✨ feat: seed custom node contracts into workflow drafts

README.md

@@ -85,7 +85,7 @@ The shell now also exposes a dedicated Nodes page for project-scoped custom cont
 The Workflows workspace now includes a template gallery. Projects can start from default or saved templates, or create a blank workflow directly.
 The workflow editor center panel now uses a real draggable node canvas with zoom, pan, mini-map, dotted background, handle-based edge creation, persisted node positions, and localized validation feedback instead of a static list of node cards.
 The workflow editor right panel now also supports saving the current workflow draft as a reusable workflow template, in addition to editing per-node runtime settings and Python hooks.
-The node library now supports both click-to-append and drag-and-drop placement into the canvas. V1 connection rules block self-edges, duplicate edges, cycles, incoming edges into source nodes, outgoing edges from export nodes, and multiple upstream edges into ordinary nodes, while allowing multi-input set nodes such as `union-assets`, `intersect-assets`, and `difference-assets`.
+The node library now supports both click-to-append and drag-and-drop placement into the canvas. When a node is inserted from the library, the editor now seeds its default runtime contract directly into the workflow draft, so custom Docker nodes keep their declared executor type and I/O contract without extra manual edits. V1 connection rules block self-edges, duplicate edges, cycles, incoming edges into source nodes, outgoing edges from export nodes, and multiple upstream edges into ordinary nodes, while allowing multi-input set nodes such as `union-assets`, `intersect-assets`, and `difference-assets` plus any custom node whose runtime contract declares `inputMode=multi_asset_set`.
 The Runs workspace now shows project-scoped run history, run-level aggregated summaries, cancel/retry controls, and run detail views with persisted task summaries, stdout/stderr sections, result previews, and artifact links into Explore.
 Selected run tasks now expose the frozen node definition id, executor config snapshot, and code-hook metadata that were captured when the run was created.
 Most built-in delivery nodes now default to `executorType=docker`. When a node uses `executorType=docker` and provides `executorConfig.image`, the worker runs a real local Docker container with mounted `input.json` / `output.json` exchange files plus read-only mounts for bound asset paths. If no image is configured, the executor falls back to the lightweight simulated behavior used by older demo tasks.

apps/web/src/runtime/workflow-editor-state.test.ts

@@ -79,6 +79,40 @@ test("add node at explicit canvas position without auto-connecting it", () => {
   );
 });
 
+test("add custom docker node seeds runtime defaults from the node definition", () => {
+  const base = createDefaultWorkflowDraft();
+  const result = addNodeToDraft(base, {
+    id: "custom-merge-assets",
+    name: "Custom Merge Assets",
+    category: "Utility",
+    defaultExecutorType: "docker",
+    defaultExecutorConfig: {
+      image: "python:3.11-alpine",
+      contract: {
+        inputMode: "multi_asset_set",
+        outputMode: "asset_set",
+        artifactType: "json",
+      },
+    },
+    allowsMultipleIncoming: true,
+    supportsCodeHook: false,
+  });
+
+  assert.equal(result.nodeId, "custom-merge-assets-1");
+  assert.equal(
+    getNodeRuntimeConfig(result.draft, "custom-merge-assets-1")?.executorType,
+    "docker",
+  );
+  assert.equal(
+    (
+      getNodeRuntimeConfig(result.draft, "custom-merge-assets-1")?.executorConfig as {
+        contract?: { inputMode?: string };
+      }
+    )?.contract?.inputMode,
+    "multi_asset_set",
+  );
+});
+
 test("remove node prunes attached edges and serialize emits workflow version payload", () => {
   const draft = workflowDraftFromVersion({
     visualGraph: { viewport: { x: 0, y: 0, zoom: 1 } },
@@ -287,3 +321,48 @@ test("allow multi-inbound utility set nodes while still blocking cycles", () =>
     "target_disallows_incoming",
   );
 });
+
+test("allow multi-inbound custom nodes when their seeded runtime contract expects multiple asset sets", () => {
+  const base = workflowDraftFromVersion({
+    logicGraph: {
+      nodes: [
+        { id: "source-a", type: "source" },
+        { id: "source-b", type: "source" },
+      ],
+      edges: [],
+    },
+    runtimeGraph: {
+      nodeBindings: {
+        "source-a": "source-asset",
+        "source-b": "source-asset",
+      },
+      nodeConfigs: {},
+    },
+  });
+  const appended = addNodeToDraftAtPosition(
+    base,
+    {
+      id: "custom-merge-assets",
+      name: "Custom Merge Assets",
+      category: "Utility",
+      defaultExecutorType: "docker",
+      defaultExecutorConfig: {
+        image: "python:3.11-alpine",
+        contract: {
+          inputMode: "multi_asset_set",
+          outputMode: "asset_set",
+          artifactType: "json",
+        },
+      },
+      allowsMultipleIncoming: true,
+    },
+    { x: 540, y: 200 },
+  );
+
+  const withFirst = connectNodesInDraft(appended.draft, "source-a", appended.nodeId);
+
+  assert.equal(
+    canConnectNodesInDraft(withFirst, "source-b", appended.nodeId).ok,
+    true,
+  );
+});

apps/web/src/runtime/workflow-editor-state.ts

@@ -85,6 +85,36 @@ const DEFAULT_NODE_LAYOUT: Record<string, WorkflowPoint> = {
 };
 const MULTI_INPUT_NODE_DEFINITION_IDS = new Set(["union-assets", "intersect-assets", "difference-assets"]);
 
+function cloneRuntimeConfig(config: WorkflowNodeRuntimeConfig) {
+  return {
+    ...config,
+    executorConfig: config.executorConfig ? { ...config.executorConfig } : undefined,
+    codeHookSpec: config.codeHookSpec ? { ...config.codeHookSpec } : undefined,
+  };
+}
+
+function createDefaultRuntimeConfigForDefinition(
+  definition: WorkflowNodeDefinitionSummary,
+): WorkflowNodeRuntimeConfig | undefined {
+  const executorType = definition.defaultExecutorType;
+  const executorConfig = definition.defaultExecutorConfig ? { ...definition.defaultExecutorConfig } : undefined;
+
+  if (!executorType && !executorConfig) {
+    return undefined;
+  }
+
+  const config: WorkflowNodeRuntimeConfig = {
+    definitionId: definition.id,
+  };
+  if (executorType) {
+    config.executorType = executorType;
+  }
+  if (executorConfig && Object.keys(executorConfig).length > 0) {
+    config.executorConfig = executorConfig;
+  }
+  return config;
+}
+
 function createDefaultNodePosition(index: number): WorkflowPoint {
   const column = index % 3;
   const row = Math.floor(index / 3);
@@ -126,11 +156,7 @@ function cloneDraft(draft: WorkflowDraft): WorkflowDraft {
       nodeConfigs: Object.fromEntries(
         Object.entries(draft.runtimeGraph.nodeConfigs ?? {}).map(([nodeId, config]) => [
           nodeId,
-          {
+          cloneRuntimeConfig(config),
-            ...config,
-            executorConfig: config.executorConfig ? { ...config.executorConfig } : undefined,
-            codeHookSpec: config.codeHookSpec ? { ...config.codeHookSpec } : undefined,
-          },
         ]),
       ),
     },
@@ -286,6 +312,10 @@ function addNodeToDraftInternal(
   next.runtimeGraph.nodeBindings ??= {};
   next.runtimeGraph.nodeBindings[nodeId] = definition.id;
   next.runtimeGraph.nodeConfigs ??= {};
+  const defaultRuntimeConfig = createDefaultRuntimeConfigForDefinition(definition);
+  if (defaultRuntimeConfig) {
+    next.runtimeGraph.nodeConfigs[nodeId] = defaultRuntimeConfig;
+  }
   next.visualGraph.nodePositions[nodeId] =
     options.position ?? createDefaultNodePosition(next.logicGraph.nodes.length - 1);
 
@@ -351,7 +381,17 @@ function nodeDisallowsIncoming(node: WorkflowLogicNode) {
 
 function allowsMultipleIncoming(draft: WorkflowDraft, nodeId: string) {
   const definitionId = resolveDefinitionIdForNode(draft, nodeId);
-  return MULTI_INPUT_NODE_DEFINITION_IDS.has(definitionId);
+  if (MULTI_INPUT_NODE_DEFINITION_IDS.has(definitionId)) {
+    return true;
+  }
+
+  const contract =
+    draft.runtimeGraph.nodeConfigs?.[nodeId]?.executorConfig &&
+    typeof draft.runtimeGraph.nodeConfigs[nodeId]?.executorConfig === "object" &&
+    !Array.isArray(draft.runtimeGraph.nodeConfigs[nodeId]?.executorConfig)
+      ? (draft.runtimeGraph.nodeConfigs[nodeId]?.executorConfig as { contract?: { inputMode?: string } }).contract
+      : undefined;
+  return contract?.inputMode === "multi_asset_set";
 }
 
 function wouldCreateCycle(draft: WorkflowDraft, sourceNodeId: string, targetNodeId: string) {

design/03-workflows/workflow-execution-model.md

@@ -104,6 +104,8 @@ V1 node categories:
 
 The current V1 runtime also supports project-level custom Docker nodes. A custom node is registered separately from the workflow graph, then exposed through the same node-definition surface as built-in nodes.
 
+When the user drops one of these node definitions into the editor, the draft should immediately inherit the node's default runtime snapshot. In practice this means the seeded `nodeConfig` already carries the declared executor type, executor config, and contract before the user opens the right-side panel.
+
 ## Node Definition Contract
 
 Each node definition must expose:
@@ -172,6 +174,8 @@ Expected output shape:
 
 If the custom node declares an `asset_set` style output contract, `result.assetIds` must be a string array. This is what allows downstream nodes to inherit the narrowed asset set.
 
+If the custom node declares `contract.inputMode = "multi_asset_set"`, the canvas should treat that node as multi-input at authoring time instead of forcing the user through single-input validation rules. The graph validator should derive this capability from the seeded runtime contract, not from a hardcoded node id list alone.
+
 The current V1 worker executes trusted-local Python hooks when a `run_task` carries a `codeHookSpec`. The hook is executed through a constrained Python harness with the task snapshot and execution context passed in as JSON. Hook stdout is captured into `stdoutLines`, hook failures populate `stderrLines`, and the returned object becomes the task artifact payload.
 
 The current V1 Docker executor now has two modes:

design/04-ui-ux/information-architecture-and-key-screens.md

@@ -187,6 +187,8 @@ The current V1 authoring rules intentionally keep the graph model constrained so
 - set-operation utility nodes may accept multiple inbound edges
 - cycles are blocked
 
+Custom nodes follow the same rule system, but the decision is contract-driven. When a custom Docker node is added to the canvas, the editor seeds its runtime defaults into the draft immediately; if that contract declares `inputMode=multi_asset_set`, the node is treated like a multi-input utility node from the first connection attempt.
+
 The current built-in node library also exposes Docker-first runtime defaults in the editor. Most built-ins now render with `docker` preselected, while still allowing the user to override the executor, image, and optional Python code hook from the right-side configuration panel.
 
 The runtime header also now exposes a visible `中文 / English` language toggle and the main shell plus workflow authoring surface are translated through a lightweight i18n layer.