Repository navigation
fix(agent-crdt): keep human-added frontend-only nodes across a workflow tab switch - #18078
Conversation
Failing repros (test.fail / it.fails) for a node added on the canvas disappearing after leaving and returning to its workflow tab while the Agent follower is bound, with passing controls beside them. The doc host rejects the add_node for any node class outside the pinned catalog once the mint rewrites widgets_values to a name-keyed object (frontend-registered nodes such as Note, Get/Set nodes from node packs, blueprint hosts with promoted widgets); the follower only logs the ok:false result. Rebinding on the tab return arms a full reconcile whose removeMissing deletes every store record the doc does not hold, and the orphan sweep removes the live node. The same reconcile arms without a tab switch when the echo of an accepted add is rejected as already registered. The agent harness now judges human doc_ops with the real applier, records every client doc frame, and can hold acknowledgements. Unit lenses rule out a serialize() drop and a missing blueprint definition in the snapshot.
|
Important Review skippedBot user detected. To trigger a single review, invoke the ⚙️ Run configurationConfiguration used: Repository: Comfy-Org/ComfyUI_frontend/.coderabbit.yaml Review profile: ASSERTIVE Plan: Advanced Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Comment |
🎨 Storybook: ✅ Built — View Storybook🎭 Playwright: ✅ 2071 passed, 0 failed📊 Browser Reports
📦 Bundle: 9.95 MB gzip ⚪ 0 BDetailsSummary
Category Glance App Entry Points — 3.81 kB (baseline 3.81 kB) • ⚪ 0 BMain entry bundles and manifests Status: 1 unchanged Graph Workspace — 1.46 MB (baseline 1.46 MB) • ⚪ 0 BGraph editor runtime, canvas, workflow orchestration Status: 3 unchanged Views & Navigation — 132 kB (baseline 132 kB) • ⚪ 0 BTop-level views, pages, and routed surfaces Status: 18 unchanged Panels & Settings — 638 kB (baseline 638 kB) • ⚪ 0 BConfiguration panels, inspectors, and settings screens Status: 28 unchanged User & Accounts — 37.9 kB (baseline 37.9 kB) • ⚪ 0 BAuthentication, profile, and account management bundles Status: 11 unchanged Editors & Dialogs — 126 kB (baseline 126 kB) • ⚪ 0 BModals, dialogs, drawers, and in-app editors Status: 8 unchanged UI Components — 124 kB (baseline 124 kB) • ⚪ 0 BReusable component library chunks Status: 16 unchanged Data & Services — 3.87 MB (baseline 3.87 MB) • ⚪ 0 BStores, services, APIs, and repositories Status: 16 unchanged Utilities & Hooks — 566 kB (baseline 566 kB) • ⚪ 0 BHelpers, composables, and utility bundles Status: 38 unchanged Vendor & Third-Party — 17.4 MB (baseline 17.4 MB) • ⚪ 0 BExternal libraries and shared vendor chunks Status: 18 unchanged Other — 17.8 MB (baseline 17.8 MB) • ⚪ 0 BBundles that do not match a named category Status: 301 unchanged ⚡ Performance Report
Show regressions
All metrics
Historical variance (last 15 runs)
Trend (last 15 commits on main)
Raw data{
"timestamp": "2026-09-20T19:13:41.782Z",
"gitSha": "44a399aef07baef767ab53b78dd7eb2ce0f8f103",
"branch": "claude/repro-node-vanishes-after-tab-switch",
"measurements": [
{
"name": "canvas-idle",
"durationMs": 2017.1549999999456,
"styleRecalcs": 9,
"styleRecalcDurationMs": 5.342999999999999,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 452.66900000000004,
"heapDeltaBytes": 10839476,
"heapUsedBytes": 72415024,
"domNodes": -263,
"jsHeapTotalBytes": 4710400,
"scriptDurationMs": 6.588,
"eventListeners": -152,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-idle",
"durationMs": 2011.1389999999574,
"styleRecalcs": 9,
"styleRecalcDurationMs": 5.088999999999998,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 308.34700000000004,
"heapDeltaBytes": -4943948,
"heapUsedBytes": 57375776,
"domNodes": -262,
"jsHeapTotalBytes": 4448256,
"scriptDurationMs": 6.901000000000001,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "canvas-mouse-sweep",
"durationMs": 1691.4290000000278,
"styleRecalcs": 73,
"styleRecalcDurationMs": 17.776,
"layouts": 14,
"layoutDurationMs": 2.7860000000000005,
"taskDurationMs": 627.224,
"heapDeltaBytes": -37196,
"heapUsedBytes": 62155500,
"domNodes": -259,
"jsHeapTotalBytes": 6283264,
"scriptDurationMs": 109.819,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "canvas-mouse-sweep",
"durationMs": 1710.062999999991,
"styleRecalcs": 73,
"styleRecalcDurationMs": 17.769,
"layouts": 14,
"layoutDurationMs": 2.82,
"taskDurationMs": 658.183,
"heapDeltaBytes": 14602048,
"heapUsedBytes": 76597596,
"domNodes": -259,
"jsHeapTotalBytes": 6021120,
"scriptDurationMs": 112.52399999999999,
"eventListeners": -152,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-zoom-sweep",
"durationMs": 1709.699999999998,
"styleRecalcs": 30,
"styleRecalcDurationMs": 11.336000000000002,
"layouts": 6,
"layoutDurationMs": 0.969,
"taskDurationMs": 269.81600000000003,
"heapDeltaBytes": 2426068,
"heapUsedBytes": 64558380,
"domNodes": 76,
"jsHeapTotalBytes": 4718592,
"scriptDurationMs": 9.119999999999997,
"eventListeners": 15,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "canvas-zoom-sweep",
"durationMs": 1710.4790000000776,
"styleRecalcs": 30,
"styleRecalcDurationMs": 11.871999999999998,
"layouts": 6,
"layoutDurationMs": 0.9579999999999999,
"taskDurationMs": 263.59900000000005,
"heapDeltaBytes": 2198848,
"heapUsedBytes": 64199320,
"domNodes": 76,
"jsHeapTotalBytes": 5242880,
"scriptDurationMs": 8.719000000000001,
"eventListeners": 15,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.670000000000012,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "dom-widget-clipping",
"durationMs": 516.7260000000056,
"styleRecalcs": 12,
"styleRecalcDurationMs": 5.653999999999999,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 331.11800000000005,
"heapDeltaBytes": 18189248,
"heapUsedBytes": 80461568,
"domNodes": 19,
"jsHeapTotalBytes": 4718592,
"scriptDurationMs": 117.68199999999999,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "dom-widget-clipping",
"durationMs": 544.4480000000453,
"styleRecalcs": 11,
"styleRecalcDurationMs": 5.3950000000000005,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 334.704,
"heapDeltaBytes": 17782284,
"heapUsedBytes": 79810364,
"domNodes": 18,
"jsHeapTotalBytes": 5767168,
"scriptDurationMs": 119.27799999999999,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "large-graph-idle",
"durationMs": 2013.3189999999672,
"styleRecalcs": 7,
"styleRecalcDurationMs": 4.5489999999999995,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 367.314,
"heapDeltaBytes": 1508488,
"heapUsedBytes": 83090672,
"domNodes": -243,
"jsHeapTotalBytes": -1019904,
"scriptDurationMs": 20.705999999999996,
"eventListeners": -152,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "large-graph-idle",
"durationMs": 2024.6940000000109,
"styleRecalcs": 8,
"styleRecalcDurationMs": 4.907,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 490.4400000000001,
"heapDeltaBytes": 634484,
"heapUsedBytes": 80063780,
"domNodes": -247,
"jsHeapTotalBytes": 516096,
"scriptDurationMs": 25.802,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "large-graph-pan",
"durationMs": 2129.551000000049,
"styleRecalcs": 68,
"styleRecalcDurationMs": 7.688000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1200.278,
"heapDeltaBytes": -1784960,
"heapUsedBytes": 80427136,
"domNodes": -245,
"jsHeapTotalBytes": -8192,
"scriptDurationMs": 584.431,
"eventListeners": -152,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "large-graph-pan",
"durationMs": 2142.831000000001,
"styleRecalcs": 69,
"styleRecalcDurationMs": 9.209999999999999,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1226.6,
"heapDeltaBytes": -13966128,
"heapUsedBytes": 68473348,
"domNodes": -219,
"jsHeapTotalBytes": 290816,
"scriptDurationMs": 605.22,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "large-graph-zoom",
"durationMs": 3099.8930000000087,
"styleRecalcs": 63,
"styleRecalcDurationMs": 8.615000000000002,
"layouts": 60,
"layoutDurationMs": 9.443999999999999,
"taskDurationMs": 1255.8920000000003,
"heapDeltaBytes": 8201268,
"heapUsedBytes": 74174092,
"domNodes": 8,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 561.1859999999999,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "large-graph-zoom",
"durationMs": 3083.77999999999,
"styleRecalcs": 63,
"styleRecalcDurationMs": 8.391,
"layouts": 60,
"layoutDurationMs": 8.902000000000001,
"taskDurationMs": 1190.4289999999999,
"heapDeltaBytes": 8029044,
"heapUsedBytes": 73814444,
"domNodes": 8,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 521.067,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "legacy-node-drag",
"durationMs": 2215.3790000000413,
"styleRecalcs": 42,
"styleRecalcDurationMs": 3.934,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1343.1349999999998,
"heapDeltaBytes": 17174016,
"heapUsedBytes": 82509452,
"domNodes": 4,
"jsHeapTotalBytes": 417792,
"scriptDurationMs": 656.8980000000001,
"eventListeners": 180,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "legacy-node-drag",
"durationMs": 2203.003000000024,
"styleRecalcs": 45,
"styleRecalcDurationMs": 6.643,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1381.73,
"heapDeltaBytes": -10903816,
"heapUsedBytes": 74437384,
"domNodes": -225,
"jsHeapTotalBytes": 253952,
"scriptDurationMs": 646.2869999999999,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "minimap-idle",
"durationMs": 2033.4960000000137,
"styleRecalcs": 8,
"styleRecalcDurationMs": 4.918,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 482.182,
"heapDeltaBytes": -8522688,
"heapUsedBytes": 79928704,
"domNodes": -250,
"jsHeapTotalBytes": 552960,
"scriptDurationMs": 25.854000000000003,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "minimap-idle",
"durationMs": 2026.9250000000056,
"styleRecalcs": 7,
"styleRecalcDurationMs": 5.129,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 471.372,
"heapDeltaBytes": 54076,
"heapUsedBytes": 82188140,
"domNodes": -248,
"jsHeapTotalBytes": -495616,
"scriptDurationMs": 24.377999999999997,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "subgraph-dom-widget-clipping",
"durationMs": 497.52300000000105,
"styleRecalcs": 46,
"styleRecalcDurationMs": 6.481999999999998,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 289.459,
"heapDeltaBytes": 13553424,
"heapUsedBytes": 76825308,
"domNodes": 18,
"jsHeapTotalBytes": 5767168,
"scriptDurationMs": 99.964,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.663333333333338,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-dom-widget-clipping",
"durationMs": 491.70699999990575,
"styleRecalcs": 47,
"styleRecalcDurationMs": 7.357000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 302.767,
"heapDeltaBytes": 13211036,
"heapUsedBytes": 76681440,
"domNodes": 20,
"jsHeapTotalBytes": 5242880,
"scriptDurationMs": 102.857,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000273
},
{
"name": "subgraph-idle",
"durationMs": 1992.4100000000067,
"styleRecalcs": 10,
"styleRecalcDurationMs": 5.621000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 374.49800000000005,
"heapDeltaBytes": -1771436,
"heapUsedBytes": 61918960,
"domNodes": -276,
"jsHeapTotalBytes": 5496832,
"scriptDurationMs": 6.107999999999999,
"eventListeners": -220,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "subgraph-idle",
"durationMs": 2010.417000000018,
"styleRecalcs": 10,
"styleRecalcDurationMs": 5.927,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 371.91299999999995,
"heapDeltaBytes": 17856812,
"heapUsedBytes": 81151076,
"domNodes": -276,
"jsHeapTotalBytes": 5234688,
"scriptDurationMs": 6.545999999999999,
"eventListeners": -198,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-mouse-sweep",
"durationMs": 1693.7389999999937,
"styleRecalcs": 77,
"styleRecalcDurationMs": 17.555,
"layouts": 18,
"layoutDurationMs": 3.1440000000000006,
"taskDurationMs": 571.893,
"heapDeltaBytes": -187492,
"heapUsedBytes": 63529132,
"domNodes": -274,
"jsHeapTotalBytes": 5758976,
"scriptDurationMs": 69.71999999999998,
"eventListeners": -220,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "subgraph-mouse-sweep",
"durationMs": 1694.9979999999982,
"styleRecalcs": 77,
"styleRecalcDurationMs": 19.139,
"layouts": 18,
"layoutDurationMs": 3.6799999999999997,
"taskDurationMs": 594.476,
"heapDeltaBytes": 11786808,
"heapUsedBytes": 75343436,
"domNodes": -275,
"jsHeapTotalBytes": 5496832,
"scriptDurationMs": 71.883,
"eventListeners": -198,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "subgraph-transition-enter",
"durationMs": 883.4400000000073,
"styleRecalcs": 17,
"styleRecalcDurationMs": 21.379,
"layouts": 5,
"layoutDurationMs": 8.472,
"taskDurationMs": 687.387,
"heapDeltaBytes": -6525436,
"heapUsedBytes": 80779332,
"domNodes": 13587,
"jsHeapTotalBytes": 11272192,
"scriptDurationMs": 22.36,
"eventListeners": 2383,
"totalBlockingTimeMs": 114,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "viewport-pan-sweep",
"durationMs": 8203.356999999982,
"styleRecalcs": 250,
"styleRecalcDurationMs": 23.337000000000003,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 4395.813999999999,
"heapDeltaBytes": 5138684,
"heapUsedBytes": 86558324,
"domNodes": -241,
"jsHeapTotalBytes": 1040384,
"scriptDurationMs": 2069.653,
"eventListeners": -152,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "viewport-pan-sweep",
"durationMs": 8135.946999999987,
"styleRecalcs": 251,
"styleRecalcDurationMs": 23.259,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 3782.681,
"heapDeltaBytes": -15030460,
"heapUsedBytes": 66568144,
"domNodes": -213,
"jsHeapTotalBytes": 2912256,
"scriptDurationMs": 1916.2920000000001,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-large-graph-idle",
"durationMs": 12141.689999999982,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 11095.73,
"heapDeltaBytes": -32814044,
"heapUsedBytes": 183917572,
"domNodes": -8061,
"jsHeapTotalBytes": -15929344,
"scriptDurationMs": 184.79600000000002,
"eventListeners": -14928,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-large-graph-idle",
"durationMs": 10713.12899999998,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 10359.211000000001,
"heapDeltaBytes": -21331096,
"heapUsedBytes": 201354548,
"domNodes": -8062,
"jsHeapTotalBytes": -10948608,
"scriptDurationMs": 130.81199999999998,
"eventListeners": -14928,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-large-graph-pan",
"durationMs": 11301.899999999989,
"styleRecalcs": 67,
"styleRecalcDurationMs": 12.337999999999989,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 10670.125,
"heapDeltaBytes": -19175100,
"heapUsedBytes": 190968368,
"domNodes": -8061,
"jsHeapTotalBytes": -12722176,
"scriptDurationMs": 643.076,
"eventListeners": -14928,
"totalBlockingTimeMs": 17,
"frameDurationMs": 17.219999999999953,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "vue-large-graph-pan",
"durationMs": 13848.013999999921,
"styleRecalcs": 68,
"styleRecalcDurationMs": 11.533000000000015,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 13254.053,
"heapDeltaBytes": -47630000,
"heapUsedBytes": 162791464,
"domNodes": -8062,
"jsHeapTotalBytes": -20779008,
"scriptDurationMs": 662.1779999999999,
"eventListeners": -14928,
"totalBlockingTimeMs": 86,
"frameDurationMs": 17.219999999999953,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "vue-node-resize-workload",
"durationMs": 4464.781000000016,
"styleRecalcs": 20,
"styleRecalcDurationMs": 59.92700000000001,
"layouts": 20,
"layoutDurationMs": 21.792999999999992,
"taskDurationMs": 4339.377,
"heapDeltaBytes": -24829760,
"heapUsedBytes": 189970980,
"domNodes": -8061,
"jsHeapTotalBytes": 8118272,
"scriptDurationMs": 76.137,
"eventListeners": -14928,
"totalBlockingTimeMs": 36,
"frameDurationMs": 17.223333333333358,
"p95FrameDurationMs": 16.80000000000109
},
{
"name": "vue-node-resize-workload",
"durationMs": 4143.006000000014,
"styleRecalcs": 20,
"styleRecalcDurationMs": 59.836,
"layouts": 20,
"layoutDurationMs": 19.622,
"taskDurationMs": 4032.1600000000003,
"heapDeltaBytes": -54027488,
"heapUsedBytes": 168590672,
"domNodes": -8062,
"jsHeapTotalBytes": 4780032,
"scriptDurationMs": 103.252,
"eventListeners": -14928,
"totalBlockingTimeMs": 7,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.80000000000109
},
{
"name": "vue-zoom-out-idle",
"durationMs": 6225.5809999999765,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 5798.181,
"heapDeltaBytes": -8844952,
"heapUsedBytes": 177426144,
"domNodes": 0,
"jsHeapTotalBytes": -23855104,
"scriptDurationMs": 100.81100000000004,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333338,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-zoom-out-idle",
"durationMs": 6031.934999999976,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 5544.500000000001,
"heapDeltaBytes": 1226604,
"heapUsedBytes": 182266224,
"domNodes": 0,
"jsHeapTotalBytes": 524288,
"scriptDurationMs": 86.70400000000001,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "workflow-execution",
"durationMs": 441.5779999999927,
"styleRecalcs": 12,
"styleRecalcDurationMs": 14.919,
"layouts": 3,
"layoutDurationMs": 0.819,
"taskDurationMs": 97.601,
"heapDeltaBytes": 5012348,
"heapUsedBytes": 70173600,
"domNodes": 124,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 10.263000000000002,
"eventListeners": 81,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "workflow-execution",
"durationMs": 447.49500000000353,
"styleRecalcs": 13,
"styleRecalcDurationMs": 16.100999999999996,
"layouts": 4,
"layoutDurationMs": 1.719,
"taskDurationMs": 98.30999999999999,
"heapDeltaBytes": 4915648,
"heapUsedBytes": 66748164,
"domNodes": 120,
"jsHeapTotalBytes": 262144,
"scriptDurationMs": 9.44,
"eventListeners": 81,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.799999999999727
}
]
}🔬 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 3f803fe1a4710a73812663950b9c9d754af2675d
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 |
…FollowerHostSocket Main extracted the routed /ws host into AgentFollowerHostSocket. The client doc_* frame recording, the humanOpsHost apply/hold switch and the applier outcomes now live in that class (default 'hold', the pre-merge behaviour for other fixtures); the conversation harness delegates to it and keeps its public lenses unchanged.
01425da to
85b966b
Compare
Frame parsing, the subscribe answer and the human-op judging each get their own function so the routed /ws host stays under the complexity gate.
… classes serializeForMint rewrote every widget-bearing node's widgets_values into a name-keyed record before minting add_node. The doc host's applier only projects name-keyed widgets for classes in its pinned catalog (the server's node definitions) and rejects any other class with uncatalogued_widget_write, while it stores a positional array for such a class opaquely. Frontend-only classes (isVirtualNode: Note, MarkdownNote, PrimitiveNode, Get/Set nodes from node packs, subgraph blueprint hosts) now keep the positional form serialize() already produced, so their add lands in the document instead of being rejected and later swept away by the first reconcile after a tab return. The unit test feeds the minted op to the real applier and asserts it is applied for a Note and for a blueprint host with a promoted widget.
With the mint keeping positional widgets_values for uncatalogued classes the host applies every add in the spec, so the test.fail markers come off: the Note and promoted-widget blueprint tab-switch cases and the no-tab-switch Note case are now regression tests, and the host lens expects applied for all four add cases.
Plan for the rest of the fixThe mint-side change in this PR (
Order: 1 and 4 share the ledger wiring and can ship together; 2 depends on 1 for the ledger-aware variant; 3 is independent and needs a package change. Generated by Claude Code |
Playwright videos for new testsDownload the recorded videos and HTML report Recorded projects: chromium cloud Recorded test titles:
|
Codecov Report✅ All modified and coverable lines are covered by tests. @@ Coverage Diff @@
## main #18078 +/- ##
==========================================
+ Coverage 82.95% 83.95% +1.00%
==========================================
Files 2956 2260 -696
Lines 145654 128706 -16948
Branches 37535 34722 -2813
==========================================
- Hits 120823 108053 -12770
+ Misses 23561 20363 -3198
+ Partials 1270 290 -980
Flags with carried forward coverage won't be shown. Click here to find out more.
... and 711 files with indirect coverage changes 🚀 New features to boost your workflow:
|
The conversation harness applied every human doc_ops batch by default, which fed page-minted ops into the host document of specs written against a host that ignored them (agentTabSwitchCatchUp's tab B MarkdownNote reached the recorded workflow's projection). The default is 'hold' again and agentHumanAddTabSwitch opts into 'apply'. The host lens accepts a batch with no rejection and at least one applied op: a blueprint paste mints more than the single add_node, and the applier answers no-op for the op that changes nothing. The chromium baseline adds nodes through the classic search box the other chromium specs use; the V2 double-click at an arbitrary position did not open the dialog in CI.
There was a problem hiding this comment.
Reviewed against the reported symptom (Get/Set node and subgraph blueprint vanishing after a workflow tab switch) and against main at the base SHA.
Fix: correct and addresses the real root cause. serializeForMint was rewriting every widget-bearing node's widgets_values into a name-keyed record before minting add_node; the doc host's applier only accepts name-keyed widgets for classes in its pinned catalog and rejects everything else with uncatalogued_widget_write, but stores a positional array for those classes opaquely. Frontend-only classes (isVirtualNode: Note/MarkdownNote in noteNode.ts, PrimitiveNode in widgetInputs.ts, SubgraphNode in subgraph/SubgraphNode.ts) have no catalog entry, so their adds were silently rejected and swallowed by the follower — then swept away by the removeMissing reconcile that fires on the first frame after a tab-return rebind. Keeping the positional form for isVirtualNode classes lets the add land in the doc, so the later reconcile has nothing to sweep. I confirmed this against main (the "before" code in the diff matches current main exactly) and against LGraphNode.isVirtualNode/SubgraphNode.isVirtualNode in the litegraph source.
Test coverage: already thorough, nothing to add. The commit history is a genuine red→green pair: a3ae029c adds test.fail repros against unmodified main, 73e53406 is the fix, 56a7196 removes the test.fail markers once they pass. The Vitest unit test feeds the minted op through the real @comfyorg/comfy-multi-player applier (not a mock) against a catalog that omits the class and asserts outcome: 'applied'. The Playwright spec (agentHumanAddTabSwitch.spec.ts) covers the host lens, four tab-switch cases, and a no-tab-switch echo-arming case, plus a non-agent baseline spec. The remaining half of the mechanism (rebind reconcile sweeping local-only nodes, rejected ops being swallowed, blueprint definitions not traveling on the human path) is correctly left as it.fails placeholders with a clear, ordered follow-up plan already posted on this PR — that's the right scope boundary for this fix, not a gap. I did not add further tests; padding an already-complete red/green suite would add noise, not confidence.
One thing to watch pre-merge: CI (Playwright shards) was still in progress at review time — worth confirming green before taking this out of draft.
🤖 Generated with Claude Code
The chromium baseline opened the search box by double-clicking the default graph, where the click hit a node. Every passing chromium spec that uses the V2 search box starts from a blank workflow and clicks at (200, 200); the baseline now does the same and returns to the tab it was on rather than assuming a single tab.
DrJKL
left a comment
There was a problem hiding this comment.
Verdict: follow-up requested
I confirmed the narrow production fix: preserving positional widgets_values for uncatalogued/virtual classes lets the real applier retain the reported frontend-only Note, KJNodes Get/Set, and blueprint-host cases. The focused mint regression fails on the pre-fix parent and passes at this head. I found no production-code blocker in that path, but I do want the test harness and coverage issues in the inline comments addressed before this is treated as approval-ready.
I also checked the earlier bot review. I agree with its root-cause analysis. I do not agree that the history demonstrates conventional red-to-green: the first repros were committed as expected failures, so the repository history shows expected-failure-to-passing unless there is a separately saved failing run. Its claim that blueprint-definition transmission is pinned by an expected failure also overstates the current three pins: they cover two reconciliation paths and rejected-op reporting, not a missing-definition case. The previously pending browser jobs are now green.
Checks and coverage
- CRDT/LGraph Vitest scope: 42 files, 622 passed, 4 expected failures.
- Focused changed tests: 19 passed, 3 expected failures.
pnpm typecheck,pnpm typecheck:browser,pnpm format:check,pnpm oxlint:main,pnpm knip, and changed-since Fallow audit passed.- Changed-file ESLint exited successfully but reported three new restricted-layer warnings in
LGraph.tabSwitchSnapshot.test.ts; see inline comment. - All current GitHub checks are successful or intentionally skipped, including cloud/chromium and new-test video jobs.
- I discovered all 12 added Playwright cases. A local cloud run could not start because this environment's backend did not register
KSampler; I treated that as an environment limitation rather than a product failure because the corresponding CI jobs passed. - I verified current KJNodes source and history: its Get/Set classes unconditionally set
isVirtualNode = true, so the reported extension case does take the fixed mint path.
Architecture context
FE-2504 proposes replacing store-first projection and compensating reconciliation with semantic graph apply plus call-carried provenance. I treated it as high-priority backlog direction, not an accepted supersession of this fix. The behavioral acceptance tests here are durable migration constraints; the fake relay/reconciliation scaffolding is temporary machinery that should be deleted or reshaped when that architecture lands. FE-2505 explicitly overlaps these acceptance pins. Conversation REST recovery is separate from graph projection and does not change this review.
| type: envelope.type, | ||
| workflowId: stringOrNull(workflow_id), | ||
| stateVector: stringOrNull(state_vector_b64), | ||
| ops: Array.isArray(ops) ? (ops as Op[]) : [] |
There was a problem hiding this comment.
This cast gives the fake host a weaker protocol boundary than the relay it represents: any JSON array is now accepted as Op[] and passed to the real applier. Please validate the operation elements (preferably through the same client-frame contract used by the relay) before treating them as Op; otherwise malformed frontend output can make this harness pass or fail for behavior production would reject earlier.
There was a problem hiding this comment.
Added isWireOp/parseWireOps to opEnvelope.ts: a runtime narrowing of an untyped JSON array to Op[] that mirrors the envelope gate the applier itself runs before dispatch (op kind in the exported FROZEN_OPS, non-empty op_id — matching dist/applier.js's validateEnvelope). agentFollowerHostSocket.ts now calls parseWireOps(ops) instead of the as Op[] cast.
Generated by Claude Code
There was a problem hiding this comment.
Tracked as FE-2558. The PR merged with this thread unresolved, so the finding is carried in Linear rather than lost here. Filed automatically; close the ticket if the finding is no longer true at main.
|
|
||
| // First half of the mechanism, on its own: the host takes the page's | ||
| // add_node op and the page keeps the node. | ||
| for (const { name, add } of [ |
There was a problem hiding this comment.
This four-case host-only matrix is a strict prefix of the four tab-return cases below: those tests repeat replay, add, host judgment, canvas, and host assertions before exercising the switch. Please remove this duplicated E2E matrix or move the host-only contract to a lower-level test so the browser suite pays for each product behavior once.
There was a problem hiding this comment.
Removed the host-only 4-case matrix. Its per-shape host verdict (Note / blueprint-with-promoted-widget → applied against the real applier) is already pinned at the unit level in mintPortWiring.test.ts (the it.for block added earlier on this branch, real mint/applyOps). The surviving tab-return cases now assert expectApplied themselves right after the add, so that coverage isn't lost, just not duplicated across a second full E2E run per shape.
Generated by Claude Code
There was a problem hiding this comment.
Tracked as FE-2558. The PR merged with this thread unresolved, so the finding is carried in Linear rather than lost here. Filed automatically; close the ticket if the finding is no longer true at main.
| vi.mock(import('@/platform/telemetry/reportError'), () => ({ | ||
| reportError: mockReportError | ||
| })) | ||
| vi.mock<unknown>(import('@/services/litegraphService'), () => ({ |
There was a problem hiding this comment.
This test places an app-integrated serialization scenario inside the vendored LiteGraph layer. Changed-file ESLint now reports three import-x/no-restricted-paths warnings here for the store, telemetry, and service imports. Please move the integrated regression to an app/workbench-owned suite and keep the LiteGraph test at the library boundary.
There was a problem hiding this comment.
Split the file: the store-divergence test (imports useNodeDataStore, reportError, litegraphService) moved to agentCrdtProjection.serializeDivergence.test.ts next to the other agent-crdt tests. LGraph.tabSwitchSnapshot.test.ts keeps only the pure litegraph test (a pasted blueprint definition surviving serialize()/reload); eslint now reports zero import-x/no-restricted-paths warnings on it.
Generated by Claude Code
There was a problem hiding this comment.
Tracked as FE-2558. The PR merged with this thread unresolved, so the finding is carried in Linear rather than lost here. Filed automatically; close the ticket if the finding is no longer true at main.
| function outcomesFor( | ||
| agentConversation: AgentConversationHarness | ||
| ): ApplyOutcome[] { | ||
| return agentConversation.humanOpOutcomes() |
There was a problem hiding this comment.
outcomesFor only renames agentConversation.humanOpOutcomes() and adds another hop to every assertion. Please call the harness method directly; this is the one-line alias-helper pattern the browser-test guidance asks us to avoid.
There was a problem hiding this comment.
Removed outcomesFor; the spec now calls agentConversation.humanOpOutcomes() directly everywhere it's needed.
Generated by Claude Code
There was a problem hiding this comment.
Tracked as FE-2558. The PR merged with this thread unresolved, so the finding is carried in Linear rather than lost here. Filed automatically; close the ticket if the finding is no longer true at main.
| const sent = JSON.parse(send.mock.calls[1][0]) as { | ||
| type: string | ||
| data: { ops: { op_id: string }[] } | ||
| } |
There was a problem hiding this comment.
This asserted parse can hide drift in the sender's wire contract: the test declares whatever shape it expects and then reads through it. Please narrow the parsed unknown at runtime or expose the typed outbound frame through the owned transport test double before extracting op_id.
There was a problem hiding this comment.
Replaced the as {...} cast on JSON.parse's result with a local sentOp() helper that narrows the sent frame through the same parseWireOps guard before reading op_id. A wire-shape drift in the outbound frame now makes the parsed op list come back empty (and the test throw), instead of silently reading op_id: undefined.
Generated by Claude Code
There was a problem hiding this comment.
Tracked as FE-2558. The PR merged with this thread unresolved, so the finding is carried in Linear rather than lost here. Filed automatically; close the ticket if the finding is no longer true at main.
| const PROMPT_NODE_ID = '6' | ||
| const ADD_POSITION: [number, number] = [400, 400] | ||
|
|
||
| interface NodeLens { |
There was a problem hiding this comment.
This spec carries roughly 330 lines of reusable fixture data, page drivers, observers, evidence capture, and helper contracts before the tests begin. That makes the scenarios difficult to audit and leaves fixture behavior owned by one spec. Please move the blueprint data to fixture data and the reusable page/host operations to the Agent conversation harness or component fixtures, leaving the spec responsible for scenarios and assertions.
There was a problem hiding this comment.
Moved the blueprint workflow-JSON factory to browser_tests/fixtures/data/agent/agentHumanAddBlueprints.ts, and the page/host drivers (add-node/add-blueprint/add-through-search-box, the tab-switch observer, the node lenses, evidence capture, the human-ops wait helpers, and the tab-switch sequence) onto AgentConversationHarness in agentConversationFixture.ts. The spec is now just the case tables, expectApplied, and the test/test.step bodies.
Generated by Claude Code
There was a problem hiding this comment.
Tracked as FE-2558. The PR merged with this thread unresolved, so the finding is carried in Linear rather than lost here. Filed automatically; close the ticket if the finding is no longer true at main.
| const node = window.LiteGraph!.createNode(nodeType) | ||
| if (!node) throw new Error(`${nodeType} is not a registered node type`) | ||
| node.pos = [pos[0], pos[1]] | ||
| window.app!.graph.add(node) |
There was a problem hiding this comment.
The Agent-bound cases call graph.add (and _deserializeItems for blueprints) directly. That proves mint/reconcile internals, but it does not prove that one real composition path—search, sidebar, or paste—still reaches the mint wiring while the follower is bound. The non-Agent baseline uses search, so please cover at least one frontend-only node through a user-facing add path in this bound suite.
There was a problem hiding this comment.
Added AgentConversationHarness.addNoteThroughSearchBox() (double-click the canvas open, type into the real search box, Enter — the same path the non-Agent baseline uses) and a fifth tab-return case, FRONTEND_ONLY_VIA_SEARCH, alongside the direct-API cases.
Generated by Claude Code
There was a problem hiding this comment.
Tracked as FE-2558. The PR merged with this thread unresolved, so the finding is carried in Linear rather than lost here. Filed automatically; close the ticket if the finding is no longer true at main.
| @@ -0,0 +1,84 @@ | |||
| import { | |||
There was a problem hiding this comment.
This workflow-tab behavior belongs with the existing topbar tab specs rather than at the browser-test root. Please place it under browser_tests/tests/topbar/ so feature ownership and targeted test selection remain discoverable.
There was a problem hiding this comment.
Moved via git mv to browser_tests/tests/topbar/workflowTabSwitchKeepsAddedNodes.spec.ts. All its imports already use the @e2e/* alias, so nothing else needed to change.
Generated by Claude Code
There was a problem hiding this comment.
Tracked as FE-2558. The PR merged with this thread unresolved, so the finding is carried in Linear rather than lost here. Filed automatically; close the ticket if the finding is no longer true at main.
| BLUEPRINT_PROMOTED, | ||
| BLUEPRINT_PLAIN | ||
| ]) { | ||
| test(`host applies the add of ${name}, and the canvas keeps the node`, async ({ |
There was a problem hiding this comment.
These long 90-second cases have distinct replay, add/ack, evidence, tab-return, and reconciliation phases but no test.step boundaries. Please name the phases with test.step; otherwise CI traces and failure reports do not identify which product phase failed.
There was a problem hiding this comment.
Wrapped each tab-return test's phases (replay, add/judge, before-switch evidence, tab switch, after-return evidence) in test.step, and did the same for the no-tab-switch reconciliation test (add the three nodes, then reconciliation/evidence).
Generated by Claude Code
There was a problem hiding this comment.
Tracked as FE-2558. The PR merged with this thread unresolved, so the finding is carried in Linear rather than lost here. Filed automatically; close the ticket if the finding is no longer true at main.
| const applied = outcomes | ||
| .filter((o) => o.outcome === 'applied') | ||
| .map((o) => o.op_id) | ||
| const skipped = outcomes |
There was a problem hiding this comment.
This does not match the real apply/relay contract. A newly consumed no-op or lww-dropped ID is reported as applied; skipped is for already-consumed duplicate IDs. Those outcomes also update applied-op bookkeeping in the Y.Doc, but this fixture neither advances seq nor emits that delta when a batch contains only these outcomes. Please derive applied/skipped and update emission from the relay contract so the replay tests do not validate against a host state machine production does not use.
There was a problem hiding this comment.
Confirmed against the source and fixed. ApplyOutcome's doc comment in @comfyorg/comfy-multi-player's dist/types.d.ts (~L456-461) says applied counts every op that consumed its op_id THIS call, including LWW-dropped writes and delete-wins no-ops (protocol-level applies), while skipped is idempotency-only — an op_id already applied before this call. dist/applier.js (~L126-151) shows why: every non-duplicate op gets mset into the __applied bookkeeping map inside its own doc.transact() regardless of whether dispatch() returned applied/no-op/lww-dropped, which is exactly what makes even an all-no-op batch produce a real Y.js delta (the bookkeeping write is the mutation).
The package's own result.applied/result.skipped (dist/applier.js ~L183-199) are non-enumerable back-compat accessors explicitly marked "Remove in 0.3", so instead of reading those, applyWire now checks hasAppliedOp(this.doc, op.op_id) (the exported ADR-004 read surface, dist/read.d.ts L133) for each op before the batch runs, and derives applied/skipped from that plus outcomes. This also fixes doc_update emission and seq advancement for a batch containing only fresh no-op/lww-dropped outcomes.
Generated by Claude Code
There was a problem hiding this comment.
Tracked as FE-2558. The PR merged with this thread unresolved, so the finding is carried in Linear rather than lost here. Filed automatically; close the ticket if the finding is no longer true at main.
…ome split - opEnvelope.ts: add isWireOp/parseWireOps, a runtime narrowing of an unknown JSON value to Op (op kind in FROZEN_OPS, non-empty op_id), mirroring the applier's own validateEnvelope gate. - agentFollowerHostSocket.ts: use parseWireOps instead of an `as Op[]` cast on the client doc_ops frame's ops array. - useAgentCrdtFollower.rejectedHumanAdd.test.ts: narrow the sent frame through parseWireOps before reading op_id, instead of an unchecked cast on JSON.parse's result. - agentConversationHostDoc.ts: applyWire now derives applied/skipped the way the real applier's ApplyOutcome contract defines them (dist/types.d.ts): applied counts every op that consumed its op_id this call, including fresh no-op/lww-dropped outcomes; skipped is only an op_id the doc had already applied before this call, checked via the exported hasAppliedOp read surface rather than the package's deprecated result.applied/result.skipped accessors. This also fixes doc_update emission and seq advancement for a batch containing only no-op/lww-dropped outcomes.
…graph LGraph.tabSwitchSnapshot.test.ts mixed a pure litegraph-boundary test with one that imports useNodeDataStore, reportError and litegraphService, tripping three import-x/no-restricted-paths warnings in src/lib (a vendored leaf that must not depend on app layers). Move the store-divergence regression to agentCrdtProjection.serializeDivergence.test.ts, next to the other agent crdt tests where those imports are legal. LGraph.tabSwitchSnapshot.test.ts keeps only the pure litegraph test: a pasted blueprint definition carried by serialize() and reloading as a subgraph node.
workflowTabSwitchKeepsAddedNodes.spec.ts belongs beside the other topbar/workflow-tab specs, not at the top level of tests/. All its imports already use the @e2e/* alias, so the move needs no import changes.
- agentHumanAddBlueprints.ts: the blueprint workflow-JSON factory,
moved out of the spec (fixtures/data, no Playwright imports).
- agentConversationFixture.ts: AgentConversationHarness gains
addNodeOfType, addBlueprint, addNoteThroughSearchBox (a real
user-facing add path: double-click, type, Enter, through the node
search box), installTabSwitchObserver, readNodeLens, attachEvidence,
waitForHumanOps, waitForPendingFrames and switchAwayAndBack, so the
spec keeps scenarios and assertions and the harness owns page/host
operations, per the fixtures/ architectural separation.
- agentHumanAddTabSwitch.spec.ts:
- Drops the host-only 4-case matrix: it was a strict prefix of the
tab-return matrix's own add+judge step, and the per-shape host
verdict (Note/blueprint host -> applied against the real applier)
is already pinned at the unit level in mintPortWiring.test.ts. The
tab-return cases now assert the host applied the add themselves,
so that coverage isn't lost, just not duplicated at the E2E layer.
- Adds a 5th tab-return case, a frontend-only node added through the
node search box (the same user-facing path the non-Agent baseline
in workflowTabSwitchKeepsAddedNodes.spec.ts uses), alongside the
direct-API add cases.
- Wraps each test's phases (replay, add/judge, evidence, tab-return,
reconciliation) in test.step so a failing trace names the phase.
- Calls agentConversation.humanOpOutcomes() directly instead of
through the one-line outcomesFor alias.
- opEnvelope.ts: isWireOp is an internal helper of parseWireOps, not a
separate export (knip flagged it unused otherwise).
…option, op coalescing) + contract e2e (Comfy-Org#18064) <!-- ccr-slack-attribution --> _Requested by **Christian Byrne** · [Slack thread](https://comfy-organization.slack.com/archives/C0BGH348Z0C/p1789670879302689?thread_ts=1789670879.302689&cid=C0BGH348Z0C)_ **Before:** Clearing a saved workflow's canvas, deleting the chat and starting a new one left the agent reasoning about the removed nodes: an N-node delete went out as N serialized single-op batches and most of them were dropped when the chat unbound. The e2e for this flow was an expected-failure repro asserting the wrong contract (a fresh `workflow_id`). The second door to the same symptom, a blank scratch tab opened after a reload adopting a persisted binding for the shared default path, landed separately in Comfy-Org#18077 (document identity gates persisted tab bindings); this PR is rebased onto it and no longer carries a gate of its own for that door. **After:** Operations minted in one tick are coalesced into one batch, so a clear made while bound is one round trip; a `doc_reset` settles pending human batches `undeliverable` at once. The e2e is a plain `@cloud` contract test: the first turn of the new chat posts to a new thread with the same `workflow_id` and an empty draft, every cleared node id reaches the host document as a `delete_node`, and after the ack and re-subscribe the canvas stays empty. Keeps a new agent chat from inheriting a stale canvas through the clear-while-bound door and pins the client side of the contract the backend change relies on. **How.** `docs/adr/CRDT-AUTHORITY-0035-human-canvas-authority-and-draft-reconciliation.md` records the decisions: the human-authored `draft` is the reconciliation primitive (conditional on the backend change that makes a posted human canvas authoritative, merged to the agent release branch and awaiting promotion to main), a new chat keeps the same `workflow_id`, ops are best-effort with same-tick coalescing (`crdt/opCoalescer.ts`, wired in `useAgentCrdtFollower`), `opSender.abortAll()` on `doc_reset`, and the blank-tab door is closed by the document-identity gate of ADR AGENT-BINDING-0035 (Comfy-Org#18077) rather than by a second gate here. The fake host's apply mode (`humanOpsHost: 'apply'`, from Comfy-Org#18078) applies human ops through the real library and answers `doc_ops_result` plus the `doc_update` echo; the contract spec opts into it, and this PR adds only the relay's `ok:false` answer for a `doc_ops` batch aimed at a workflow the host does not serve. Coalescing only narrows the clear-while-bound ordering; edits made while unbound reach the document only through the next turn's draft. **Commits, red to green.** `test:` reworked spec + failing unit tests (red: one op reaches the never-answering host; `abortAll` missing; eight batches instead of one; no settle on `doc_reset`) → `fix:` coalescer, `abortAll` + reset abort (units green; spec still red at the host-document outcome because all eight ops arrive but the fake host does not apply them) → `docs:` ADR → `test:` spec opts into the fake host's apply mode, host answers foreign-workflow `doc_ops` (spec green; `agentConversationReplay`, `agentTabSwitchCatchUp`, `load3dAgentUpdate` stay green) → `test:` flush the coalescer before asserting in the R-73 cross-workflow tests → `refactor:` drop the unused coalescer `pending()`. **Verified locally** on the rebased head: eslint, oxfmt check, `tsc --noEmit`, `vue-tsc` for `browser_tests`, knip, `fallow audit --changed-since origin/main` pass (0 introduced), `adr:check`, Vitest for `src/workbench/extensions/agent` and `src/platform/workflow/persistence` (128 files, 1976 passed, 5 expected fail). Playwright `cloud` project on an earlier head of this branch: the new spec, `agentConversationReplay` (all recordings), `agentTabSwitchCatchUp`, `load3dAgentUpdate`, 19 passed; not re-run after the rebase. The one local failure there was the pre-existing `@screenshot` replay case, identical 25279-pixel diff on the base commit (local Chromium revision drift, Linux CI owns the baseline). **Still open.** Backend promotion to main (until then the saved-target door still fails in production). Follow-ups: withhold `draft` while a bound follower is disconnected or in a gap; surface `undeliverable` human ops to the user (`detach()` drops the queue with no event today); the target-keyed follower of GRAPH-DOCUMENT-0024; a server-side thread delete so the client tombstone can go; two browser tabs on one document are out of scope. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01LZ1S2w986aWaX4XTw2ynmo --------- Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Christian Byrne <cbyrne@comfy.org>
…raph (Comfy-Org#18109) <!-- ccr-slack-attribution --> _Requested by **Christian Byrne** · [Slack thread](https://comfy-organization.slack.com/archives/C0BGH348Z0C/p1789709550015299?thread_ts=1789709550.015299&cid=C0BGH348Z0C)_ ## Summary **Before:** While the Agent panel is bound to a workflow, switching to another workflow tab reuses the same in-memory root graph object for the new tab's nodes before the panel's document binding has caught up. Every node LiteGraph attaches during that window queues a layout `createNode` change, and the mint port did not check which graph a change belonged to, so those changes for the new tab were minted as `add_node` operations and sent into the *previous* tab's still-bound document. **After:** The layout mint port now compares every root-scoped `createNode`/`deleteNode`/`clearGraph` change against the bound document's own root graph id, sourced from a plain accessor derived from the bound workflow's own serialized `activeState.id` (not the live canvas graph, and not a memoized `computed`, so it can never latch a value that goes stale while the same workflow's own graph id rotates in place). A change naming a different graph is dropped and reported once (`agent_crdt_op_for_unbound_graph`) instead of being minted, so a tab switch can no longer leak one workflow's nodes into another workflow's document. Mechanism: `app.ts` reuses one root graph instance across every workflow and reconfigures it (rewriting its stored id) before the panel's `isBoundWorkflowActive` flag has a chance to flip, so the layout store's queued `createNode` notifications for the newly-loaded workflow flush on a microtask while the mint port still believes the previous workflow's document is bound. ## Changes - **What**: - `AgentPanelRoot.vue` derives `boundRootGraphId` as a plain function reading the bound workflow's own serialized `activeState.id` (via `boundOrOpenWorkflowFor(boundWorkflowId)`), converted with `toRootGraphId`, and passes the accessor into `attachMintPortWiring`. Reading it as a plain accessor (not a `computed`) matters because the workflow's `changeTracker` is `markRaw`'d: a `computed` reading through it has no reactive dependency on `activeState.id` and would keep serving a stale id after the same bound workflow's own graph id rotates without a rebind (e.g. `LGraph.clear()`). - `mintPortWiring.ts` forwards the `boundRootGraphId` accessor (typed `RootGraphId | null`) through to the layout mint port. - `layoutMintPort.ts` drops a root-scoped `createNode`/`deleteNode`/`clearGraph` whose `graphId` does not match the bound root graph id, reporting once per action and graph pair per tick via `reportError` (`agent_crdt_op_for_unbound_graph`). Its own mint-dedupe bookkeeping is keyed by each operation's own `graphId` (one id set per root graph, looked up by the value the op itself carries — every real op's `graphId` comes from the live `LGraph`, which always has one) rather than by the `boundRootGraphId()` accessor, so a graph's bucket is reachable only through its own ops: a rebind, a still-in-flight foreign load, or a document whose id has not hydrated yet can never share or evict another graph's bookkeeping. `clearGraph` only forgets a graph's bucket for an intentional (human-confirmed) clear captured via `runIntentionalClear`, never for an incidental clear (a tab switch's in-place `LGraph.clear()`), so a return to a root whose nodes are still in the document does not re-mint them. `deleteNode`'s forget-side bookkeeping only fires for a genuinely same-graph, root-owned delete (`ownerGraphId === graphId`), so a subgraph-interior delete sharing a node id with a root-level node cannot corrupt the root's own entry. ## Review Focus - `src/renderer/core/layout/store/layoutStoreMintDelivery.test.ts` reproduces the exact microtask race against the real layout store (teardown bracket open → configure the new workflow's nodes → teardown bracket close, all synchronous → flush). It failed before this change (an `add_node` for the other workflow's node was minted) and passes after. - `layoutMintPort.test.ts` covers the dedupe-bookkeeping regressions: a root-scoped delete or clear for a foreign graph is dropped and reported without touching the bound graph's own dedupe entry or an active intentional-clear capture; an incidental (non-intentional) same-graph clear — the shape a tab switch's in-place `LGraph.clear()` emits before rebinding — also leaves the bound graph's dedupe entry untouched; two unidentified documents (accessor reporting null for both) no longer share a dedupe bucket; a subgraph-interior delete colliding on node id with a root-level node cannot clear the root's own entry; and a rebind to a different root, including an A→B→A round trip that actually delivers B's own create while B is bound, mints/suppresses correctly on each side. - For `createNode` and `deleteNode` the scope check only applies once a change is already root-scoped (`ownerGraphId === graphId`); `clearGraph` has no owner-graph check because it is always root-scoped by construction (`layoutStore.clearGraph(rootGraphId)` is its only call site) — it is classified directly by its `graphId`. The existing subgraph-interior handling in `reportUnrepresentableInteriorChange` is untouched, as is `serializeForMint`. - `boundRootGraphId()` returning `null` (no workflow bound, or a workflow bound before its `activeState` has hydrated) is treated as permissive rather than restrictive, matching every other existing caller's fixtures: failing closed there would turn an ordinary, transient post-bind race into permanently dropped user edits, which is worse than the narrow foreign-load race it would guard against — a race the dedupe bucket's own per-graphId keying (above) already removes the practical risk from, and one still independently covered by the teardown/session brackets. ### Verification - `pnpm vitest run src/workbench/extensions/agent/crdt src/renderer/core/layout/store/layoutStoreMintDelivery.test.ts`: 727 passed, 2 pre-existing expected fails (unrelated). - `pnpm vitest run src/workbench/extensions/agent`: 1803 passed, 4 pre-existing expected fails (unrelated). - `pnpm vitest run src/workbench/extensions/agent/AgentPanelRoot.test.ts`: included above. - `pnpm typecheck`, `pnpm lint` (targeted to changed files plus the full pre-commit hook), `pnpm format`, `pnpm knip`: clean. Related: Comfy-Org#18078, Comfy-Org#18071 --- _Generated by [Claude Code](https://claude.ai/code/session_01NQVcfKhCKjLstqo3mJS7fc)_ --------- Co-authored-by: Claude <noreply@anthropic.com>
|
|
|
|
…nodes across a workflow tab switch (#18378) <!-- ccr-slack-attribution --> _Requested by **Christian Byrne** · [Slack thread](https://comfy-organization.slack.com/archives/C0BGH348Z0C/p1789709550015299?thread_ts=1789709550.015299&cid=C0BGH348Z0C)_ Backport of #18078 to `core/1.54`. ## Summary **Before:** With the Agent panel bound to a workflow, a user adds a Get/Set node (or any other frontend-registered node such as Note) or a subgraph blueprint to the canvas, switches to another workflow tab, and switches back. The node is gone. Nodes from the server's node catalog (KSampler, CLIPTextEncode) survive the same round trip. **After:** The add of a frontend-only node or blueprint host lands in the document like any catalogued node's, so it survives the tab return. `serializeForMint` (`src/workbench/extensions/agent/crdt/mintPortWiring.ts`) no longer rewrites a frontend-only node's `widgets_values` into the name-keyed object the document applier rejects for uncatalogued classes; it keeps the positional form `serialize()` already produced. ## Changes - **What**: Cherry-picked the squash-merge commit `963d1ee947179cf2124bead716610cf8e6fed52f` (#18078) onto a new branch from `origin/core/1.54`. The fix itself (`mintPortWiring.ts`) and most of the new Vitest regression tests applied cleanly. - **Conflicts resolved faithfully**: `mintPortWiring.test.ts` and `agentConversationHostDoc.ts` conflicted because `core/1.54`'s copies predate this PR's changes; both were resolved by combining `core/1.54`'s existing content with #18078's additions (the `applyClient` → `applyWire` rewrite in `HostDoc`, and the two new Note/blueprint mint cases in the unit test), with no logic invented beyond what #18078 wrote. - **Adapted one pre-existing call site**: `agentAutogrowTabSwitchReconcile.spec.ts` (already on `core/1.54`, untouched by #18078 on `main` because an intervening, unbackported refactor already migrated it there) called `HostDoc.applyClient(...)`, which this PR renames to `applyWire` with a different return shape. Updated that one call site to the new `{ result, update }` shape with identical frame-send behavior; no test assertions changed. - **Omitted three files whose subject module/infra is entirely absent from `core/1.54`** (not part of #18078's own new work — each belongs to a separate, unrelated, unbackported main PR that this branch never got): - `browser_tests/fixtures/agentConversationFixture.ts` and `browser_tests/fixtures/agentFollowerHostSocket.ts` — the "replay recorded agent conversations" Playwright harness these belong to was added by #16764 and #17537, neither backported to `core/1.54`; their dependency chain (`agentConversation.ts`, `agentReplayNodeDefs.ts`, `agentConversationExpectations.ts`, `agentReplayNodeContract.ts`) is also absent. Porting them would mean re-implementing those unrelated PRs wholesale, so they were dropped along with the new spec that exclusively depends on them (`agentHumanAddTabSwitch.spec.ts`), its dedicated unit test (`agentFollowerHostSocket.test.ts`), and the now-orphaned blueprint data file (`agentHumanAddBlueprints.ts`) and the `TabSwitchLens`/`__tabSwitchLens` type addition in `browser_tests/types/globals.d.ts` that only that spec used. - `browser_tests/tests/agent/agentPastedNodeOrphanedOnRetarget.spec.ts` — this pre-existing spec (added by #18071, not on `core/1.54`) doesn't exist to receive #18078's one-line `test.use({ humanOpsHost: 'apply' })` tweak, so there is nothing to apply it to. - `src/workbench/extensions/agent/crdt/agentCrdtProjection.tabReturn.test.ts` — imports `AgentCrdtProjection` from `./agentCrdtProjection`, a class introduced by the unrelated refactor #16908 ("split CRDT follower layers"), which is not on `core/1.54`. The sibling test that doesn't need that module, `agentCrdtProjection.serializeDivergence.test.ts`, was kept. - The still-applicable e2e coverage (`workflowTabSwitchKeepsAddedNodes.spec.ts`, the non-Agent baseline) and Vitest coverage (`mintPortWiring.test.ts`, `agentCrdtProjection.serializeDivergence.test.ts`, `useAgentCrdtFollower.rejectedHumanAdd.test.ts`, `LGraph.tabSwitchSnapshot.test.ts`) were kept as-is. ## Review Focus - The three omissions above are infra/coverage gaps caused by branch divergence, not behavior changes — the production fix in `mintPortWiring.ts` is unconditional and applies exactly as on `main`. - `agentAutogrowTabSwitchReconcile.spec.ts`'s adapted call site is mechanical (same two frames sent in the same order); worth a second look given it wasn't part of #18078's own diff. ## Local verification - `pnpm vitest run src/workbench/extensions/agent`: 1697 passed, 4 expected fail, 4 failed. 3 of the 4 failures (`AgentPanelRoot.test.ts` "Media-card" attach/upload cases) confirmed pre-existing by running the same file against unmodified `origin/core/1.54`. The 4th (`Composer.test.ts` "closes when the caret moves out of the token") passes when run in isolation on this branch and passed on unmodified `origin/core/1.54`; it is a pre-existing flake unrelated to this change (the file is untouched by this backport). - `pnpm lint`, `pnpm typecheck`, `pnpm typecheck:browser`: all clean (ran automatically via the pre-commit hook). - `pnpm format:check` (oxfmt) on changed files: clean. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01NQVcfKhCKjLstqo3mJS7fc --- _Generated by [Claude Code](https://claude.ai/code/session_01NQVcfKhCKjLstqo3mJS7fc)_ Co-authored-by: Claude <noreply@anthropic.com>
… nodes across a workflow tab switch (#18382) <!-- ccr-slack-attribution --> _Requested by **Christian Byrne** · [Slack thread](https://comfy-organization.slack.com/archives/C0BGH348Z0C/p1789709550015299?thread_ts=1789709550.015299&cid=C0BGH348Z0C)_ Backport of #18078 to `cloud/1.54`. ## Summary **Before:** With the Agent panel bound to a workflow, a user adds a Get/Set node (or any other frontend-registered node such as Note) or a subgraph blueprint to the canvas, switches to another workflow tab, and switches back. The node is gone. Nodes from the server's node catalog (KSampler, CLIPTextEncode) survive the same round trip. **After:** The add of a frontend-only node or blueprint host lands in the document like any catalogued node's, so it survives the tab return. `serializeForMint` (`src/workbench/extensions/agent/crdt/mintPortWiring.ts`) no longer rewrites a frontend-only node's `widgets_values` into the name-keyed object the document applier rejects for uncatalogued classes; it keeps the positional form `serialize()` already produced. ## Changes - **What**: Cherry-picked the squash-merge commit `963d1ee947179cf2124bead716610cf8e6fed52f` (#18078) onto a new branch from `origin/cloud/1.54`. The fix itself (`mintPortWiring.ts`) and most of the new Vitest regression tests applied cleanly. - **Conflicts resolved faithfully**: `mintPortWiring.test.ts` and `agentConversationHostDoc.ts` conflicted because `cloud/1.54`'s copies (byte-identical to `core/1.54`'s) predate this PR's changes; both were resolved by combining the branch's existing content with #18078's additions (the `applyClient` → `applyWire` rewrite in `HostDoc`, and the two new Note/blueprint mint cases in the unit test), with no logic invented beyond what #18078 wrote. - **Adapted one pre-existing call site**: `agentAutogrowTabSwitchReconcile.spec.ts` (already on `cloud/1.54`, untouched by #18078 on `main` because an intervening, unbackported refactor already migrated it there) called `HostDoc.applyClient(...)`, which this PR renames to `applyWire` with a different return shape. Updated that one call site to the new `{ result, update }` shape with identical frame-send behavior; no test assertions changed. - **Omitted the same three files as the `core/1.54` backport (#18378), whose subject module/infra is checked and confirmed independently absent from `cloud/1.54`** (not part of #18078's own new work — each belongs to a separate, unrelated, unbackported main PR that this branch never got either): - `browser_tests/fixtures/agentConversationFixture.ts` and `browser_tests/fixtures/agentFollowerHostSocket.ts` — the "replay recorded agent conversations" Playwright harness these belong to was added by #16764 and #17537, neither backported to `cloud/1.54`; their dependency chain (`agentConversation.ts`, `agentReplayNodeDefs.ts`, `agentConversationExpectations.ts`, `agentReplayNodeContract.ts`) is also absent. Porting them would mean re-implementing those unrelated PRs wholesale, so they were dropped along with the new spec that exclusively depends on them (`agentHumanAddTabSwitch.spec.ts`), its dedicated unit test (`agentFollowerHostSocket.test.ts`), the now-orphaned blueprint data file (`agentHumanAddBlueprints.ts`), and the `TabSwitchLens`/`__tabSwitchLens` type addition in `browser_tests/types/globals.d.ts` that only that spec used. - `browser_tests/tests/agent/agentPastedNodeOrphanedOnRetarget.spec.ts` — this pre-existing spec (added by #18071, not on `cloud/1.54`) doesn't exist to receive #18078's one-line `test.use({ humanOpsHost: 'apply' })` tweak, so there is nothing to apply it to. - `src/workbench/extensions/agent/crdt/agentCrdtProjection.tabReturn.test.ts` — imports `AgentCrdtProjection` from `./agentCrdtProjection`, a class introduced by the unrelated refactor #16908 ("split CRDT follower layers"), which is not on `cloud/1.54`. The sibling test that doesn't need that module, `agentCrdtProjection.serializeDivergence.test.ts`, was kept. - The still-applicable e2e coverage (`workflowTabSwitchKeepsAddedNodes.spec.ts`, the non-Agent baseline) and Vitest coverage (`mintPortWiring.test.ts`, `agentCrdtProjection.serializeDivergence.test.ts`, `useAgentCrdtFollower.rejectedHumanAdd.test.ts`, `LGraph.tabSwitchSnapshot.test.ts`) were kept as-is. ## Review Focus - The three omissions above are infra/coverage gaps caused by branch divergence, not behavior changes — the production fix in `mintPortWiring.ts` is unconditional and applies exactly as on `main`. - `agentAutogrowTabSwitchReconcile.spec.ts`'s adapted call site is mechanical (same two frames sent in the same order); worth a second look given it wasn't part of #18078's own diff. ## Local verification - `pnpm vitest run src/workbench/extensions/agent`: 1732 passed, 4 expected fail, 3 failed. All 3 failures (`AgentPanelRoot.test.ts` "Media-card" attach/upload cases) confirmed pre-existing by running the same file against unmodified `origin/cloud/1.54`, where they fail identically. - `pnpm lint`, `pnpm typecheck`, `pnpm typecheck:browser`: all clean (ran automatically via the pre-commit hook). - `pnpm format:check` (oxfmt) on changed files: clean. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01NQVcfKhCKjLstqo3mJS7fc --- _Generated by [Claude Code](https://claude.ai/code/session_01NQVcfKhCKjLstqo3mJS7fc)_ Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com> Co-authored-by: Claude <noreply@anthropic.com>
|
|
|
|
|
|
Requested by Christian Byrne · Slack thread
Before: With the Agent panel bound to a workflow, a user adds a Get/Set node (or any other frontend-registered node such as Note) or a subgraph blueprint to the canvas, switches to another workflow tab, and switches back. The node is gone. Nodes from the server's node catalog (KSampler, CLIPTextEncode) survive the same round trip. Nothing in the UI says why.
After: The add of a frontend-only node or blueprint host lands in the document like any catalogued node's, so it survives the tab return. The e2e repros that pinned the symptom are now regression tests; the unit repros for the second half of the mechanism (a never-landed add is still swept by the first reconcile after a rebind, and a rejected human op is still swallowed) stay
it.failsand are the acceptance tests for the follow-up work in the plan comment.Summary
serializeForMint(src/workbench/extensions/agent/crdt/mintPortWiring.ts) rewrote every widget-bearing node'swidgets_valuesinto a name-keyed object before mintingadd_node. The document applier (@comfyorg/comfy-multi-player) accepts name-keyed widgets only for classes in its pinned catalog (the server's node definitions) and rejects everything else withuncatalogued_widget_write, while it stores a positional array for such a class opaquely. Frontend-only classes (isVirtualNode: Note, MarkdownNote, PrimitiveNode, Get/Set nodes from node packs, subgraph blueprint hosts) now keep the positional formserialize()already produced.Mechanism the tests pin down
Two halves, both required for the symptom:
ok: falseresult as a dev event; no revert, no toast. The node stayed on the canvas but was permanently absent from the document.removeMissingagainst the document's node set, deleting the store record, and the orphan sweep removes the live node. The same reconcile also arms without any tab switch: the echo of an accepted human add is rejected locally as "already registered", and the next frame runs it.Tests
Vitest:
src/workbench/extensions/agent/crdt/mintPortWiring.test.ts:add_nodefor a Note and for a blueprint host with a promoted widget keeps positionalwidgets_values, and the real applier applies the minted op against a catalog that lacks the class. The existing name-keyed case (LoadImage, a catalogued class) still holds.src/workbench/extensions/agent/crdt/agentCrdtProjection.tabReturn.test.ts:it.failsa node whoseadd_nodenever reached the doc survives a rebind + catch-up;it.failsthe same without a tab return after an accepted add's echo; two passing controls showing a node that did land in the doc survives, and that the echo of the page's own add is rejected locally ("already registered").src/workbench/extensions/agent/crdt/useAgentCrdtFollower.rejectedHumanAdd.test.ts:it.failsthe follower surfaces a rejected humanadd_nodethroughreportErrorinstead of swallowing it.src/workbench/extensions/agent/crdt/agentCrdtProjection.serializeDivergence.test.ts: the app-integrated lens —serialize()still emits a live node whose store record is gone, and reports the divergence. Moved out of the vendored litegraph layer so its store/telemetry imports don't cross that boundary.src/lib/litegraph/src/LGraph.tabSwitchSnapshot.test.ts: the pure litegraph lens that remains — a just-pasted blueprint definition is carried byserialize()and reloads as a subgraph node.Playwright,
browser_tests/tests/agent/agentHumanAddTabSwitch.spec.ts(cloud project, agent harness with the real@comfyorg/comfy-multi-playerapplier as the host), six cases total:serialize(), trackeractiveState) and an in-page observer recording the node set right after the canvas rebuild and every removal since are attached; each case also asserts the host applied the add before the switch runs.Playwright,
browser_tests/tests/topbar/workflowTabSwitchKeepsAddedNodes.spec.ts(chromium project, no Agent): KSampler, Note and the devtools "test blueprint" added through the search box survive a tab switch. Baseline for the Agent-bound spec.Fixture changes:
agentConversationHostDoc.tsgainsapplyWire(judge a client-minted batch with the real applier and answerdoc_ops_result+doc_update);agentFollowerHostSocket.ts(shared with the Load3D agent fixture) records every clientdoc_*frame and applies or holds humandoc_ops(humanOpsHost, defaulthold, the previous behaviour);agentWireFrame.tsholds the wire-batch parser (parseWireOps, validating a batch as one unit rather than filtering individual members) and thedoc_ops-specific non-empty/unique-id gate (isValidDocOpsBatch) — both test-only, moved out of the productionopEnvelope.ts, which no consumer outside fixtures/tests used;agentConversationFixture.tsowns the blueprint/search-box add drivers, the tab-switch sequence, node lenses and evidence capture;browser_tests/fixtures/data/agent/agentHumanAddBlueprints.tsholds the blueprint workflow-JSON factory.browser_tests/types/globals.d.tstypes the in-page observer.Review Focus
isVirtualNodeas the client-side proxy for "absent from the pinned catalog". The node-def store cannot tell: system defs (Note, Reroute, PrimitiveNode) and blueprint defs are injected withpython_module: 'nodes'. A custom node installed after the document was minted is still in the store and not in the catalog; that case still needs the follow-up in the plan comment (surface rejected ops).AgentFollowerHostSocketrejects a malformed, empty, or duplicate-iddoc_opsframe asinvalid_framebefore it ever reachesHostDoc— mirroring the real relay's own pre-application gate on the frame as a whole. A frame that clears that gate is judged byapplyWire, which projects the real applier's per-opApplyOutcomes into the relay'sdoc_ops_resultsplit (applied vs skipped, first rejection underfailed); the follower's result handling reads onlyok,applied,skippedandfailed.op_id.Local verification
src/workbench/extensions/agent/crdt/directory plus the LGraph lens: 730 passed, 5 expected fail (3 of them this PR's acceptance tests; the other 2 are unrelated repros from test(agent): failing repro for agent connect skipping slot type check #18107 and fix(agent): land agent-inserted template nodes near existing canvas content #18084, picked up by the merge withmain).pnpm typecheck,pnpm typecheck:browser, eslint on the changed files,pnpm format:check,pnpm knip,pnpm fallow audit --changed-since origin/main: clean.🤖 Generated with Claude Code
https://claude.ai/code/session_01NQVcfKhCKjLstqo3mJS7fc