onbuzz 4.4.0 → 4.6.0

package/src/core/__tests__/flowEndToEnd.test.js

@@ -0,0 +1,565 @@
+/**
+ * End-to-end integration tests for Flows v2.
+ *
+ * These exercise FULL pipelines (multiple agent nodes, edges, typed I/O)
+ * through executeFlow → assembleNodeInputs → flowContextService →
+ * messageProcessor → jobdone-completion → resumeFlow, asserting the
+ * exact handoff content one agent passes to the next.
+ *
+ * The unit tests prove each helper works alone. THESE prove the helpers
+ * COMPOSE correctly — which is where bugs like "the outputs bag was
+ * silently dropped between nodes" actually live.
+ */
+
+import { jest, describe, test, expect, beforeEach } from '@jest/globals';
+import path from 'path';
+import os from 'os';
+import { promises as fs } from 'fs';
+import { createMockLogger, createMockConfig, createMockStateManager, createMockAgentPool } from '../../__test-utils__/mockFactories.js';
+
+jest.unstable_mockModule('../../utilities/constants.js', () => ({
+  AGENT_MODES: { CHAT: 'chat', AGENT: 'agent' },
+}));
+
+const { default: FlowExecutor } = await import('../flowExecutor.js');
+const { default: FlowContextService } = await import('../../services/flowContextService.js');
+const { FlowCheckpointStore } = await import('../flowCheckpointStore.js');
+
+// Helper: build a 3-stage v2 typed flow (researcher → writer → reviewer)
+function buildResearchFlow() {
+  return {
+    id: 'flow-e2e', name: 'E2E research flow', schemaVersion: 2,
+    nodes: [
+      { id: 'in', type: 'input', data: {},
+        inputs: [], outputs: [{ name: 'topic', type: 'text' }] },
+      { id: 'researcher', type: 'agent',
+        data: { agentId: 'researcher', label: 'Researcher',
+          promptTemplate: 'Research: {{topic}}' },
+        inputs:  [{ name: 'topic', type: 'text', required: true }],
+        outputs: [
+          { name: 'findings', type: 'json' },
+          { name: 'citations', type: 'list<text>' },
+        ] },
+      { id: 'writer', type: 'agent',
+        data: { agentId: 'writer', label: 'Writer',
+          promptTemplate: 'Write about {{topic}} using findings: {{findings}}' },
+        inputs: [
+          { name: 'topic',    type: 'text', required: true },
+          { name: 'findings', type: 'json', required: true },
+        ],
+        outputs: [{ name: 'draft', type: 'text' }] },
+      { id: 'reviewer', type: 'agent',
+        data: { agentId: 'reviewer', label: 'Reviewer',
+          promptTemplate: 'Review: {{draft}} (citations: {{citations}})' },
+        inputs: [
+          { name: 'draft',     type: 'text',       required: true },
+          { name: 'citations', type: 'list<text>', required: true },
+        ],
+        outputs: [{ name: 'verdict', type: 'text' }] },
+      { id: 'out', type: 'output', data: { outputFormat: 'markdown' },
+        inputs: [{ name: 'context', type: 'text', required: true }], outputs: [] },
+    ],
+    edges: [
+      { source: 'in',         sourceField: 'topic',     target: 'researcher', targetField: 'topic' },
+      { source: 'in',         sourceField: 'topic',     target: 'writer',     targetField: 'topic' },
+      { source: 'researcher', sourceField: 'findings',  target: 'writer',     targetField: 'findings' },
+      { source: 'researcher', sourceField: 'citations', target: 'reviewer',   targetField: 'citations' },
+      { source: 'writer',     sourceField: 'draft',     target: 'reviewer',   targetField: 'draft' },
+      { source: 'reviewer',   sourceField: 'verdict',   target: 'out',        targetField: 'context' },
+    ],
+    variables: {},
+  };
+}
+
+describe('End-to-end: 3-agent typed flow', () => {
+  let fe, config, logger, stateManager, agentPool, messageProcessor;
+  let promptsByAgent;       // agentId → array of prompt strings sent
+  let completionsByAgent;   // agentId → number of times agent completed
+
+  beforeEach(() => {
+    config = createMockConfig();
+    logger = createMockLogger();
+    stateManager = createMockStateManager();
+    agentPool = createMockAgentPool();
+    promptsByAgent = {};
+    completionsByAgent = {};
+
+    messageProcessor = {
+      processMessage: jest.fn().mockImplementation(async (agentId, prompt) => {
+        promptsByAgent[agentId] = promptsByAgent[agentId] || [];
+        promptsByAgent[agentId].push(prompt);
+
+        // Simulate the agent "doing work" and emitting structured outputs
+        // matching the node's declared contract.
+        completionsByAgent[agentId] = (completionsByAgent[agentId] || 0) + 1;
+        let outputs;
+        if (agentId === 'researcher') {
+          outputs = {
+            findings: { topic: 'AI safety', summary: 'Safety is hard' },
+            citations: ['Bostrom 2014', 'Russell 2019'],
+          };
+        } else if (agentId === 'writer') {
+          outputs = { draft: 'A 500-word article on AI safety...' };
+        } else if (agentId === 'reviewer') {
+          outputs = { verdict: 'Approved with minor notes' };
+        }
+        setImmediate(() => fe.notifyAgentCompletion(agentId, {
+          summary: `${agentId} completed`,
+          success: true,
+          outputs,
+        }));
+      }),
+    };
+
+    fe = new FlowExecutor(config, logger, stateManager, agentPool, messageProcessor);
+
+    // Stub agentPool to "have" all three agents already loaded.
+    agentPool.getAgent = jest.fn().mockImplementation(async (id) => ({
+      id, name: id, mode: 'chat',
+      conversations: { full: { messages: [] } },
+    }));
+    agentPool.persistAgentState = jest.fn().mockResolvedValue();
+    agentPool.clearConversation = jest.fn().mockResolvedValue();
+  });
+
+  test('runs all three agents in topological order', async () => {
+    const flow = buildResearchFlow();
+    stateManager.getFlow = jest.fn().mockResolvedValue(flow);
+    stateManager.createFlowRun = jest.fn().mockResolvedValue({ id: 'run-e2e' });
+    stateManager.updateFlowRun = jest.fn().mockResolvedValue();
+    stateManager.getFlowRun = jest.fn().mockResolvedValue({ id: 'run-e2e', nodeStates: {} });
+
+    const result = await fe.executeFlow('flow-e2e', { userInput: 'AI safety' });
+
+    expect(result.status).toBe('completed');
+    expect(completionsByAgent.researcher).toBe(1);
+    expect(completionsByAgent.writer).toBe(1);
+    expect(completionsByAgent.reviewer).toBe(1);
+  });
+
+  test('writer prompt contains researcher outputs assembled by NAME', async () => {
+    const flow = buildResearchFlow();
+    stateManager.getFlow = jest.fn().mockResolvedValue(flow);
+    stateManager.createFlowRun = jest.fn().mockResolvedValue({ id: 'run-1' });
+    stateManager.updateFlowRun = jest.fn().mockResolvedValue();
+    stateManager.getFlowRun = jest.fn().mockResolvedValue({ id: 'run-1', nodeStates: {} });
+
+    await fe.executeFlow('flow-e2e', { userInput: 'AI safety' });
+
+    // Writer's prompt template was: "Write about {{topic}} using findings: {{findings}}"
+    // {{topic}} ← in.topic = 'AI safety'
+    // {{findings}} ← researcher.findings = {topic, summary} (JSON-stringified)
+    const writerPrompt = promptsByAgent.writer?.[0];
+    expect(writerPrompt).toBeDefined();
+    expect(writerPrompt).toContain('AI safety');
+    // The findings object should be present in some form (stringified)
+    expect(writerPrompt).toMatch(/Safety is hard/);
+  });
+
+  test('reviewer prompt contains BOTH writer.draft AND researcher.citations (fan-in)', async () => {
+    const flow = buildResearchFlow();
+    stateManager.getFlow = jest.fn().mockResolvedValue(flow);
+    stateManager.createFlowRun = jest.fn().mockResolvedValue({ id: 'run-2' });
+    stateManager.updateFlowRun = jest.fn().mockResolvedValue();
+    stateManager.getFlowRun = jest.fn().mockResolvedValue({ id: 'run-2', nodeStates: {} });
+
+    await fe.executeFlow('flow-e2e', { userInput: 'AI safety' });
+
+    const reviewerPrompt = promptsByAgent.reviewer?.[0];
+    expect(reviewerPrompt).toBeDefined();
+    // {{draft}} from writer
+    expect(reviewerPrompt).toContain('500-word article');
+    // {{citations}} from researcher (passed through writer's stage as a fan-in)
+    expect(reviewerPrompt).toContain('Bostrom 2014');
+    expect(reviewerPrompt).toContain('Russell 2019');
+  });
+
+  test('flowMetadata is correctly numbered for each step', async () => {
+    const flow = buildResearchFlow();
+    stateManager.getFlow = jest.fn().mockResolvedValue(flow);
+    stateManager.createFlowRun = jest.fn().mockResolvedValue({ id: 'run-meta' });
+    stateManager.updateFlowRun = jest.fn().mockResolvedValue();
+    stateManager.getFlowRun = jest.fn().mockResolvedValue({ id: 'run-meta', nodeStates: {} });
+
+    const calls = [];
+    messageProcessor.processMessage = jest.fn().mockImplementation(async (agentId, _p, opts) => {
+      calls.push({ agentId, flowMetadata: opts.flowMetadata, nodeContract: opts.nodeContract });
+      // Provide outputs so the contract validator passes
+      let outputs;
+      if (agentId === 'researcher') outputs = { findings: {}, citations: [] };
+      else if (agentId === 'writer') outputs = { draft: 'x' };
+      else if (agentId === 'reviewer') outputs = { verdict: 'ok' };
+      setImmediate(() => fe.notifyAgentCompletion(agentId, { summary: `${agentId} done`, success: true, outputs }));
+    });
+
+    await fe.executeFlow('flow-e2e', { userInput: 'AI safety' });
+
+    expect(calls).toHaveLength(3);
+    expect(calls[0].flowMetadata).toMatchObject({ nodeName: 'Researcher', nodePosition: 2, totalNodes: 5 });
+    expect(calls[1].flowMetadata).toMatchObject({ nodeName: 'Writer',     nodePosition: 3, totalNodes: 5 });
+    expect(calls[2].flowMetadata).toMatchObject({ nodeName: 'Reviewer',   nodePosition: 4, totalNodes: 5 });
+    // Node contracts forwarded so the system prompt can render REQUIRED OUTPUTS
+    expect(calls[0].nodeContract.outputs).toHaveLength(2);
+    expect(calls[1].nodeContract.outputs).toHaveLength(1);
+  });
+
+  test('previousAgentData carries the structured outputs end-to-end', async () => {
+    const flow = buildResearchFlow();
+    stateManager.getFlow = jest.fn().mockResolvedValue(flow);
+    stateManager.createFlowRun = jest.fn().mockResolvedValue({ id: 'run-pad' });
+    stateManager.updateFlowRun = jest.fn().mockResolvedValue();
+    stateManager.getFlowRun = jest.fn().mockResolvedValue({ id: 'run-pad', nodeStates: {} });
+
+    const calls = [];
+    messageProcessor.processMessage = jest.fn().mockImplementation(async (agentId, _p, opts) => {
+      calls.push({ agentId, prev: opts.previousAgentData });
+      let outputs;
+      if (agentId === 'researcher') outputs = { findings: { x: 1 }, citations: ['c1'] };
+      else if (agentId === 'writer') outputs = { draft: 'd' };
+      else if (agentId === 'reviewer') outputs = { verdict: 'v' };
+      setImmediate(() => fe.notifyAgentCompletion(agentId, { summary: 'ok', success: true, outputs }));
+    });
+
+    await fe.executeFlow('flow-e2e', { userInput: 'topic' });
+
+    // Researcher: no upstream agent → previousAgentData null
+    expect(calls[0].prev).toBeNull();
+    // Writer: receives researcher's outputs in prev.outputs
+    expect(calls[1].prev?.outputs).toEqual({ findings: { x: 1 }, citations: ['c1'] });
+    // Reviewer: receives outputs from BOTH researcher + writer (fan-in merge)
+    expect(calls[2].prev?.outputs).toEqual({ findings: { x: 1 }, citations: ['c1'], draft: 'd' });
+    expect(calls[2].prev?.contributors?.length).toBe(2);
+  });
+
+  test('Phase 7: writer system prompt contains description + example verbatim from declared contract', async () => {
+    // Producer (researcher) declares an output with description + example.
+    // Consumer (writer) declares the corresponding input ALSO with description + example.
+    // The writer's system prompt should render BOTH the writer's own input contract
+    // AND see the upstream STRUCTURED HANDOFF.
+    const fcs = new FlowContextService({}, logger);
+    const flow = {
+      id: 'rich-e2e', name: 'rich e2e', schemaVersion: 2,
+      description: 'Produce a fact-checked article.',
+      nodes: [
+        { id: 'in', type: 'input', data: {},
+          inputs: [],
+          outputs: [{ name: 'topic', type: 'text', description: 'Topic.' }] },
+        { id: 'researcher', type: 'agent',
+          data: { agentId: 'researcher', label: 'Researcher',
+            instructions: 'Search peer-reviewed sources.',
+            promptTemplate: 'Research {{topic}}' },
+          inputs: [{ name: 'topic', type: 'text', required: true, description: 'Topic to research.' }],
+          outputs: [{
+            name: 'findings', type: 'json',
+            description: 'Structured research with title and citations.',
+            example: { title: 'AI', citations: ['Bostrom 2014'] },
+          }] },
+        { id: 'writer', type: 'agent',
+          data: { agentId: 'writer', label: 'Writer',
+            instructions: 'Write a 500-word article.',
+            promptTemplate: 'Write about {{topic}} using {{findings}}' },
+          inputs: [
+            { name: 'topic',    type: 'text', required: true,
+              description: 'Article topic.' },
+            { name: 'findings', type: 'json', required: true,
+              description: 'Research findings to cite from. Must include title and citations.',
+              example: { title: '...', citations: ['...'] } },
+          ],
+          outputs: [{ name: 'draft', type: 'text', description: 'The article draft.' }] },
+      ],
+      edges: [
+        { source: 'in',         sourceField: 'topic',    target: 'researcher', targetField: 'topic' },
+        { source: 'in',         sourceField: 'topic',    target: 'writer',     targetField: 'topic' },
+        { source: 'researcher', sourceField: 'findings', target: 'writer',     targetField: 'findings' },
+      ],
+    };
+    stateManager.getFlow = jest.fn().mockResolvedValue(flow);
+    stateManager.createFlowRun = jest.fn().mockResolvedValue({ id: 'run-rich' });
+    stateManager.updateFlowRun = jest.fn().mockResolvedValue();
+    stateManager.getFlowRun = jest.fn().mockResolvedValue({ id: 'run-rich', nodeStates: {} });
+
+    const captured = [];
+    messageProcessor.processMessage = jest.fn().mockImplementation(async (agentId, _p, opts) => {
+      const addendum = fcs.buildFlowAgentContext(opts.flowMetadata, opts.previousAgentData, opts.nodeContract);
+      captured.push({ agentId, addendum });
+      let outputs;
+      if (agentId === 'researcher') outputs = { findings: { title: 'AI Safety', citations: ['Bostrom 2014'] } };
+      else if (agentId === 'writer') outputs = { draft: 'Article draft' };
+      setImmediate(() => fe.notifyAgentCompletion(agentId, { summary: 'ok', success: true, outputs }));
+    });
+
+    await fe.executeFlow('rich-e2e', { userInput: 'AI Safety' });
+
+    const writerAddendum = captured.find(c => c.agentId === 'writer').addendum;
+
+    // Flow goal rendered
+    expect(writerAddendum).toContain('FLOW GOAL');
+    expect(writerAddendum).toContain('Produce a fact-checked article');
+
+    // Node instructions rendered
+    expect(writerAddendum).toContain('NODE INSTRUCTIONS');
+    expect(writerAddendum).toContain('Write a 500-word article');
+
+    // Writer's INPUT description + example are rendered
+    expect(writerAddendum).toContain('Article topic');
+    expect(writerAddendum).toContain('Research findings to cite from');
+
+    // Writer's OUTPUT description rendered (draft)
+    expect(writerAddendum).toContain('The article draft');
+
+    // Structured handoff from upstream (researcher's findings) is present
+    expect(writerAddendum).toContain('STRUCTURED HANDOFF');
+    expect(writerAddendum).toContain('"AI Safety"');
+    expect(writerAddendum).toContain('"Bostrom 2014"');
+  });
+
+  test('system prompt rendered for writer contains STRUCTURED HANDOFF section', async () => {
+    // Verify that flowContextService + buildPreviousAgentData compose
+    // correctly: when the data arrives at the prompt-injection layer,
+    // the structured handoff section IS produced.
+    const fcs = new FlowContextService({}, logger);
+    const flow = buildResearchFlow();
+    stateManager.getFlow = jest.fn().mockResolvedValue(flow);
+    stateManager.createFlowRun = jest.fn().mockResolvedValue({ id: 'run-x' });
+    stateManager.updateFlowRun = jest.fn().mockResolvedValue();
+    stateManager.getFlowRun = jest.fn().mockResolvedValue({ id: 'run-x', nodeStates: {} });
+
+    const captured = [];
+    messageProcessor.processMessage = jest.fn().mockImplementation(async (agentId, _p, opts) => {
+      // Render the system prompt addendum the scheduler WOULD produce
+      const addendum = fcs.buildFlowAgentContext(opts.flowMetadata, opts.previousAgentData, opts.nodeContract);
+      captured.push({ agentId, addendum });
+      let outputs;
+      if (agentId === 'researcher') outputs = { findings: { topic: 'AI', summary: 'Hard' }, citations: ['c'] };
+      else if (agentId === 'writer') outputs = { draft: 'd' };
+      else if (agentId === 'reviewer') outputs = { verdict: 'v' };
+      setImmediate(() => fe.notifyAgentCompletion(agentId, { summary: 'ok', success: true, outputs }));
+    });
+
+    await fe.executeFlow('flow-e2e', { userInput: 'AI' });
+
+    const writerAddendum = captured.find(c => c.agentId === 'writer').addendum;
+    expect(writerAddendum).toContain('STRUCTURED HANDOFF');
+    expect(writerAddendum).toContain('findings');
+    expect(writerAddendum).toContain('citations');
+    expect(writerAddendum).toContain('REQUIRED OUTPUTS');
+    expect(writerAddendum).toContain('draft');
+
+    const reviewerAddendum = captured.find(c => c.agentId === 'reviewer').addendum;
+    expect(reviewerAddendum).toContain('ALL UPSTREAM CONTRIBUTORS');
+    // contributor.agentName comes from nodeOutput.agentName which falls
+    // back to the agent's name property (here lowercase 'researcher'/'writer').
+    expect(reviewerAddendum).toMatch(/researcher/i);
+    expect(reviewerAddendum).toMatch(/writer/i);
+    // Both contributors' structured outputs are included
+    expect(reviewerAddendum).toContain('findings');
+    expect(reviewerAddendum).toContain('draft');
+  });
+});
+
+describe('End-to-end: re-prompt loop content', () => {
+  let fe, config, logger, stateManager, agentPool, messageProcessor;
+  beforeEach(() => {
+    config = createMockConfig();
+    logger = createMockLogger();
+    stateManager = createMockStateManager();
+    agentPool = createMockAgentPool();
+    messageProcessor = { processMessage: jest.fn() };
+    fe = new FlowExecutor(config, logger, stateManager, agentPool, messageProcessor);
+    agentPool.getAgent = jest.fn().mockResolvedValue({ id: 'w', mode: 'chat' });
+    agentPool.persistAgentState = jest.fn().mockResolvedValue();
+    agentPool.clearConversation = jest.fn().mockResolvedValue();
+  });
+
+  test('corrective re-prompt message names every missing field + shows example shape', async () => {
+    const node = {
+      id: 'w', type: 'agent',
+      data: { agentId: 'w', promptTemplate: '{{input}}' },
+      inputs: [{ name: 'input', type: 'text', required: true }],
+      outputs: [
+        { name: 'draft',     type: 'text' },
+        { name: 'wordCount', type: 'number' },
+      ],
+    };
+    const flow = {
+      id: 'f', name: 'reprompt content', schemaVersion: 2,
+      nodes: [
+        { id: 'in', type: 'input', data: {}, inputs: [], outputs: [{ name: 'topic', type: 'text' }] },
+        node,
+      ],
+      edges: [{ source: 'in', sourceField: 'topic', target: 'w', targetField: 'input' }],
+    };
+    const context = {
+      input: 'AI', nodeOutputs: { in: { type: 'input', outputs: { topic: 'AI' } } },
+      variables: {}, sortedNodes: flow.nodes, flow,
+    };
+    stateManager.updateFlowRun = jest.fn().mockResolvedValue();
+    stateManager.getFlowRun = jest.fn().mockResolvedValue({ id: 'r', nodeStates: {} });
+
+    let call = 0;
+    messageProcessor.processMessage = jest.fn().mockImplementation(async (id) => {
+      call++;
+      if (call === 1) {
+        // First job-done: missing both required fields
+        setImmediate(() => fe.notifyAgentCompletion(id, { summary: 'done', success: true, outputs: {} }));
+      } else {
+        // Re-prompt response: full bag
+        setImmediate(() => fe.notifyAgentCompletion(id, {
+          summary: 'fixed', success: true,
+          outputs: { draft: 'D', wordCount: 1 },
+        }));
+      }
+    });
+
+    await fe.executeAgentNode(node, context, 'r', null, flow);
+
+    // Inspect the corrective message content (2nd call)
+    const repromptMsg = messageProcessor.processMessage.mock.calls[1][1];
+    expect(repromptMsg).toContain('Missing field(s): draft, wordCount');
+    expect(repromptMsg).toContain('draft: text');
+    expect(repromptMsg).toContain('wordCount: number');
+    // Shows the example JSON shape
+    expect(repromptMsg).toContain('"toolId": "jobdone"');
+    expect(repromptMsg).toContain('"action": "complete"');
+    // Tells the agent NOT to redo the work
+    expect(repromptMsg).toMatch(/do not redo|don't redo/i);
+    // 2nd call is flagged as a reprompt
+    expect(messageProcessor.processMessage.mock.calls[1][2].isReprompt).toBe(true);
+  });
+
+  // -- Phase 7: re-prompt includes description + example --
+
+  test('re-prompt includes description + example for missing fields when declared', async () => {
+    const node = {
+      id: 'w', type: 'agent',
+      data: { agentId: 'w', promptTemplate: '{{input}}' },
+      inputs: [{ name: 'input', type: 'text', required: true }],
+      outputs: [
+        {
+          name: 'draft', type: 'text',
+          description: 'The article draft, ≥500 words, plain prose.',
+          example: 'Once upon a time...',
+        },
+        {
+          name: 'wordCount', type: 'number',
+          description: 'Total word count of the produced draft.',
+          example: 850,
+        },
+      ],
+    };
+    const flow = {
+      id: 'f', name: 'rich reprompt', schemaVersion: 2,
+      nodes: [
+        { id: 'in', type: 'input', data: {}, inputs: [], outputs: [{ name: 'topic', type: 'text' }] },
+        node,
+      ],
+      edges: [{ source: 'in', sourceField: 'topic', target: 'w', targetField: 'input' }],
+    };
+    const context = {
+      input: 'AI', nodeOutputs: { in: { type: 'input', outputs: { topic: 'AI' } } },
+      variables: {}, sortedNodes: flow.nodes, flow,
+    };
+    stateManager.updateFlowRun = jest.fn().mockResolvedValue();
+    stateManager.getFlowRun = jest.fn().mockResolvedValue({ id: 'r2', nodeStates: {} });
+
+    let call = 0;
+    messageProcessor.processMessage = jest.fn().mockImplementation(async (id) => {
+      call++;
+      if (call === 1) {
+        setImmediate(() => fe.notifyAgentCompletion(id, { summary: 'done', success: true, outputs: {} }));
+      } else {
+        setImmediate(() => fe.notifyAgentCompletion(id, {
+          summary: 'fixed', success: true,
+          outputs: { draft: 'D', wordCount: 1 },
+        }));
+      }
+    });
+
+    await fe.executeAgentNode(node, context, 'r2', null, flow);
+    const repromptMsg = messageProcessor.processMessage.mock.calls[1][1];
+
+    // Description for each missing field
+    expect(repromptMsg).toContain('The article draft, ≥500 words');
+    expect(repromptMsg).toContain('Total word count of the produced draft');
+    // Example for each missing field
+    expect(repromptMsg).toContain('Once upon a time');
+    expect(repromptMsg).toContain('850');
+  });
+});
+
+describe('End-to-end: checkpoint + resume preserves typed outputs', () => {
+  let tmpRoot;
+  let fe, config, logger, stateManager, agentPool, messageProcessor;
+
+  beforeEach(async () => {
+    tmpRoot = await fs.mkdtemp(path.join(os.tmpdir(), 'flow-e2e-cp-'));
+    config = createMockConfig();
+    logger = createMockLogger();
+    stateManager = createMockStateManager();
+    agentPool = createMockAgentPool();
+    messageProcessor = { processMessage: jest.fn() };
+    fe = new FlowExecutor(config, logger, stateManager, agentPool, messageProcessor);
+    fe.setCheckpointStore(new FlowCheckpointStore({ baseDir: tmpRoot }));
+    agentPool.getAgent = jest.fn().mockImplementation(async (id) => ({ id, name: id, mode: 'chat' }));
+    agentPool.persistAgentState = jest.fn().mockResolvedValue();
+    agentPool.clearConversation = jest.fn().mockResolvedValue();
+  });
+
+  afterEach(async () => {
+    try { await fs.rm(tmpRoot, { recursive: true, force: true }); } catch {}
+  });
+
+  test('resume rehydrates typed outputs from checkpoints into downstream prompts', async () => {
+    // Stage 1: full successful run, persisting outputs to disk.
+    // Stage 2: simulate a "fresh" executor (e.g. after restart) and
+    // resumeFlow — it should rehydrate the persisted outputs and feed
+    // them as typed inputs into the second agent's prompt.
+    const flow = buildResearchFlow();
+    stateManager.getFlow = jest.fn().mockResolvedValue(flow);
+    stateManager.createFlowRun = jest.fn().mockResolvedValue({ id: 'run-cp' });
+    stateManager.updateFlowRun = jest.fn().mockResolvedValue();
+    stateManager.getFlowRun = jest.fn().mockResolvedValue({
+      id: 'run-cp', flowId: 'flow-e2e', status: 'failed',
+      initialInput: { userInput: 'AI safety' },
+      nodeStates: {},
+    });
+
+    // First, manually persist a researcher checkpoint to disk so we
+    // can resume past it. The result shape mirrors what executeAgentNode
+    // would persist.
+    await fe.checkpointStore.saveNodeResult('run-cp', 'in', {
+      type: 'input', output: 'AI safety', outputs: { topic: 'AI safety' },
+    });
+    await fe.checkpointStore.saveNodeResult('run-cp', 'researcher', {
+      type: 'agent', agentId: 'researcher', agentName: 'Researcher',
+      output: 'researcher done',
+      outputs: {
+        findings: { topic: 'AI', summary: 'Hard' },
+        citations: ['Bostrom 2014'],
+      },
+    });
+
+    const promptsSeen = {};
+    messageProcessor.processMessage = jest.fn().mockImplementation(async (id, prompt) => {
+      promptsSeen[id] = prompt;
+      let outputs;
+      if (id === 'writer')   outputs = { draft: 'D' };
+      if (id === 'reviewer') outputs = { verdict: 'V' };
+      setImmediate(() => fe.notifyAgentCompletion(id, { summary: 'ok', success: true, outputs }));
+    });
+
+    const result = await fe.resumeFlow('run-cp');
+    expect(result.status).toBe('completed');
+
+    // Researcher was checkpointed → never re-invoked
+    expect(messageProcessor.processMessage).not.toHaveBeenCalledWith('researcher', expect.anything(), expect.anything());
+
+    // Writer's prompt should have the researcher's persisted outputs
+    // expanded by name — proving rehydration worked.
+    expect(promptsSeen.writer).toContain('AI safety');
+    expect(promptsSeen.writer).toContain('Hard');
+
+    // Reviewer's prompt should have writer's draft + researcher's citations
+    expect(promptsSeen.reviewer).toContain('Bostrom 2014');
+  });
+});