Repository navigation
test(agent): repro canvas blanking for a run turn's duration after a mid-turn doc reset - #18179
Conversation
|
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 |
🎭 Playwright: ✅ 2066 passed, 0 failed📊 Browser Reports
🌾 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 b36caf43edd3acf99b3707f6db41f428a0349e1a
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 📦 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": "2026-09-20T16:40:32.669Z",
"gitSha": "4ac27a49c9ac7ee96fd037db37ca3b3be531aea7",
"branch": "test/agent-run-turn-canvas-blank-repro",
"measurements": [
{
"name": "canvas-idle",
"durationMs": 2033.6980000000153,
"styleRecalcs": 7,
"styleRecalcDurationMs": 8.063,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 484.864,
"heapDeltaBytes": 14430972,
"heapUsedBytes": 76507724,
"domNodes": -264,
"jsHeapTotalBytes": 4710400,
"scriptDurationMs": 7.366999999999998,
"eventListeners": -152,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-idle",
"durationMs": 2005.0889999999981,
"styleRecalcs": 9,
"styleRecalcDurationMs": 7.6819999999999995,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 455.02,
"heapDeltaBytes": -70020,
"heapUsedBytes": 61949064,
"domNodes": -264,
"jsHeapTotalBytes": 4710400,
"scriptDurationMs": 9.312000000000001,
"eventListeners": -152,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-mouse-sweep",
"durationMs": 1781.9670000000087,
"styleRecalcs": 76,
"styleRecalcDurationMs": 25.506,
"layouts": 14,
"layoutDurationMs": 3.6380000000000003,
"taskDurationMs": 810.4549999999999,
"heapDeltaBytes": 16703848,
"heapUsedBytes": 78776516,
"domNodes": -262,
"jsHeapTotalBytes": 5234688,
"scriptDurationMs": 157.97,
"eventListeners": -152,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-mouse-sweep",
"durationMs": 1843.8619999999446,
"styleRecalcs": 76,
"styleRecalcDurationMs": 30.097,
"layouts": 14,
"layoutDurationMs": 3.985,
"taskDurationMs": 860.5480000000001,
"heapDeltaBytes": 3096748,
"heapUsedBytes": 64827496,
"domNodes": -265,
"jsHeapTotalBytes": 5758976,
"scriptDurationMs": 165.488,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "canvas-zoom-sweep",
"durationMs": 1737.7819999999815,
"styleRecalcs": 32,
"styleRecalcDurationMs": 14.985,
"layouts": 6,
"layoutDurationMs": 1.3470000000000002,
"taskDurationMs": 324.46299999999997,
"heapDeltaBytes": 2175048,
"heapUsedBytes": 64102660,
"domNodes": 75,
"jsHeapTotalBytes": 5242880,
"scriptDurationMs": 10.077,
"eventListeners": 15,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "canvas-zoom-sweep",
"durationMs": 1723.4930000000759,
"styleRecalcs": 31,
"styleRecalcDurationMs": 15.046999999999999,
"layouts": 6,
"layoutDurationMs": 1.277,
"taskDurationMs": 325.59299999999996,
"heapDeltaBytes": 2185260,
"heapUsedBytes": 64057408,
"domNodes": 75,
"jsHeapTotalBytes": 4980736,
"scriptDurationMs": 10.609,
"eventListeners": 15,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "dom-widget-clipping",
"durationMs": 608.8040000000206,
"styleRecalcs": 11,
"styleRecalcDurationMs": 7.410000000000002,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 426.7460000000001,
"heapDeltaBytes": 18169916,
"heapUsedBytes": 80243524,
"domNodes": 18,
"jsHeapTotalBytes": 5242880,
"scriptDurationMs": 161.54,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "dom-widget-clipping",
"durationMs": 643.3769999999868,
"styleRecalcs": 11,
"styleRecalcDurationMs": 8.187999999999999,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 450.604,
"heapDeltaBytes": 18594652,
"heapUsedBytes": 80251364,
"domNodes": 18,
"jsHeapTotalBytes": 5505024,
"scriptDurationMs": 168.898,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "large-graph-idle",
"durationMs": 2027.1339999999896,
"styleRecalcs": 5,
"styleRecalcDurationMs": 3.5509999999999984,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 524.8439999999999,
"heapDeltaBytes": 3170420,
"heapUsedBytes": 83474724,
"domNodes": -249,
"jsHeapTotalBytes": -794624,
"scriptDurationMs": 30.485,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "large-graph-idle",
"durationMs": 2033.6450000000923,
"styleRecalcs": 4,
"styleRecalcDurationMs": 3.8639999999999994,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 538.8009999999999,
"heapDeltaBytes": 2576564,
"heapUsedBytes": 83781000,
"domNodes": -253,
"jsHeapTotalBytes": -1019904,
"scriptDurationMs": 32.81199999999999,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "large-graph-pan",
"durationMs": 2242.8350000000137,
"styleRecalcs": 64,
"styleRecalcDurationMs": 8.748999999999997,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1585.422,
"heapDeltaBytes": 15703908,
"heapUsedBytes": 79850992,
"domNodes": -255,
"jsHeapTotalBytes": 1011712,
"scriptDurationMs": 831.323,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "large-graph-pan",
"durationMs": 2292.2509999999647,
"styleRecalcs": 66,
"styleRecalcDurationMs": 10.891000000000002,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1708.94,
"heapDeltaBytes": 1734472,
"heapUsedBytes": 83879312,
"domNodes": -255,
"jsHeapTotalBytes": -8192,
"scriptDurationMs": 862.6429999999999,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "large-graph-zoom",
"durationMs": 3401.929999999993,
"styleRecalcs": 61,
"styleRecalcDurationMs": 9.599000000000004,
"layouts": 60,
"layoutDurationMs": 12.108,
"taskDurationMs": 1464.6659999999997,
"heapDeltaBytes": 8092008,
"heapUsedBytes": 74075140,
"domNodes": 2,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 762.806,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.669999999999952,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "large-graph-zoom",
"durationMs": 3570.0789999999643,
"styleRecalcs": 60,
"styleRecalcDurationMs": 9.557000000000002,
"layouts": 60,
"layoutDurationMs": 12.620999999999999,
"taskDurationMs": 1688.378,
"heapDeltaBytes": 9845368,
"heapUsedBytes": 73612320,
"domNodes": -263,
"jsHeapTotalBytes": 0,
"scriptDurationMs": 787.032,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333338,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "legacy-node-drag",
"durationMs": 3217.0249999999783,
"styleRecalcs": 45,
"styleRecalcDurationMs": 10.023000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1991.491,
"heapDeltaBytes": -4517600,
"heapUsedBytes": 86445904,
"domNodes": -258,
"jsHeapTotalBytes": 2019328,
"scriptDurationMs": 1040.514,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "legacy-node-drag",
"durationMs": 3112.6629999999977,
"styleRecalcs": 40,
"styleRecalcDurationMs": 4.615000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 1927.975,
"heapDeltaBytes": 22019560,
"heapUsedBytes": 87741624,
"domNodes": 0,
"jsHeapTotalBytes": 417792,
"scriptDurationMs": 983.0889999999998,
"eventListeners": 180,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "minimap-idle",
"durationMs": 2053.6410000000274,
"styleRecalcs": 6,
"styleRecalcDurationMs": 5.614000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 556.5879999999999,
"heapDeltaBytes": -5431784,
"heapUsedBytes": 82894732,
"domNodes": -252,
"jsHeapTotalBytes": -233472,
"scriptDurationMs": 31.426999999999996,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "minimap-idle",
"durationMs": 2022.5960000000214,
"styleRecalcs": 6,
"styleRecalcDurationMs": 7.116999999999998,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 547.032,
"heapDeltaBytes": -3652104,
"heapUsedBytes": 84324792,
"domNodes": -248,
"jsHeapTotalBytes": -495616,
"scriptDurationMs": 28.483999999999995,
"eventListeners": -152,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333335,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "subgraph-dom-widget-clipping",
"durationMs": 563.808999999992,
"styleRecalcs": 48,
"styleRecalcDurationMs": 10.791,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 393.38699999999994,
"heapDeltaBytes": 14551444,
"heapUsedBytes": 77885436,
"domNodes": 22,
"jsHeapTotalBytes": 5767168,
"scriptDurationMs": 141.456,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "subgraph-dom-widget-clipping",
"durationMs": 574.561000000017,
"styleRecalcs": 47,
"styleRecalcDurationMs": 9.826,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 389.6820000000001,
"heapDeltaBytes": 13665676,
"heapUsedBytes": 76974604,
"domNodes": 20,
"jsHeapTotalBytes": 5767168,
"scriptDurationMs": 139.202,
"eventListeners": 6,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "subgraph-idle",
"durationMs": 2022.1230000000219,
"styleRecalcs": 11,
"styleRecalcDurationMs": 9.399,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 433.421,
"heapDeltaBytes": -8253956,
"heapUsedBytes": 55062024,
"domNodes": -276,
"jsHeapTotalBytes": 4972544,
"scriptDurationMs": 6.739999999999999,
"eventListeners": -198,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "subgraph-idle",
"durationMs": 2022.5580000000036,
"styleRecalcs": 10,
"styleRecalcDurationMs": 9.507000000000001,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 437.799,
"heapDeltaBytes": 10918496,
"heapUsedBytes": 74429252,
"domNodes": -276,
"jsHeapTotalBytes": 4972544,
"scriptDurationMs": 7.187000000000002,
"eventListeners": -198,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "subgraph-mouse-sweep",
"durationMs": 1706.912999999986,
"styleRecalcs": 76,
"styleRecalcDurationMs": 24.353999999999996,
"layouts": 18,
"layoutDurationMs": 4.961,
"taskDurationMs": 737.832,
"heapDeltaBytes": 18622968,
"heapUsedBytes": 81947108,
"domNodes": -278,
"jsHeapTotalBytes": 5496832,
"scriptDurationMs": 103.64899999999999,
"eventListeners": -198,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.699999999999818
},
{
"name": "subgraph-mouse-sweep",
"durationMs": 1738.4359999999788,
"styleRecalcs": 78,
"styleRecalcDurationMs": 27.442,
"layouts": 18,
"layoutDurationMs": 4.857,
"taskDurationMs": 783.601,
"heapDeltaBytes": -7256156,
"heapUsedBytes": 56174188,
"domNodes": -275,
"jsHeapTotalBytes": 4972544,
"scriptDurationMs": 106.08500000000001,
"eventListeners": -198,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.800000000000182
},
{
"name": "subgraph-transition-enter",
"durationMs": 1418.5580000000755,
"styleRecalcs": 16,
"styleRecalcDurationMs": 30.12,
"layouts": 5,
"layoutDurationMs": 13.517000000000001,
"taskDurationMs": 957.1650000000001,
"heapDeltaBytes": -6805224,
"heapUsedBytes": 86490056,
"domNodes": 13587,
"jsHeapTotalBytes": 10747904,
"scriptDurationMs": 31.596999999999994,
"eventListeners": 2383,
"totalBlockingTimeMs": 144,
"frameDurationMs": 16.66333333333332,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "viewport-pan-sweep",
"durationMs": 8350.555999999982,
"styleRecalcs": 247,
"styleRecalcDurationMs": 32.739,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 5746.777,
"heapDeltaBytes": -3189896,
"heapUsedBytes": 77964028,
"domNodes": -251,
"jsHeapTotalBytes": 516096,
"scriptDurationMs": 2819.191,
"eventListeners": -174,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "viewport-pan-sweep",
"durationMs": 8209.492000000067,
"styleRecalcs": 247,
"styleRecalcDurationMs": 34.132999999999996,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 5056.185,
"heapDeltaBytes": -4801520,
"heapUsedBytes": 76059916,
"domNodes": -248,
"jsHeapTotalBytes": 221184,
"scriptDurationMs": 2699.723,
"eventListeners": -152,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-large-graph-idle",
"durationMs": 16615.245000000017,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 15931.274,
"heapDeltaBytes": -32844312,
"heapUsedBytes": 183244928,
"domNodes": -8061,
"jsHeapTotalBytes": -8167424,
"scriptDurationMs": 241.046,
"eventListeners": -14928,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.80000000000291
},
{
"name": "vue-large-graph-idle",
"durationMs": 16507.87600000001,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 15700.444999999998,
"heapDeltaBytes": -58054756,
"heapUsedBytes": 177427040,
"domNodes": -8061,
"jsHeapTotalBytes": -12001280,
"scriptDurationMs": 267.344,
"eventListeners": -14928,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-large-graph-pan",
"durationMs": 20336.553000000036,
"styleRecalcs": 165,
"styleRecalcDurationMs": 17.380000000000006,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 20048.507999999998,
"heapDeltaBytes": -54828396,
"heapUsedBytes": 167706848,
"domNodes": -8061,
"jsHeapTotalBytes": -16109568,
"scriptDurationMs": 855.07,
"eventListeners": -14928,
"totalBlockingTimeMs": 15,
"frameDurationMs": 17.776666666666763,
"p95FrameDurationMs": 16.80000000000291
},
{
"name": "vue-large-graph-pan",
"durationMs": 19952.51500000006,
"styleRecalcs": 169,
"styleRecalcDurationMs": 16.368999999999996,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 19560.443999999996,
"heapDeltaBytes": -24031204,
"heapUsedBytes": 180444712,
"domNodes": -8061,
"jsHeapTotalBytes": -12652544,
"scriptDurationMs": 859.624,
"eventListeners": -14928,
"totalBlockingTimeMs": 45,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "vue-node-resize-workload",
"durationMs": 6123.466000000008,
"styleRecalcs": 20,
"styleRecalcDurationMs": 86.32300000000004,
"layouts": 20,
"layoutDurationMs": 25.06800000000001,
"taskDurationMs": 6032.581,
"heapDeltaBytes": -35434348,
"heapUsedBytes": 180573256,
"domNodes": -8061,
"jsHeapTotalBytes": 18046976,
"scriptDurationMs": 115.56500000000001,
"eventListeners": -14928,
"totalBlockingTimeMs": 454,
"frameDurationMs": 17.779999999999927,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-node-resize-workload",
"durationMs": 5907.293000000095,
"styleRecalcs": 20,
"styleRecalcDurationMs": 76.815,
"layouts": 20,
"layoutDurationMs": 22.131999999999998,
"taskDurationMs": 5806.598999999999,
"heapDeltaBytes": -40909700,
"heapUsedBytes": 175511596,
"domNodes": -8061,
"jsHeapTotalBytes": 12558336,
"scriptDurationMs": 101.49900000000001,
"eventListeners": -14928,
"totalBlockingTimeMs": 310,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.700000000000728
},
{
"name": "vue-zoom-out-idle",
"durationMs": 8314.76200000003,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 8153.628,
"heapDeltaBytes": -19603468,
"heapUsedBytes": 164756544,
"domNodes": 0,
"jsHeapTotalBytes": -10223616,
"scriptDurationMs": 125.14599999999999,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.776666666666642,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "vue-zoom-out-idle",
"durationMs": 8277.389999999968,
"styleRecalcs": 0,
"styleRecalcDurationMs": 0,
"layouts": 0,
"layoutDurationMs": 0,
"taskDurationMs": 7937.371000000001,
"heapDeltaBytes": -9676672,
"heapUsedBytes": 175276116,
"domNodes": 0,
"jsHeapTotalBytes": -14942208,
"scriptDurationMs": 143.73699999999994,
"eventListeners": 0,
"totalBlockingTimeMs": 0,
"frameDurationMs": 17.779999999999927,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "workflow-execution",
"durationMs": 482.407999999964,
"styleRecalcs": 18,
"styleRecalcDurationMs": 25.182999999999996,
"layouts": 4,
"layoutDurationMs": 1.9709999999999996,
"taskDurationMs": 139.76999999999998,
"heapDeltaBytes": 5120664,
"heapUsedBytes": 66759112,
"domNodes": 149,
"jsHeapTotalBytes": 262144,
"scriptDurationMs": 13.774999999999999,
"eventListeners": 81,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.666666666666668,
"p95FrameDurationMs": 16.799999999999272
},
{
"name": "workflow-execution",
"durationMs": 453.44199999999546,
"styleRecalcs": 13,
"styleRecalcDurationMs": 17.932999999999996,
"layouts": 2,
"layoutDurationMs": 0.6169999999999999,
"taskDurationMs": 102.35,
"heapDeltaBytes": 4858944,
"heapUsedBytes": 66599428,
"domNodes": 121,
"jsHeapTotalBytes": 524288,
"scriptDurationMs": 9.450999999999999,
"eventListeners": 81,
"totalBlockingTimeMs": 0,
"frameDurationMs": 16.66999999999998,
"p95FrameDurationMs": 16.699999999999818
}
]
}🔬 E2E Coverage
Top 10 files by uncovered lines
🎨 Storybook: ✅ Built — View Storybook |
…mid-turn doc reset (PM-1355-shaped) Intentionally-failing repro for the nightly bug report (Jo Zhang, 2026-09-20): asking the in-app agent to run the canvas workflow blanks the canvas the moment the agent starts working, for the whole run, with every node reappearing at once when the run finishes. No fix is attempted here. Two independently-confirmed frontend mechanisms combine to produce exactly that timeline: - A `doc_reset`/`follower_replaced` frame synchronously sweeps every live materialized node off the canvas (`AgentCrdtProjection.clearForReset`) -- the only path that actively removes already-rendered nodes. - The resubscribe that follows can be acknowledged (`doc_subscribed:ok`) without its catch-up `doc_update`. This exact anomaly is visible in the bug report's own attached debug log, on a different (later, smaller) workflow in the same session, self-healing only via the passive 30-second `STALE_AFTER_MS` stale-probe heartbeat (`agentCrdtDocLifecycle.ts`) -- there is no faster active recovery. Together, whatever triggers the reset, the canvas can sit blank for up to 30 real seconds with zero active recovery -- comfortably spanning one generation run (this report's own run + wait_for_job took 15386ms). Step 0 (existing PR check): searched open PRs across ComfyUI_frontend/cloud for "canvas blank"/"run"/"agent run"/"crdt-authority" during this run. #18099 ("repro schema-gate latch leaving canvas at zero nodes") is the closest sibling but covers a PERMANENT latch (until an explicit doc_reset) triggered by an unreadable schema_version -- not this transient, self-healing, run-duration-scoped blank. #18064 (new-chat stale canvas) is a different trigger (thread binding, not a mid-turn reset). No existing PR covers this. Open half of the RCA, stated plainly: the backend condition that would emit a doc_reset for a plain, edit-free "run the workflow" turn was NOT found. `run` / `wait_for_job` / `get_output` (cloud/services/agent/internal/loop/ tools.go) make no calls into the CRDT doc store at all, and the turn-start reconciliation that can emit a reset (`ensureDoc`/`prepareWorkflow` in cloud/services/agent/internal/loop/crdt.go) is heavily unit-tested against firing on an unmodified canvas. Which backend path actually emitted the reset in production is unconfirmed; this repro pins what the user sees once one is emitted, for whatever reason, and gives backend engineers a concrete signature (a `doc_reset` frame, or a `doc_subscribed:ok` with no following `doc_update`) to search production logs for around the incident. Verification: - pnpm exec eslint / oxlint / oxfmt --check on the new file -- clean - pnpm exec vue-tsc --project browser_tests/tsconfig.json --noEmit -- clean - pnpm exec playwright test --list -- collects the spec correctly under the cloud project with the right tags - pnpm exec vitest run src/workbench/extensions/agent/crdt/ useAgentCrdtFollower.test.ts -- 61/61 passed, no regressions - Known limitation: this sandbox's egress policy blocks cdn.playwright.dev (403), so the pinned Chromium build could not be downloaded and the spec could not be executed against a real browser here (same limitation noted on sibling repro PRs #18099/#18096 in this program). CI's Playwright run on this PR is the first real execution; treat its result, not this description, as the authoritative pass/fail signal. Related Slack thread: #comfy-agent-user-feedback, Jo Zhang's nightly report, 2026-09-20 03:24 UTC.
293c120 to
0b2202e
Compare
Playwright videos for new testsDownload the recorded videos and HTML report Recorded projects: cloud Recorded test titles:
|
Codecov Report✅ All modified and coverable lines are covered by tests. @@ Coverage Diff @@
## main #18179 +/- ##
==========================================
+ Coverage 82.86% 83.93% +1.07%
==========================================
Files 2950 2260 -690
Lines 145272 128628 -16644
Branches 37414 35055 -2359
==========================================
- Hits 120373 107963 -12410
+ Misses 23630 20375 -3255
+ Partials 1269 290 -979
Flags with carried forward coverage won't be shown. Click here to find out more.
... and 753 files with indirect coverage changes 🚀 New features to boost your workflow:
|
agentCanvasBlanksDuringRunTurn.spec.ts followed its test.fail()-guarded assertion with more code (a further clock fast-forward and two recovery assertions) that could never run: if the guarded assertion throws, as intended, execution stops there; if it doesn't, Playwright already flags "expected fail but passed" before reaching the rest. Extract the shared arrange (drive the run turn to the blanked-canvas point) into a helper, keep test.fail() as the last thing the repro test does, and move the recovery check into its own test in the same describe, matching the pattern in agentZeroNodesOnUnreadableSchema.spec.ts and agentTurnSurvivesSocketDrop.spec.ts.
|
@/tmp/claude-0/comment_body.md |
…ng the full stale budget
PM-1355 RCA: a confirmed subscribe (`doc_subscribed: {ok: true}`) does not
guarantee its catch-up `doc_update` arrives - the two are separate frames,
and the second can be silently dropped. Until now the only recovery was the
passive 30s stale-probe heartbeat, leaving the canvas blank for up to
STALE_AFTER_MS after a mid-turn doc_reset.
AgentCrdtDocLifecycle now arms the recency probe with a short
SUBSCRIBE_CATCHUP_GRACE_MS (2s) delay right after a subscribe is confirmed,
instead of the full budget. A resubscribe fired from this shorter window is a
no-op on a channel that had nothing to catch up on, and the active repair on
one whose catch-up went missing. Any doc-scoped frame that lands first
(catch-up or otherwise) still re-arms the probe at the full STALE_AFTER_MS
budget as before.
Updates useAgentCrdtFollower.test.ts for the new behavior, including the
previously passive-only "a confirmed subscribe clears the retry timer" case.
…fast recovery Adds a third E2E test modeling a ONE-OFF (transient) dropped subscribe catch-up after a mid-turn doc_reset, rather than the permanent drop the first two tests use. It asserts the canvas recovers within SUBSCRIBE_CATCHUP_GRACE_MS (2s), well inside the run's own 15,386ms duration -- the concrete before/after Christian asked for. Refactors the shared arrange step (driveThroughDocReset) to take a per-resubscribe catch-up-drop policy instead of a single permanent boolean, so the existing two tests and the new one can share setup without duplicating ~150 lines of route/socket wiring.
…plexity Splits the mock /ws onMessage handler's frame parsing and field validation into two named helpers (parseDocSubscribeData, parseDocSubscribeFields), leaving the handler's own branching to the post-parse reset/catch-up decisions. Pure refactor, resolves the Fallow complexity gate finding on the arrow function; no behavior change.
Resolves the armStaleProbe() conflict between this branch's catch-up grace probe (SUBSCRIBE_CATCHUP_GRACE_MS, fired after every confirmed subscribe) and main's ackTimer/gaveUp guards: the grace probe's timer callback now checks gaveUp and ackTimer exactly like the recurring stale probe does, so an early catch-up retry can't fire a spurious resubscribe on a channel that already gave up or has an ack in flight. Updates the "a confirm disarms the ack timer and hands the channel to the recency probe" regression test (new on main) to account for the grace probe firing before the full recency budget, and to isolate the probe-handoff behavior from the ack-timeout retry cascade (which has its own dedicated tests) by using a bare resubscribe double instead of the auto-chaining one.
agentCanvasBlanksDuringRunTurn.spec.ts asserted the seed's two nodes were visible right after selecting the workflow as the agent's target, before any message was sent. boundWorkflowId (useAgentSession.ts) only changes via bindWorkflow, which fires from the agent_active_tab push handled in AgentPanelRoot.vue's onAgentActiveTab -- so the doc never subscribes, and the baseline assertion times out waiting for content that nothing has requested yet. Move the baseline check after the agent_active_tab frame that actually binds the follower.
christian-byrne
left a comment
There was a problem hiding this comment.
The fast retry repeats indefinitely on acknowledged subscriptions without content. The incident tracker reference is also incorrect.
Full context for agent readers
Reviewed the six-file diff at this head and all 18 messages in the original Slack thread. The isolated lifecycle suite passes its 13 existing cases, but a new acknowledged-retry case makes 15 resubscriptions by 31,999ms instead of one fast probe followed by the documented 30-second cadence. No browser rerun or original-incident root-cause proof is claimed. The completion lane retains the real-fix and acceptance follow-through under canvas disappearance and run-duration acceptance, including release qualification; neither is finished by a partial merge.
Glossary: acknowledgment confirms the subscription, not document delivery; catch-up is the missing document content.
…episode onSubscribeConfirmed re-armed the fast SUBSCRIBE_CATCHUP_GRACE_MS probe on every confirmed subscribe, including one triggered by the probe's own resubscribe. A backend that keeps acking a resubscribe without ever minting content made this loop at the fast interval instead of falling back to the 30s heartbeat. Track whether the current gap episode already spent its one fast shot and use the full budget for the rest of the episode, resetting on real content or a fresh bind/reconnect. Also corrects this branch's earlier PM-1355 tracker references (an unrelated report) to the actual pair: PM-1405 for the underlying fix/mechanism, PM-1406 for the specific canvas-blank-during-run-turn repro/acceptance test.
christian-byrne
left a comment
There was a problem hiding this comment.
Bounded retry verified. Transient-drop recovery verified. Original incident remains open.
Full context for agent readers
Reviewed the corrected head. The acknowledgment-only path now uses one two-second probe, then the thirty-second interval until content or a lifecycle restart. The 81 lifecycle/follower tests pass. The isolated Cloud browser suite passes all three cases, including its explicitly expected failure for permanently withheld content. I inspected the transient-drop recording: both nodes disappear and reappear. A rebuilt control restoring the parent's thirty-second-only confirmation behavior fails the fast-recovery assertion at line 549; restoring and rebuilding this head passes again. This control restores the changed behavior, not the entire parent tree.
Both prior review findings are addressed. This is a scoped recovery mitigation, not a root-cause fix for the reported run-time reset. The source Slack debug attachments show an acknowledgment with no sequence followed 30,001ms later by a stale probe and later document content; they do not establish the initial reset cause. The frontend still deliberately clears the canvas on an explicit document reset.
The canvas-disappearance fix and original-run acceptance stay open under Christian's completion lane. Required hosted checks remain merge gates. No production deployment or release acceptance is implied.
Glossary: acknowledgment confirms subscription, not receipt of document content; catch-up is the missing document content; control means the same test run with the old behavior restored.
…loss (PM-1496) The per-slot swap guard added in this PR for the disconnect/reconnect race also fixes a second known bug: LGraph.convertToSubgraph reconnects grouped output links with a synchronous connectSlots call per slot, and the old node-wide swappingConnection guard swallowed the genuine connect events for slots 2+ while slot 1's swap was in flight. The test.fail()-marked repro for that bug (added in #18179) now passes for real, so CI failed with "Expected to fail, but passed." Removing the stale annotation turns the repro into a real regression test. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01CoKpwnbrjRvnAyVd6dnPiq
…loss (PM-1496) The per-slot swap guard added in this PR for the disconnect/reconnect race also fixes a second known bug: LGraph.convertToSubgraph reconnects grouped output links with a synchronous connectSlots call per slot, and the old node-wide swappingConnection guard swallowed the genuine connect events for slots 2+ while slot 1's swap was in flight. The test.fail()-marked repro for that bug (added in #18179) now passes for real, so CI failed with "Expected to fail, but passed." Removing the stale annotation turns the repro into a real regression test. Claude-Session: https://claude.ai/code/session_01CoKpwnbrjRvnAyVd6dnPiq
…1575/PM-1576) (#18391) <!-- ccr-slack-attribution --> _Requested by **Christian Byrne** · [Slack thread](https://comfy-organization.slack.com/archives/C0BGH348Z0C/p1790054750260419?thread_ts=1790054750.260419&cid=C0BGH348Z0C)_ ## Summary Fixes PM-1575/PM-1576: the agent's chat-side tool-call-completion affordance ("done" spinner, "Worked for Xs" summary) settled well before the resulting graph change actually rendered on canvas, with nothing in the UI indicating the canvas hadn't caught up. This PR now contains both the failing repro test (red) and the minimal client-side fix (green). ## Changes **Before (bug):** an agent tool call that touches the canvas (e.g. adding a node) reports done — the tool's own row flips straight to its settled icon — while the canvas itself still shows nothing for the new content, for a stretch that (per the stagingcloud report) is close to 10 real seconds. `agentEventTransport.ts`'s `ingest()` flipped a tool part's `state` to `'done'` purely from the chat-channel `agent_tool_call` frame's own `status` field, with zero coupling to whether the corresponding CRDT `doc_update` had actually reached and been projected onto the canvas (`useAgentCrdtFollower.ts` / `LayoutFollowerBridge`) — two structurally disjoint listeners, never consulting each other. **After (fixed):** `createAgentEventTransport` now takes an optional `shouldAwaitCanvasSync` gate. While it returns `true`, a tool call that reports done holds its part's displayed `state` at `'streaming'` (the same spinner glyph it showed while running) instead of flipping straight to `'done'`, until either `notifyCanvasCaughtUp()` is called or `STALE_AFTER_MS` (30s) elapses — reusing the CRDT follower's own passive-heartbeat bound (`agentCrdtDocLifecycle.ts`) rather than inventing a new magic number. The gate defaults to always-`false`, so any caller that never wires it up (including every pre-existing unit test) keeps the old immediate-done behavior unchanged. `agentConversationStore.ts` threads this gate through to every transport it creates and exposes `setCanvasSyncGate` / `notifyCanvasCaughtUp` so a host can register the gate and forward catch-up signals to the active turn and any stashed background ones. `AgentPanelRoot.vue` wires the gate to `agentPanelStore.enabled` (the same flag that gates the CRDT follower itself) and calls `notifyCanvasCaughtUp()` whenever `useAgentCrdtFollower`'s `outcomes.applied` counter increases for the bound workflow — i.e. the moment a fresh doc update actually lands. **Relationship to PM-1355 (`#18179`, `agentCanvasBlanksDuringRunTurn.spec.ts`):** this mirrors that fix's *shape* (a bounded catch-up-confirmation signal) rather than reusing its primitive directly. PM-1355's `AgentCrdtDocLifecycle`/`SUBSCRIBE_CATCHUP_GRACE_MS` answers "is THIS subscribe's own catch-up frame overdue" — a question scoped to one resubscribe. This bug needed a different question answered: "has some later doc update landed since this tool call reported done", which lives one level up, at the chat-transport/composable boundary rather than inside the subscribe lifecycle. Both do reuse the same `STALE_AFTER_MS` bound as their fallback timeout, per the design guidance, rather than adding a second magic number. **Out of scope:** the actual backend ~10s doc-broadcast propagation delay itself is not addressed here — this fix only makes the client accurately reflect canvas state instead of eliminating the underlying lag. Root-causing that lag is still open (see the file-header RCA in the spec). ## Test Plan - `browser_tests/tests/agent/agentCanvasLagsToolCallCompletion.spec.ts`'s first test (previously `test.fail()`) now asserts the fix directly: after the `add_node` tool call reports `success`, its row (expanded from the "Worked for Xs" summary) stays on the spinner glyph — not the settled one — while the canvas still has zero nodes; once the delayed `doc_update` lands, the row flips to done in the same moment the node appears. - The second test (node eventually lands once the delayed `doc_update` arrives) is unchanged and still passes. - `eslint`, `vue-tsc --noEmit`, and `playwright test --list` (both tests listed and named correctly) all pass. Every existing agent-suite unit test (125 files, ~1956 tests) still passes unchanged, since the new gate defaults to off unless a caller opts in. - Could not run a real browser in this sandbox (Playwright's Chromium/headless-shell download is CDN-blocked here, same known limitation as before). ## Screenshots (if applicable) N/A — this is a Playwright spec + client-side logic change; no visual/UI screenshot to attach. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01VR48ZgDstU3S7MKeC1q3bH --- _Generated by [Claude Code](https://claude.ai/code/session_01VR48ZgDstU3S7MKeC1q3bH)_ --------- Co-authored-by: Claude <noreply@anthropic.com>
…1575/PM-1576) (#18391) <!-- ccr-slack-attribution --> _Requested by **Christian Byrne** · [Slack thread](https://comfy-organization.slack.com/archives/C0BGH348Z0C/p1790054750260419?thread_ts=1790054750.260419&cid=C0BGH348Z0C)_ Fixes PM-1575/PM-1576: the agent's chat-side tool-call-completion affordance ("done" spinner, "Worked for Xs" summary) settled well before the resulting graph change actually rendered on canvas, with nothing in the UI indicating the canvas hadn't caught up. This PR now contains both the failing repro test (red) and the minimal client-side fix (green). **Before (bug):** an agent tool call that touches the canvas (e.g. adding a node) reports done — the tool's own row flips straight to its settled icon — while the canvas itself still shows nothing for the new content, for a stretch that (per the stagingcloud report) is close to 10 real seconds. `agentEventTransport.ts`'s `ingest()` flipped a tool part's `state` to `'done'` purely from the chat-channel `agent_tool_call` frame's own `status` field, with zero coupling to whether the corresponding CRDT `doc_update` had actually reached and been projected onto the canvas (`useAgentCrdtFollower.ts` / `LayoutFollowerBridge`) — two structurally disjoint listeners, never consulting each other. **After (fixed):** `createAgentEventTransport` now takes an optional `shouldAwaitCanvasSync` gate. While it returns `true`, a tool call that reports done holds its part's displayed `state` at `'streaming'` (the same spinner glyph it showed while running) instead of flipping straight to `'done'`, until either `notifyCanvasCaughtUp()` is called or `STALE_AFTER_MS` (30s) elapses — reusing the CRDT follower's own passive-heartbeat bound (`agentCrdtDocLifecycle.ts`) rather than inventing a new magic number. The gate defaults to always-`false`, so any caller that never wires it up (including every pre-existing unit test) keeps the old immediate-done behavior unchanged. `agentConversationStore.ts` threads this gate through to every transport it creates and exposes `setCanvasSyncGate` / `notifyCanvasCaughtUp` so a host can register the gate and forward catch-up signals to the active turn and any stashed background ones. `AgentPanelRoot.vue` wires the gate to `agentPanelStore.enabled` (the same flag that gates the CRDT follower itself) and calls `notifyCanvasCaughtUp()` whenever `useAgentCrdtFollower`'s `outcomes.applied` counter increases for the bound workflow — i.e. the moment a fresh doc update actually lands. **Relationship to PM-1355 (`#18179`, `agentCanvasBlanksDuringRunTurn.spec.ts`):** this mirrors that fix's *shape* (a bounded catch-up-confirmation signal) rather than reusing its primitive directly. PM-1355's `AgentCrdtDocLifecycle`/`SUBSCRIBE_CATCHUP_GRACE_MS` answers "is THIS subscribe's own catch-up frame overdue" — a question scoped to one resubscribe. This bug needed a different question answered: "has some later doc update landed since this tool call reported done", which lives one level up, at the chat-transport/composable boundary rather than inside the subscribe lifecycle. Both do reuse the same `STALE_AFTER_MS` bound as their fallback timeout, per the design guidance, rather than adding a second magic number. **Out of scope:** the actual backend ~10s doc-broadcast propagation delay itself is not addressed here — this fix only makes the client accurately reflect canvas state instead of eliminating the underlying lag. Root-causing that lag is still open (see the file-header RCA in the spec). - `browser_tests/tests/agent/agentCanvasLagsToolCallCompletion.spec.ts`'s first test (previously `test.fail()`) now asserts the fix directly: after the `add_node` tool call reports `success`, its row (expanded from the "Worked for Xs" summary) stays on the spinner glyph — not the settled one — while the canvas still has zero nodes; once the delayed `doc_update` lands, the row flips to done in the same moment the node appears. - The second test (node eventually lands once the delayed `doc_update` arrives) is unchanged and still passes. - `eslint`, `vue-tsc --noEmit`, and `playwright test --list` (both tests listed and named correctly) all pass. Every existing agent-suite unit test (125 files, ~1956 tests) still passes unchanged, since the new gate defaults to off unless a caller opts in. - Could not run a real browser in this sandbox (Playwright's Chromium/headless-shell download is CDN-blocked here, same known limitation as before). N/A — this is a Playwright spec + client-side logic change; no visual/UI screenshot to attach. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01VR48ZgDstU3S7MKeC1q3bH --- _Generated by [Claude Code](https://claude.ai/code/session_01VR48ZgDstU3S7MKeC1q3bH)_ ---------
The cherry-picked PM-1575 fix imports STALE_AFTER_MS from agentCrdtDocLifecycle.ts, a module split out of useAgentCrdtFollower.ts by PM-1355 (#18179) on main. That split has not been backported to cloud/1.54, where STALE_AFTER_MS is still exported directly from useAgentCrdtFollower.ts, so point both the fix and its test at the constant's current location on this branch.
…s materialization (PM-1575/PM-1576) (#18593) <!-- ccr-slack-attribution --> _Requested by **Christian Byrne** · [Slack thread](https://comfy-organization.slack.com/archives/C0BGH348Z0C/p1790054937773959?thread_ts=1790054937.773959&cid=C0BGH348Z0C)_ ## Before An agent tool call that touches the canvas (e.g. adding a node) reports done — the tool's own row flips straight to its settled icon — while the canvas itself still shows nothing for the new content, for a stretch that can run close to 10 real seconds. `agentEventTransport.ts`'s `ingest()` flipped a tool part's `state` to `'done'` purely from the chat-channel `agent_tool_call` frame's own `status` field, with zero coupling to whether the corresponding CRDT `doc_update` had actually reached and been projected onto the canvas. ## After `createAgentEventTransport` now takes an optional `shouldAwaitCanvasSync` gate. While it returns `true`, a tool call that reports done holds its part's displayed `state` at `'streaming'` instead of flipping straight to `'done'`, until either `notifyCanvasCaughtUp()` is called or `STALE_AFTER_MS` (30s) elapses. `agentConversationStore.ts` threads this gate through every transport it creates and exposes `setCanvasSyncGate` / `notifyCanvasCaughtUp`; `AgentPanelRoot.vue` wires the gate to `agentPanelStore.enabled` and calls `notifyCanvasCaughtUp()` whenever the CRDT follower's `outcomes.appliedLive` counter increases for the bound workflow. The gate defaults to always-`false`, so any caller that never wires it up keeps the old immediate-done behavior unchanged. ## How Cherry-picked #18391's merge commit onto `cloud/1.54` (stacked on #18586, which backports the JSON fixture corpus this PR's new regression test reads off disk). `cloud/1.54` has diverged architecturally from `main` in this area since PM-1355/#18179: `useAgentCrdtFollower.ts` inlines outcome tracking in its `onUpdate` handler instead of the `incrementOutcome`/`applyAndReconcile` helpers #18391 was written against, so the two real conflicts there were resolved by porting #18391's actual upstream diff (the `appliedLive`/`!update.catchUp` classification) into `cloud/1.54`'s current handler rather than reintroducing stale helper functions. Three follow-up fixes for other main-only references that only surfaced once the build/tests ran (not as git conflicts): `STALE_AFTER_MS` now imports from `useAgentCrdtFollower.ts` (its current home on this branch, not the `agentCrdtDocLifecycle.ts` split that hasn't landed here), and the new e2e spec's `UserDataFullInfo` import, open-agent button i18n key, and `HostDoc` node-presence check were pointed at their current cloud/1.54 equivalents. `pnpm lint`, `pnpm typecheck`, `pnpm typecheck:browser`, and the full targeted unit suite (`agentEventTransport.test.ts`, `agentConversationStore.test.ts`, `useAgentCrdtFollower.test.ts` — 155/155) all pass. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01NQDb2pk4BuLEgzkga3GhEs --- _Generated by [Claude Code](https://claude.ai/code/session_01NQDb2pk4BuLEgzkga3GhEs)_ --------- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com> Co-authored-by: Claude <noreply@anthropic.com>
<!-- ccr-slack-attribution -->
Requested by Christian Byrne · Slack thread (#comfy-agent-user-feedback, Jo Zhang's nightly report, 2026-09-20 03:24 UTC)
Summary
Intentionally-failing repro for Jo Zhang's nightly bug report: asking the in-app agent to run the canvas workflow (ask mode) makes the canvas go fully blank the moment the agent starts working, blank for the whole run, with every node reappearing at once once the run finishes. Ask-mode confirmation UI and the run itself both worked correctly (
print_workflow/validate/run/wait_for_job/get_outputallok,wait_for_job15386ms) — only the canvas rendering is affected. No fix is attempted here.The spec is now two tests sharing one arrange helper: the first pins the known defect (
test.fail(), canvas still blank the instant the run reports done), the second pins that the passive stale-probe genuinely recovers once given the chance — the "all nodes came back" half of the report is real, intended behavior, not part of the bug.Mechanism (confirmed)
Two independently-confirmed frontend behaviors combine to produce exactly this timeline:
doc_reset/follower_replacedframe makesAgentCrdtProjection.clearForReset(src/workbench/extensions/agent/crdt/agentCrdtProjection.ts) synchronously sweep every live materialized node off the canvas, before any replacement content exists. This is the only path that actively removes already-rendered nodes — already pinned by an existing test (useAgentCrdtFollower.test.ts: "reconciles after a doc_reset clear, without waiting for another frame").doc_subscribed: {ok: true}) without its catch-updoc_update. This exact anomaly is visible in the bug report's own attached debug log, on a different (later, smaller) workflow in the same session: a content-lessokack, followed exactly 30 seconds later —STALE_AFTER_MSinagentCrdtDocLifecycle.ts— by the passive stale-probe's forced resubscribe, which is the one that actually delivers content and materializes the nodes. There is no faster active recovery (already pinned by "a confirmed subscribe clears the retry timer", which asserts nothing fires beforeSTALE_AFTER_MS).Together: whatever triggers the reset, the canvas can sit fully blank for up to 30 real seconds afterward with zero active recovery — comfortably spanning one generation run (this report's own run + wait_for_job took 15386ms) — sufficient by itself to produce the reported timeline, independent of whether the run itself actually finished at that moment.
Open half of the RCA (stated plainly)
The backend condition that would emit a
doc_resetfor a plain, edit-free "run the workflow" turn was not found:run/validate/wait_for_job/get_output(cloud/services/agent/internal/loop/tools.go) make no calls into the CRDT doc store at all (docDeps(),replaceDocFromClient,.pub.Reset— zero hits).ensureDoc/prepareWorkflowincloud/services/agent/internal/loop/crdt.go) is heavily unit-tested against firing on an unmodified canvas (TestHumanDraftWithoutAcceptanceSnapshotResyncsFromDocument,TestClientSeedDoesNotOverwriteANewerClientSeed) — both the doc projection and the posted draft pass through the samedraft.NormalizeGraph.answerAsk(approving the run) posts to a separate REST endpoint (/agent/threads/{id}/asks/{askId}/answer) and does not repostdraft, ruling out a confirmation-time draft/CAS race.workflow_docsauthoritative at turn start" family doesn't cleanly fit either: its promotion PR (cloud#9718) merged tomainon 2026-09-19T20:10:38Z, but the commit nightly.engcomfy.com actually ran at the incident (ad2efe2, committed 2026-09-20T02:59:08Z) is confirmed not a descendant of that merge, yet already carries mature pre-1.39.2 #8447 CAS/HumanAuthoredreconciliation guards.Which backend path actually emitted the reset in production is unconfirmed. This repro pins what the user sees once one is emitted, for whatever reason, and gives backend engineers a concrete signature to search production logs for: a
doc_resetframe, or adoc_subscribed:okwith no followingdoc_update, around aruntool call.Step 0 — checked for an existing PR first
doc_reset) triggered by an unreadableschema_version— not this transient, self-healing, run-duration-scoped blank.main.Test structure fix
CI failed with "expected test to fail, but it passed". Reading the spec found the real defect: the
test.fail()-guarded assertion was followed by more code (a furtherpage.clock.fastForwardand two recovery.toBeVisible()assertions) that could never actually run — if the guarded assertion throws as intended, execution stops there; if it doesn't throw, Playwright already reports the expected-fail-but-passed failure before reaching the rest. Either way the recovery half was dead code and never verified anything.Fixed by extracting the shared arrange (drive the run turn to the blanked-canvas point) into a
driveRunTurnUntilCanvasIsBlankhelper and splitting into two tests in the samedescribe, matching the pattern already used byagentZeroNodesOnUnreadableSchema.spec.ts(#18099) andagentTurnSurvivesSocketDrop.spec.ts:test.fail()is now the last thing the first test does, and the recovery check (not a bug — genuinely intended, if slow, behavior) is its own second test with notest.fail().Production-fix experiment (added at Christian Byrne's request)
A follow-up commit (
2b0b4a9) adds a real, scoped production fix for the RCA's "no active recovery" gap:AgentCrdtDocLifecyclenow also arms a short (2s) active catch-up probe right after a subscribe is confirmed, instead of relying solely on the passive 30s stale-probe heartbeat, so a missing catch-updoc_updateis retried well before the full budget. The repro spec itself (test.fail()/assertions) is untouched — this exists to see whether it changes this PR's e2e outcome, not to make the test pass by construction.Verification
pnpm exec oxfmt --check,pnpm exec oxlint,pnpm exec eslinton the changed file — cleanpnpm typecheck(fullvue-tsc+ workspace packages) — cleanpnpm exec playwright test --project=cloud --list browser_tests/tests/agent/agentCanvasBlanksDuringRunTurn.spec.ts— now collects two tests under thecloudproject with@cloud @agenttags🤖 Generated with Claude Code
https://claude.ai/code/session_012sDXswotMsabTTHot1LTTu
Generated by Claude Code