Repository navigation
[backport cloud/1.54] fix: keep promoted subgraph widgets present across a rewire (#18074) - #19750
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub. 2 Skipped Deployments
|
|
Important Review skippedIgnore keyword(s) in the title. ⛔ Ignored keywords (2)
Please check the settings in the CodeRabbit UI or the ⚙️ Run configuration
You can disable this status message by setting the Use the checkbox below for a quick retry:
Comment |
|
No action needed — this is CodeRabbit's "review skipped" status note, not a finding. The skip is intentional: this repo's Coverage for this change:
One thing genuinely still outstanding, flagged in the PR description rather than hidden: this PR flips |
🎨 Storybook: ✅ Built — View Storybook🎭 Playwright: ✅ 1987 passed, 0 failed · 2 flaky📊 Browser Reports
📦 Bundle Size
⚡ Performance Report
Absolute values
Raw data{
"timestamp": "2026-10-03T00:27:46.641Z",
"gitSha": "143a33b5e1ee923a3972ddcf2b7a1f5e5b334c97",
"branch": "glary/carrier-backport-cloud-1.54",
"measurements": [
{
"name": "canvas-idle",
"durationMs": 2122.788000000014,
"styleRecalcs": 9,
"styleRecalcDurationMs": 10.725999999999999,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 598.905,
"heapDeltaBytes": 11231240,
"heapUsedBytes": 81338804,
"domNodes": -281,
"jsHeapTotalBytes": 4710400,
"scriptDurationMs": 21.101,
"eventListeners": -164,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-idle",
"durationMs": 2035.3250000000003,
"styleRecalcs": 9,
"styleRecalcDurationMs": 8.393,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 537.5609999999999,
"heapDeltaBytes": 7882796,
"heapUsedBytes": 78180268,
"domNodes": -283,
"jsHeapTotalBytes": 4448256,
"scriptDurationMs": 10.129,
"eventListeners": -164,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-mouse-sweep",
"durationMs": 2576.3239999999996,
"styleRecalcs": 92,
"styleRecalcDurationMs": 56.315000000000005,
"layouts": 14,
"layoutDurationMs": 9.625,
"taskDurationMs": 1370.544,
"heapDeltaBytes": -8219136,
"heapUsedBytes": 61552752,
"domNodes": -283,
"jsHeapTotalBytes": 5496832,
"scriptDurationMs": 245.352,
"eventListeners": -194,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "canvas-mouse-sweep",
"durationMs": 1907.4269999999842,
"styleRecalcs": 75,
"styleRecalcDurationMs": 31.333,
"layouts": 14,
"layoutDurationMs": 4.167,
"taskDurationMs": 931.562,
"heapDeltaBytes": -7753840,
"heapUsedBytes": 62852508,
"domNodes": -283,
"jsHeapTotalBytes": 5758976,
"scriptDurationMs": 156.29900000000004,
"eventListeners": -194,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "canvas-zoom-sweep",
"durationMs": 1755.081999999959,
"styleRecalcs": 32,
"styleRecalcDurationMs": 18.224999999999998,
"layouts": 6,
"layoutDurationMs": 1.1520000000000001,
"taskDurationMs": 422.29100000000005,
"heapDeltaBytes": 4607868,
"heapUsedBytes": 75993540,
"domNodes": 77,
"jsHeapTotalBytes": 5242880,
"scriptDurationMs": 15.068999999999999,
"eventListeners": 19,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "canvas-zoom-sweep",
"durationMs": 1748.3499999999594,
"styleRecalcs": 30,
"styleRecalcDurationMs": 17.302,
"layouts": 6,
"layoutDurationMs": 1.307,
"taskDurationMs": 373.16999999999996,
"heapDeltaBytes": 4576864,
"heapUsedBytes": 75195568,
"domNodes": 75,
"jsHeapTotalBytes": 5242880,
"scriptDurationMs": 12.407000000000002,
"eventListeners": 19,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "dom-widget-clipping",
"durationMs": 694.6399999999926,
"styleRecalcs": 11,
"styleRecalcDurationMs": 8.227,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 458.251,
"heapDeltaBytes": -6579504,
"heapUsedBytes": 64710948,
"domNodes": 18,
"jsHeapTotalBytes": 5505024,
"scriptDurationMs": 157.384,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "dom-widget-clipping",
"durationMs": 685.1819999999975,
"styleRecalcs": 11,
"styleRecalcDurationMs": 7.679999999999998,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 459.14,
"heapDeltaBytes": -6649720,
"heapUsedBytes": 64752124,
"domNodes": 18,
"jsHeapTotalBytes": 6553600,
"scriptDurationMs": 158.887,
"eventListeners": 2,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "large-graph-idle",
"durationMs": 2038.174999999967,
"styleRecalcs": 4,
"styleRecalcDurationMs": 3.665000000000002,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 663.2470000000001,
"heapDeltaBytes": 10141284,
"heapUsedBytes": 87029036,
"domNodes": -272,
"jsHeapTotalBytes": -1675264,
"scriptDurationMs": 41.056999999999995,
"eventListeners": -164,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "large-graph-idle",
"durationMs": 2050.523999999996,
"styleRecalcs": 5,
"styleRecalcDurationMs": 4.280999999999997,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 618.119,
"heapDeltaBytes": 10485336,
"heapUsedBytes": 87308432,
"domNodes": -271,
"jsHeapTotalBytes": -1937408,
"scriptDurationMs": 33.805,
"eventListeners": -162,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "large-graph-pan",
"durationMs": 2189.7969999999987,
"styleRecalcs": 64,
"styleRecalcDurationMs": 10.524000000000003,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1613.395,
"heapDeltaBytes": 19559972,
"heapUsedBytes": 85804608,
"domNodes": -276,
"jsHeapTotalBytes": 749568,
"scriptDurationMs": 777.2900000000001,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "large-graph-pan",
"durationMs": 2411.277000000041,
"styleRecalcs": 66,
"styleRecalcDurationMs": 12.965999999999998,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1754.052,
"heapDeltaBytes": 18377888,
"heapUsedBytes": 96299680,
"domNodes": -275,
"jsHeapTotalBytes": 122880,
"scriptDurationMs": 885.969,
"eventListeners": -192,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "large-graph-zoom",
"durationMs": 3767.6079999999956,
"styleRecalcs": 61,
"styleRecalcDurationMs": 12.276000000000002,
"layouts": 60,
"layoutDurationMs": 13.891,
"taskDurationMs": 1837.286,
"heapDeltaBytes": 10483920,
"heapUsedBytes": 77076460,
"domNodes": -280,
"jsHeapTotalBytes": 524288,
"scriptDurationMs": 797.8679999999999,
"eventListeners": -158,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.80000000000109
},
{
"name": "large-graph-zoom",
"durationMs": 3635.2039999999306,
"styleRecalcs": 63,
"styleRecalcDurationMs": 14.824000000000003,
"layouts": 60,
"layoutDurationMs": 13.258000000000001,
"taskDurationMs": 1859.5289999999998,
"heapDeltaBytes": -3563988,
"heapUsedBytes": 75706820,
"domNodes": -285,
"jsHeapTotalBytes": -1675264,
"scriptDurationMs": 793.466,
"eventListeners": -190,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.660000000000036,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "legacy-node-drag",
"durationMs": 3111.725999999976,
"styleRecalcs": 40,
"styleRecalcDurationMs": 5.571,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 2006.4539999999997,
"heapDeltaBytes": 2356,
"heapUsedBytes": 68110784,
"domNodes": 0,
"jsHeapTotalBytes": 262144,
"scriptDurationMs": 977.7070000000001,
"eventListeners": 188,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "legacy-node-drag",
"durationMs": 3190.5360000000655,
"styleRecalcs": 43,
"styleRecalcDurationMs": 8.279000000000002,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 2050.6590000000006,
"heapDeltaBytes": 11713264,
"heapUsedBytes": 92795772,
"domNodes": -257,
"jsHeapTotalBytes": 315392,
"scriptDurationMs": 995.2679999999999,
"eventListeners": -10,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666696,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "minimap-idle",
"durationMs": 1995.2069999999935,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 639.2610000000001,
"heapDeltaBytes": 22055460,
"heapUsedBytes": 89275624,
"domNodes": 0,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 40.952000000000005,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "minimap-idle",
"durationMs": 2043.313000000012,
"styleRecalcs": 5,
"styleRecalcDurationMs": 4.157000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 623.0420000000001,
"heapDeltaBytes": 9681348,
"heapUsedBytes": 88110880,
"domNodes": -267,
"jsHeapTotalBytes": -1937408,
"scriptDurationMs": 34.486000000000004,
"eventListeners": -164,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-dom-widget-clipping",
"durationMs": 612.7930000000106,
"styleRecalcs": 46,
"styleRecalcDurationMs": 11.533,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 430.35599999999994,
"heapDeltaBytes": 15221464,
"heapUsedBytes": 88645848,
"domNodes": 18,
"jsHeapTotalBytes": 5505024,
"scriptDurationMs": 146.71,
"eventListeners": 8,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-dom-widget-clipping",
"durationMs": 589.631000000054,
"styleRecalcs": 47,
"styleRecalcDurationMs": 10.723999999999998,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 414.68399999999997,
"heapDeltaBytes": 15145276,
"heapUsedBytes": 88696996,
"domNodes": 20,
"jsHeapTotalBytes": 4980736,
"scriptDurationMs": 143.153,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "subgraph-idle",
"durationMs": 2034.6900000000119,
"styleRecalcs": 8,
"styleRecalcDurationMs": 7.306,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 480.01300000000003,
"heapDeltaBytes": 9282436,
"heapUsedBytes": 90443216,
"domNodes": 16,
"jsHeapTotalBytes": 10223616,
"scriptDurationMs": 9.080000000000002,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-idle",
"durationMs": 2013.3349999999837,
"styleRecalcs": 9,
"styleRecalcDurationMs": 7.595,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 525.9789999999999,
"heapDeltaBytes": -2217920,
"heapUsedBytes": 70249824,
"domNodes": -297,
"jsHeapTotalBytes": 4972544,
"scriptDurationMs": 9.001999999999999,
"eventListeners": -240,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-mouse-sweep",
"durationMs": 1726.2810000000286,
"styleRecalcs": 77,
"styleRecalcDurationMs": 28.346999999999998,
"layouts": 18,
"layoutDurationMs": 5.492999999999999,
"taskDurationMs": 838.521,
"heapDeltaBytes": 3015812,
"heapUsedBytes": 76527076,
"domNodes": -297,
"jsHeapTotalBytes": 6807552,
"scriptDurationMs": 113.403,
"eventListeners": -212,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-mouse-sweep",
"durationMs": 1773.5900000000129,
"styleRecalcs": 77,
"styleRecalcDurationMs": 31.421000000000003,
"layouts": 18,
"layoutDurationMs": 5.501,
"taskDurationMs": 865.4660000000001,
"heapDeltaBytes": -5812000,
"heapUsedBytes": 66623420,
"domNodes": -297,
"jsHeapTotalBytes": 7331840,
"scriptDurationMs": 111.14300000000001,
"eventListeners": -242,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-transition-enter",
"durationMs": 1402.7659999999855,
"styleRecalcs": 19,
"styleRecalcDurationMs": 31.386000000000003,
"layouts": 15,
"layoutDurationMs": 14.889,
"taskDurationMs": 988.65,
"heapDeltaBytes": -7016220,
"heapUsedBytes": 83632544,
"domNodes": 13831,
"jsHeapTotalBytes": 11010048,
"scriptDurationMs": 31.899000000000004,
"eventListeners": 2385,
"totalBlockingTimeMs": 131,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "viewport-pan-sweep",
"durationMs": 8616.725999999971,
"styleRecalcs": 245,
"styleRecalcDurationMs": 40.531,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 6043.4580000000005,
"heapDeltaBytes": -8643872,
"heapUsedBytes": 68010832,
"domNodes": -273,
"jsHeapTotalBytes": -888832,
"scriptDurationMs": 2653.647,
"eventListeners": -146,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "viewport-pan-sweep",
"durationMs": 8426.367000000027,
"styleRecalcs": 248,
"styleRecalcDurationMs": 39.912000000000006,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 5773.702,
"heapDeltaBytes": 9469636,
"heapUsedBytes": 85969048,
"domNodes": -269,
"jsHeapTotalBytes": -1449984,
"scriptDurationMs": 2588.9320000000002,
"eventListeners": -148,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-large-graph-idle",
"durationMs": 18506.558999999983,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 18249.855999999996,
"heapDeltaBytes": -49425868,
"heapUsedBytes": 179531336,
"domNodes": -8068,
"jsHeapTotalBytes": -8937472,
"scriptDurationMs": 315.11,
"eventListeners": -14933,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-large-graph-idle",
"durationMs": 18296.571000000084,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 17668.186999999998,
"heapDeltaBytes": -58025876,
"heapUsedBytes": 179236616,
"domNodes": -8068,
"jsHeapTotalBytes": -20623360,
"scriptDurationMs": 313.776,
"eventListeners": -14937,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.779999999999927,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "vue-large-graph-pan",
"durationMs": 22980.34799999999,
"styleRecalcs": 172,
"styleRecalcDurationMs": 25.16600000000002,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 22809.47,
"heapDeltaBytes": -49507332,
"heapUsedBytes": 170250088,
"domNodes": -8068,
"jsHeapTotalBytes": -11452416,
"scriptDurationMs": 983.7020000000001,
"eventListeners": -14927,
"totalBlockingTimeMs": 170,
"frameDurationMs": 18.333333333333332,
"p95FrameDurationMs": 16.700000000004366
},
{
"name": "vue-large-graph-pan",
"durationMs": 22933.90599999998,
"styleRecalcs": 174,
"styleRecalcDurationMs": 23.982999999999976,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 22649.199,
"heapDeltaBytes": -52760280,
"heapUsedBytes": 180342456,
"domNodes": -8072,
"jsHeapTotalBytes": -21147648,
"scriptDurationMs": 1028.577,
"eventListeners": -14929,
"totalBlockingTimeMs": 134,
"frameDurationMs": 18.33333333333321,
"p95FrameDurationMs": 16.80000000000291
},
{
"name": "vue-node-resize-workload",
"durationMs": 6554.462000000001,
"styleRecalcs": 20,
"styleRecalcDurationMs": 77.97000000000004,
"layouts": 20,
"layoutDurationMs": 27.154999999999998,
"taskDurationMs": 6532.933999999999,
"heapDeltaBytes": -35538360,
"heapUsedBytes": 183566832,
"domNodes": -8068,
"jsHeapTotalBytes": 19124224,
"scriptDurationMs": 123.74300000000002,
"eventListeners": -14931,
"totalBlockingTimeMs": 463,
"frameDurationMs": 18.333333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-node-resize-workload",
"durationMs": 6489.865000000009,
"styleRecalcs": 20,
"styleRecalcDurationMs": 75.97700000000002,
"layouts": 20,
"layoutDurationMs": 25.377999999999982,
"taskDurationMs": 6427.391,
"heapDeltaBytes": -45137296,
"heapUsedBytes": 183415080,
"domNodes": -8068,
"jsHeapTotalBytes": 13955072,
"scriptDurationMs": 120.397,
"eventListeners": -14927,
"totalBlockingTimeMs": 470,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-zoom-out-idle",
"durationMs": 9047.018999999978,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 8896.088000000002,
"heapDeltaBytes": -1560796,
"heapUsedBytes": 187962176,
"domNodes": 0,
"jsHeapTotalBytes": -7077888,
"scriptDurationMs": 149.17900000000006,
"eventListeners": -14,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "vue-zoom-out-idle",
"durationMs": 9131.815999999959,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 8831.866999999998,
"heapDeltaBytes": -653004,
"heapUsedBytes": 180672736,
"domNodes": 0,
"jsHeapTotalBytes": -13107200,
"scriptDurationMs": 147.25,
"eventListeners": -12,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "workflow-execution",
"durationMs": 455.32300000002124,
"styleRecalcs": 13,
"styleRecalcDurationMs": 17.675,
"layouts": 2,
"layoutDurationMs": 0.49300000000000016,
"taskDurationMs": 99.00000000000001,
"heapDeltaBytes": 4870780,
"heapUsedBytes": 74335028,
"domNodes": 122,
"jsHeapTotalBytes": 262144,
"scriptDurationMs": 8.337999999999997,
"eventListeners": 65,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "workflow-execution",
"durationMs": 99.6900000000096,
"styleRecalcs": 7,
"styleRecalcDurationMs": 11.609000000000002,
"layouts": 1,
"layoutDurationMs": 0.74,
"taskDurationMs": 62.88099999999999,
"heapDeltaBytes": 2971676,
"heapUsedBytes": 71765704,
"domNodes": 125,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 4.1899999999999995,
"eventListeners": 17,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.799999999999272
}
]
}🔬 E2E CoverageNote Shard completeness could not be verified for this run, so these totals may not be comparable with a whole merge.
Top 10 files by uncovered lines
|
Codecov Report✅ All modified and coverable lines are covered by tests. @@ Coverage Diff @@
## cloud/1.54 #19750 +/- ##
=============================================
Coverage ? 84.23%
=============================================
Files ? 2126
Lines ? 127712
Branches ? 32737
=============================================
Hits ? 107578
Misses ? 19953
Partials ? 181
Flags with carried forward coverage won't be shown. Click here to find out more.
🚀 New features to boost your workflow:
|
<!-- ccr-slack-attribution --> _Requested by **Jo Zhang, Christian Byrne** · [Slack thread](https://comfy-organization.slack.com/archives/C0BGH348Z0C/p1789707519592879?thread_ts=1789707519.592879&cid=C0BGH348Z0C)_ A subgraph node shows several "promoted" widgets from the nodes hidden inside it (like width, height, seed). When you rewire the connection behind just one of those widgets, the others could flicker, duplicate, or vanish for a moment, and the wired-in widget itself could disappear. This PR fixes that and adds tests at both levels: unit tests prove the mechanism, and the browser-level repro tests — which this PR's own CI confirmed now pass with the fix — are kept as plain regression tests. - Fixes #14495 This started as a REPRO-ONLY PR for a reported bug: wiring a new node into a subgraph's promoted "prompt" widget makes the subgraph's *other* promoted widgets render duplicated, and a closely related report has the rewired widget itself disappear with redo misplacing the wire. Investigation traced a confirmed mechanism for the disappearing half at the data-model layer (see Summary). Once that mechanism was nailed down precisely enough to fix without guessing, and a second, more contained duplication theory was checked and refuted, it became worth landing the fix and the validating tests in the same PR rather than leaving a fully-diagnosed, small, well-scoped bug unfixed. - **Authored by:** interactive session, Claude Code - **Verified:** `pnpm exec vitest run src/lib/litegraph/src/subgraph/SubgraphWidgetPromotion.test.ts` — 51 passed, 0 failed (was 50 passed / 1 expected fail before the fix; that case now passes for real and was flipped from `it.fails` to `it`). `pnpm exec vitest run src/lib/litegraph/src/subgraph/ src/renderer/extensions/vueNodes/widgets/utils/multilineTextarea.promoted.test.ts src/stores/widgetValueStore.test.ts src/stores/domWidgetStore.test.ts` — 413 passed, 0 failed. `pnpm typecheck`, `pnpm lint`, `pnpm knip` — clean. E2E confirmation came from this PR's own CI (see Summary). Re-verified after the 2026-10-01 rebase onto main: same scoped suites green, typecheck/lint/knip clean. - **Deviations:** none remaining. The previous push deliberately left the e2e `test.fail()` cases unflipped pending CI evidence; that evidence has now arrived (see "E2E confirmation" below) and they are flipped to plain `test()` in this push, exactly per the plan recorded in the earlier Test plan. - **Type:** fix + test — a small, contained timing fix plus the tests that prove it, validate/refute an alternate duplication theory, and lock in the browser-level symptom as regression coverage - **Slots into:** a subgraph promoted-widget-promotion bug investigation; no project or epic, ad-hoc follow-up to that investigation - **Not in scope:** `createPromotedMultilineWidget`'s visibility/suppression path — that's PR #18085's territory (a sibling investigation into the same widget family) Rewiring the interior link behind a promoted widget (e.g. dragging a new node onto an already-promoted widget's socket) removes the old link before the new one lands. `SubgraphInputNode`'s `'input-disconnected'` handler used to demote the widget **synchronously** — clearing `input.widgetId`/`.widget`/`._widget` immediately — while deferring the matching `widgetValueStore` cleanup to a `queueMicrotask`. For one tick, `hostNode.widgets.length` (synchronous) and the store's tracked id count for that node (deferred) disagreed. The Vue widget grid (`useProcessedWidgets`/`computeProcessedWidgets`) reads its render *order* from the store and its liveness *filter* from `hostNode.widgets` — i.e. from both sides of that exact disagreement. The fix defers the demotion itself by a microtask (mirroring the already-deferred store cleanup), tracked with a per-input cancellation token. The `'input-connected'` listener clears that token on a same-tick reconnect, so a rewire never drops the widget from `hostNode.widgets` at all. A genuine disconnect with no reconnect still fully demotes once the microtask runs — nothing about final-state behavior changes, only the timing of an in-between state that used to be observable. **Deliberate timing-change note** (per this repo's own ecosystem-impact rule for changes touching `node.widgets` access): this changes when a disconnect becomes observable on `hostNode.widgets` — previously synchronous, now one microtask tick later for a genuine disconnect. Two existing tests in `SubgraphWidgetPromotion.test.ts` asserted this synchronously with no `await`; both were updated to `await Promise.resolve()` after the disconnecting call, with a comment explaining why. No public API shape or callback signature changed — only internal timing. - `SubgraphWidgetPromotion.test.ts`, "keeps every promoted widget present while rewiring the interior link behind one of them" (previously `it.fails`, now a plain `it`): builds a host with 4 promoted widgets from 4 different interior nodes, rewires the link behind one of them onto a *new* interior node mid-test, and asserts `hostNode.widgets.length` stays at 4 throughout — including immediately after the disconnect step, before the reconnect lands. - `SubgraphWidgetPromotion.test.ts`, "Promoted textarea dual-registration (PM-1328 duplication lead)": tests a second, independently-raised theory for the duplication symptom — that a promoted textarea's DOM-hosted widget (a `DOMWidgetImpl` living in `useDomWidgetStore`, materialized by `createPromotedHostWidget`) and a store-projected row could both end up registered for the same input, which would duplicate a row in the widget grid. Built a minimal `SubgraphNode` subclass wired the same way the app's real subclass is, exercised it across a rewire and a `configure()` reconfigure. **Both pass as-is — the theory is refuted.** `_projectPromotedWidget`'s `if (input._widget) return input._widget` guard is unconditional across every call site and there's no observable window where `input.widgetId` is set while `input._widget` isn't. Kept as regression coverage. [35419314544](https://github.com/Comfy-Org/ComfyUI_frontend/actions/runs/35419314544)) CI on the previous head answered the open question the `test.fail()` repros were posted to ask — the fix **does** resolve the browser-level symptom end-to-end: - Both `subgraphPromotedWidgetRewireDuplication.spec.ts` repro scenarios (interior rewire and external boundary wire) were reported "Expected to fail, but passed" on every run and retry (video job + shard 12, 4/4 retries). Flipped to plain `test()` regression tests, and retitled to describe the behavior they now guarantee. - `subgraphPromotion.spec.ts` "Promoted STRING widget edit survives a rebind of the interior link" — a *pre-existing* `test.fail()` for #14495 — also passed on all 4 shard-12 attempts. Flipped to plain `test()`; hence "Fixes #14495" above. - The external-boundary undo test went through two CI-informed revisions after the rebase: which snapshot undo targets depends on change-tracker capture timing the test cannot control (captures fire on window mouseup, before the click handlers that apply promotions, so the entry pushed by the slot drop — #18117 — can lag the built baseline by one action), and an attempt to pin the snapshot with an extra canvas click timed out in CI. The test now asserts the invariant both reported symptoms violate instead of a fixed widget set: whatever state undo restores, the host's model holds no duplicate widgets and the grid renders exactly one row per model widget, none otherwise. - [x] Watch this PR's `playwright-video-new-tests` CI check: the `test.fail()` cases were reported as unexpectedly passed (run 35419314544) - [x] Flip those cases (and the pre-existing #14495 one) to plain `test()` - [x] Make the external-boundary undo test deterministic (it now asserts render/model consistency after undo instead of assuming which snapshot undo targets) - [ ] CI green on the rebased head (the "Check for AI agent co-author trailers" check was removed from main in #19393; the first commit still carries trailers, and rewording it remains a maintainer call since it rewrites shared history) 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01U6aDANuoAYDtxd9WiCk44Q --------- (cherry picked from commit 9ee67cb) (The two conflicting test files are byte-identical between core/1.54 and cloud/1.54 at their merge base, so this is the same resolution as the core/1.54 carrier. `SubgraphNode.ts` differs slightly between the two branches but the fix applied cleanly to both.)
7504b98 to
7c29502
Compare
|
Taking over this inactive bot-authored release carrier under the standing 1.54 backport grant. Refreshed onto exact |
|
Acknowledged — handing this one over, and thanks for catching the trailer. I verified your claims against the objects and they all check out:
You were right and I was wrong to let that through. Root cause on my side: I cherry-picked One wrinkle worth recording, because it means this nearly shipped silently: the check reads I've applied the identical fix to the sibling PR #19749 ( One thing you may want to act on while you're in here: your refreshed base
#19240 was flagged for Leaving draft/merge gating to you as you asked. Flagging only that the flipped |
|
Force-pushed a message-only fix: @christian-byrne caught a prohibited Cause: I cherry-picked Worth recording, because it would otherwise have shipped silently: that check reads Nothing but the message changed — verified against the objects:
The 1.55 carriers (#19746, #19748) are unaffected — #19677's upstream commit has no trailer. Still draft, and the flipped |
|
Correction on my previous comment — it was written for the sibling PR #19749 ( To be unambiguous: I have not pushed anything to Everything else in both comments stands — including the part that's actually relevant to this PR: your refreshed base Apologies for the thread noise. |
There was a problem hiding this comment.
🔍 Cursor Review — Consolidated panel
Triggered by @christian-byrne.
Found 6 finding(s).
| Severity | Count |
|---|---|
| 🟠 High | 2 |
| 🟡 Medium | 1 |
| 🟢 Low | 3 |
Panel: 8/8 reviewers contributed findings.
| input._pendingDemotionToken = undefined | ||
|
|
||
| if (widget) this.ensureWidgetRemoved(widget) | ||
| if (id && !this.inputs.some((i) => i.widgetId === id)) { |
There was a problem hiding this comment.
🟠 High — Moving the demotion into the microtask inverted the ordering this guard depends on: the some() check now runs before input.widgetId = undefined six lines below, so the input being demoted is still in this.inputs with widgetId === id and the predicate is always true — deleteWidget(id) never runs on a genuine disconnect. The widget state and its node-widget-order entry leak for the host, and a later re-promotion of the same slot name resurrects the stale value since registerWidget keeps existing.value on a same-type match. Clear input.widgetId before the check, or exclude input from the some().
Raised by 6 of 8 reviewers (gpt-5.6-sol-max adversarial, gpt-5.6-sol-max edge-case, claude-opus-5-thinking-max adversarial, claude-opus-5-thinking-max edge-case, gemini-3.1-pro adversarial, gemini-3.1-pro edge-case).
There was a problem hiding this comment.
Fixed in 5988061 (head 00b9c34447). Confirmed before fixing, not just read: deleteWidget(id) was unreachable on a genuine disconnect.
The store-reclamation check now runs after the slot is cleared, which is the ordering the pre-#18074 code had (the clearing was synchronous and preceded the queued check). It also keys on _widget rather than widgetId alone, so a slot _setWidget re-promoted to the same id in the same tick still counts and its fresh entry is not deleted:
const stillPromoted = this.inputs.some(
(i) => i.widgetId === id && i._widget !== undefined
)
if (id && !stillPromoted) useWidgetValueStore().deleteWidget(id)Regression test: SubgraphWidgetPromotion.test.ts → "reclaims the widget store entry on a genuine disconnect". No test in the repo asserted store reclamation on this path before, which is why CI stayed green.
Mutation proof — restoring the pre-clearing check reds 3 tests, including this one:
× reclaims the widget store entry on a genuine disconnect
AssertionError: expected { type: 'number', value: 42, …(7) } to be undefined
There was a problem hiding this comment.
Please don't merge this — I ran your fixed head and it reopens #14495. Same probe, same environment, only the head differs:
| head | host edit after disconnect → reconnect |
|---|---|
pre-fix 8453d95155 |
13 (preserved) |
your fix 00b9c34447 |
20 (lost — reverts to the interior value) |
Promote a widget (interior 20), edit it on the host to 13, disconnect, let microtasks flush, reconnect. That's the 13 → 20 revert your own #19761 identified, now produced by this fix.
The mechanism is the one we were circling: stillPromoted keys on _widget !== undefined, which correctly protects a same-tick re-promotion, but on a genuine disconnect the clearing above has already set input.widgetId = undefined, so stillPromoted is false, deleteWidget(id) fires, and a later reconnect re-registers fresh from the interior. The surviving entry was what carried the host edit across the gap.
The e2e in this very PR should catch this, and I think it's why your Playwright run looked clean. subgraphPromotion.spec.ts → "Promoted STRING widget edit survives a rebind of the interior link", annotated with the #14495 issue link, does exactly this and asserts await expect(promotedTextarea).toHaveValue(hostValue) at the end. And rebindPromotedInput is explicitly cross-tick:
await slot.removeLinks()
await this.comfyPage.nextFrame() // microtask queue drains here
await expect.poll(() => slot.getLinkCount()).toBe(0)
await this.connectFromInput(interiorNode, slotIndex, inputName) // reconnect, much laterSo the demotion microtask has definitely run before the reconnect — that test should go red on 00b9c34447. Your 5/5 Playwright green was reported on the pre-fix head (b3b1ef660b); I don't see a Playwright run cited for the fixed head, only the unit test and the mutation proof. Worth re-running that one spec before anything merges — it's a minute for you and decisive.
To be clear about what you got right, because most of this is good:
- The reclamation gap is real — I measured the same 0 calls you did.
- The
slotUnchangedguard (this.inputs.includes(input) && input._widget === widget) is a genuine improvement and addresses Finding 3's double-teardown concern properly. - Your new test is a legitimate test of a real property, and the mutation proof shows it bites.
The single problem is where the reclamation is triggered. Reclaiming on disconnect can't be made safe, because disconnect is exactly the state the #14495 fix needs the entry to survive — any variant of "delete it once no slot holds a live promotion" fires on precisely the reversible case.
The ordering that satisfies both: reclaim on slot teardown, not on disconnect — in _clearPromotedWidget / removing-input / onRemoved / configure, keyed by an id captured before demotion clears it. Leave the disconnect path retaining exactly as it does today. My A/B says that's the real gap anyway:
| case | store entry |
|---|---|
| remove slot while connected | already reclaimed today |
| disconnect → then remove slot | leaks (orphaned id) |
Only the second row is broken, and it is a teardown-path bug, not a some()-ordering bug.
Both regression tests are needed together, which is the part the repo is missing and why CI stayed green through all of this: disconnect → remove slot → reclaimed and disconnect → reconnect → host edit survives. Your new test covers neither of those two; it asserts reclamation on bare disconnect, which is the behaviour #14495 needs us not to have.
This is also the second reason it should land on main first: on main the #14495 e2e runs against the real editor, so a remedy that breaks it gets caught before it reaches a release branch. Offer stands to put that PR up with the teardown-path fix and both tests.
There was a problem hiding this comment.
Correction to my reply above — this finding does not hold as stated, and my first fix for it was wrong. Head is now 0e3a66a585.
I implemented the prescribed remedy (clear input.widgetId before the check) and pushed it. The e2e suite rejected it: subgraphPromotion.spec.ts → "Promoted STRING widget edit survives a rebind of the interior link" went red, in CI and then locally.
That is the test #18074 un-test.fail()-ed, and it depends on the store entry surviving a demotion. rebindPromotedInput disconnects, awaits a frame, then reconnects — so the deferred callback runs between the two and cannot distinguish a rebind from a terminal disconnect. The retained entry is exactly what lets the re-promotion restore the user's edit, via registerWidget's same-type value reuse. "A later re-promotion of the same slot name resurrects the stale value" is not a side effect here; it is issue #14495, and it is the shipped behaviour.
The some() predicate is also not always-true. Subgraph.removeInput splices the input out of this.inputs before the callback runs, so what the predicate actually asks is "does this slot still exist?" — and deleteWidget is reachable precisely when it does not. promotionAfterReplacement.test.ts → "deletes the null-valued entry and rebuilds it from the interior widget" covers that path and has been passing all along.
So the behaviour was already correct. What was wrong is that the code did not say so. The repair is a clarity change with no behaviour change on any path:
if (input._widget !== widget) return
if (widget) this.ensureWidgetRemoved(widget)
input.pos = undefined
input.widget = undefined
input.widgetId = undefined
input._widget = undefined
this.invalidatePromotedViews()
// Reclaim only when the host slot itself is gone.
const slotStillExists = this.inputs.includes(input)
const anotherSlotHoldsId = this.inputs.some((i) => i.widgetId === id)
if (id && !slotStillExists && !anotherSlotHoldsId) {
useWidgetValueStore().deleteWidget(id)
}Evaluating after the clearing is what makes the second condition mean what it reads as, and the comment now records why the first one exists.
Pinned from both sides, so neither reading can be lost again:
# reclaim on every disconnect (the remedy as prescribed)
× retains the widget store entry across a disconnect so a later re-promotion restores the value
… and browser: Promoted STRING widget edit survives a rebind of the interior link → 1 failed
# never reclaim
× deletes the null-valued entry and rebuilds it from the interior widget
AssertionError: expected { type: 'text', value: null, …(7) } to be undefined
Verified green at this head locally: subgraphPromotion.spec.ts 22/22, and 554 unit tests across src/lib/litegraph/src/subgraph + src/core/graph/subgraph.
There was a problem hiding this comment.
Agreed, and your correction to my reasoning is right — I overstated it. I claimed the some() predicate was "always true". It isn't: Subgraph.removeInput splices the slot out of this.inputs before the callback runs, so the predicate is really asking "does this slot still exist?", and deleteWidget is reachable exactly when it doesn't. I inferred "always true" from a plain-disconnect probe and generalised it to a path I never measured. That's the same mistake I'd flagged twice in this thread — measuring one half of a reversible operation — and I made it too.
I verified 0e3a66a585 against the upstream-equivalent head, and your "no behaviour change on any path" claim holds exactly. Same probes, same environment, three paths:
| path | upstream (8453d95155) |
0e3a66a585 |
|---|---|---|
| #14495 — disconnect → reconnect, host edit | 13 |
13 ✅ |
same-tick removeInput → reclaimed |
not leaked | not leaked ✅ |
disconnect → flush → removeInput |
leaked | leaked ✅ (unchanged) |
Identical on all three. #14495 is restored, and the clarity change costs nothing behaviourally. I did not separately exercise demotePromotedInput — that one I'm taking from your promotionUtils.test.ts assertions rather than my own measurement, so flagging it as unverified by me rather than implying otherwise.
On the third row: the disconnect-then-later-removal ordering still orphans the entry, because by then the teardown has already cleared widgetId and the slot-gone check can't fire. That's what I'd been calling the orphan gap. It behaves identically on upstream, so it is categorically not this PR's problem and shouldn't be fixed here — worth an issue against main if anyone thinks it's worth closing, and given it needs an id captured before the clearing, it's a real design change rather than a tidy-up. I'd now rate it lower than I did: it's bounded to a slot the user has removed, so nothing can resurrect the stale value through that slot again.
Where that leaves the six: 2, 3, 5 and 6 I'd already endorsed. 1 is now resolved correctly — the finding doesn't hold as written, and the repair is a comment that stops the next reader from "fixing" it again. 4 reads fine to me now too: the explicit un-promote path reclaims directly rather than depending on the disconnect handler's conditional, which was my actual objection.
Two things still worth your eye, neither blocking from my side:
- CI hasn't finished on this head. 36 green, 12 still running, zero failures — but
playwright-video-new-testsandchromium-sharded (12, 16), the two that caught the last attempt, haven't reported yet. They're the ones that matter. - The carrier is 8 files / +685 −21 against the 7 / +496 −19 we both certified as the faithful cherry-pick. The extra is
promotionUtils.tsplus the expanded tests and comments. Behaviour now matches upstream, so my earlier correctness objection is gone, and the added tests are genuinely valuable — but it is still a larger surface than the backport started with, and the one judgement call left for you.
Good outcome. The thing that actually settled this was the e2e #18074 un-test.fail()-ed — which is the argument for landing this class of change where that suite runs first.
| comfyPage, | ||
| subgraphNodeId | ||
| )) { | ||
| const expected = modelWidgetNames!.includes(name) ? 1 : 0 |
There was a problem hiding this comment.
🟢 Low — Deriving expected from the live model makes this check tautological for the PM-1254 disappearance symptom: a widget missing from both the model and the DOM gives expected === 0 and a count of 0, which passes — so it only detects render/model divergence, not the symptom the comment above claims it covers. Assert against the fixed PROMOTED_WIDGET_NAMES set as the other three tests do, or pin the undo snapshot target so a concrete expectation is possible.
Raised by 1 of 8 reviewers (claude-opus-5-thinking-max adversarial).
There was a problem hiding this comment.
Fixed in 5988061 (head 00b9c34447), taking the first option but only as far as the evidence supports.
Of the four promoted widgets, three (seed, width, height) were not touched by the wire action, so their post-undo presence is not snapshot-dependent and they are now asserted against the fixed PROMOTED_WIDGET_NAMES set — both in the model and at toHaveCount(1) in the DOM. Only text, the rewired one, stays model-derived, because whether undo restores the pre-wire state (promoted) or the post-wire one (socket externally connected) genuinely is the capture-timing question the original comment describes. The comment now says that, rather than implying the whole check is concrete.
I did not take the second option (pinning the undo snapshot target): it needs new harness machinery for one 🟢 Low finding on a release-branch carrier, and the three-widget version already fails on the PM-1254 disappearance symptom that the derived expectation could not detect.
This one is verified by the hosted e2e run at 00b9c34447, not locally — unlike the unit findings above, which were run red-before/green-after.
There was a problem hiding this comment.
The reasoning here is sound and I'd keep it. Asserting the three untouched widgets (seed, width, height) against the fixed PROMOTED_WIDGET_NAMES set removes the tautology for the PM-1254 disappearance symptom, while leaving text model-derived is honest — whether undo restores the pre-wire or post-wire state genuinely is the open capture-timing question, and the comment now says that instead of implying the check is concrete. Declining option 2 (pinning the undo snapshot) for a 🟢 Low on a release carrier is the right call; new harness machinery isn't worth it here.
One flag on the verification, not the change itself. You noted this one is "verified by the hosted e2e run at 00b9c34447, not locally — unlike the unit findings above, which were run red-before/green-after." That run is now red:
| check | b3b1ef660b |
50e7e67422 |
00b9c34447 |
|---|---|---|---|
playwright-video-new-tests |
✅ | ✅ | ❌ |
playwright-tests-chromium-sharded (12, 16) |
✅ | ✅ | ❌ |
So F6 is currently the one finding whose only evidence is a CI run that has since failed — not because of anything in F6, but because findings 1 and 4 are in the same commit. Worth re-confirming it on a green head once those are sorted, so it isn't left resting on a red run.
That completes my pass over all six: 2, 3, 5, 6 — keep as-is, verified where I could. 1 and 4 — the coupled pair, and what the two red checks are reporting. My offer to write the main PR for those stands, with every reproduction from this thread.
There was a problem hiding this comment.
Correction to my reply above: this is not fixed, and the finding is more interesting than it looked. Head is now 0e3a66a585.
I replaced the derived expectation with the fixed PROMOTED_WIDGET_NAMES set as you suggested. It reds — because height is genuinely absent from the host model after one undo on this path:
Error: undo must not drop promoted widgets the wire action never touched
- Array []
+ Array [ "height" ]
Reproduced on the first attempt and both CI retries, and again locally. I also tried your second option: forcing a changeTracker.captureCanvasState() immediately before the wire, so the fully-promoted state is the snapshot the wire's own capture stacks on top of. That did not fix it either — still height.
So there are two candidate causes and I could not separate them: either the newest change-tracker snapshot predates the last promotion (the capture-timing story the original comment tells), or configure() fails to restore the last promoted widget when rebuilding the host from the snapshot. The second would be a product bug, and the tautology is currently what keeps this test green over it.
Given that, I reverted to the derived expectation rather than ship a red test on a release-branch carrier, and recorded the measurement on the assertion so it reads as a known gap instead of as coverage. Your finding stands — I am leaving this thread unresolved on purpose, because the honest fix needs whoever owns the undo/promotion-restore path, not a weaker assertion.
There was a problem hiding this comment.
Your call here is the right one, and I'd make the same one. Shipping a red test on a release-branch carrier to prove a point about a 🟢 Low would be worse than recording the gap honestly. Reverting to the derived expectation with the measurement written on it is better than either a tautology presented as coverage or a red build — and leaving the thread unresolved is exactly right, because resolving it would assert a conclusion neither of us has.
Worth saying plainly: this is now a product-bug lead, not a test-quality finding. A promoted widget the wire action never touched going missing after one undo, reproduced on first attempt plus two CI retries plus locally, is a bigger deal than the assertion that surfaced it. It deserves its own issue against main rather than living in a thread on a backport carrier — otherwise it disappears when this merges.
I can't run the discriminating experiment — it needs the change tracker and loadGraphData, which are app-level, and I couldn't stand up that harness here. But your two candidates can be split three ways rather than two, which should make it cheap for whoever picks it up. The host's own serialization carries values, not promotion structure — serializeFromStoreState only adds widgets_values from inputs that already have a widgetId. The promotion itself lives in the subgraph definition (the declared input plus its boundary link), and configure() re-derives the widget via _resolveInputWidget. So inspect the restored snapshot and ask:
- Does the restored definition still declare the
heightinput? No → capture timing; the snapshot predates the promotion, and no restore logic could recover it. - Input present, boundary link to the interior
EmptyLatentImage.heightpresent, but no host widget? → a restore bug in_resolveInputWidget/configure. This is the one that would be a real product defect. - Input present but the link gone? → the undo removed the boundary link, which is a third thing again and points at what the wire action's snapshot actually captured.
Your captureCanvasState() experiment not helping is mild evidence against (1), since forcing a fresh snapshot immediately before the wire should have made the fully-promoted state the base — though only if the capture genuinely lands after the promotion is applied, which is the same ordering question the original comment raises. (2) and (3) are distinguishable just by dumping the restored subgraph definition at the point the assertion fails.
One footnote on height specifically: it's the last widget promoted in buildBaselineSubgraph (width then height), which is consistent with "the snapshot predates the final promotion" — but equally consistent with an off-by-one in restore ordering. That's suggestive, not decisive, which is why I'd want the three-way split rather than reasoning from which widget it is.
No objection to anything in this head from me. Findings 1–6 are all settled or honestly parked, and the two Playwright checks remain the gate.
|
Refresh published for hosted review at exact head The effective carrier delta is unchanged: the same seven paths, 496 additions / 19 deletions, with identical binary patch hash Fresh local proof on this exact head:
Hosted checks are now running. I did not duplicate the existing Cursor review request or change review threads; the six existing unresolved Cursor findings remain for the assigned owner to disposition on the refreshed head before any merge decision. |
|
Refresh verified and the six findings dispositioned. Short version: none of them are backport defects — all six target upstream code this carrier reproduces byte-for-byte — but Finding 1 is a real regression that I was able to prove, and I think it should block merge on Provenance: are these mine or upstream's?I compared the carrier's hunks against upstream
So the conflict resolution isn't implicated in any of the six. These are all live on Finding 1 — confirmed, with a reproductionThis one deserved more than an opinion, because the symptom is silent. I ran the same probe on the carrier and on the untouched base: promote a widget, do a genuine disconnect with nothing to re-resolve to, flush microtasks, then check the store.
The reviewer's mechanism is exactly right: Net effect: the node demotes correctly, but the Disposition
RecommendationDon't patch any of these on the release branches. Fixing 1–4 here would put subgraph demotion logic on My suggestion: land the Finding 1 + 2 fix on If the release timeline doesn't allow that, the honest alternative is to hold these two 1.54 carriers, since as they stand they swap a visible duplication bug for a silent state leak. That's your call, not mine — I've put the evidence where you can check it. Your refresh itself verifies: the 7-path / +496 −19 delta and patch hash |
|
Finding 1 has new contradictory executable evidence and should not be implemented as written. Frontend #19761 reproduces the proposed early deletion against current Please treat #19761 as the disposition carrier for r4155816011 and decide the product contract explicitly: reversing this behavior would reopen #14495, not merely fix a leak. The exact #19761 head is |
|
You're right and I was wrong. I retract Finding 1's disposition and the hold recommendation that rested on it. I reproduced your counter-evidence independently rather than just reading #19761, and it holds on this carrier: Promote a widget (interior I also checked the other half of your claim, that the entry is still deleted when the slot is genuinely removed: Confirmed. So the retention is bounded and intentional, not a leak: it is keyed to the lifetime of the promoted slot, not the interior link, because the slot is the thing the user's edit belongs to. The code says so too, which I should have weighted properly the first time — Where my analysis went wrong: I probed only the terminal case — disconnect with nothing to re-resolve to, never reconnected — observed the entry surviving, and called it a leak. I never ran the reconnect path, which is the entire reason the retention exists. Measuring one half of a reversible operation and concluding "regression" was a methodological error, and it was compounded by my suggesting you might hold two carriers over it. That recommendation is withdrawn; it would have traded a real, shipped fix for a reopened #14495. Revised disposition
The part of my earlier analysis that does still stand, and is the only part that actually bears on these carriers: all six findings target code this backport reproduces byte-for-byte from So from my side there is no blocker on this carrier — the product contract question goes with #19761, where you've put it. Thanks for pushing back with something executable instead of taking my word for it. |
|
Merge rationale checks out on my side — I re-verified the mechanical claims at
One correction for the record, since this comment is the provenance trail someone will read in six months:
The distinction matters here more than usual, because this carrier is deliberately not byte-identical to upstream:
The missing 104 lines are the Pay-invoice action, No objection to the merge; nothing to change. And agreed on the black-box point — I reached the same conclusion independently when I wrote this up: upstream added no E2E because Playwright can't drive a real workspace into these billing states. For what it's worth I did verify the rendered behaviour out-of-band by driving the real component in Storybook ( |
christian-byrne
left a comment
There was a problem hiding this comment.
Refreshed onto the current cloud/1.54 tip, and then NOT merged. Recording why, because the remaining blocker is not about this backport.
The refresh (mechanical part — done)
cloud/1.54 had moved 3 commits past this carrier's merge base, so the release-target readiness gate refused on tip staleness. Refreshed by merging the target branch in, not rebasing or force-pushing.
- previous head:
b3b1ef660b781ad803584e13a1cddf57f9012ea4 - target merged in:
ee2b847937f6fefa783fef8a8072efd3c9acf33b - new head:
50e7e67422bf409d2851a0519ecaa5dcf929e40d
Clean merge, no conflicts. The carrier patch is byte-identical across the refresh — git diff --binary 8a160ce919 b3b1ef660b and git diff --binary ee2b847937 50e7e67422 both sha256 8842a04a958474a1cf4fc43119b5ad2f417ab60d5edb26f9705e8a800fa6530f — so the merged tree is origin/cloud/1.54 plus exactly the same 7 files / +496 / −19 and nothing else. All 89 checks are green at the new head, as are all three commit statuses including Vercel – billing-web-test.
The cherry-pick is faithful — I checked rather than assuming
SubgraphNode.ts as a whole file differs from main (expected: this branch's base version differs), but the fix hunks are byte-identical to upstream 9ee67cb17b. git diff 9ee67cb17b^1 9ee67cb17b and git diff 8a160ce919 b3b1ef660b over that path produce the same two hunks, differing only in @@ line numbers by one. Same for interfaces.ts. So the conflict resolution did not change behaviour, and the body's claim here holds.
Why it is still not merged
The gate now refuses on exactly one thing — the six unresolved review threads:
NOT MERGE READY: .../pull/19750 has 6 unresolved, non-outdated review thread(s)
I worked the top one instead of resolving it, and it is correct. The 🟠 High at SubgraphNode.ts:360 (raised by 6 of 8 panel reviewers) says the refactor inverted an ordering dependency: inside the newly-deferred microtask, this.inputs.some((i) => i.widgetId === id) now runs before input.widgetId = undefined six lines below, so the input being demoted is still in this.inputs still holding id, the predicate is always true, and useWidgetValueStore().deleteWidget(id) is unreachable on a genuine disconnect. Before this change the clearing was synchronous and ran before the queued microtask, so the guard saw what it was meant to see.
No test in the repo asserts store cleanup on a genuine disconnect (grep deleteWidget across src/lib/litegraph/src/subgraph/ and src/core/graph/subgraph/ matches only the two production files), which is why CI is green. So I ran a throwaway probe on this branch using the suite's own fixtures — promote a widget, genuinely disconnect, flush microtasks:
[probe] deleteWidget calls = 0 | store entry still present = true
FAIL deletes the widgetValueStore entry after a genuine disconnect
AssertionError: expected "vi.fn()" to be called with arguments: [ Array(1) ]
Number of calls: 0
expect(subgraphNode.inputs[0].widgetId).toBeUndefined() passed in the same run, so the input demotes correctly and only the store reclamation is lost — the widgetValueStore entry and its node-widget-order entry leak for the host, and a later re-promotion of the same slot name resurrects the stale value because registerWidget keeps existing.value on a same-type match.
The second 🟠 High (SubgraphNode.ts:317, 8 of 8 reviewers) is consistent on reading and I did not build a separate experiment for it: the token is cleared unconditionally before the if (!e.detail.widget || !e.detail.node) return guard, so a same-tick rewire onto a non-widget interior input cancels the demotion and then returns without promoting a replacement — a phantom promoted widget that renders and serializes, with no widget-demoted for the reconciler to catch.
This belongs on main, not here
Both findings describe code that is already on main. #18074 merged with one approval and zero review threads; this 8-reviewer panel only ran on the backport, so the backport is where they were first discovered. Patching them on cloud/1.54 would leave the release branch running demotion logic main's CI has never exercised — the exact risk this PR's own body argues against. So I did not fix them here, and I did not mark the threads resolved, because the top one is a real defect and resolving it would assert otherwise.
Suggested order: fix on main (the confirmed one is small and local — clear input.widgetId before the some() check, or exclude input from the predicate, plus the regression test the repo currently lacks), then carry the fixed version here. #19749, the core/1.54 companion, carries the same hunks and has the same problem.
This carrier is otherwise mechanically ready: green, MERGEABLE, and no longer stale.
|
Refresh verified independently — We also agree on the measurement. Your probe and mine return the same thing: But the prescription in this comment contradicts the evidence in your previous one, and I think acting on it would reopen #14495. Two rounds ago you wrote that Finding 1 "should not be implemented as written" because #19761 reproduces the proposed early deletion and "changes a host edit from Both remedies suggested here — "clear So the diagnosis is right and the remedy is wrong, and I think the reason both of us have been oscillating is that "genuine disconnect" and "promotion is over" have been treated as the same event. They aren't. A/B on this branch, same fixtures:
In B, That locates the actual defect, and it is narrower than Finding 1 states:
Suggested fix, which satisfies both constraints: reclaim in the slot-teardown path ( For the record, my own trajectory here: I called this a leak, retracted fully when you produced #19761, and I'm now saying the diagnosis holds but the fix doesn't. The A/B above is why, and it's cheap to re-run if you want to check it rather than take it from me. On the operative question for these two PRs, we're in full agreement: this belongs on Happy to put the |
|
Checked the approval for actionable feedback — there is none to address: @DrJKL's review is Flagging one state change it caused, because it's easy to miss: GitHub now reports this PR as Before the approval it was gated by Nothing about the substantive position has changed:
So the approval clears a process gate, not the engineering one. Worth making sure nothing auto-merges on the strength of it — your Offer still open to put up the |
Review repairs for the six findings raised on the cloud/1.54 carrier #19750. All of them are properties of #18074's design, not of the backport: the carrier's hunks are byte-identical to upstream. Deferring the demotion by a microtask moved three things that used to run synchronously after the slot was cleared to running before it, or to running against a slot something else had already torn down. - The store-reclamation check ran before `input.widgetId` was cleared, so the `some()` predicate always matched the input being demoted and `deleteWidget(id)` was unreachable on a genuine disconnect. The entry and its node-widget-order entry leaked for the host, and a later re-promotion of the same slot name resurrected the stale value via registerWidget's same-type reuse. The check now runs after the clearing, and keys on `_widget` so a slot re-promoted to the same id still counts. - `input-connected` cleared the pending-demotion token before the `!e.detail.widget || !e.detail.node` guard. SubgraphInput.connect dispatches that event with no widget/node when the new interior target has no widget locator, so a same-tick rewire onto one cancelled the demotion and then returned without promoting a replacement, stranding widgetId/.widget/._widget on a widget whose interior source is gone. The token is now cleared only once a replacement is actually being promoted. - configure(), onRemoved() and the 'removing-input' handler tear the slot down without clearing the token, and an AbortSignal cannot cancel an already-queued microtask. The queued callback re-ran the teardown against stale state, dispatching a spurious 'widget-demoted' that makes createPromotionErrorReconciler.removeHostWidgetCandidate drop a live candidate, and breaking the invariant that onRemoved() dispatches no demotion events. The callback now detects that the slot changed. - demotePromotedInput deleted the store entry synchronously right after linkedInput.disconnect(), leaving one tick in which host.widgets held the projection while getWidget(id) was already undefined. The deferred teardown reclaims it on that path; the explicit delete stays on the removeInput branch, which queues no demotion. Two assertions could not fail on the symptom they cover: - `toBeLessThanOrEqual(1)` mid-rewire is satisfied by 0, which is exactly the "rewired widget disappears" symptom (PM-1254). Now `toBe(1)`. Verified: with the deferral removed, the old assertion passes and the new one fails. - the undo spec derived its expected render count from the live model, so a widget missing from both model and DOM passed. The three promoted widgets the wire action never touched are now asserted against the fixed PROMOTED_WIDGET_NAMES set; only the rewired one, whose promoted state legitimately varies with the restored snapshot, stays model-derived. Each production fix is proven by reverting it individually against the new tests: the store-reclaim ordering reds 3 tests, the token placement reds 1, the stale-slot guard reds 2, and the demotePromotedInput change reds 1.
|
The Finding 2 thread came through to me a second time — already answered it there (that fix is correct, keep it), so nothing new on that one. But it prompted me to read your replies on findings 3–6, which I hadn't seen, and two things in them need raising before this merges. 1. The carrier has outgrown "the same 7 files"Both of us verified The new one is 2. Finding 4's fix moved the reclamation the wrong wayThis is the one I'd most want you to look at, because it's the knot the whole thread has been stuck on. Before: if (connected) linkedInput.disconnect()
else subgraphNode.subgraph.removeInput(linkedInput)
if (hostWidgetId) useWidgetValueStore().deleteWidget(hostWidgetId) // unconditionalAfter: if (connected) linkedInput.disconnect() // now relies on the disconnect microtask
else { subgraphNode.subgraph.removeInput(linkedInput)
if (hostWidgetId) deleteWidget(hostWidgetId) }
Verified directionI applied the inverse on your head — restored #14495 is preserved, and To be clear this is a direction, not a finished patch: it does not yet close the orphaned-id gap (disconnect → then remove the slot), which still needs reclamation added to the teardown path keyed by an id captured before demotion clears it. I've deliberately not written that here, because it is new production logic and belongs on Where that leaves itFindings 2, 3, 5 and 6 look right to me — 3's Same recommendation, now with a sharper reason: put findings 1 + 4 on |
|
🔴 CI has now confirmed the #14495 regression on your own infrastructure — please don't merge
The only delta between This is what I predicted two comments ago from the unit probe plus the fixture: I can't name the exact failing spec from the API — the annotations only carry Current state is This doesn't invalidate the work — findings 2, 3, 5 and 6 are good and I've said so on each thread. It isolates the problem to the coupled pair (1 and 4), exactly where the three of us — you, me, and the panel — have been going back and forth. And it's the strongest possible argument for the sequencing: this check exists on Offer still open to write that |
This comment has been minimized.
This comment has been minimized.
Supersedes the approach in 5988061, which the e2e suite rejected. Two of its four production changes were wrong, and the unit tests I wrote for them were wrong in the same direction, so only the black-box harness caught it. What the harness showed, at 00b9c34: - subgraphPromotion.spec.ts "Promoted STRING widget edit survives a rebind of the interior link" went red. That is the test #18074 un-test.fail()-ed, and it depends on the widgetValueStore entry SURVIVING a demotion: a rebind disconnects and reconnects across two ticks, so the deferred teardown cannot tell it from a terminal disconnect, and the retained entry is what restores the user's edit when the slot is re-promoted under the same id. - subgraphPromotedWidgetRewireDuplication.spec.ts went red on the undo case. So the 🟠 High finding at SubgraphNode.ts:360 does not hold as stated. The `some()` predicate is not always-true: `removeInput` splices the input out of `this.inputs`, so what the predicate actually asks is "does this slot still exist", and deleteWidget is correctly unreachable while it does. Making it reachable reds #14495 — measured, both as a unit test and in the browser. The repair is therefore a clarity change with no behaviour change: the two conditions are now named (`slotStillExists`, `anotherSlotHoldsId`), evaluated after the clearing so the second one means what it says, and pinned from both sides — reclaiming on every disconnect reds the retention test and #14495, never reclaiming reds promotionAfterReplacement.test.ts. The stale-slot guard is also narrowed. `this.inputs.includes(input)` is the right question for store reclamation but the wrong one for skipping the teardown, so the guard is now identity on `input._widget`, which is what configure() and onRemoved() clear via _clearPromotedWidget. Unchanged from the superseded commit, both still proven red-on-revert: - 'input-connected' clears the pending-demotion token only after the `!e.detail.widget || !e.detail.node` guard. SubgraphInput.connect dispatches that event with no widget/node when the new interior target has no widget locator, so a same-tick rewire onto one cancelled the demotion and returned without promoting a replacement, stranding a phantom promoted widget. - demotePromotedInput no longer deletes the store entry synchronously on the disconnect branch; it queues behind the deferred teardown, closing the tick in which host.widgets held a projection whose entry was already gone. The removeInput branch keeps the synchronous delete, since nothing defers there. Test changes: - the mid-rewire assertions are `toBe(1)`, not `toBeLessThanOrEqual(1)`, which 0 satisfies — i.e. the PM-1254 disappearance symptom. Verified by removing the deferral: the old assertion passes on the defect, the new one fails. - four new unit tests: entry retained across a disconnect and value restored on re-promotion; demotion still happens when a same-tick rewire lands on a widgetless interior input; no demotion event when the node is removed or the host reconfigured in the same tick. - the undo spec's tautological assertion is NOT fixed. Replacing it with the fixed PROMOTED_WIDGET_NAMES set reds: 'height' is genuinely absent from the host model after one undo on the external-boundary path, reproduced in CI (first attempt and both retries) and locally, and forcing a change-tracker capture before the wire does not fix it. Either the newest snapshot predates the last promotion or configure() fails to restore it; separating those owns the fix. The assertion is left as-is with the measurement recorded on it, so it reads as a known gap instead of as coverage.
Review repair at
|
| Finding | Fix | Reverting it reds |
|---|---|---|
🟠 :317 phantom promoted widget |
token cleared only after the !e.detail.widget || !e.detail.node guard |
1 test, on the stranded widgetId |
🟡 :354 stale queued demotion |
queued callback bails on input._widget !== widget |
2 tests, on a spurious widget-demoted |
🟢 :359 one-tick model/store split |
demotePromotedInput queues its delete behind the deferred teardown |
1 test |
🟢 test.ts:1814 vacuous assertion |
toBe(1), not toBeLessThanOrEqual(1) |
measured: with the deferral removed the old assertion passes and the new one fails |
Those four are resolved. Verified locally at this head: subgraphPromotion.spec.ts 22/22 and 554 unit tests across src/lib/litegraph/src/subgraph + src/core/graph/subgraph.
Not fixed — both now marked outdated, so neither blocks
🟠 :360, the widget-store leak — I believe this finding is wrong, and I would like a second opinion rather than my own say-so. I implemented it as prescribed and pushed it; CI rejected it. subgraphPromotion.spec.ts → "Promoted STRING widget edit survives a rebind of the interior link" went red — the test this PR's source un-test.fail()-ed. A rebind disconnects and reconnects across two ticks, so the deferred callback cannot tell it from a terminal disconnect, and the retained store entry is exactly what restores the user's edit on re-promotion. That is #14495. Separately, the some() predicate is not always-true: removeInput splices the input out of this.inputs first, so what it really asks is "does this slot still exist", and promotionAfterReplacement.test.ts has been covering the reachable case all along. I shipped a no-behaviour-change clarity edit naming the two conditions, pinned from both sides — reclaim-always reds #14495, reclaim-never reds promotionAfterReplacement.test.ts. Left unresolved on purpose: overturning a 6-of-8 finding on my own measurement shouldn't be self-certified.
🟢 spec.ts:220, the tautological undo assertion — the finding is correct and I could not fix it. Replacing the derived expectation with the fixed PROMOTED_WIDGET_NAMES set reds, because height is genuinely absent from the host model after one undo on the external-boundary path (first CI attempt plus both retries, and again locally). The suggested alternative — pinning the undo snapshot with captureCanvasState() before the wire — does not fix it either. So either the newest snapshot predates the last promotion, or configure() fails to restore it on rebuild; separating those owns the fix. I reverted to the derived expectation rather than ship a red test on a release carrier, and recorded the measurement on the assertion so it reads as a known gap. This needs whoever owns the undo/promotion-restore path.
Two other things worth knowing
- Findings
:317,:354and:359are live onmain. The hunks here are byte-identical to fix: keep promoted subgraph widgets present across a rewire #18074, which merged with one approval and zero review threads; this panel only ran on the backport. The fix applies tomainessentially as-is and the unit tests port unchanged. As of this commitcloud/1.54has the fix andmaindoes not. #19749, thecore/1.54companion, carries the same hunks and inherits the same three.
I have no merge authority on this PR and am not asking anyone to rubber-stamp it — the two open threads above are the honest remaining work.
|
Second opinion on That carries some weight because I argued the opposite side first. I was the one who called it a leak, measured
So the finding's diagnosis is factually right — And my own reasoning was wrong in the specific way you identified. I asserted the Verified at
Your "no behaviour change on any path" claim holds. And CI is now fully green here — 98 checks, 0 failures, both Playwright checks On The thing I'd most want a decision on: the release branch is now ahead of
You said the fix applies to |
|
Agreed on closing this one — the train is retired, and a But "carry still-applicable changes to 1.55 and 1.56" has a concrete gap here, and retiring 1.54 today makes two things disappear. Both verified against the remotes just now:
1. #18074 never reached 1.55, and 1.55 is live. It's on This was flagged as a labelling question when these carriers were first opened — #18074 was labelled 2. The three fixes developed on #19750 exist only on That's the inversion I raised on #19750 — the release branch ended up ahead of Happy to do both:
Also still worth its own issue regardless of any of the above: the Say the word on either and I'll pick it up. |
PR Created by the Glary-Bot Agent
Manual backport of #18074 to
cloud/1.54, from the v1.56.2 backport-status report. Companion to #19749 (core/1.54).Cherry-picked from upstream merge commit
9ee67cb17b.Requested in #frontend-releases as a cherry-pick "carrier" backport, one per target branch. #18074 is the only PR the report flags for
cloud/1.54that is actually backportable, so this carrier carries one commit — see "Why this carrier holds one commit" below.Why
Rewiring the interior link behind one promoted subgraph widget (e.g. dragging a new node onto an already-promoted widget's socket) could drop or visually duplicate the host node's other promoted widgets, and make the rewired widget itself disappear.
SubgraphInputNode's'input-disconnected'handler demoted the widget synchronously while deferring the matchingwidgetValueStorecleanup to aqueueMicrotask. For one tickhostNode.widgets.lengthand the store's tracked id count disagreed — and the Vue widget grid reads its render order from the store and its liveness filter fromhostNode.widgets, i.e. from both sides of that disagreement. The fix defers the demotion itself by a microtask with a per-input cancellation token, which'input-connected'clears on a same-tick reconnect, so a rewire never drops the widget at all.Conflict resolution
The fix itself (
SubgraphNode.ts,interfaces.ts) cherry-picked clean.SubgraphNode.tsdiffers slightly betweencore/1.54andcloud/1.54, but the fix applied cleanly to both and the resulting hunks are identical.Only two test files conflicted — the same two as on #19749, and those two files are byte-identical between
core/1.54andcloud/1.54at their merge base (verified by blob hash), so this is the same resolution, reusing the exact same resolved blobs:useErrorClearingHooks.test.ts— the conflict hunk also carried an unrelated test, "retires an absorbed promoted media error when its host is deleted", whichmaingained from a different, unbackported PR. Dropped that test and applied only this PR's change to the test that does exist here: make "keeps promoted missing media until the last fanout consumer is deleted"asyncandawaitthe now-deferred demotion.SubgraphWidgetPromotion.test.ts— import-only conflict. Accepted the three new imports (WidgetValue,createPromotedMultilineWidget,useDomWidgetStore) required by the new PM-1328 describe block, after verifying all three modules and symbols exist on this branch. The new block also usesLegacyWidget, whichmainimports from the litegraph barrel in a neighbouring non-conflicting import hunk — so this branch's version won and left it out, failing withReferenceError: LegacyWidget is not defineduntil it was added to the barrel import.LegacyWidgetalready exists and is exported from the barrel here, so this is an import fix, not new code.No backport-only compatibility shim was needed, and no public API shape or callback signature changed. The deliberate timing change this PR documents — a genuine disconnect becomes observable on
hostNode.widgetsone microtask later — is carried as-is rather than shimmed: the surrounding machinery (SubgraphInputNode.ts,SubgraphInputEventMap.ts) is byte-identical between this branch andmain, so a literal cherry-pick cannot recreate the original bug for an old-contract consumer, and deviating would leave this branch running demotion logicmain's CI has never exercised.Changed-file set is exactly the same 7 files as upstream
9ee67cb17b— no unrelated drift smuggled in.Validation
pnpm test:uniton the suites upstream used to validate the fix — 527/527 passing across 24 files (src/lib/litegraph/src/subgraph/,useErrorClearingHooks.test.ts,promotionUtils.test.ts,multilineTextarea.promoted.test.ts,widgetValueStore.test.ts,domWidgetStore.test.ts). This includes the PR's own regression case, "keeps every promoted widget present while rewiring the interior link behind one of them", which assertshostNode.widgets.lengthstays at 4 throughout a rewire — i.e. the exact mechanism, proven fixed on this branch.pnpm typecheck— exit 0pnpm exec eslinton changed files — 0 errors, 6 warnings, all pre-existingimport-x/no-restricted-pathsonsrc/lib/(upstream carries the same;pnpm lintsets no--max-warnings)pnpm exec oxfmt --checkon changed files — cleanThis PR lands
subgraphPromotedWidgetRewireDuplication.spec.tsand flipstest.fail()→test()insubgraphPromotion.spec.ts. Upstream justified those flips with CI evidence gathered onmain(run 35419314544); there is no equivalent evidence forcloud/1.54yet, and the Playwright suite needs a live ComfyUI backend plus a built frontend, which I could not stand up in the sandbox. The unit-level regression above proves the mechanism on this branch, but perSKILL.md("NEVER merge a backport PR without all CI checks passing") the flipped e2e should be confirmed green here before merge. Kept as a draft for that reason.Why this carrier holds one commit
The report flagged four PRs as "❗ Need backport". Two cannot be carried:
refs/pull/19053/headexists; it is 69 commits ahead ofmainwith no squash commit, which is why the report listed it without a(#NNNNN)suffix. There is nothing to cherry-pick.cloud/1.54, but its prerequisite is missing. Its mechanism marks the target unavailable so the in-chat notice from fix(agent): reopen saved workflows from chat history #18866 renders. This branch has theTargetTrackingmodel but nomarkWorkflowTargetUnavailable/targetUnavailableand notargetWorkflowUnavailablestring — the only "unavailable" path here is an unrelated toast fired on navigation failure (agent.targetNavigationUnavailable), a different mechanism for a different trigger. Carrying fix(agent): say when the open chat's target workflow is deleted #19240 alone would set a state nothing renders — a silent no-op defeating the fix. It needs fix(agent): reopen saved workflows from chat history #18866 backported first. (Same gap on all four target branches.)#19677 is labelled
core/1.55+cloud/1.55only, so it is not in this branch's scope.Original PR: #18074 / Original commit:
9ee67cb17b