a problem hiding this comment.
Fixed in 229d0ca51d, same change as the first thread on this line: the brief now cuts a release with --watch and stops at deployable: true.
| function canLoadModels(nodeType: string): boolean | undefined { | ||
| const definitions = nodeDefStore.nodeDefsByName | ||
| if (!Object.hasOwn(definitions, nodeType)) return undefined | ||
| return Object.values(definitions[nodeType].inputs).some(isComboInputSpec) |
There was a problem hiding this comment.
[P2] Still open on the merged head: COMBO means a generic dropdown, not a model input. A node with an unrelated dropdown and free text ending in .pt passes this node-wide predicate, and unknown classes still suffix-scan every string. Please classify the specific model-capable widget, or limit unknown classes to properties.models.
There was a problem hiding this comment.
Fixed in 229d0ca51d, same change as the first thread on this line: models come only from a known loader's model input, per the model-folder registry. Other nodes contribute only properties.models.
| } | ||
| } | ||
|
|
||
| /** The nodes of every subgraph definition the root graph reaches, each once. */ |
There was a problem hiding this comment.
[P3] Nit: still open on the merged head. This duplicates reachable-subgraph traversal already owned by flattenWorkflowNodes() / buildSubgraphExecutionPaths(). Since the returned classes, packs, and models are deduplicated, per-instance repeats do not affect the result. Please reuse the existing traversal after the tolerant outer parse.
There was a problem hiding this comment.
No change, same reason as the first thread on this line: both existing traversals are per-path and exponential for reused definitions. This walker reads each definition once.
| } | ||
|
|
||
| const collator = new Intl.Collator('en') | ||
| const sorted = (values: Set<string>) => |
There was a problem hiding this comment.
[P3] Nit: still open on the merged head. Repository guidance requires a function declaration when one is possible; this helper has no lexical reason to be an arrow expression.
There was a problem hiding this comment.
Fixed in 229d0ca51d. sorted is now a function declaration.
| }) | ||
| } | ||
|
|
||
| describe('useAgentHandoff', () => { |
There was a problem hiding this comment.
[P2] Still open on the merged head: no composable test exercises the real node-definition model filter. Replacing canLoadModels with unconditional undefined leaves the composable and utility suites green. Please seed real node definitions for a text-only node and loader, then assert the copied brief excludes prompt-like .pt text and retains the loader model.
There was a problem hiding this comment.
Fixed in 229d0ca51d. A composable test uses real node definitions and fails if the registry wiring is replaced.
| expect(order).toEqual(['copy', 'download']) | ||
| }) | ||
|
|
||
| it('downloads nothing off Cloud, where the agent reads the install', async () => { |
There was a problem hiding this comment.
[P2] Still open on the merged head: the composition suite covers Cloud and localhost but not Desktop. Routing Desktop through localhost leaves the composable and document suites green. Please table this case for localhost and Desktop, asserting comfy which versus --from-snapshot and no download.
There was a problem hiding this comment.
Fixed in 229d0ca51d. A localhost/Desktop table asserts comfy which versus --from-snapshot and no download.
|
@DrJKL round 2 is in
Checks: 108 tests across the deploy, clipboard, and workflow-menu suites pass. Typecheck, eslint, oxlint, format, knip, and fallow are clean. |
DrJKL
left a comment
There was a problem hiding this comment.
Verdict: changes requested because the handoff can still omit a required model and cut a release that can never become deployable.
The green-release lifecycle, function-declaration convention, real model-store composition coverage, and Desktop composition coverage are fixed in this head. The generic-COMBO false positive is also fixed; the remaining model issue is the one-to-many registry loss described inline.
Still open from my prior review:
- [P2] Deploy with your agent still promises deployment, while the action copies a build brief and the brief intentionally stops before deployment. The PR body says design is reviewing the label, so I am carrying it until that decision lands.
- [P3] The duplicate reachable-subgraph traversal remains; I re-anchored it inline.
New body-only finding:
- [P3] Nit: the PR body still says the branch “is being rebased” onto the gated base, but this head already descends from the merge that brought in that gate. Please remove the stale sentence.
I ran the focused deploy, clipboard, and workflow-menu suites: 8 files and 108 tests passed. Focused ESLint passed, and the full typecheck passed.
| const { copyToClipboard } = useCopyToClipboard() | ||
|
|
||
| function modelWidgetFor(nodeType: string): string | undefined { | ||
| const loaders = modelToNodeStore.getRegisteredNodeTypes() |
There was a problem hiding this comment.
[P1] This reverse lookup cannot represent every model input of a loader. MODEL_NODE_MAPPINGS registers INPAINT_LoadFooocusInpaint with both head and patch, but getRegisteredNodeTypes() builds a Record<string, string> with Object.fromEntries, so patch overwrites head. For object-form widgets_values, the generated brief then omits the head model and can produce an incomplete Build. Please preserve all registered keys per node type and add the two-input loader as the regression case.
There was a problem hiding this comment.
Fixed in 080d9f54d2. The composable now reads the registry's providers instead of the collapsed node-type record, so it keeps every model input per class. INPAINT_LoadFooocusInpaint lists both head and patch. Tests cover it in buildInputs and at the composable with real node definitions, and collapsing to one input fails the composable test.
|
|
||
| \`\`\`bash | ||
| comfy build push ${directory} | ||
| comfy build release create ${directory} --target <os>/<gpu> --watch |
There was a problem hiding this comment.
[P1] The official build contract says only a ready linux/nvidia artifact makes a release deployable: true. This placeholder permits a valid windows-only target; that release reaches complete with deployable: false, and the following recovery text only handles failed targets. Require linux/nvidia for this API flow, with any other targets as optional additions, so following the brief can reach its stated terminal condition.
There was a problem hiding this comment.
Fixed in 080d9f54d2. The cut now uses --target linux/nvidia as required, with other targets optional. The recovery text says a complete release with deployable: false has no ready linux/nvidia artifact, and that the fix is to cut it or repair it.
| - Set the ComfyUI version: \`comfy build update . --comfy-version <ref>\` | ||
| - Resolve every model the report lists: \`comfy build refs resolve '<filename>'\` | ||
| - Pin every pack in \`comfy-build.yaml\`. Each \`customNodes\` entry needs either | ||
| \`registryVersion\` set to the version listed below, with the pack's registry |
There was a problem hiding this comment.
[P2] The version listed below is workflow ver metadata, not a registry result. The official authoring rules require registryVersion to come from the registry row's latest_version.version and explicitly say not to write an unverified version. A stale or fabricated workflow value can make push fail or select the wrong pack. Please tell the agent to verify the recorded ID/version against registry results and use only a returned version, otherwise use a verified repository commit or leave the pack unresolved.
There was a problem hiding this comment.
Fixed in 080d9f54d2. The pin step says the listed ids and versions are workflow metadata. It has the agent check each against nodes/search, take the slug and latest_version.version (or another returned version) only from the row whose repository matches, and otherwise pin a repository commit or report the pack unresolved. It never writes an unreturned version.
| const push = document.indexOf(`comfy build push ${directory}\n`) | ||
|
|
||
| expect(disclosure).toBeGreaterThan(-1) | ||
| expect(push).toBeGreaterThan(disclosure) |
There was a problem hiding this comment.
[P2] This proves only that push follows consent. Moving release create above “Wait for a yes” while leaving push here keeps every assertion green, although the release cut is the user-authorized side effect this test is meant to protect. Please compare all three indices and require disclosure < push < release create.
There was a problem hiding this comment.
Fixed in 080d9f54d2. The test now asserts disclosure < push < release create for every distribution.
| } | ||
|
|
||
| /** The nodes of every subgraph definition the root graph reaches, each once. */ | ||
| function reachableDefinitionNodes( |
There was a problem hiding this comment.
[P3] Nit: this still duplicates reachable-subgraph traversal already owned by flattenWorkflowNodes() / buildSubgraphExecutionPaths(). The returned classes, packs, and models are deduplicated, so the per-instance repeats from the existing traversal do not change the result. Please reuse or expose the existing owner rather than maintaining a second cycle and reachability algorithm.
There was a problem hiding this comment.
Fixed in 080d9f54d2. Reachability now lives in workflowFlattening as collectReachableSubgraphDefinitions, beside buildSubgraphExecutionPaths, and buildInputs uses it. I did not reuse the per-path traversal because it recurses once per instantiation path, which is 2^n for a definition reused twice at each of n levels. A new test with 24 such levels checks that each definition is visited once.
| "readDocs": "Read the docs", | ||
| "deployOnPlatform": "Deploy on Platform" | ||
| "deployOnPlatform": "Deploy on Platform", | ||
| "deployWithAgent": "Deploy with your agent", |
There was a problem hiding this comment.
[P2] This still promises deployment, but clicking it only copies a build handoff (and may download the workflow), and the corrected brief explicitly stops at a green release before deployment. Please use copy that names the action it performs, such as Build with your agent, unless the flow is extended to create a serving deployment.
There was a problem hiding this comment.
No change. The CTA copy is held for the designer, and the operator is deciding it.
There was a problem hiding this comment.
Renamed in eb4270a04e: the button now reads "Build with your agent".
DrJKL
left a comment
There was a problem hiding this comment.
Verdict: changes requested because the Cloud recipe still leaves every model out of the spec, and retry cuts can spend another build without renewed consent.
Pre-merge blockers:
- [P1]
refs resolveonly returns candidates. The brief never writes the selected model source and digest intocomfy-build.yaml, which starts withmodels: []. - [P1] After a failed release, the brief says to tell the user about an edit but does not wait for another yes before pushing and cutting again.
Prior findings
- Fixed: all registered inputs of a multi-input loader are retained and covered through the real store.
- Fixed: every release now includes
linux/nvidia. - Fixed: the test now orders first consent, push, and release.
- Fixed: reachable-definition traversal moved to its shared owner.
- Still open: [P2] the supplied registry search exposes
latest_version.version; it cannot verify an older recorded version as the PR body claims. - Still open: [P2] Deploy with your agent copies/downloads a handoff whose instructions stop before deployment. I am carrying this while design decides the label.
Body-only findings
- [P3] Nit: the PR body still says this branch is being rebased onto the gated base. Head
754a531062already merges that base. - [P3] Nit: the verification count says 885 passing tests. The exact-head command now reports 886 passed and 1 expected failure after the latest base merge.
Checked and cleared
- The latest base merge is faithful: its tree exactly matches Git's automatic merge result. It preserves the billing and customer mocks added to the Playwright fixture.
- I rejected the placeholder-test candidate because the path placeholders are asserted in their concrete commands; the remaining placeholders are refs or targets rather than paths requiring shell quoting.
- I kept the loose linear getter bound in the traversal test because it catches exponential growth without freezing an exact access count.
- The docs-link test loop is inherited unchanged from the current base, so it is outside this PR.
Coverage: standard correctness, maintainability, idiomaticity, tests, Playwright, and documentation lanes. pnpm vitest run src/platform/workflow src/composables/useCopyToClipboard.test.ts src/composables/useWorkflowActionsMenu.test.ts passed 53 files and 886 tests with 1 expected failure. Focused ESLint and the full typecheck passed.
| things need settling by hand: | ||
|
|
||
| - Set the ComfyUI version: \`comfy build update . --comfy-version <ref>\` | ||
| - Resolve every model the report lists: \`comfy build refs resolve '<filename>'\` |
There was a problem hiding this comment.
[P1] comfy build refs resolve is a read-only candidate lookup. The workflow importer starts the definition with models: [], and this recipe goes straight from the lookup to validation without writing the chosen candidate into comfy-build.yaml. An agent can therefore produce a green image that does not contain the workflow's weights. Please tell it to add the selected model's type, filename, sourceUri, and sha256 to the spec before validation, with the authoring skill as the procedure.
There was a problem hiding this comment.
Fixed in eb4270a04e. After refs resolve, the brief has the agent write each chosen candidate into definition.models in comfy-build.yaml as type (the loader's directory under models/), filename, sourceUri and sha256, before validate. It prefers a candidate with a digest, reports models with none, and points to comfy-build-authoring for the full rules.
| cut, and then the fix is to cut it. When a target fails, read its | ||
| \`artifacts[].failureReason\`, then | ||
| \`comfy build release logs --target <os>/<gpu>\`. Fix one cause per cut, tell the | ||
| user each edit before the next cut, and stop after three cuts.` |
There was a problem hiding this comment.
[P1] The first consent covers the disclosed definition and first cut, but this retry instruction only says to tell the user about later edits. The official build procedure requires describing the cause and exact edit, then waiting before each new cut. Please require a fresh yes before every changed retry push and release, rather than allowing two more paid cuts after notification alone.
There was a problem hiding this comment.
Fixed in eb4270a04e. Before every new push and cut, the agent names the cause, the exact edit and which cut it is, then waits for a new yes. The brief says the first yes does not cover a retry, and a test pins that wording for every distribution.
| the file. The pack ids and versions listed below are what the workflow | ||
| recorded, not registry results, so check each one first with | ||
| \`curl -s 'https://api.comfy.org/nodes/search?search=<id>'\`. Take the slug and | ||
| \`latest_version.version\`, or another version the registry returns for that |
There was a problem hiding this comment.
[P2] This search response gives the row's latest_version.version; it does not list every published version. The instruction to use “another version the registry returns” therefore cannot verify an older ver recorded by the workflow, despite the PR body saying each recorded version is verified. Please either direct the agent to use the returned latest version, or provide the supported endpoint/command that verifies the recorded version.
There was a problem hiding this comment.
Fixed in eb4270a04e. The step says the search returns only the newest version. It writes latest_version.version from the matching row, tells the user when it differs from the recorded ver, and pins a repository commit to keep the recorded one. It never claims a recorded version was verified. The PR body claim is corrected.
| } | ||
|
|
||
| function isFlattenableWorkflowNode( | ||
| export function isFlattenableWorkflowNode( |
There was a problem hiding this comment.
[P2] This newly exported guard claims FlattenableWorkflowNode, but it verifies only id and type. Values such as properties: 'bad', widgets_values: 42, or a non-number mode cross this serialized-data boundary as trusted optional fields. Please validate every field represented by the return type, or return a narrower type containing only the two fields this guard proves and parse the optional fields separately.
There was a problem hiding this comment.
Fixed in eb4270a04e. The guard is private again. buildInputs parses every node with zod: id and type are required, and widgets_values and properties are kept only in the right shape, otherwise dropped. A test covers properties: 'bad' and widgets_values: 42.
| "readDocs": "Read the docs", | ||
| "deployOnPlatform": "Deploy on Platform" | ||
| "deployOnPlatform": "Deploy on Platform", | ||
| "deployWithAgent": "Deploy with your agent", |
There was a problem hiding this comment.
[P2] This still promises deployment, but clicking it only copies a handoff and may download the workflow; the handoff explicitly stops at a release before deployment. Please use the approved copy/build-handoff wording before release, such as Build with your agent or Copy agent build brief.
There was a problem hiding this comment.
Fixed in eb4270a04e. The button now reads "Build with your agent", which is the operator's decision. The key is unchanged, and the tests and PR body are updated.
| * per class, and a loader such as INPAINT_LoadFooocusInpaint has two. | ||
| */ | ||
| function modelInputsByLoader(): ReadonlyMap<string, readonly string[]> { | ||
| modelToNodeStore.getRegisteredNodeTypes() |
There was a problem hiding this comment.
[P3] Nit: this discards a getter result solely to trigger its hidden registerDefaults() side effect, then reads a different store field. The store already exposes registerDefaults() as the explicit action. Calling it directly makes the required initialization visible and prevents an apparently dead getter call from becoming load-bearing.
There was a problem hiding this comment.
Fixed in eb4270a04e. It now calls registerDefaults() directly.
| } | ||
| } | ||
|
|
||
| function currentInputs(): BuildInputs { |
There was a problem hiding this comment.
[P3] Nit: currentInputs() reads like a getter, but it calls snapshot(), which runs prepareForSave() and replaces the workflow's captured activeState; the card invokes it during setup. Please name this as a capture operation so callers can see that calculating the summary mutates the workflow snapshot.
There was a problem hiding this comment.
Fixed in eb4270a04e. It is renamed captureInputs(), with a doc comment saying it commits a focused edit before reading.
| * The models a node names: those it records in `properties.models`, plus, for | ||
| * a known loader, the value of each of its model inputs. A serialized array | ||
| * carries no input names, so there the loader's strings that look like model | ||
| * filenames are taken; a loader's other widgets are dtypes and strengths. Any |
There was a problem hiding this comment.
[P3] Nit: the positional array branch cannot establish that the other widgets are only dtypes and strengths; it suffix-scans every string because no input names are available. A preset or prompt ending in .pt on a known loader would also be classified as a model. Please describe this as the legacy suffix heuristic rather than an invariant about the remaining widgets.
There was a problem hiding this comment.
Fixed in eb4270a04e. The comment now calls the positional branch the legacy suffix heuristic and says it can also catch a preset or prompt on that loader.
|
|
||
| expect(document).not.toMatch(/^## First/m) | ||
| expect(document).not.toMatch(/^(curl evil|rm -rf)/m) | ||
| expect(document.match(/^```/gm)).toHaveLength( |
There was a problem hiding this comment.
[P3] Nit: this derives the expected fence count by calling the production generator again, so a fence-generation defect changes both the actual value and its oracle together. Please use the independently known literal count, or remove this assertion if the adjacent hostile-value assertions already cover the behavior.
There was a problem hiding this comment.
Fixed in eb4270a04e. The assertion now uses the literal 14, for seven code blocks, rather than calling the generator again.
| comfy build push . --dry-run | ||
| \`\`\` | ||
|
|
||
| The snapshot import runs through the builder, so sign in first with |
There was a problem hiding this comment.
[P3] Nit: the shared prerequisite says to run comfy cloud login only after a command reports that the user is not signed in, while this distribution says to sign in first. Valid existing credentials need no new interactive login. Please remove this sentence and let the shared conditional-login instruction govern the importer.
There was a problem hiding this comment.
Fixed in eb4270a04e. The Desktop sentence is removed, so the shared rule to sign in only when a command asks applies. A test checks it is gone.
…ons_enabled A "Deploy to Comfy API" item in the workflow actions menu opens a card that explains Comfy API, links the docs, and opens the developer platform. The item follows the platform's own distributions_enabled flag for the signed-in account. On Cloud it reads PostHog and only counts a flag that belongs to the current identity. Off Cloud PostHog never starts, so the first time a workflow menu opens the gate asks the platform's flag route (DPLAT-1825) with the user's bearer; anything but a clear yes keeps the item hidden. Signing out or switching accounts hides it at once, and a stale answer cannot restore it. An open menu holds the row in place, disables it if access is revoked, and the command rechecks access around the workflow switch. Browser specs answer the platform flag route by default, so the lazy check never reaches the network under isolation. DES-1085
Build with your agent copies a brief for a coding agent that drives comfy-cli from the open workflow to a green release, per distribution: from the downloaded workflow file on Cloud, a scan of the install on localhost, and Desktop's snapshot. The brief lists the node classes, the packs the workflow records and the models its loaders read. Every value from the workflow is data: shell metacharacters are left out, paths are quoted, and names render literally. It asks the user before every push and release, and stops before deployment. useCopyToClipboard now reports whether the copy succeeded.
DrJKL
left a comment
There was a problem hiding this comment.
Verdict: changes requested because Cloud and Desktop send workflow data before consent, and localhost failure recovery loses the Build it needs to inspect.
Pre-merge blockers:
- [P1]
--from-workflowand--from-snapshotsend the workflow or snapshot to the builder before the brief discloses that transmission or asks for consent. - [P1] The recovery commands omit both the release id and the localhost Build scope, so a failed localhost cut cannot be inspected from the directory where the brief starts.
Prior findings
- Fixed: selected model candidates are written into
definition.modelsbefore validation. - Fixed: every changed retry push and cut requires a fresh yes.
- Fixed: the registry instructions now use
latest_version.versionand no longer claim to verify the recorded older version. - Fixed for this PR boundary: the broad node guard is private, and
buildInputsreparses consumed nodes while dropping malformed optional fields. - Fixed: the button now says Build with your agent.
- Fixed: default model registration uses the explicit action.
- Fixed:
captureInputsnames the snapshot side effect. - Fixed: the positional-array comment now calls the behavior a legacy suffix heuristic and states its false-positive limit.
- Fixed: the fence-count expectation is now an independent literal.
- Fixed: the unconditional Desktop login instruction is gone.
- Still open. Nit: [P3] The PR body says the branch "is being rebased" onto the gated base, but the gated base has now been merged and is the current merge-base.
- Still open. Nit: [P3] After that base merge, the PR body's exact command no longer reports 53 files and 888 passing tests; it now reports 56 files and 969 passing tests, with one expected failure.
Checked and cleared
- I rejected the remaining shared type-guard candidate because the same guards and exported flattening functions exist on the base branch. The new deploy consumer reparses their nodes before reading optional fields.
- I kept the literal fence count. It is the independent injection oracle requested in the prior round, and hostile input must not add a code fence.
- I traced Fallow's three public-contract decisions. Outside consumers remain compatible with the clipboard result,
useJobErrorReporting, and existing workflow-flattening exports. - Fallow reports 15 dev-dependency findings for
@pinia/testing,@tailwindcss/vite,jsonc-parser,markdown-table,mixpanel-browser,monocart-coverage-reports,openai,picocolors,pretty-bytes,unplugin-icons,unplugin-vue-components,uuid,yaml,zip-dir, andzod-to-json-schema. Its JSON marks every oneintroduced: false; this PR does not changepackage.json. - Fallow also reports the existing
icon-[lucide--square-arrow-out-up-right],max-w-[640px], andmax-h-[85dvh]token warnings asintroduced: false.
Coverage
Standard correctness, maintainability, idiomaticity, tests, and documentation lanes reviewed all ten follow-up files. The tests lane applied its unit-test, suite-design, mock-boundary, setup, and proof specialists. No browser_tests file changed, so the Playwright procedures were not applicable.
At feature commit eb4270a04e4777a9678738040166b984259b568d, focused ESLint, format, full typecheck, oxlint, knip, and Fallow completed successfully, and focused mutations confirmed the two test gaps posted inline. I compared merge head 1e4093e1fb70d7c64440922f9b625b71396f16d8 with Git's automatic merge of that feature commit and current base 55b0c288323f09581a715b6f34661cb197364478: the only additional change makes the existing clipboard mock return the new boolean result. On that merge head, full typecheck and focused ESLint pass. The PR body's exact Vitest command passes 56 files and 969 tests; the broader focused run including the adjusted dialog-service test passes 57 files and 990 tests. Both runs retain one expected failure.
| function cut(directory: string): string { | ||
| return `## Cut the release | ||
|
|
||
| Nothing has been sent yet. Before anything is, tell the user what goes: the |
There was a problem hiding this comment.
[P1] --from-workflow and --from-snapshot authenticate and POST the full JSON to the builder importer, so this statement is already false when the agent reaches it. Please disclose the workflow or snapshot transmission and wait for an explicit yes before either importer command. Keep the later cut consent separate so it can include the final targets, upload bound, release slot, retry limit, and immutable policy before push and release create.
There was a problem hiding this comment.
Fixed in dd60b54234. The Cloud and Desktop recipes now say the importer uploads the whole workflow or snapshot JSON to the builder, and wait for a yes before --from-workflow or --from-snapshot runs. The cut has its own consent before push and release create: what goes with the dry-run upload bound, the targets with linux/nvidia required, the release slot, the three-cut limit, and the policy the release records for good.
| candidates: \`comfy build refs resolve '<filename>'\`. Then write each chosen | ||
| candidate into \`definition.models\` in \`comfy-build.yaml\`, as \`type\` (the | ||
| directory under \`models/\` the loader reads), \`filename\`, \`sourceUri\` and | ||
| \`sha256\`. A model left out is not in the build. Prefer a candidate with a |
There was a problem hiding this comment.
[P2] A resolver candidate may omit sha256; the current comfy-cli contract makes that field optional and calls such a candidate an unpinned fetch. Requiring the field leaves the agent unable to follow this list when no digest-bearing candidate exists. Please say to write sha256 when supplied, and require the agent to disclose and get agreement before selecting a digestless source.
There was a problem hiding this comment.
Fixed in dd60b54234. sha256 is written when the candidate has one. A candidate without it is an unpinned fetch: the agent tells the user and gets agreement before choosing it.
| no ready \`linux/nvidia\` artifact: either that target failed, or it was not | ||
| cut, and then the fix is to cut it. When a target fails, read its | ||
| \`artifacts[].failureReason\`, then | ||
| \`comfy build release logs --target <os>/<gpu>\`. Fix one cause per cut. Before |
There was a problem hiding this comment.
[P1] These bare release show and release logs commands default to the current directory. The localhost recipe passes "<install>" to release create but never changes into that directory, so the first failed localhost cut cannot be resolved here. Preserve the release id returned by release create and pass it to both commands, or otherwise keep the correct Build scope explicitly. Please cover the localhost recovery commands in the distribution table.
There was a problem hiding this comment.
Fixed in dd60b54234. The agent notes the release id release create prints. Every later command takes it: release show <release-id> and release logs <release-id> --target linux/nvidia, so recovery does not depend on the working directory. That covers localhost at "<install>". A table test over all three distributions checks it, and dropping the id fails all three rows.
| comfy build release create ${directory} --target linux/nvidia --watch | ||
| \`\`\` | ||
|
|
||
| \`--watch\` polls until every target finishes. The release is green when |
There was a problem hiding this comment.
[P2] --watch has no timeout, so a queued or stuck build can hold the coding agent indefinitely. The current comfy-build procedure stops after 30 minutes and reports the release as still running. Please carry that bound and the release id into this brief, with instructions for resuming inspection later.
There was a problem hiding this comment.
Fixed in dd60b54234. --watch is gone. The agent checks with release show <release-id>, and after 30 minutes without a deployable release it stops and reports the id as still running, with comfy build release show <release-id> to resume. The test asserts the bound and that --watch is absent.
| ) | ||
| expect(document).not.toContain('--from-workflow') | ||
| expect(document).not.toContain('comfy which') | ||
| expect(document).not.toContain('sign in first') |
There was a problem hiding this comment.
[P2] This only excludes the old phrase. Deleting the replacement rule at prerequisites() line 109 still leaves both focused suites green. Please positively assert that the generated brief runs comfy cloud login only after a command reports not signed in, so the intended recovery instruction cannot disappear unnoticed.
There was a problem hiding this comment.
Fixed in dd60b54234. A test now asserts the positive rule: run comfy cloud login only when a command answers not signed in. Deleting that line fails it.
| 'write each chosen\n candidate into `definition.models` in `comfy-build.yaml`, as `type`' | ||
| ) | ||
| expect(document).toContain('`filename`, `sourceUri` and\n `sha256`') | ||
| expect(document).toContain('do not claim a recorded version was verified') |
There was a problem hiding this comment.
[P2] The regression test still does not pin the response field used for registry versions. Replacing latest_version.version with the plausible wrong field version leaves both focused suites green. Please assert that the matching row's latest_version.version is written as registryVersion.
There was a problem hiding this comment.
Fixed in dd60b54234. The Cloud test asserts that the matching row's latest_version.version is written as registryVersion. Swapping in version fails it.
| 'takes $distribution to a green release and stops before deploying', | ||
| ({ distribution, directory }) => { | ||
| const document = buildAgentHandoffDocument({ distribution, inputs }) | ||
| const disclosure = document.indexOf('Wait for\na yes.') |
There was a problem hiding this comment.
[P2] This ordering check starts at the cut disclosure, so it cannot detect the current Cloud and Desktop importer commands sending JSON before consent. Add the remote importer and its own disclosure/yes to the ordering oracle, then keep the cut disclosure ordered before push and release create.
There was a problem hiding this comment.
Fixed in dd60b54234. The ordering test now covers the importer: for Cloud and Desktop, the upload disclosure comes before the importer command, the importer before the cut disclosure, the cut disclosure before push, and push before release create. Localhost has no importer, and the test asserts it discloses no upload.
55b0c28 to
0bd8905
Compare
1e4093e to
370157c
Compare
|
DrJKL's latest review: the Verification section now reports the current 7 files and 116 tests at |
DrJKL
left a comment
There was a problem hiding this comment.
Verdict: ready to approve
3a91486 clarifies why copyBrief writes the clipboard before starting the workflow download. The PR Verification section now records the current 7-file/116-test result, so both non-blocking documentation items from the prior round are closed. I found no new issues.
Review record
Automation findings dispositioned before the first reviewer round
- Workflow-value Markdown injection: fixed by literal rendering and unsafe-value omission; the remaining heading syntax was separately fixed.
- Export cancellation and filename drift: fixed by direct download and then by using one normalized filename in both brief and download.
- Preparation rejection outside the error boundary: fixed by moving capture and document construction inside the reported/toasted path.
- Shell metacharacters and control characters in workflow values: fixed by folding controls and omitting unsafe values.
- Clipboard activation lost before download: fixed by copying first.
- Unresolved Desktop install path: fixed by explicit Desktop path guidance and quoted placeholders.
- Exponential subgraph flattening: fixed by traversing collected definitions once.
- Model detection over arbitrary strings: fixed by registry-backed named inputs; unknown loaders now contribute only
properties.models. - Discarded pack version: fixed by retaining and rendering
properties.ver. - Stale graph captured when the dialog opened: fixed by capturing on click.
- Preparation still outside the catch: fixed.
- Clipboard false result lacked telemetry: fixed with
reportError. comfy-coretreated as an external pack: fixed by excluding the sentinel.- Windows interpreter path absent: fixed with Unix, venv, and portable paths and explicit POSIX-command framing.
- Doubled snapshot suffix: fixed.
- Focused-field state read before save preparation: implementation and regression test fixed.
- Unused subgraph definitions included in requirements: fixed by reachable-definition traversal.
- Build output written into the caller directory: fixed with an isolated
comfy-builddirectory. - Copied state surviving a later failure: fixed by clearing before each attempt and testing success-then-failure.
Review 5326998744
- [P1] Ungated action: fixed when the gated entry branch was merged.
- [P1] Backtick caused brief/download filename disagreement: fixed with
handoffFileNameas the single owner. - [P1] Unquoted path placeholders: fixed by quoting every placeholder and instructing the agent to preserve quotes.
- [P2] Markdown in the workflow heading rendered as syntax: fixed with adaptive literal code spans.
- [P2] Full-document equality duplicated implementation prose: fixed with focused behavioral assertions.
- [P3] Copied-label measurement machinery: fixed by using the short label.
- [P3] Comment claimed rejected shell values were illegitimate: fixed with policy-accurate wording.
- [P2] Parsed workflow was trusted as a domain type: fixed by narrowing root nodes and dropping malformed optional fields.
- [P3] Distribution mock duplicated the union: fixed by importing and using the authoritative type.
- [P2] Clipboard boolean was not asserted across failure/fallback branches: fixed with branch coverage.
- [P2] Save-order test did not distinguish pre/post-prepare values: fixed with a fresh-state assertion.
- [P2] Pack pinning lacked an actionable file/field: fixed with registry/repository pin instructions.
- [P3] Verification described obsolete tests and counts: updated, then tracked again after later base merges below.
Review 5327155060 and merge re-anchor 5327171733
- [P1] Brief stopped at
pushwhile promising a green artifact: fixed by creating and checking a release. - [P2] Generic
COMBOcaused model false positives: fixed by mapping specific registered model inputs. - [P3] Duplicate reachable-subgraph traversal: fixed by moving reachability to the shared workflow-flattening owner.
- [P3] Arrow helper violated function-declaration guidance: fixed.
- [P2] No composition test for the real model registry: fixed with real loader/text-node wiring.
- [P2] Desktop could be routed through localhost without a failing composition test: fixed with distribution-specific assertions.
- [P2] CTA promised deployment while the brief stopped before deployment: fixed by renaming it Build with your agent.
- [P3] PR body remained stale in the first response: fixed before the merged-head review.
Review 5327258969
- [P1] Reverse lookup retained only one model input per loader: fixed by preserving every registered input, including both Fooocus inpaint models.
- [P1] Release could omit required
linux/nvidia: fixed by requiring that target. - [P2] Workflow pack version was mistaken for a verified registry version: fixed by using
latest_version.versionor a verified repository commit. - [P2] Consent-order test did not order the release cut: fixed by ordering disclosure, push, and create.
- [P3] Shared traversal remained duplicated: fixed.
- [P2] Deployment wording remained open: fixed by the approved build wording.
- [P3] "being rebased" body text was stale: fixed.
- [P3] Verification count drifted after the base merge: updated, with later recurrences tracked below.
Review 5327331991
- [P1] Resolved model candidates were never written into the build spec: fixed by populating
definition.modelsbefore validation. - [P1] Retry cuts lacked renewed consent: fixed by requiring a new yes after each cause/edit disclosure.
- [P2] Registry search could not verify the claimed older recorded version: fixed by using the returned latest version and stating the limit.
- [P2] Exported node guard promised optional fields it did not validate: fixed for the new boundary by reparsing consumed nodes and dropping malformed optionals; the broad base-owned guard remained unchanged.
- [P2] CTA still promised deployment: fixed.
- [P3] Hidden initialization through a discarded getter: fixed by calling
registerDefaults()explicitly. - [P3]
currentInputshid snapshot mutation: fixed by naming itcaptureInputs. - [P3] Positional-array comment overstated model certainty: fixed by documenting it as a legacy suffix heuristic and its false-positive limit.
- [P3] Fence oracle called the generator twice: fixed at the time with a literal count; that replacement is now reopened below as an implementation-coupled change detector.
- [P3] Desktop forced an unnecessary login: fixed by using the shared conditional-login rule.
- [P3] Stale rebase wording: fixed.
- [P3] Verification count drift after another base merge: updated.
Review 5341471633
- [P1] Cloud/Desktop uploaded JSON before disclosure and consent: fixed with separate importer consent before transmission.
- [P1] Localhost recovery lost release/build scope: fixed by recording and passing the release id.
- [P2] Digestless resolver candidates were impossible to select: fixed with disclosure and agreement for unpinned sources.
- [P2] Unbounded
--watchcould hold the agent forever: fixed with a 30-minute bound and resumable release id. - [P2] Conditional login rule was tested only by excluding old prose: fixed with a positive assertion.
- [P2] Registry response field was not pinned by a test: fixed with an assertion for
latest_version.versiontoregistryVersion. - [P2] Consent-order test ignored pre-cut importer transmission: fixed by ordering importer disclosure/yes separately from cut consent.
- [P3] PR body still described an old rebase state: fixed.
- [P3] Verification counts were stale after the merge: updated.
- Literal fence count was initially cleared as an independent oracle; this disposition is withdrawn in the current round because a differential safe-input oracle is independent without freezing unrelated document structure.
Review 5342187678
- [P3] Consent-order table branched inside the test body: fixed by splitting Cloud/Desktop, localhost, and common cut-order behavior into branch-free tests.
- [P3] Verification count reopened after the base merge: updated.
- Previously reopened consent, release-id, digestless-candidate, timeout, login, registry-field, and ordering findings were confirmed fixed.
Review 5342308689
- [P3] Localhost test name claimed no upload while proving only no hosted importer: fixed by naming the importer-specific invariant.
- [P3] Verification count reopened after the base merge: updated.
Review 5342385867
- Localhost test naming: confirmed fixed.
- Verification count: confirmed fixed at that head.
Review 5342460111
- [P3] Verification count reopened after the base merge: updated.
- Merge resolution: confirmed as Git's faithful automatic merge with no code/test finding reopened.
Review 5342741308
- [P3] Claimed touched-only, base-stable verification scope: superseded by the broader finding in the next round.
- Merge resolution: confirmed as Git's faithful automatic merge.
Review 5357359300
- [P3] Touched-test verification scope omitted one touched file, included six untouched files, and had a stale count: fixed at
0969b60. The PR body now lists exactly seven touched test files, and the command passes 7 files and 111 tests. The author included this disposition in their reply. - [P3] Literal count of 16 fence lines coupled the injection test to unrelated static blocks: fixed at
8202060with a same-shape safe control. The dangerous-text assertions remain. The author answered in the inline thread. - Nit: [P3]
fenceLinesuses a function expression where engineering guideline 9 requires a declaration: open in the current inline review. - Nit: [P3] The
subgraphfixture helper uses the same prohibited function-expression form: open in the current inline review. - [P2] Importer-consent disclosure test does not pin that Cloud uploads the whole workflow JSON and Desktop uploads the whole snapshot JSON: confirmed in CodeRabbit's existing thread. The implementation is currently correct; this is a regression-test gap.
- Loose linear traversal bound: retained. Review
5327331991already recorded that the margin catches exponential traversal without freezing exact getter access, which would test implementation structure rather than behavior. - Current base merge: confirmed as Git's faithful automatic merge. The head tree exactly matches the merge of
8202060and45d1748. - Checks at
0969b60: 7 files and 111 tests passed; typecheck, changed-file ESLint, oxlint, formatting, and knip passed. Oxlint emitted only unrelated existing warnings. Correctness, maintainability, idiomaticity, tests, and documentation lanes covered all 15 files; no Playwright file changed.
Review 5357708471 and current round
- [P2] Importer-consent test did not pin that Cloud uploads the whole workflow JSON and Desktop uploads the whole snapshot JSON: fixed at
16bb5cf. Each table row now supplies its exact payload, and the test orders payload disclosure, consent, importer command, and cut disclosure. The author confirmed the mutation in their reply. - Nit: [P3]
fenceLinesused a function expression: fixed at16bb5cfwith a module-level declaration; author reply, verification reply. - Nit: [P3]
subgraphused a function expression: fixed at16bb5cfwith a module-level declaration; author reply, verification reply. - Conditional icon class in
LoopingVideo.vue: not a PR finding. The binding comes unchanged from current base680c336; this branch only consumes the inherited component. - Manual conflict in
DeployToComfyApiCard.vue: cleared at3e4103c. The resolution preserves the handoff controls and incomingLoopingVideobehavior. - Checks at
16bb5cf: 7 files and 111 tests passed; typecheck, changed-file ESLint, oxlint, formatting, and knip passed. Oxlint emitted only unrelated existing warnings. Playwright listed 3 deploy-dialog tests. All standard lanes covered the 15-file PR diff and six merge-delta files.
Review 5393907279
- [P2] Conflicting recorded pack versions were order-dependent: fixed at
92617b1by retaining every distinct version and asking the user to choose. - [P2] Frontend prose copied version-sensitive CLI policy: fixed across
92617b1and126e681by delegating build mechanics to thecomfy-buildrecipe. - [P2] Cloud download failure after a successful copy lacked regression coverage: fixed at
92617b1. - Nit: [P3] Redundant clipboard fallback cross-product row: deleted at
92617b1. - Nit: [P3] Redundant unknown-node model row: deleted at
92617b1. - Nit: [P3] Markdown tests pinned exact wrapping: fixed at
92617b1with whitespace-tolerant behavioral checks. - Nit: [P3] Card test repeated placeholder coverage: deleted at
92617b1. - [P2] PR body misstated malformed optional-field behavior: fixed in the PR description after this review.
- Nit: [P3] Unsupported Fallow verification claim: removed from the PR description after this review.
Review 5394077367
- [P2] The brief still copied commands and target policy despite naming the recipe as owner: fixed at
126e681. - Nit: [P3] Derived build-input arrays were mutable: fixed at
126e681with readonly contracts. - [P2] Recipe-order assertion passed when the recipe marker was absent: fixed at
126e681. - Nit: [P3] Distribution-independent recipe and retry assertions were repeated: consolidated at
126e681. - Nit: [P3] Pack-order table branched in its body: fixed at
126e681by moving complete inputs into rows. - Nit: [P3] Review Focus overstated recipe ownership of targets: fixed in the PR description after this review.
Review 5394198011
- [P2] Stale shell filtering removed valid workflow facts: fixed at
6eda6ed. The author answered in the thread. - [P2] No-build-command guard missed root options: fixed at
6eda6ed. The author answered in the thread. - Nit: [P3] Review Focus overstated policy ownership: fixed in the PR description and recorded in the discussion.
Review 5394355695
- [P2] One malformed subgraph node discarded valid siblings: fixed at
a29b454by parsing members independently. The author answered in the thread.
Review 5394483555
- [P1] Malformed non-list
nodesbecame an empty valid definition: fixed atdba59a4. Non-list collections remain invalid; the author answered in the thread. - [P2] Recursive normalization and collection overflowed on cycles or depth: fixed at
dba59a4with object-identity and iterator worklists. The author answered in the thread. - [P2] The malformed-sibling regression did not exercise nesting: fixed at
dba59a4. The author answered in the thread.
Current round
- [P2] PR description called workflow-value normalization lossless: fixed in the current PR body. It now names control-character replacement, whitespace collapse and trimming, retained values, and filename filtering. The author recorded the change in the discussion.
- Nit: [P3] PR description overstated the push/cut consent frequency: fixed in the current PR body. It now says one yes covers one push/cut attempt and each retry needs fresh consent. The author recorded the change in the discussion.
- Maintainability candidate to collapse deploy parsing into the shared collector: withdrawn. The current split keeps tolerant external-input policy out of the strict core utility.
- Test-quality audit: all four changed regression tests are behavioral and distinct. None is tautological, implementation-coupled, or redundant.
- Checks at
dba59a4: head unchanged; prior results remain 7 focused files and 112 passing tests, plus passing typecheck, changed-file ESLint, repository oxlint, formatting, and knip. This round rechecked the changed PR-description claims against the implementation. No owner feedback exposed a lane-brief gap.
Review 5395766315
- [P1] Distribution mock omitted
isCloud, souseAgentHandoff.test.tscould not collect: fixed atef2dd99. The seven focused files now pass 116 tests; verification reply. - [P2] PR Verification section still reports only the historical green run at
6eda6ed: open, non-blocking. Current evidence is 7 files and 116 tests atef2dd99. - Nit: [P3]
copyBriefdocblock has an ambiguous long sentence: open, non-blocking in the thread. - The new retry-consent policy was cleared as an explicit product change. Import uploads and the first cut remain gated; retries follow the recipe's limits and deployment needs a separate request.
- All changed tests were classified as behavioral with distinct falsifiers. None was tautological, implementation-coupled, redundant, or a change detector.
Current round
- Checks at
ef2dd99: 7 focused files and 116 tests passed; typecheck, changed-file ESLint, repository oxlint, formatting, and knip passed. - Standard correctness, maintainability, idiomaticity, tests, and documentation review covered the one-file fix and carried both non-blocking documentation items forward. No owner feedback exposed a lane-brief gap.
Review 5396205020
- [P2] PR Verification section reported only the historical 108-test run: fixed in the current PR body, which records 7 files and 116 tests at
3a91486. - Nit: [P3]
copyBriefdocblock had an ambiguous long sentence: fixed at3a91486. The author answered in the thread.
Current round
- Correctness, maintainability, idiomaticity, tests, and documentation review found no new issue in the one-file documentation-only delta.
- The owner test-quality request remains covered by the tests lane; no owner feedback exposed a lane-brief gap.
- Checks at
3a91486: 7 focused files and 116 tests passed; typecheck, changed-file ESLint, repository oxlint, formatting, and knip passed.oxlint:mainis not defined in this checkout, so repositorypnpm oxlintwas used and passed with pre-existing warnings.
|
This head has green checks, but GitHub reports merge conflicts against the declared base |
…rols Its browser tests pin the original silent looping design with no play or pause button; the deploy card keeps its control.
…' into HEAD # Conflicts: # src/platform/workflow/deploy/components/DeployToComfyApiCard.vue
DrJKL
left a comment
There was a problem hiding this comment.
Verdict: ready to approve
f137c80 closes the final owner-review round. The PR now parses untrusted workflow data once into typed recursive definitions, keeps unknown at that external boundary, and accesses mocked composables through their public functions instead of storing module-scoped mock references. The full local verification suite and every required GitHub check pass. I found no open findings.
Review record
Automation findings dispositioned before the first reviewer round
- Workflow-value Markdown injection: fixed by literal rendering and unsafe-value omission; the remaining heading syntax was separately fixed.
- Export cancellation and filename drift: fixed by direct download and then by using one normalized filename in both brief and download.
- Preparation rejection outside the error boundary: fixed by moving capture and document construction inside the reported/toasted path.
- Shell metacharacters and control characters in workflow values: fixed by folding controls and omitting unsafe values.
- Clipboard activation lost before download: fixed by copying first.
- Unresolved Desktop install path: fixed by explicit Desktop path guidance and quoted placeholders.
- Exponential subgraph flattening: fixed by traversing collected definitions once.
- Model detection over arbitrary strings: fixed by registry-backed named inputs; unknown loaders now contribute only
properties.models. - Discarded pack version: fixed by retaining and rendering
properties.ver. - Stale graph captured when the dialog opened: fixed by capturing on click.
- Preparation still outside the catch: fixed.
- Clipboard false result lacked telemetry: fixed with
reportError. comfy-coretreated as an external pack: fixed by excluding the sentinel.- Windows interpreter path absent: fixed with Unix, venv, and portable paths and explicit POSIX-command framing.
- Doubled snapshot suffix: fixed.
- Focused-field state read before save preparation: implementation and regression test fixed.
- Unused subgraph definitions included in requirements: fixed by reachable-definition traversal.
- Build output written into the caller directory: fixed with an isolated
comfy-builddirectory. - Copied state surviving a later failure: fixed by clearing before each attempt and testing success-then-failure.
Review 5326998744
- [P1] Ungated action: fixed when the gated entry branch was merged.
- [P1] Backtick caused brief/download filename disagreement: fixed with
handoffFileNameas the single owner. - [P1] Unquoted path placeholders: fixed by quoting every placeholder and instructing the agent to preserve quotes.
- [P2] Markdown in the workflow heading rendered as syntax: fixed with adaptive literal code spans.
- [P2] Full-document equality duplicated implementation prose: fixed with focused behavioral assertions.
- [P3] Copied-label measurement machinery: fixed by using the short label.
- [P3] Comment claimed rejected shell values were illegitimate: fixed with policy-accurate wording.
- [P2] Parsed workflow was trusted as a domain type: fixed by narrowing root nodes and dropping malformed optional fields.
- [P3] Distribution mock duplicated the union: fixed by importing and using the authoritative type.
- [P2] Clipboard boolean was not asserted across failure/fallback branches: fixed with branch coverage.
- [P2] Save-order test did not distinguish pre/post-prepare values: fixed with a fresh-state assertion.
- [P2] Pack pinning lacked an actionable file/field: fixed with registry/repository pin instructions.
- [P3] Verification described obsolete tests and counts: updated, then tracked again after later base merges below.
Review 5327155060 and merge re-anchor 5327171733
- [P1] Brief stopped at
pushwhile promising a green artifact: fixed by creating and checking a release. - [P2] Generic
COMBOcaused model false positives: fixed by mapping specific registered model inputs. - [P3] Duplicate reachable-subgraph traversal: fixed by moving reachability to the shared workflow-flattening owner.
- [P3] Arrow helper violated function-declaration guidance: fixed.
- [P2] No composition test for the real model registry: fixed with real loader/text-node wiring.
- [P2] Desktop could be routed through localhost without a failing composition test: fixed with distribution-specific assertions.
- [P2] CTA promised deployment while the brief stopped before deployment: fixed by renaming it Build with your agent.
- [P3] PR body remained stale in the first response: fixed before the merged-head review.
Review 5327258969
- [P1] Reverse lookup retained only one model input per loader: fixed by preserving every registered input, including both Fooocus inpaint models.
- [P1] Release could omit required
linux/nvidia: fixed by requiring that target. - [P2] Workflow pack version was mistaken for a verified registry version: fixed by using
latest_version.versionor a verified repository commit. - [P2] Consent-order test did not order the release cut: fixed by ordering disclosure, push, and create.
- [P3] Shared traversal remained duplicated: fixed.
- [P2] Deployment wording remained open: fixed by the approved build wording.
- [P3] "being rebased" body text was stale: fixed.
- [P3] Verification count drifted after the base merge: updated, with later recurrences tracked below.
Review 5327331991
- [P1] Resolved model candidates were never written into the build spec: fixed by populating
definition.modelsbefore validation. - [P1] Retry cuts lacked renewed consent: fixed by requiring a new yes after each cause/edit disclosure.
- [P2] Registry search could not verify the claimed older recorded version: fixed by using the returned latest version and stating the limit.
- [P2] Exported node guard promised optional fields it did not validate: fixed for the new boundary by reparsing consumed nodes and dropping malformed optionals; the broad base-owned guard remained unchanged.
- [P2] CTA still promised deployment: fixed.
- [P3] Hidden initialization through a discarded getter: fixed by calling
registerDefaults()explicitly. - [P3]
currentInputshid snapshot mutation: fixed by naming itcaptureInputs. - [P3] Positional-array comment overstated model certainty: fixed by documenting it as a legacy suffix heuristic and its false-positive limit.
- [P3] Fence oracle called the generator twice: fixed at the time with a literal count; that replacement is now reopened below as an implementation-coupled change detector.
- [P3] Desktop forced an unnecessary login: fixed by using the shared conditional-login rule.
- [P3] Stale rebase wording: fixed.
- [P3] Verification count drift after another base merge: updated.
Review 5341471633
- [P1] Cloud/Desktop uploaded JSON before disclosure and consent: fixed with separate importer consent before transmission.
- [P1] Localhost recovery lost release/build scope: fixed by recording and passing the release id.
- [P2] Digestless resolver candidates were impossible to select: fixed with disclosure and agreement for unpinned sources.
- [P2] Unbounded
--watchcould hold the agent forever: fixed with a 30-minute bound and resumable release id. - [P2] Conditional login rule was tested only by excluding old prose: fixed with a positive assertion.
- [P2] Registry response field was not pinned by a test: fixed with an assertion for
latest_version.versiontoregistryVersion. - [P2] Consent-order test ignored pre-cut importer transmission: fixed by ordering importer disclosure/yes separately from cut consent.
- [P3] PR body still described an old rebase state: fixed.
- [P3] Verification counts were stale after the merge: updated.
- Literal fence count was initially cleared as an independent oracle; this disposition is withdrawn in the current round because a differential safe-input oracle is independent without freezing unrelated document structure.
Review 5342187678
- [P3] Consent-order table branched inside the test body: fixed by splitting Cloud/Desktop, localhost, and common cut-order behavior into branch-free tests.
- [P3] Verification count reopened after the base merge: updated.
- Previously reopened consent, release-id, digestless-candidate, timeout, login, registry-field, and ordering findings were confirmed fixed.
Review 5342308689
- [P3] Localhost test name claimed no upload while proving only no hosted importer: fixed by naming the importer-specific invariant.
- [P3] Verification count reopened after the base merge: updated.
Review 5342385867
- Localhost test naming: confirmed fixed.
- Verification count: confirmed fixed at that head.
Review 5342460111
- [P3] Verification count reopened after the base merge: updated.
- Merge resolution: confirmed as Git's faithful automatic merge with no code/test finding reopened.
Review 5342741308
- [P3] Claimed touched-only, base-stable verification scope: superseded by the broader finding in the next round.
- Merge resolution: confirmed as Git's faithful automatic merge.
Review 5357359300
- [P3] Touched-test verification scope omitted one touched file, included six untouched files, and had a stale count: fixed at
0969b60. The PR body now lists exactly seven touched test files, and the command passes 7 files and 111 tests. The author included this disposition in their reply. - [P3] Literal count of 16 fence lines coupled the injection test to unrelated static blocks: fixed at
8202060with a same-shape safe control. The dangerous-text assertions remain. The author answered in the inline thread. - Nit: [P3]
fenceLinesuses a function expression where engineering guideline 9 requires a declaration: open in the current inline review. - Nit: [P3] The
subgraphfixture helper uses the same prohibited function-expression form: open in the current inline review. - [P2] Importer-consent disclosure test does not pin that Cloud uploads the whole workflow JSON and Desktop uploads the whole snapshot JSON: confirmed in CodeRabbit's existing thread. The implementation is currently correct; this is a regression-test gap.
- Loose linear traversal bound: retained. Review
5327331991already recorded that the margin catches exponential traversal without freezing exact getter access, which would test implementation structure rather than behavior. - Current base merge: confirmed as Git's faithful automatic merge. The head tree exactly matches the merge of
8202060and45d1748. - Checks at
0969b60: 7 files and 111 tests passed; typecheck, changed-file ESLint, oxlint, formatting, and knip passed. Oxlint emitted only unrelated existing warnings. Correctness, maintainability, idiomaticity, tests, and documentation lanes covered all 15 files; no Playwright file changed.
Review 5357708471 and current round
- [P2] Importer-consent test did not pin that Cloud uploads the whole workflow JSON and Desktop uploads the whole snapshot JSON: fixed at
16bb5cf. Each table row now supplies its exact payload, and the test orders payload disclosure, consent, importer command, and cut disclosure. The author confirmed the mutation in their reply. - Nit: [P3]
fenceLinesused a function expression: fixed at16bb5cfwith a module-level declaration; author reply, verification reply. - Nit: [P3]
subgraphused a function expression: fixed at16bb5cfwith a module-level declaration; author reply, verification reply. - Conditional icon class in
LoopingVideo.vue: not a PR finding. The binding comes unchanged from current base680c336; this branch only consumes the inherited component. - Manual conflict in
DeployToComfyApiCard.vue: cleared at3e4103c. The resolution preserves the handoff controls and incomingLoopingVideobehavior. - Checks at
16bb5cf: 7 files and 111 tests passed; typecheck, changed-file ESLint, oxlint, formatting, and knip passed. Oxlint emitted only unrelated existing warnings. Playwright listed 3 deploy-dialog tests. All standard lanes covered the 15-file PR diff and six merge-delta files.
Review 5393907279
- [P2] Conflicting recorded pack versions were order-dependent: fixed at
92617b1by retaining every distinct version and asking the user to choose. - [P2] Frontend prose copied version-sensitive CLI policy: fixed across
92617b1and126e681by delegating build mechanics to thecomfy-buildrecipe. - [P2] Cloud download failure after a successful copy lacked regression coverage: fixed at
92617b1. - Nit: [P3] Redundant clipboard fallback cross-product row: deleted at
92617b1. - Nit: [P3] Redundant unknown-node model row: deleted at
92617b1. - Nit: [P3] Markdown tests pinned exact wrapping: fixed at
92617b1with whitespace-tolerant behavioral checks. - Nit: [P3] Card test repeated placeholder coverage: deleted at
92617b1. - [P2] PR body misstated malformed optional-field behavior: fixed in the PR description after this review.
- Nit: [P3] Unsupported Fallow verification claim: removed from the PR description after this review.
Review 5394077367
- [P2] The brief still copied commands and target policy despite naming the recipe as owner: fixed at
126e681. - Nit: [P3] Derived build-input arrays were mutable: fixed at
126e681with readonly contracts. - [P2] Recipe-order assertion passed when the recipe marker was absent: fixed at
126e681. - Nit: [P3] Distribution-independent recipe and retry assertions were repeated: consolidated at
126e681. - Nit: [P3] Pack-order table branched in its body: fixed at
126e681by moving complete inputs into rows. - Nit: [P3] Review Focus overstated recipe ownership of targets: fixed in the PR description after this review.
Review 5394198011
- [P2] Stale shell filtering removed valid workflow facts: fixed at
6eda6ed. The author answered in the thread. - [P2] No-build-command guard missed root options: fixed at
6eda6ed. The author answered in the thread. - Nit: [P3] Review Focus overstated policy ownership: fixed in the PR description and recorded in the discussion.
Review 5394355695
- [P2] One malformed subgraph node discarded valid siblings: fixed at
a29b454by parsing members independently. The author answered in the thread.
Review 5394483555
- [P1] Malformed non-list
nodesbecame an empty valid definition: fixed atdba59a4. Non-list collections remain invalid; the author answered in the thread. - [P2] Recursive normalization and collection overflowed on cycles or depth: fixed at
dba59a4with object-identity and iterator worklists. The author answered in the thread. - [P2] The malformed-sibling regression did not exercise nesting: fixed at
dba59a4. The author answered in the thread.
Current round
- [P2] PR description called workflow-value normalization lossless: fixed in the current PR body. It now names control-character replacement, whitespace collapse and trimming, retained values, and filename filtering. The author recorded the change in the discussion.
- Nit: [P3] PR description overstated the push/cut consent frequency: fixed in the current PR body. It now says one yes covers one push/cut attempt and each retry needs fresh consent. The author recorded the change in the discussion.
- Maintainability candidate to collapse deploy parsing into the shared collector: withdrawn. The current split keeps tolerant external-input policy out of the strict core utility.
- Test-quality audit: all four changed regression tests are behavioral and distinct. None is tautological, implementation-coupled, or redundant.
- Checks at
dba59a4: head unchanged; prior results remain 7 focused files and 112 passing tests, plus passing typecheck, changed-file ESLint, repository oxlint, formatting, and knip. This round rechecked the changed PR-description claims against the implementation. No owner feedback exposed a lane-brief gap.
Review 5395766315
- [P1] Distribution mock omitted
isCloud, souseAgentHandoff.test.tscould not collect: fixed atef2dd99. The seven focused files now pass 116 tests; verification reply. - [P2] PR Verification section still reports only the historical green run at
6eda6ed: open, non-blocking. Current evidence is 7 files and 116 tests atef2dd99. - Nit: [P3]
copyBriefdocblock has an ambiguous long sentence: open, non-blocking in the thread. - The new retry-consent policy was cleared as an explicit product change. Import uploads and the first cut remain gated; retries follow the recipe's limits and deployment needs a separate request.
- All changed tests were classified as behavioral with distinct falsifiers. None was tautological, implementation-coupled, redundant, or a change detector.
Current round
- Checks at
ef2dd99: 7 focused files and 116 tests passed; typecheck, changed-file ESLint, repository oxlint, formatting, and knip passed. - Standard correctness, maintainability, idiomaticity, tests, and documentation review covered the one-file fix and carried both non-blocking documentation items forward. No owner feedback exposed a lane-brief gap.
Review 5396205020
- [P2] PR Verification section reported only the historical 108-test run: fixed in the current PR body, which records 7 files and 116 tests at
3a91486. - Nit: [P3]
copyBriefdocblock had an ambiguous long sentence: fixed at3a91486. The author answered in the thread.
Current round
- Correctness, maintainability, idiomaticity, tests, and documentation review found no new issue in the one-file documentation-only delta.
- The owner test-quality request remains covered by the tests lane; no owner feedback exposed a lane-brief gap.
- Checks at
3a91486: 7 focused files and 116 tests passed; typecheck, changed-file ESLint, repository oxlint, formatting, and knip passed.oxlint:mainis not defined in this checkout, so repositorypnpm oxlintwas used and passed with pre-existing warnings.
Final owner-review round
- Typed traversal: fixed at
f137c80.unknown[]remains only at the untrusted workflow boundary; parsing produces recursiveFlattenableSubgraphDefinitionvalues before collection and reachability traversal. - Mock ownership: fixed at
f137c80. Tests configure and inspectuseAgentHandoff().copyBriefanduseCopyToClipboard().copyToClipboardthrough the mocked modules instead of retaining separate aliases. - The 15 direct owner comments on fixture order, hoisted state, table typing, comments, helper naming, constant-returning helpers, tagged-template composition, and redundant structure were fixed at
784c611and rechecked at the final head. - Final checks at
f137c80: 7 focused files and 116 tests passed; the missing-model blast-radius run passed 2 files and 103 tests; typecheck, changed-file ESLint, type-aware oxlint, formatting, knip, and pre-push Fallow passed. Every required GitHub check is green.
# Conflicts: # src/locales/en/main.json # src/platform/workflow/deploy/components/DeployToComfyApiCard.test.ts # src/platform/workflow/deploy/components/DeployToComfyApiCard.vue
…#20938) ## Summary Turns review findings that keep recurring by hand into lint rules and tooling tests. Existing violations are baselined in `oxlint-suppressions.json`, so only new code is held. Fixes [FE-3420](https://linear.app/comfyorg/issue/FE-3420) ```diff tools/oxlint-plugins/ +├── noReexports.ts # comfy/no-reexports: export … from, or import-then-export (named or default), of ./, ../, @/ modules in src +├── playwrightPoll.ts # comfy/no-throw-in-expect-poll: a throw escapes poll retry (browser_tests, apps/*/e2e) +├── positionalBoolean.ts # comfy/no-positional-boolean-params: exported fns + public methods, 2+ params +├── arrayContaining.ts # comfy/prefer-array-containing: sorting an expect() subject or a literal array to compare without order +├── vitestHoisted.ts # comfy/no-unread-hoisted-mock-state: vi.hoisted binding no vi.mock factory or other vi.hoisted callback reads +├── astTypes.ts # node types derived from oxlint's visitor signatures └── comfy.ts # rule registry: static .ts imports replace the createRequire indirection tools/repo-checks/ +├── docsPackageManager.test.ts # no npm/npx commands in tracked Markdown (agent skill dirs excluded) +└── vitestGuards.test.ts # every Vitest project's setupFiles loads the network guard; timer-guard exceptions pinned per project .oxlintrc.json + jsPlugins: eslint-plugin-regexp — no-dupe-disjunctions, no-dupe-characters-character-class, + no-empty-alternative, no-useless-{assertions,lazy,quantifier} ``` Baselined counts (per rule, from `--suppress-all`): reexports 138, hoisted 161, array-containing 52, positional booleans 112 (burn-down: [FE-3418](https://linear.app/comfyorg/issue/FE-3418)), regexp 19. The regexp autofixes `lint:fix` applies in CI are committed here: duplicate `a-z` ranges under the `i` flag and useless lazy quantifiers in five files. The rules found two live violations, fixed here: a `throw` inside `expect.poll` in `pointerGestures.spec.ts` (now returns `null` so the poll retries), and a sorted comparison in `agentApiSchema.test.ts`, which now uses `expect.arrayContaining`. `no-reexports` only covers source modules (`./`, `../`, `@/`). Re-exporting a package such as `@comfyorg/ingest-types` stays allowed, which matches the existing "allows generated contracts" probe. `regexp/no-super-linear-backtracking` is not enabled: CI's `lint:fix` applies its autofixes, which rewrite unrelated regexes into unreadable forms. Before it was dropped, it found exponential backtracking in `src/platform/telemetry/redactTelemetryUrls.ts`: a 58-character dotted token takes 3.1 s. Tracked in [FE-3422](https://linear.app/comfyorg/issue/FE-3422). `pnpm-lock.yaml` only adds the plugin's packages. `jsdoc-type-pratt-parser` is pinned to 9.2.0 in the lockfile, because 9.2.2 depends on `@types/node ^26`, which re-resolved vite and vitest onto `@types/node` 26 and broke the website typecheck. 9.2.3 drops that dependency but is newer than the minimum release age. ## Rules Each block gives the rule's rationale, why fixing a violation is worth it, and a violation from this repo (or a minimal one when the repo has none). <details> <summary><code>comfy/no-reexports</code>: import from the defining module, not a barrel</summary> **Rationale.** `docs/guidance/engineering.md` bans barrel files inside `src`. A re-export adds a second import path for the same symbol, so two modules can reach one class through different paths. It hides the real dependency, which makes import cycles and layer violations harder to see. Owners raised this by hand on Comfy-Org#15884. Re-exports of packages (`@comfyorg/ingest-types`) are still allowed. **Why fix.** Callers that import the defining module show their actual dependency, and dropping the barrel removes the import order the barrel was quietly enforcing. Baseline: 138 sites, 54 in `litegraph.ts` and 36 in `src/types/index.ts`. ```ts // src/lib/litegraph/src/litegraph.ts export { LinkConnector } from './canvas/LinkConnector' // src/types/index.ts export type { ComfyApi } from '@/scripts/api' ``` ```ts // caller, after import { LinkConnector } from '@/lib/litegraph/src/canvas/LinkConnector' ``` Also reported: `import { a } from './a'` followed by `export { a }` or `export default a`. </details> <details> <summary><code>comfy/no-throw-in-expect-poll</code>: return a sentinel so the poll retries</summary> **Rationale.** Playwright's `expect.poll` retries only when the matcher fails. If the callback throws, the error escapes on the first tick and the test fails without waiting, which is the race the poll was there to absorb (Comfy-Org#17412). **Why fix.** The throw makes a timing-dependent test flaky in exactly the state it should tolerate: the element is not rendered yet. Returning `null` keeps the retry, and the matcher still fails with a clear diff if the state never arrives. This PR fixes the only live site. ```ts // browser_tests/tests/vueNodes/interactions/canvas/pointerGestures.spec.ts, before await expect.poll(async () => { const aAfter = await a.boundingBox() const bAfter = await b.boundingBox() if (!aAfter || !bAfter) throw new Error('Both nodes must be rendered') return { x: aAfter.x - bAfter.x, y: aAfter.y - bAfter.y } }).toEqual(expected) // after if (!aAfter || !bAfter) return null ``` </details> <details> <summary><code>comfy/no-positional-boolean-params</code>: pass flags through an options object</summary> **Rationale.** A boolean argument at a call site carries no name, so `setDirty(true, true)` cannot be read without opening the declaration. Adding a third flag silently shifts the meaning of every existing call. Scope: exported functions and public methods with two or more parameters. A single setter-style flag (`setVisible(visible)`) is allowed, and so are `boolean | Options` unions, so deprecated overloads keep working (Comfy-Org#20098). **Why fix.** Named options are self-documenting at every call site and can grow without breaking callers. Baseline: 112 sites, burned down in [FE-3418](https://linear.app/comfyorg/issue/FE-3418). Extension-facing classes keep the old signature as a deprecated overload. ```ts // src/lib/litegraph/src/LGraphCanvas.ts setDirty(fgcanvas: boolean, bgcanvas?: boolean): void this.setDirty(true, true) // after setDirty({ foreground, background }: { foreground?: boolean; background?: boolean }): void this.setDirty({ foreground: true, background: true }) ``` </details> <details> <summary><code>comfy/prefer-array-containing</code>: compare without order through <code>expect.arrayContaining</code></summary> **Rationale.** Sorting the actual value, or a literal, just to compare it without regard to order encodes a detail the test does not care about. It also hides what the assertion means. `expect.arrayContaining` states the intent directly; add `toHaveLength` when extra items must fail. **Why fix.** The sorted form fails on a non-default comparator (numbers sort as strings) and adds noise on both sides. The intent-revealing matcher gives a clearer diff on failure. Baseline: 52 sites. ```ts // src/platform/cloud/onboarding/survey/surveySchema.test.ts expect(values.slice(-2).sort()).toEqual(['not_sure', 'other']) // after expect(values.slice(-2)).toEqual(expect.arrayContaining(['not_sure', 'other'])) ``` </details> <details> <summary><code>comfy/no-unread-hoisted-mock-state</code>: hoist only what hoisted code reads</summary> **Rationale.** `vi.hoisted` exists so a `vi.mock` factory can read a value before imports run. A binding no factory and no other `vi.hoisted` callback reads gains nothing from hoisting. It still pays the costs: no imports inside the callback (hence `require` plus lint suppressions) and state that lives outside the test lifecycle (Comfy-Org#17966). **Why fix.** A plain `const` with normal imports is shorter, type-checked without casts, and needs no suppressions. Tests that configure the mock can use `vi.mocked(importedFn)` on the real module. Baseline: 161 bindings. ```ts // src/platform/workspace/composables/useTeamPlan.test.ts, today const { mockIsTeamPlan, ... } = vi.hoisted(() => { // oxlint-disable-next-line typescript/no-require-imports, typescript/consistent-type-imports const { ref } = require('vue') as typeof import('vue') return { mockIsTeamPlan: ref(true), ... } }) vi.mock(import('@/composables/billing/useBillingContext')) // after: the factory is automock, so nothing needs hoisting import { ref } from 'vue' const mockIsTeamPlan = ref(true) ``` </details> <details> <summary><code>regexp/*</code> (eslint-plugin-regexp): redundant or ambiguous regex parts</summary> **Rationale.** These rules find parts of a regex that can never change the result, or that make the engine try the same match several ways (Comfy-Org#20033). Enabled: `no-dupe-disjunctions`, `no-dupe-characters-character-class`, `no-empty-alternative`, `no-useless-assertions`, `no-useless-lazy`, `no-useless-quantifier`. `no-super-linear-backtracking` is not enabled; see FE-3422. **Why fix.** Redundant alternatives mislead the reader about what the pattern accepts and can add backtracking. An empty alternative silently accepts the empty string. Baseline: 19 sites. The autofixes CI applies are committed here. ```ts // src/workbench/extensions/agent/crdt/crdtDebugReport.ts: `apikey` is already matched by `api[-_]?key` /(token|secret|password|passwd|credential|api[-_]?key|apikey|auth|bearer|session|cookie|private)/i // src/platform/telemetry/utils/surveyNormalization.ts: the trailing `|` also matches '' industry.match(/^(other|none|undefined|unknown|n\/a|not applicable|-|)$/) // src/platform/telemetry/redactTelemetryUrls.ts: `a-z` duplicates `A-Z` under the `i` flag /[A-Za-z0-9._~%-]+/i ``` </details> <details> <summary>Repo checks: <code>tools/repo-checks/*.test.ts</code></summary> **`docsPackageManager`.** Tracked Markdown outside agent-skill directories must not show `npm`/`npx` commands, including `npm --prefix … install`. The repo is pnpm-only, and an `npm install` in a package README sent a reviewer back on Comfy-Org#19872. **`vitestGuards`.** Every Vitest project's `setupFiles` must load the network guard (`vitest.network.setup.ts`, or `vitest.setup.ts`, which imports it). A missing guard let tests make live calls to docs.comfy.org on Comfy-Org#19713. The timer guard is required except for six pinned projects. The list fails when it changes, so it can only shrink deliberately. </details> ## Evidence - `pnpm exec vitest run --project tooling tools/oxlint-plugins tools/repo-checks`: 245 passed. Mutating each rule (inverting the hoisted check, dropping the `protected` guard, reversing the sort order) fails its test. - `pnpm exec oxlint . --type-aware`: exit 0 with the baseline. Before baselining it reported 138 / 164 / 112 / 79 new violations. - New code still fails. Appending `export { LGraph as LGraphAlias } from './LGraph'` to `litegraph.ts` fails `comfy/no-reexports`, and this scratch file fails both rules: ```ts export class Canvas { deselectAll(keep?: object, sendEvent?: boolean) {} } export const LINK = /(?:https?:\/\/)?(?:[a-z]|[a-z0-9])+\.com/i // → comfy(no-positional-boolean-params), regexp(no-dupe-disjunctions) ``` - `pnpm typecheck:tools`, `pnpm typecheck:browser`, and `pnpm typecheck:website`: clean. `pnpm install --frozen-lockfile` and `--lockfile-only` leave the lockfile unchanged. - Guard checks fail when the network guard is removed from `apps/website/vitest.config.ts` or `setupFiles` is removed from `packages/design-system/vitest.config.ts`. ## Merge Danger **Door:** two-way Revert the config entries to turn the rules off. Suppression counts only ratchet down. **Blast Radius:** contributors New code that adds a re-export, a positional boolean, or an unread `vi.hoisted` binding now fails lint. Fixing a baselined site makes its suppression stale, and the fix is `pnpm oxlint --prune-suppressions` (documented in TROUBLESHOOTING.md). https://ampcode.com/threads/T-155ce223-5111-4d69-a3a8-5d90c12378bb --------- Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Summary
Adds the coding-agent lane to the Deploy to Comfy API card: a Build with your agent button that copies a comfy-cli brief for the open workflow, built from what the graph contains.
Stacked on #17964, which adds the only entry point and shows it to every user.
Changes
utils/buildInputs.ts: reads node classes, node packs (cnr_id/aux_idwith every distinctverthe nodes record,comfy-coreleft out) and model filenames off the serialized workflow, includingproperties.models. The workflow is parsed at the boundary, node by node:idandtypeare required; a malformed optional field is ignored and the node still counts. Only an invalid outer shape or an invalid node contributes nothing. Only subgraph definitions the graph instantiates count, each read once. A string is a model only in a model input of a known loader, per the model-folder registry the model browser and missing-model scan use, and every input of a multi-input loader counts; any other node contributes onlyproperties.models. Reachable definitions come fromcollectReachableSubgraphDefinitionsinworkflowFlattening, which visits each definition once.utils/agentHandoff.ts: one markdown brief per distribution. It has the agent read and followcomfy skills show comfy-build(installing the current comfy-cli only if that is missing), which owns every command, the pins, models, targets, the cut, release watching and failures. The brief adds only what the editor knows and where the user decides: which source to build from, described rather than named by the recipe's path letters (the downloaded workflow file on Cloud; on localhost the install when ComfyUI is on the agent's own machine, otherwise the downloaded workflow file, asking the user when unsure; the newest snapshot on Desktop), the Build name, where the file or snapshot is, the classes, packs and models the workflow records, a yes before an import uploads the workflow or snapshot, one yes before the first cut, after which the agent fixes and re-cuts on its own within the recipe's limits and tells the user what each retry changed, that the user means to deploy later, and a stop at the green release. It contains nocomfy buildcommand. A pack whose nodes record different versions lists them all and asks the user to choose. Every value read off the workflow, its name included, renders as literal code on one line: control characters become spaces, runs of spaces collapse and the ends are trimmed, so a newline or Markdown in a value cannot add a heading or command. No value is dropped for its characters. The downloaded file name additionally drops backticks and path separators.composables/useAgentHandoff.ts: at the click, captures any edit still in a focused field, derives the inputs and brief, copies first while the click's activation still counts, then, everywhere but Desktop, downloads the same graph. The download name comes fromhandoffFileName, the same value the brief shows. A failure is reported withreportErrorand toasted.useCopyToClipboard:{ toastOnSuccess }and a boolean result.CopyHandlerinuseJobErrorReportingwidened to accept it.Review Focus
Comfy.PromptFilenameprompt on purpose: the brief names the file.properties.models; the brief asks the agent to check the definition accounts for everything listed. Asset-backed widget values are still out of scope.comfy … buildcommand or a recipe path letter reappears in the brief.Verification
pnpm vitest run src/composables/useCopyToClipboard.test.ts src/platform/workflow/core/utils/workflowFlattening.test.ts src/platform/workflow/deploy/components/DeployToComfyApiCard.test.ts src/platform/workflow/deploy/composables/useAgentHandoff.test.ts src/platform/workflow/deploy/utils/agentHandoff.test.ts src/platform/workflow/deploy/utils/buildInputs.test.ts src/services/dialogService.errorCatalog.test.ts: the seven test files this PR touches. Test Files 7 passed (7), Tests 116 passed (116) at3a91486.DeployToComfyApiCard.test.tsis shared with the base branch, so a base merge that adds card tests moves the count. Mutation checks: moving the Cloud download outside the error handling fails the download-failure test, and dropping the recipe command fails the recipe-ordering test.pnpm typecheck, eslint on the touched files,pnpm format,pnpm knip: clean.