Merge pull request #35 from madhura68/feat/sprint-batch-mcp

PBI-50: SPRINT_IMPLEMENTATION single-session sprint runner (MCP-side)

README.md

@@ -32,6 +32,9 @@ activity and create todos via native tool calls instead of curl.
 | `check_queue_empty` | Synchronous, non-blocking count of active jobs (QUEUED/CLAIMED/RUNNING); optional `product_id` scope | no |
 | `set_pbi_pr` | Write `pr_url` on a PBI and clear `pr_merged_at`. Idempotent: re-calling overwrites `pr_url` and resets `pr_merged_at` to null | no |
 | `mark_pbi_pr_merged` | Set `pr_merged_at = now()` on a PBI. Requires `pr_url` to already be set. Idempotent: re-calling overwrites the timestamp | no |
+| `verify_sprint_task` | SPRINT_IMPLEMENTATION-flow: compare a `SprintTaskExecution`'s frozen `plan_snapshot` against `git diff <base_sha>...HEAD`. Returns `verify_result` + `allowed_for_done`. For `task[1..N]` zonder base_sha vult de tool die in op basis van de head_sha van de vorige DONE-execution | yes (read-only) |
+| `update_task_execution` | SPRINT_IMPLEMENTATION-flow: mutate `SprintTaskExecution.status` (PENDING/RUNNING/DONE/FAILED/SKIPPED). Token must own the parent SPRINT-job. Idempotent | no |
+| `job_heartbeat` | Extend `claude_jobs.lease_until` by 5 min. For SPRINT-jobs: response includes `sprint_run_status` + `sprint_run_pause_reason` so the worker can break its task-loop on UI-side cancel/pause | no |
 
 Demo accounts may read but writes return `PERMISSION_DENIED`.
 

__tests__/job-heartbeat.test.ts

@@ -0,0 +1,137 @@
+import { describe, it, expect, vi, beforeEach } from 'vitest'
+
+vi.mock('../src/prisma.js', () => ({
+  prisma: {
+    $queryRaw: vi.fn(),
+    sprintRun: { findUnique: vi.fn() },
+  },
+}))
+
+vi.mock('../src/auth.js', async (importOriginal) => {
+  const original = await importOriginal<typeof import('../src/auth.js')>()
+  return { ...original, requireWriteAccess: vi.fn() }
+})
+
+import { prisma } from '../src/prisma.js'
+import { requireWriteAccess } from '../src/auth.js'
+import { registerJobHeartbeatTool } from '../src/tools/job-heartbeat.js'
+import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
+
+const mockPrisma = prisma as unknown as {
+  $queryRaw: ReturnType<typeof vi.fn>
+  sprintRun: { findUnique: ReturnType<typeof vi.fn> }
+}
+const mockAuth = requireWriteAccess as ReturnType<typeof vi.fn>
+
+const TOKEN_ID = 'tok-owner'
+
+function makeServer() {
+  let handler: (args: Record<string, unknown>) => Promise<unknown>
+  const server = {
+    registerTool: vi.fn((_name: string, _meta: unknown, fn: typeof handler) => {
+      handler = fn
+    }),
+    call: (args: Record<string, unknown>) => handler(args),
+  }
+  registerJobHeartbeatTool(server as unknown as McpServer)
+  return server
+}
+
+beforeEach(() => {
+  vi.clearAllMocks()
+  mockAuth.mockResolvedValue({
+    userId: 'u-1',
+    tokenId: TOKEN_ID,
+    username: 'agent',
+    isDemo: false,
+  })
+})
+
+describe('job_heartbeat', () => {
+  it('returns 403-style error when no row matched (token mismatch / terminal)', async () => {
+    mockPrisma.$queryRaw.mockResolvedValue([])
+    const server = makeServer()
+    const result = (await server.call({ job_id: 'job-x' })) as {
+      content: { text: string }[]
+      isError?: boolean
+    }
+    expect(result.isError).toBe(true)
+    expect(result.content[0].text).toMatch(/not found|terminal|claimed_by/i)
+  })
+
+  it('non-SPRINT job returns ok + lease_until without sprint fields', async () => {
+    const lease = new Date()
+    mockPrisma.$queryRaw.mockResolvedValue([
+      {
+        id: 'job-1',
+        lease_until: lease,
+        kind: 'TASK_IMPLEMENTATION',
+        sprint_run_id: null,
+      },
+    ])
+    const server = makeServer()
+    const result = (await server.call({ job_id: 'job-1' })) as {
+      content: { text: string }[]
+    }
+    const body = JSON.parse(result.content[0].text)
+    expect(body).toEqual({
+      ok: true,
+      job_id: 'job-1',
+      lease_until: lease.toISOString(),
+      sprint_run_status: null,
+      sprint_run_pause_reason: null,
+    })
+    expect(mockPrisma.sprintRun.findUnique).not.toHaveBeenCalled()
+  })
+
+  it('SPRINT job returns sprint_run_status from sibling lookup', async () => {
+    const lease = new Date()
+    mockPrisma.$queryRaw.mockResolvedValue([
+      {
+        id: 'job-2',
+        lease_until: lease,
+        kind: 'SPRINT_IMPLEMENTATION',
+        sprint_run_id: 'sr-1',
+      },
+    ])
+    mockPrisma.sprintRun.findUnique.mockResolvedValue({
+      status: 'PAUSED',
+      pause_context: { pause_reason: 'QUOTA_DEPLETED' },
+    })
+
+    const server = makeServer()
+    const result = (await server.call({ job_id: 'job-2' })) as {
+      content: { text: string }[]
+    }
+    const body = JSON.parse(result.content[0].text)
+    expect(body).toMatchObject({
+      ok: true,
+      sprint_run_status: 'PAUSED',
+      sprint_run_pause_reason: 'QUOTA_DEPLETED',
+    })
+  })
+
+  it('SPRINT job tolerates missing pause_context', async () => {
+    const lease = new Date()
+    mockPrisma.$queryRaw.mockResolvedValue([
+      {
+        id: 'job-3',
+        lease_until: lease,
+        kind: 'SPRINT_IMPLEMENTATION',
+        sprint_run_id: 'sr-2',
+      },
+    ])
+    mockPrisma.sprintRun.findUnique.mockResolvedValue({
+      status: 'RUNNING',
+      pause_context: null,
+    })
+
+    const server = makeServer()
+    const result = (await server.call({ job_id: 'job-3' })) as {
+      content: { text: string }[]
+    }
+    const body = JSON.parse(result.content[0].text)
+    expect(body.sprint_run_status).toBe('RUNNING')
+    expect(body.sprint_run_pause_reason).toBeNull()
+  })
+})

__tests__/update-job-status-sprint-gate.test.ts

@@ -0,0 +1,192 @@
+import { describe, it, expect, vi, beforeEach } from 'vitest'
+
+vi.mock('../src/prisma.js', () => ({
+  prisma: {
+    sprintTaskExecution: {
+      findMany: vi.fn(),
+    },
+    sprintRun: {
+      findUnique: vi.fn(),
+      update: vi.fn(),
+    },
+    story: {
+      count: vi.fn(),
+    },
+  },
+}))
+
+import { prisma } from '../src/prisma.js'
+import {
+  checkSprintVerifyGate,
+  finalizeSprintRunOnDone,
+} from '../src/tools/update-job-status.js'
+
+type MockedPrisma = {
+  sprintTaskExecution: { findMany: ReturnType<typeof vi.fn> }
+  sprintRun: {
+    findUnique: ReturnType<typeof vi.fn>
+    update: ReturnType<typeof vi.fn>
+  }
+  story: { count: ReturnType<typeof vi.fn> }
+}
+
+const mocked = prisma as unknown as MockedPrisma
+
+const LONG_SUMMARY = 'Refactor touched extra files for type narrowing.'
+
+function execRow(overrides: Record<string, unknown>) {
+  return {
+    id: 'exec-' + Math.random().toString(36).slice(2, 8),
+    task_id: 't1',
+    order: 0,
+    status: 'DONE',
+    verify_result: 'ALIGNED',
+    verify_summary: null,
+    verify_required_snapshot: 'ALIGNED_OR_PARTIAL',
+    verify_only_snapshot: false,
+    task: { code: 'TASK-1', title: 'Sample task' },
+    ...overrides,
+  }
+}
+
+describe('checkSprintVerifyGate', () => {
+  beforeEach(() => {
+    vi.clearAllMocks()
+  })
+
+  it('rejects when no executions exist (claim-bug)', async () => {
+    mocked.sprintTaskExecution.findMany.mockResolvedValue([])
+    const r = await checkSprintVerifyGate('job-x')
+    expect(r.allowed).toBe(false)
+    if (!r.allowed) expect(r.error).toMatch(/geen SprintTaskExecution-rows/i)
+  })
+
+  it('blocks PENDING/RUNNING executions', async () => {
+    mocked.sprintTaskExecution.findMany.mockResolvedValue([
+      execRow({ status: 'PENDING' }),
+      execRow({ status: 'RUNNING' }),
+    ])
+    const r = await checkSprintVerifyGate('job-x')
+    expect(r.allowed).toBe(false)
+    if (!r.allowed) {
+      expect(r.error).toMatch(/PENDING/)
+      expect(r.error).toMatch(/RUNNING/)
+    }
+  })
+
+  it('blocks FAILED executions', async () => {
+    mocked.sprintTaskExecution.findMany.mockResolvedValue([
+      execRow({ status: 'FAILED' }),
+    ])
+    const r = await checkSprintVerifyGate('job-x')
+    expect(r.allowed).toBe(false)
+    if (!r.allowed) expect(r.error).toMatch(/FAILED/)
+  })
+
+  it('blocks SKIPPED unless verify_required_snapshot=ANY', async () => {
+    mocked.sprintTaskExecution.findMany.mockResolvedValue([
+      execRow({ status: 'SKIPPED', verify_required_snapshot: 'ALIGNED' }),
+    ])
+    const r = await checkSprintVerifyGate('job-x')
+    expect(r.allowed).toBe(false)
+    if (!r.allowed) expect(r.error).toMatch(/SKIPPED/)
+  })
+
+  it('allows SKIPPED when verify_required_snapshot=ANY', async () => {
+    mocked.sprintTaskExecution.findMany.mockResolvedValue([
+      execRow({ status: 'SKIPPED', verify_required_snapshot: 'ANY' }),
+    ])
+    expect((await checkSprintVerifyGate('job-x')).allowed).toBe(true)
+  })
+
+  it('runs per-row gate for DONE executions', async () => {
+    // PARTIAL zonder summary onder ALIGNED_OR_PARTIAL → blocker
+    mocked.sprintTaskExecution.findMany.mockResolvedValue([
+      execRow({
+        status: 'DONE',
+        verify_result: 'PARTIAL',
+        verify_summary: null,
+        verify_required_snapshot: 'ALIGNED_OR_PARTIAL',
+      }),
+    ])
+    const r = await checkSprintVerifyGate('job-x')
+    expect(r.allowed).toBe(false)
+    if (!r.allowed) expect(r.error).toMatch(/DONE-gate/)
+  })
+
+  it('passes when all DONE rows pass per-row gate', async () => {
+    mocked.sprintTaskExecution.findMany.mockResolvedValue([
+      execRow({ verify_result: 'ALIGNED' }),
+      execRow({
+        verify_result: 'PARTIAL',
+        verify_summary: LONG_SUMMARY,
+        verify_required_snapshot: 'ALIGNED_OR_PARTIAL',
+      }),
+    ])
+    expect((await checkSprintVerifyGate('job-x')).allowed).toBe(true)
+  })
+
+  it('aggregates multiple blockers in one error message', async () => {
+    mocked.sprintTaskExecution.findMany.mockResolvedValue([
+      execRow({ status: 'FAILED', task: { code: 'A', title: 'a' } }),
+      execRow({ status: 'PENDING', task: { code: 'B', title: 'b' } }),
+    ])
+    const r = await checkSprintVerifyGate('job-x')
+    expect(r.allowed).toBe(false)
+    if (!r.allowed) {
+      expect(r.error).toMatch(/2 task\(s\) blokkeren/)
+      expect(r.error).toMatch(/A: a/)
+      expect(r.error).toMatch(/B: b/)
+    }
+  })
+})
+
+describe('finalizeSprintRunOnDone', () => {
+  beforeEach(() => {
+    vi.clearAllMocks()
+  })
+
+  it('no-op when SprintRun already DONE (idempotent)', async () => {
+    mocked.sprintRun.findUnique.mockResolvedValue({
+      id: 'sr-1',
+      status: 'DONE',
+      sprint_id: 's1',
+    })
+    await finalizeSprintRunOnDone('sr-1')
+    expect(mocked.sprintRun.update).not.toHaveBeenCalled()
+  })
+
+  it('no-op when SprintRun does not exist', async () => {
+    mocked.sprintRun.findUnique.mockResolvedValue(null)
+    await finalizeSprintRunOnDone('sr-x')
+    expect(mocked.sprintRun.update).not.toHaveBeenCalled()
+  })
+
+  it('no-op when stories still open', async () => {
+    mocked.sprintRun.findUnique.mockResolvedValue({
+      id: 'sr-1',
+      status: 'RUNNING',
+      sprint_id: 's1',
+    })
+    mocked.story.count.mockResolvedValue(2)
+    await finalizeSprintRunOnDone('sr-1')
+    expect(mocked.sprintRun.update).not.toHaveBeenCalled()
+  })
+
+  it('sets SprintRun → DONE when all stories DONE/FAILED', async () => {
+    mocked.sprintRun.findUnique.mockResolvedValue({
+      id: 'sr-1',
+      status: 'RUNNING',
+      sprint_id: 's1',
+    })
+    mocked.story.count.mockResolvedValue(0)
+    await finalizeSprintRunOnDone('sr-1')
+    expect(mocked.sprintRun.update).toHaveBeenCalledWith({
+      where: { id: 'sr-1' },
+      data: expect.objectContaining({
+        status: 'DONE',
+        finished_at: expect.any(Date),
+      }),
+    })
+  })
+})

__tests__/update-task-execution.test.ts

@@ -0,0 +1,199 @@
+import { describe, it, expect, vi, beforeEach } from 'vitest'
+
+vi.mock('../src/prisma.js', () => ({
+  prisma: {
+    sprintTaskExecution: {
+      findUnique: vi.fn(),
+      update: vi.fn(),
+    },
+  },
+}))
+
+vi.mock('../src/auth.js', async (importOriginal) => {
+  const original = await importOriginal<typeof import('../src/auth.js')>()
+  return { ...original, requireWriteAccess: vi.fn() }
+})
+
+import { prisma } from '../src/prisma.js'
+import { requireWriteAccess } from '../src/auth.js'
+import { registerUpdateTaskExecutionTool } from '../src/tools/update-task-execution.js'
+import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
+
+const mockPrisma = prisma as unknown as {
+  sprintTaskExecution: {
+    findUnique: ReturnType<typeof vi.fn>
+    update: ReturnType<typeof vi.fn>
+  }
+}
+const mockAuth = requireWriteAccess as ReturnType<typeof vi.fn>
+
+const TOKEN_ID = 'tok-owner'
+
+function makeServer() {
+  let handler: (args: Record<string, unknown>) => Promise<unknown>
+  const server = {
+    registerTool: vi.fn((_name: string, _meta: unknown, fn: typeof handler) => {
+      handler = fn
+    }),
+    call: (args: Record<string, unknown>) => handler(args),
+  }
+  registerUpdateTaskExecutionTool(server as unknown as McpServer)
+  return server
+}
+
+function execRecord(overrides: Record<string, unknown> = {}) {
+  return {
+    id: 'exec-1',
+    sprint_job_id: 'job-1',
+    sprint_job: {
+      claimed_by_token_id: TOKEN_ID,
+      status: 'CLAIMED',
+      kind: 'SPRINT_IMPLEMENTATION',
+    },
+    ...overrides,
+  }
+}
+
+beforeEach(() => {
+  vi.clearAllMocks()
+  mockAuth.mockResolvedValue({
+    userId: 'u-1',
+    tokenId: TOKEN_ID,
+    username: 'agent',
+    isDemo: false,
+  })
+})
+
+describe('update_task_execution', () => {
+  it('rejects when execution not found', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(null)
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'missing',
+      status: 'RUNNING',
+    })) as { content: { text: string }[]; isError?: boolean }
+    expect(result.isError).toBe(true)
+    expect(result.content[0].text).toMatch(/not found/i)
+  })
+
+  it('rejects wrong job-kind', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(
+      execRecord({
+        sprint_job: { claimed_by_token_id: TOKEN_ID, status: 'CLAIMED', kind: 'TASK_IMPLEMENTATION' },
+      }),
+    )
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'exec-1',
+      status: 'RUNNING',
+    })) as { content: { text: string }[]; isError?: boolean }
+    expect(result.isError).toBe(true)
+    expect(result.content[0].text).toMatch(/SPRINT_IMPLEMENTATION/)
+  })
+
+  it('rejects when token does not own the job', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(
+      execRecord({
+        sprint_job: { claimed_by_token_id: 'other-token', status: 'CLAIMED', kind: 'SPRINT_IMPLEMENTATION' },
+      }),
+    )
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'exec-1',
+      status: 'RUNNING',
+    })) as { content: { text: string }[]; isError?: boolean }
+    expect(result.isError).toBe(true)
+    expect(result.content[0].text).toMatch(/Forbidden/)
+  })
+
+  it('rejects when job is in terminal state', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(
+      execRecord({
+        sprint_job: { claimed_by_token_id: TOKEN_ID, status: 'DONE', kind: 'SPRINT_IMPLEMENTATION' },
+      }),
+    )
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'exec-1',
+      status: 'DONE',
+    })) as { content: { text: string }[]; isError?: boolean }
+    expect(result.isError).toBe(true)
+    expect(result.content[0].text).toMatch(/terminal/)
+  })
+
+  it('writes started_at on RUNNING', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(execRecord())
+    mockPrisma.sprintTaskExecution.update.mockResolvedValue({
+      id: 'exec-1',
+      status: 'RUNNING',
+      base_sha: null,
+      head_sha: null,
+      verify_result: null,
+      verify_summary: null,
+      skip_reason: null,
+      started_at: new Date(),
+      finished_at: null,
+    })
+
+    const server = makeServer()
+    await server.call({ execution_id: 'exec-1', status: 'RUNNING' })
+
+    const updateCall = mockPrisma.sprintTaskExecution.update.mock.calls[0][0]
+    expect(updateCall.data.status).toBe('RUNNING')
+    expect(updateCall.data.started_at).toBeInstanceOf(Date)
+    expect(updateCall.data.finished_at).toBeUndefined()
+  })
+
+  it('writes finished_at on DONE/FAILED/SKIPPED', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(execRecord())
+    mockPrisma.sprintTaskExecution.update.mockResolvedValue({
+      id: 'exec-1',
+      status: 'DONE',
+      base_sha: 'sha-base',
+      head_sha: 'sha-head',
+      verify_result: null,
+      verify_summary: null,
+      skip_reason: null,
+      started_at: new Date(),
+      finished_at: new Date(),
+    })
+
+    const server = makeServer()
+    await server.call({
+      execution_id: 'exec-1',
+      status: 'DONE',
+      head_sha: 'sha-head',
+    })
+
+    const updateCall = mockPrisma.sprintTaskExecution.update.mock.calls[0][0]
+    expect(updateCall.data.status).toBe('DONE')
+    expect(updateCall.data.finished_at).toBeInstanceOf(Date)
+    expect(updateCall.data.head_sha).toBe('sha-head')
+  })
+
+  it('persists skip_reason on SKIPPED', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(execRecord())
+    mockPrisma.sprintTaskExecution.update.mockResolvedValue({
+      id: 'exec-1',
+      status: 'SKIPPED',
+      base_sha: null,
+      head_sha: null,
+      verify_result: null,
+      verify_summary: null,
+      skip_reason: 'no-op task',
+      started_at: null,
+      finished_at: new Date(),
+    })
+
+    const server = makeServer()
+    await server.call({
+      execution_id: 'exec-1',
+      status: 'SKIPPED',
+      skip_reason: 'no-op task',
+    })
+
+    const updateCall = mockPrisma.sprintTaskExecution.update.mock.calls[0][0]
+    expect(updateCall.data.skip_reason).toBe('no-op task')
+    expect(updateCall.data.finished_at).toBeInstanceOf(Date)
+  })
+})

__tests__/verify-sprint-task.test.ts

@@ -0,0 +1,216 @@
+import { describe, it, expect, vi, beforeEach } from 'vitest'
+
+vi.mock('../src/prisma.js', () => ({
+  prisma: {
+    sprintTaskExecution: {
+      findUnique: vi.fn(),
+      findFirst: vi.fn(),
+      update: vi.fn(),
+    },
+  },
+}))
+
+vi.mock('../src/auth.js', async (importOriginal) => {
+  const original = await importOriginal<typeof import('../src/auth.js')>()
+  return { ...original, requireWriteAccess: vi.fn() }
+})
+
+vi.mock('../src/verify/classify.js', () => ({
+  classifyDiffAgainstPlan: vi.fn(),
+}))
+
+vi.mock('node:child_process', () => ({
+  execFile: vi.fn(),
+}))
+
+import { prisma } from '../src/prisma.js'
+import { requireWriteAccess } from '../src/auth.js'
+import { classifyDiffAgainstPlan } from '../src/verify/classify.js'
+import { execFile } from 'node:child_process'
+import { registerVerifySprintTaskTool } from '../src/tools/verify-sprint-task.js'
+import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
+
+const mockPrisma = prisma as unknown as {
+  sprintTaskExecution: {
+    findUnique: ReturnType<typeof vi.fn>
+    findFirst: ReturnType<typeof vi.fn>
+    update: ReturnType<typeof vi.fn>
+  }
+}
+const mockAuth = requireWriteAccess as ReturnType<typeof vi.fn>
+const mockClassify = classifyDiffAgainstPlan as ReturnType<typeof vi.fn>
+const mockExecFile = execFile as unknown as ReturnType<typeof vi.fn>
+
+const TOKEN_ID = 'tok-owner'
+
+function makeServer() {
+  let handler: (args: Record<string, unknown>) => Promise<unknown>
+  const server = {
+    registerTool: vi.fn((_name: string, _meta: unknown, fn: typeof handler) => {
+      handler = fn
+    }),
+    call: (args: Record<string, unknown>) => handler(args),
+  }
+  registerVerifySprintTaskTool(server as unknown as McpServer)
+  return server
+}
+
+function stubGitDiff(stdout: string) {
+  // promisify(execFile) calls (cmd, args, opts, cb)
+  mockExecFile.mockImplementation(
+    (
+      _cmd: string,
+      _args: string[],
+      _opts: unknown,
+      cb: (err: null, result: { stdout: string; stderr: string }) => void,
+    ) => {
+      cb(null, { stdout, stderr: '' })
+    },
+  )
+}
+
+function execRecord(overrides: Record<string, unknown> = {}) {
+  return {
+    id: 'exec-1',
+    sprint_job_id: 'job-1',
+    order: 0,
+    base_sha: 'sha-base',
+    plan_snapshot: 'frozen plan',
+    verify_required_snapshot: 'ALIGNED_OR_PARTIAL',
+    verify_only_snapshot: false,
+    sprint_job: {
+      claimed_by_token_id: TOKEN_ID,
+      status: 'CLAIMED',
+      kind: 'SPRINT_IMPLEMENTATION',
+    },
+    ...overrides,
+  }
+}
+
+beforeEach(() => {
+  vi.clearAllMocks()
+  mockAuth.mockResolvedValue({
+    userId: 'u-1',
+    tokenId: TOKEN_ID,
+    username: 'agent',
+    isDemo: false,
+  })
+})
+
+describe('verify_sprint_task', () => {
+  it('rejects when execution not found', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(null)
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'missing',
+      worktree_path: '/tmp/wt',
+    })) as { content: { text: string }[]; isError?: boolean }
+    expect(result.isError).toBe(true)
+    expect(result.content[0].text).toMatch(/not found/i)
+  })
+
+  it('rejects wrong token', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(
+      execRecord({
+        sprint_job: { claimed_by_token_id: 'other', status: 'CLAIMED', kind: 'SPRINT_IMPLEMENTATION' },
+      }),
+    )
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'exec-1',
+      worktree_path: '/tmp/wt',
+    })) as { content: { text: string }[]; isError?: boolean }
+    expect(result.isError).toBe(true)
+    expect(result.content[0].text).toMatch(/Forbidden/)
+  })
+
+  it('PARTIAL with summary returns allowed_for_done=true under ALIGNED_OR_PARTIAL', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(execRecord())
+    stubGitDiff('diff --git a/x b/x\n+ change\n')
+    mockClassify.mockReturnValue({ result: 'PARTIAL', reasoning: 'extra files' })
+
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'exec-1',
+      worktree_path: '/tmp/wt',
+      summary: 'Refactor touched extra files for type narrowing.',
+    })) as { content: { text: string }[] }
+    const body = JSON.parse(result.content[0].text)
+    expect(body.result).toBe('partial')
+    expect(body.allowed_for_done).toBe(true)
+    expect(body.reason).toBeNull()
+  })
+
+  it('PARTIAL without summary returns allowed_for_done=false', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(execRecord())
+    stubGitDiff('diff --git a/x b/x\n')
+    mockClassify.mockReturnValue({ result: 'PARTIAL', reasoning: 'r' })
+
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'exec-1',
+      worktree_path: '/tmp/wt',
+    })) as { content: { text: string }[] }
+    const body = JSON.parse(result.content[0].text)
+    expect(body.result).toBe('partial')
+    expect(body.allowed_for_done).toBe(false)
+    expect(body.reason).toMatch(/summary/i)
+  })
+
+  it('DIVERGENT with strict ALIGNED returns allowed_for_done=false', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(
+      execRecord({ verify_required_snapshot: 'ALIGNED' }),
+    )
+    stubGitDiff('diff --git a/x b/x\n')
+    mockClassify.mockReturnValue({ result: 'DIVERGENT', reasoning: 'r' })
+
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'exec-1',
+      worktree_path: '/tmp/wt',
+      summary: 'Long enough summary describing the deviation rationale clearly.',
+    })) as { content: { text: string }[] }
+    const body = JSON.parse(result.content[0].text)
+    expect(body.allowed_for_done).toBe(false)
+    expect(body.reason).toMatch(/ALIGNED/)
+  })
+
+  it('auto-fills base_sha from previous DONE execution head_sha', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(
+      execRecord({ order: 1, base_sha: null }),
+    )
+    mockPrisma.sprintTaskExecution.findFirst.mockResolvedValue({
+      head_sha: 'prev-head-sha',
+    })
+    stubGitDiff('diff\n')
+    mockClassify.mockReturnValue({ result: 'ALIGNED', reasoning: 'ok' })
+
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'exec-1',
+      worktree_path: '/tmp/wt',
+    })) as { content: { text: string }[] }
+    const body = JSON.parse(result.content[0].text)
+    expect(body.base_sha).toBe('prev-head-sha')
+
+    // Persisted back to row
+    const updateCalls = mockPrisma.sprintTaskExecution.update.mock.calls
+    const baseShaPersist = updateCalls.find((c) => c[0].data.base_sha === 'prev-head-sha')
+    expect(baseShaPersist).toBeDefined()
+  })
+
+  it('errors when base_sha cannot be derived (no prior DONE)', async () => {
+    mockPrisma.sprintTaskExecution.findUnique.mockResolvedValue(
+      execRecord({ order: 2, base_sha: null }),
+    )
+    mockPrisma.sprintTaskExecution.findFirst.mockResolvedValue(null)
+
+    const server = makeServer()
+    const result = (await server.call({
+      execution_id: 'exec-1',
+      worktree_path: '/tmp/wt',
+    })) as { content: { text: string }[]; isError?: boolean }
+    expect(result.isError).toBe(true)
+    expect(result.content[0].text).toMatch(/MISSING_BASE_SHA/)
+  })
+})

prisma/schema.prisma

@@ -87,6 +87,7 @@ enum SprintRunStatus {
 enum PrStrategy {
   SPRINT
   STORY
+  SPRINT_BATCH
 }
 
 enum IdeaStatus {
@@ -105,6 +106,15 @@ enum ClaudeJobKind {
   IDEA_GRILL
   IDEA_MAKE_PLAN
   PLAN_CHAT
+  SPRINT_IMPLEMENTATION
+}
+
+enum SprintTaskExecutionStatus {
+  PENDING
+  RUNNING
+  DONE
+  FAILED
+  SKIPPED
 }
 
 enum IdeaLogType {
@@ -299,24 +309,27 @@ model Sprint {
 }
 
 model SprintRun {
-  id             String          @id @default(cuid())
+  id              String          @id @default(cuid())
-  sprint         Sprint          @relation(fields: [sprint_id], references: [id], onDelete: Cascade)
+  sprint          Sprint          @relation(fields: [sprint_id], references: [id], onDelete: Cascade)
-  sprint_id      String
+  sprint_id       String
-  started_by     User            @relation("SprintRunStartedBy", fields: [started_by_id], references: [id])
+  started_by      User            @relation("SprintRunStartedBy", fields: [started_by_id], references: [id])
-  started_by_id  String
+  started_by_id   String
-  status         SprintRunStatus @default(QUEUED)
+  status          SprintRunStatus @default(QUEUED)
-  pr_strategy    PrStrategy
+  pr_strategy     PrStrategy
-  branch         String?
+  branch          String?
-  pr_url         String?
+  pr_url          String?
-  started_at     DateTime?
+  started_at      DateTime?
-  finished_at    DateTime?
+  finished_at     DateTime?
-  failure_reason String?
+  failure_reason  String?
-  failed_task    Task?           @relation("SprintRunFailedTask", fields: [failed_task_id], references: [id], onDelete: SetNull)
+  failed_task     Task?           @relation("SprintRunFailedTask", fields: [failed_task_id], references: [id], onDelete: SetNull)
-  failed_task_id String?
+  failed_task_id  String?
-  pause_context  Json?
+  pause_context   Json?
-  created_at     DateTime        @default(now())
+  previous_run_id String?         @unique
-  updated_at     DateTime        @updatedAt
+  previous_run    SprintRun?      @relation("SprintRunChain", fields: [previous_run_id], references: [id], onDelete: SetNull)
-  jobs           ClaudeJob[]
+  next_run        SprintRun?      @relation("SprintRunChain")
+  created_at      DateTime        @default(now())
+  updated_at      DateTime        @updatedAt
+  jobs            ClaudeJob[]
 
   @@index([sprint_id, status])
   @@index([started_by_id, status])
@@ -324,32 +337,33 @@ model SprintRun {
 }
 
 model Task {
-  id                  String           @id @default(cuid())
+  id                     String                @id @default(cuid())
-  story               Story            @relation(fields: [story_id], references: [id], onDelete: Cascade)
+  story                  Story                 @relation(fields: [story_id], references: [id], onDelete: Cascade)
-  story_id            String
+  story_id               String
-  product             Product          @relation(fields: [product_id], references: [id], onDelete: Cascade)
+  product                Product               @relation(fields: [product_id], references: [id], onDelete: Cascade)
-  product_id          String
+  product_id             String
-  sprint              Sprint?          @relation(fields: [sprint_id], references: [id])
+  sprint                 Sprint?               @relation(fields: [sprint_id], references: [id])
-  sprint_id           String?
+  sprint_id              String?
-  code                String           @db.VarChar(30)
+  code                   String                @db.VarChar(30)
-  title               String
+  title                  String
-  description         String?
+  description            String?
-  implementation_plan String?
+  implementation_plan    String?
-  priority            Int
+  priority               Int
-  sort_order          Float
+  sort_order             Float
-  status              TaskStatus       @default(TO_DO)
+  status                 TaskStatus            @default(TO_DO)
-  verify_only         Boolean          @default(false)
+  verify_only            Boolean               @default(false)
-  verify_required     VerifyRequired   @default(ALIGNED_OR_PARTIAL)
+  verify_required        VerifyRequired        @default(ALIGNED_OR_PARTIAL)
   // Override product.repo_url for branch/worktree/push purposes. Set when
   // a task targets a different repo than its parent product (e.g. an
   // MCP-server task tracked under the main product's PBI). Falls back to
   // product.repo_url when null.
-  repo_url            String?
+  repo_url               String?
-  created_at          DateTime         @default(now())
+  created_at             DateTime              @default(now())
-  updated_at          DateTime         @updatedAt
+  updated_at             DateTime              @updatedAt
-  claude_questions    ClaudeQuestion[]
+  claude_questions       ClaudeQuestion[]
-  claude_jobs         ClaudeJob[]
+  claude_jobs            ClaudeJob[]
-  sprint_run_failures SprintRun[]      @relation("SprintRunFailedTask")
+  sprint_run_failures    SprintRun[]           @relation("SprintRunFailedTask")
+  sprint_task_executions SprintTaskExecution[]
 
   @@unique([product_id, code])
   @@index([story_id, priority, sort_order])
@@ -359,20 +373,20 @@ model Task {
 }
 
 model ClaudeJob {
-  id                  String          @id @default(cuid())
+  id                  String                @id @default(cuid())
-  user                User            @relation(fields: [user_id], references: [id], onDelete: Cascade)
+  user                User                  @relation(fields: [user_id], references: [id], onDelete: Cascade)
   user_id             String
-  product             Product         @relation(fields: [product_id], references: [id], onDelete: Cascade)
+  product             Product               @relation(fields: [product_id], references: [id], onDelete: Cascade)
   product_id          String
-  task                Task?           @relation(fields: [task_id], references: [id], onDelete: Cascade)
+  task                Task?                 @relation(fields: [task_id], references: [id], onDelete: Cascade)
   task_id             String?
-  idea                Idea?           @relation(fields: [idea_id], references: [id], onDelete: Cascade)
+  idea                Idea?                 @relation(fields: [idea_id], references: [id], onDelete: Cascade)
   idea_id             String?
-  sprint_run          SprintRun?      @relation(fields: [sprint_run_id], references: [id], onDelete: SetNull)
+  sprint_run          SprintRun?            @relation(fields: [sprint_run_id], references: [id], onDelete: SetNull)
   sprint_run_id       String?
-  kind                ClaudeJobKind   @default(TASK_IMPLEMENTATION)
+  kind                ClaudeJobKind         @default(TASK_IMPLEMENTATION)
-  status              ClaudeJobStatus @default(QUEUED)
+  status              ClaudeJobStatus       @default(QUEUED)
-  claimed_by_token    ApiToken?       @relation(fields: [claimed_by_token_id], references: [id], onDelete: SetNull)
+  claimed_by_token    ApiToken?             @relation(fields: [claimed_by_token_id], references: [id], onDelete: SetNull)
   claimed_by_token_id String?
   claimed_at          DateTime?
   started_at          DateTime?
@@ -391,9 +405,11 @@ model ClaudeJob {
   pr_url              String?
   summary             String?
   error               String?
-  retry_count         Int             @default(0)
+  retry_count         Int                   @default(0)
-  created_at          DateTime        @default(now())
+  lease_until         DateTime?
-  updated_at          DateTime        @updatedAt
+  task_executions     SprintTaskExecution[] @relation("SprintJobExecutions")
+  created_at          DateTime              @default(now())
+  updated_at          DateTime              @updatedAt
 
   @@index([user_id, status])
   @@index([task_id, status])
@@ -401,9 +417,41 @@ model ClaudeJob {
   @@index([sprint_run_id, status])
   @@index([status, claimed_at])
   @@index([status, finished_at])
+  @@index([status, lease_until])
   @@map("claude_jobs")
 }
 
+// PBI-50: frozen scope-snapshot per SPRINT_IMPLEMENTATION-claim. Bij claim
+// wordt voor elke TO_DO-task in scope één PENDING-record gemaakt met
+// implementation_plan + verify_required gesnapshot. Worker en gate werken
+// uitsluitend op deze rows; latere wijzigingen aan Task hebben geen
+// invloed op de lopende batch.
+model SprintTaskExecution {
+  id                       String                    @id @default(cuid())
+  sprint_job               ClaudeJob                 @relation("SprintJobExecutions", fields: [sprint_job_id], references: [id], onDelete: Cascade)
+  sprint_job_id            String
+  task                     Task                      @relation(fields: [task_id], references: [id], onDelete: Cascade)
+  task_id                  String
+  order                    Int
+  plan_snapshot            String                    @db.Text
+  verify_required_snapshot VerifyRequired
+  verify_only_snapshot     Boolean                   @default(false)
+  base_sha                 String?
+  head_sha                 String?
+  status                   SprintTaskExecutionStatus @default(PENDING)
+  verify_result            VerifyResult?
+  verify_summary           String?                   @db.Text
+  skip_reason              String?                   @db.Text
+  started_at               DateTime?
+  finished_at              DateTime?
+  created_at               DateTime                  @default(now())
+  updated_at               DateTime                  @updatedAt
+
+  @@unique([sprint_job_id, task_id])
+  @@index([sprint_job_id, order])
+  @@map("sprint_task_executions")
+}
+
 model ModelPrice {
   id                       String   @id @default(cuid())
   model_id                 String   @unique

src/index.ts

@@ -29,6 +29,10 @@ import { registerUpdateIdeaPlanMdTool } from './tools/update-idea-plan-md.js'
 import { registerLogIdeaDecisionTool } from './tools/log-idea-decision.js'
 import { registerGetWorkerSettingsTool } from './tools/get-worker-settings.js'
 import { registerWorkerHeartbeatTool } from './tools/worker-heartbeat.js'
+// PBI-50: SPRINT_IMPLEMENTATION-tools
+import { registerVerifySprintTaskTool } from './tools/verify-sprint-task.js'
+import { registerUpdateTaskExecutionTool } from './tools/update-task-execution.js'
+import { registerJobHeartbeatTool } from './tools/job-heartbeat.js'
 import { registerImplementNextStoryPrompt } from './prompts/implement-next-story.js'
 import { getAuth } from './auth.js'
 import { registerWorker } from './presence/worker.js'
@@ -92,6 +96,10 @@ async function main() {
   // M13: worker quota-gate tools
   registerGetWorkerSettingsTool(server)
   registerWorkerHeartbeatTool(server)
+  // PBI-50: SPRINT_IMPLEMENTATION-tools
+  registerVerifySprintTaskTool(server)
+  registerUpdateTaskExecutionTool(server)
+  registerJobHeartbeatTool(server)
   registerImplementNextStoryPrompt(server)
 
   // Presence bootstrap MUST run before server.connect — the stdio transport

src/lib/tasks-status-update.ts

@@ -38,6 +38,11 @@ export async function propagateStatusUpwards(
   taskId: string,
   newStatus: TaskStatus,
   client?: Prisma.TransactionClient,
+  // PBI-50: optionele expliciete sprint_run_id voor SPRINT_IMPLEMENTATION
+  // (waar geen ClaudeJob.task_id-koppeling bestaat). Wanneer afwezig valt
+  // de helper terug op de lookup via ClaudeJob.task_id, met als laatste
+  // fallback Story → Sprint → SprintRun.findFirst({ status: active }).
+  sprintRunId?: string,
 ): Promise<PropagationResult> {
   const run = async (tx: Prisma.TransactionClient): Promise<PropagationResult> => {
     const task = await tx.task.update({
@@ -151,18 +156,43 @@ export async function propagateStatusUpwards(
       }
     }
 
-    // SprintRun herevalueren — via ClaudeJob.sprint_run_id van deze task
+    // SprintRun herevalueren. Resolve sprint_run_id in volgorde:
+    //   1. Expliciete sprintRunId-arg (PBI-50: SPRINT_IMPLEMENTATION-pad).
+    //   2. ClaudeJob.task_id-lookup (PER_TASK-flow).
+    //   3. Story → Sprint → SprintRun.findFirst({ status: active }) (geen
+    //      task-job, bv. handmatige task-statuswijziging via UI).
     let sprintRunChanged = false
     if (nextSprintStatus === 'FAILED' || nextSprintStatus === 'COMPLETED') {
-      const job = await tx.claudeJob.findFirst({
+      let resolvedRunId: string | null = sprintRunId ?? null
-        where: { task_id: taskId, sprint_run_id: { not: null } },
+      let cancelExceptJobId: string | null = null
-        orderBy: { created_at: 'desc' },
-        select: { id: true, sprint_run_id: true },
-      })
 
-      if (job?.sprint_run_id) {
+      if (!resolvedRunId) {
+        const job = await tx.claudeJob.findFirst({
+          where: { task_id: taskId, sprint_run_id: { not: null } },
+          orderBy: { created_at: 'desc' },
+          select: { id: true, sprint_run_id: true },
+        })
+        if (job?.sprint_run_id) {
+          resolvedRunId = job.sprint_run_id
+          cancelExceptJobId = job.id
+        }
+      }
+
+      if (!resolvedRunId && story.sprint_id) {
+        const activeRun = await tx.sprintRun.findFirst({
+          where: {
+            sprint_id: story.sprint_id,
+            status: { in: ['QUEUED', 'RUNNING', 'PAUSED'] },
+          },
+          orderBy: { created_at: 'desc' },
+          select: { id: true },
+        })
+        if (activeRun) resolvedRunId = activeRun.id
+      }
+
+      if (resolvedRunId) {
         const sprintRun = await tx.sprintRun.findUnique({
-          where: { id: job.sprint_run_id },
+          where: { id: resolvedRunId },
           select: { id: true, status: true },
         })
         if (
@@ -180,11 +210,16 @@ export async function propagateStatusUpwards(
                 failed_task_id: taskId,
               },
             })
+            // Cancel sibling-jobs binnen dezelfde SprintRun behalve de
+            // huidige task-job (als die er is). Voor SPRINT_IMPLEMENTATION
+            // is cancelExceptJobId null en hebben we geen siblings om te
+            // cancellen — de SPRINT-job zelf blijft actief en de worker
+            // detecteert dit via job_heartbeat.
             await tx.claudeJob.updateMany({
               where: {
                 sprint_run_id: sprintRun.id,
                 status: { in: ['QUEUED', 'CLAIMED', 'RUNNING'] },
-                id: { not: job.id },
+                ...(cancelExceptJobId ? { id: { not: cancelExceptJobId } } : {}),
               },
               data: {
                 status: 'CANCELLED',
@@ -230,14 +265,16 @@ export interface UpdateTaskStatusResult {
   task: PropagationResult['task']
   storyStatusChange: StoryStatusChange
   storyId: string
+  sprintRunChanged: boolean
 }
 
 export async function updateTaskStatusWithStoryPromotion(
   taskId: string,
   newStatus: TaskStatus,
   client?: Prisma.TransactionClient,
+  sprintRunId?: string,
 ): Promise<UpdateTaskStatusResult> {
-  const result = await propagateStatusUpwards(taskId, newStatus, client)
+  const result = await propagateStatusUpwards(taskId, newStatus, client, sprintRunId)
   let storyStatusChange: StoryStatusChange = null
   if (result.storyChanged) {
     storyStatusChange = newStatus === 'DONE' ? 'promoted' : 'demoted'
@@ -246,5 +283,6 @@ export async function updateTaskStatusWithStoryPromotion(
     task: result.task,
     storyStatusChange,
     storyId: result.storyId,
+    sprintRunChanged: result.sprintRunChanged,
   }
 }

src/tools/job-heartbeat.ts

@@ -0,0 +1,81 @@
+// PBI-50 F3-T3: job_heartbeat
+//
+// Verlengt ClaudeJob.lease_until met 5 min zodat resetStaleClaimedJobs een
+// long-running job (bv. SPRINT_IMPLEMENTATION over 30+ min) niet ten onrechte
+// als stale markt. Worker draait een achtergrond-loop elke 60s.
+//
+// Voor SPRINT-jobs: respons bevat sprint_run_status zodat de worker zijn
+// loop kan breken bij ≠ RUNNING (bv. UI-side cancel of MERGE_CONFLICT-pause).
+
+import { z } from 'zod'
+import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
+import { prisma } from '../prisma.js'
+import { requireWriteAccess } from '../auth.js'
+import { toolError, toolJson, withToolErrors } from '../errors.js'
+
+const inputSchema = z.object({
+  job_id: z.string().min(1),
+})
+
+export function registerJobHeartbeatTool(server: McpServer) {
+  server.registerTool(
+    'job_heartbeat',
+    {
+      title: 'Job heartbeat',
+      description:
+        'Extend the lease on a CLAIMED/RUNNING job by 5 minutes. Token must own the job. ' +
+        'For SPRINT_IMPLEMENTATION jobs: response includes sprint_run_status so the worker ' +
+        'can break its task-loop on UI-side cancel/pause without an extra query. ' +
+        'Worker should call this every ~60s during long-running batches. ' +
+        'Forbidden for demo accounts.',
+      inputSchema,
+    },
+    async ({ job_id }) =>
+      withToolErrors(async () => {
+        const auth = await requireWriteAccess()
+
+        // Atomic conditional UPDATE so a non-owner / non-active job is rejected
+        // without a separate read.
+        const updated = await prisma.$queryRaw<
+          Array<{ id: string; lease_until: Date; kind: string; sprint_run_id: string | null }>
+        >`
+          UPDATE claude_jobs
+          SET lease_until = NOW() + INTERVAL '5 minutes'
+          WHERE id = ${job_id}
+            AND claimed_by_token_id = ${auth.tokenId}
+            AND status IN ('CLAIMED', 'RUNNING')
+          RETURNING id, lease_until, kind::text AS kind, sprint_run_id
+        `
+        if (updated.length === 0) {
+          return toolError(
+            `Job ${job_id} not found, not claimed by your token, or in terminal state`,
+          )
+        }
+        const row = updated[0]
+
+        let sprint_run_status: string | null = null
+        let sprint_run_pause_reason: string | null = null
+        if (row.kind === 'SPRINT_IMPLEMENTATION' && row.sprint_run_id) {
+          const sprintRun = await prisma.sprintRun.findUnique({
+            where: { id: row.sprint_run_id },
+            select: { status: true, pause_context: true },
+          })
+          sprint_run_status = sprintRun?.status ?? null
+          // Extract pause_reason from pause_context Json (best-effort)
+          const ctx = sprintRun?.pause_context as
+            | { pause_reason?: string }
+            | null
+            | undefined
+          sprint_run_pause_reason = ctx?.pause_reason ?? null
+        }
+
+        return toolJson({
+          ok: true,
+          job_id: row.id,
+          lease_until: row.lease_until.toISOString(),
+          sprint_run_status,
+          sprint_run_pause_reason,
+        })
+      }),
+  )
+}

src/tools/update-job-status.ts

@@ -225,6 +225,106 @@ export function checkVerifyGate(
   return { allowed: true }
 }
 
+// PBI-50 F4-T1: aggregate verify-gate voor SPRINT_IMPLEMENTATION DONE.
+// Bron: alleen SprintTaskExecution-rows voor deze job. Per row:
+//   DONE     → checkVerifyGate met snapshot-velden (gate per row)
+//   SKIPPED  → alleen toegestaan als verify_required_snapshot === 'ANY'
+//   FAILED/PENDING/RUNNING → blocker (sprint mag niet DONE met openstaand werk)
+// Bij overall pass → { allowed: true }; anders error met opsomming.
+export async function checkSprintVerifyGate(
+  sprintJobId: string,
+): Promise<{ allowed: true } | { allowed: false; error: string }> {
+  const executions = await prisma.sprintTaskExecution.findMany({
+    where: { sprint_job_id: sprintJobId },
+    orderBy: { order: 'asc' },
+    select: {
+      id: true,
+      task_id: true,
+      order: true,
+      status: true,
+      verify_result: true,
+      verify_summary: true,
+      verify_required_snapshot: true,
+      verify_only_snapshot: true,
+      task: { select: { code: true, title: true } },
+    },
+  })
+  if (executions.length === 0) {
+    return {
+      allowed: false,
+      error:
+        'Sprint-job heeft geen SprintTaskExecution-rows. ' +
+        'Dit duidt op een claim-bug; reclaim de sprint.',
+    }
+  }
+
+  const blockers: string[] = []
+  for (const exec of executions) {
+    const taskLabel = `${exec.task.code}: ${exec.task.title}`
+    if (exec.status === 'PENDING' || exec.status === 'RUNNING') {
+      blockers.push(`[${exec.status}] ${taskLabel} — onafgemaakt werk`)
+      continue
+    }
+    if (exec.status === 'FAILED') {
+      blockers.push(`[FAILED] ${taskLabel}`)
+      continue
+    }
+    if (exec.status === 'SKIPPED') {
+      if (exec.verify_required_snapshot !== 'ANY') {
+        blockers.push(
+          `[SKIPPED] ${taskLabel} — alleen toegestaan bij verify_required=ANY`,
+        )
+      }
+      continue
+    }
+    // DONE: per-row gate
+    const gate = checkVerifyGate(
+      exec.verify_result,
+      exec.verify_only_snapshot,
+      exec.verify_required_snapshot,
+      exec.verify_summary ?? undefined,
+    )
+    if (!gate.allowed) {
+      blockers.push(`[DONE-gate] ${taskLabel}: ${gate.error}`)
+    }
+  }
+
+  if (blockers.length === 0) return { allowed: true }
+  return {
+    allowed: false,
+    error:
+      `Sprint kan niet DONE — ${blockers.length} task(s) blokkeren:\n` +
+      blockers.map((b) => `  - ${b}`).join('\n'),
+  }
+}
+
+// PBI-50 F4-T2: idempotent SprintRun-finalisering.
+// Invariant: alleen aanroepen wanneer alle stories in de sprint status
+// DONE/FAILED/CANCELLED hebben. Effect: SprintRun.status → DONE +
+// finished_at = NOW(). Idempotent — bij al-DONE: no-op.
+export async function finalizeSprintRunOnDone(sprintRunId: string): Promise<void> {
+  const sprintRun = await prisma.sprintRun.findUnique({
+    where: { id: sprintRunId },
+    select: { id: true, status: true, sprint_id: true },
+  })
+  if (!sprintRun) return
+  if (sprintRun.status === 'DONE') return // idempotent
+
+  // Check alle stories in deze sprint zijn klaar
+  const openStories = await prisma.story.count({
+    where: {
+      sprint_id: sprintRun.sprint_id,
+      status: { notIn: ['DONE', 'FAILED'] },
+    },
+  })
+  if (openStories > 0) return // nog werk over — niet finaliseren
+
+  await prisma.sprintRun.update({
+    where: { id: sprintRunId },
+    data: { status: 'DONE', finished_at: new Date() },
+  })
+}
+
 const DB_STATUS_MAP = {
   running: 'RUNNING',
   done: 'DONE',
@@ -332,6 +432,68 @@ export async function maybeCreateAutoPr(opts: {
   return null
 }
 
+// PBI-50 F4-T2: SPRINT_BATCH PR-flow. Eén draft-PR voor de hele sprint,
+// title = sprint.sprint_goal. Mens reviewt + mergt zelf — geen auto-merge.
+// Lijkt op de SPRINT-mode van maybeCreateAutoPr maar zonder task-context.
+export async function maybeCreateSprintBatchPr(opts: {
+  jobId: string
+  productId: string
+  worktreePath: string
+  branchName: string
+  summary: string | undefined
+}): Promise<string | null> {
+  const { jobId, productId, worktreePath, branchName, summary } = opts
+
+  const product = await prisma.product.findUnique({
+    where: { id: productId },
+    select: { auto_pr: true },
+  })
+  if (!product?.auto_pr) return null
+
+  const job = await prisma.claudeJob.findUnique({
+    where: { id: jobId },
+    select: {
+      sprint_run_id: true,
+      sprint_run: {
+        select: { id: true, sprint: { select: { sprint_goal: true } } },
+      },
+    },
+  })
+  if (!job?.sprint_run) return null
+
+  // Resume-pad: oude SprintRun heeft mogelijk al een PR via vorige run-job.
+  // Lookup via SprintRunChain (previous_run_id) of via sibling-SPRINT-job.
+  const previousRun = await prisma.sprintRun.findUnique({
+    where: { id: job.sprint_run.id },
+    select: { previous_run_id: true },
+  })
+  if (previousRun?.previous_run_id) {
+    const prevPr = await prisma.claudeJob.findFirst({
+      where: { sprint_run_id: previousRun.previous_run_id, pr_url: { not: null } },
+      select: { pr_url: true },
+    })
+    if (prevPr?.pr_url) return prevPr.pr_url
+  }
+
+  const goal = job.sprint_run.sprint.sprint_goal
+  const sprintTitle = `Sprint: ${goal}`.slice(0, 200)
+  const body = summary
+    ? `${summary}\n\n---\n\n*Draft PR voor sprint-batch \`${job.sprint_run.id}\` (single-session). Wordt ready-for-review zodra alle tasks DONE zijn.*`
+    : `*Draft PR voor sprint-batch \`${job.sprint_run.id}\` (single-session). Wordt ready-for-review zodra alle tasks DONE zijn.*`
+
+  const result = await createPullRequest({
+    worktreePath,
+    branchName,
+    title: sprintTitle,
+    body,
+    draft: true,
+    enableAutoMerge: false,
+  })
+  if ('url' in result) return result.url
+  console.warn(`[update_job_status] sprint-batch draft-PR skipped for job ${jobId}:`, result.error)
+  return null
+}
+
 export function registerUpdateJobStatusTool(server: McpServer) {
   server.registerTool(
     'update_job_status',
@@ -379,6 +541,7 @@ export function registerUpdateJobStatusTool(server: McpServer) {
             product_id: true,
             task_id: true,
             idea_id: true,
+            sprint_run_id: true,
             kind: true,
             verify_result: true,
             task: { select: { verify_only: true, verify_required: true } },
@@ -419,6 +582,19 @@ export function registerUpdateJobStatusTool(server: McpServer) {
             actualStatus = 'done'
             // pushedAt blijft undefined, branch/error overrides ook
             skipWorktreeCleanup = true
+          } else if (job.kind === 'SPRINT_IMPLEMENTATION') {
+            // PBI-50 F4-T2: aggregate verify-gate via SprintTaskExecution-rows.
+            // Geen single-task verify_result op de SPRINT-job zelf.
+            const gate = await checkSprintVerifyGate(job_id)
+            if (!gate.allowed) return toolError(gate.error)
+
+            const plan = await prepareDoneUpdate(job_id, branch)
+            actualStatus = plan.dbStatus === 'DONE' ? 'done' : 'failed'
+            pushedAt = plan.pushedAt
+            if (plan.branchOverride !== undefined) branchToWrite = plan.branchOverride
+            if (plan.errorOverride !== undefined) errorToWrite = plan.errorOverride
+            skipWorktreeCleanup = plan.skipWorktreeCleanup
+            headShaToWrite = plan.headSha
           } else {
             const gate = checkVerifyGate(
               job.verify_result ?? null,
@@ -440,6 +616,7 @@ export function registerUpdateJobStatusTool(server: McpServer) {
 
         // Auto-PR: best-effort, only when push actually happened.
         // M12: idee-jobs hebben geen task_id en geen branch — skip auto-PR.
+        // PBI-50: SPRINT_IMPLEMENTATION krijgt een eigen PR-flow (sprint-goal als title).
         let prUrl: string | null = null
         if (
           actualStatus === 'done' &&
@@ -460,6 +637,23 @@ export function registerUpdateJobStatusTool(server: McpServer) {
             console.warn(`[update_job_status] auto-PR error for job ${job_id}:`, err)
             return null
           })
+        } else if (
+          actualStatus === 'done' &&
+          pushedAt &&
+          branchToWrite &&
+          job.kind === 'SPRINT_IMPLEMENTATION'
+        ) {
+          const worktreeDir = getWorktreeRoot()
+          prUrl = await maybeCreateSprintBatchPr({
+            jobId: job_id,
+            productId: job.product_id,
+            worktreePath: path.join(worktreeDir, job_id),
+            branchName: branchToWrite,
+            summary,
+          }).catch((err) => {
+            console.warn(`[update_job_status] sprint-batch PR error for job ${job_id}:`, err)
+            return null
+          })
         }
 
         const dbStatus = DB_STATUS_MAP[actualStatus as keyof typeof DB_STATUS_MAP]
@@ -493,6 +687,7 @@ export function registerUpdateJobStatusTool(server: McpServer) {
             error: true,
             started_at: true,
             finished_at: true,
+            head_sha: true,
           },
         })
 
@@ -694,10 +889,88 @@ export function registerUpdateJobStatusTool(server: McpServer) {
         // cancel all queued/claimed/running siblings under the same PBI and
         // undo any pushed commits (close open PRs / open revert-PRs for
         // already-merged ones). Idempotent + non-blocking — never throws.
+        // PBI-50: SPRINT_IMPLEMENTATION SKIPS this — cascade naar tasks/stories/
+        // PBIs is al gebeurd via per-task update_task_status('failed')-calls
+        // van de worker. Sprint-job heeft geen task_id; cancelPbi-flow past niet.
         if (actualStatus === 'failed' && job.kind === 'TASK_IMPLEMENTATION' && job.task_id) {
           await cancelPbiOnFailure(job_id)
         }
 
+        // PBI-50 F4-T2: SPRINT_IMPLEMENTATION DONE → finalize SprintRun.
+        if (
+          actualStatus === 'done' &&
+          job.kind === 'SPRINT_IMPLEMENTATION' &&
+          job.sprint_run_id
+        ) {
+          try {
+            await finalizeSprintRunOnDone(job.sprint_run_id)
+            // Mark draft-PR ready-for-review als de SprintRun nu DONE is
+            const finalRun = await prisma.sprintRun.findUnique({
+              where: { id: job.sprint_run_id },
+              select: { status: true },
+            })
+            if (finalRun?.status === 'DONE' && updated.pr_url) {
+              try {
+                const ready = await markPullRequestReady({ prUrl: updated.pr_url })
+                if ('error' in ready) {
+                  console.warn(
+                    `[update_job_status] sprint-batch markPullRequestReady failed for ${updated.pr_url}: ${ready.error}`,
+                  )
+                }
+              } catch (err) {
+                console.warn(`[update_job_status] sprint-batch markPullRequestReady error:`, err)
+              }
+            }
+          } catch (err) {
+            console.warn(`[update_job_status] finalizeSprintRunOnDone error:`, err)
+          }
+        }
+
+        // PBI-50 F4-T3: SPRINT_IMPLEMENTATION FAILED →
+        //  - Detect QUOTA_PAUSE: error-prefix → PAUSED met pause_context.
+        //  - Anders: vul SprintRun.failure_reason + failed_task_id (uit error).
+        if (actualStatus === 'failed' && job.kind === 'SPRINT_IMPLEMENTATION' && job.sprint_run_id) {
+          const isQuotaPause = (errorToWrite ?? '').startsWith('QUOTA_PAUSE:')
+          if (isQuotaPause) {
+            // Vind laatst-DONE execution voor pause-context
+            const lastDone = await prisma.sprintTaskExecution.findFirst({
+              where: { sprint_job_id: job_id, status: 'DONE' },
+              orderBy: { order: 'desc' },
+              select: { id: true, order: true, task_id: true },
+            })
+            await prisma.sprintRun.update({
+              where: { id: job.sprint_run_id },
+              data: {
+                status: 'PAUSED',
+                pause_context: {
+                  pause_reason: 'QUOTA_DEPLETED',
+                  paused_at: new Date().toISOString(),
+                  resume_instructions:
+                    'Wacht tot quota is gereset, dan resume de SprintRun via de UI. Een nieuwe SprintRun wordt gemaakt met previous_run_id en branch hergebruik.',
+                  last_completed_execution_id: lastDone?.id ?? null,
+                  last_completed_order: lastDone?.order ?? null,
+                  last_completed_task_id: lastDone?.task_id ?? null,
+                  pr_url: updated.pr_url ?? null,
+                  pr_head_sha: updated.head_sha ?? null,
+                  conflict_files: [],
+                  claude_question_id: '',
+                } as any,
+              },
+            })
+          } else {
+            const failedTaskId = (errorToWrite ?? '').match(/task[:\s]+([a-z0-9]+)/i)?.[1] ?? null
+            await prisma.sprintRun.update({
+              where: { id: job.sprint_run_id },
+              data: {
+                status: 'FAILED',
+                failure_reason: errorToWrite?.slice(0, 500) ?? null,
+                failed_task_id: failedTaskId,
+                finished_at: new Date(),
+              },
+            })
+          }
+        }
+
         // PBI-9: release product-worktree locks on terminal transitions.
         // No-op for jobs without registered locks (i.e. TASK_IMPLEMENTATION).
         if (actualStatus === 'done' || actualStatus === 'failed') {

src/tools/update-task-execution.ts

@@ -0,0 +1,110 @@
+// PBI-50 F3-T2: update_task_execution
+//
+// SPRINT_IMPLEMENTATION-flow lifecycle-tool. Worker roept dit aan voor elke
+// task in de batch om de SprintTaskExecution-row te muteren:
+//   PENDING → RUNNING → DONE/FAILED/SKIPPED
+// Idempotent: dezelfde call kan veilig herhaald worden.
+
+import { z } from 'zod'
+import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
+import { prisma } from '../prisma.js'
+import { requireWriteAccess } from '../auth.js'
+import { toolError, toolJson, withToolErrors } from '../errors.js'
+
+const inputSchema = z.object({
+  execution_id: z.string().min(1),
+  status: z.enum(['PENDING', 'RUNNING', 'DONE', 'FAILED', 'SKIPPED']),
+  base_sha: z.string().optional(),
+  head_sha: z.string().optional(),
+  skip_reason: z.string().max(2000).optional(),
+})
+
+export function registerUpdateTaskExecutionTool(server: McpServer) {
+  server.registerTool(
+    'update_task_execution',
+    {
+      title: 'Update SprintTaskExecution status',
+      description:
+        'Mutate a SprintTaskExecution row in a SPRINT_IMPLEMENTATION batch. ' +
+        'Status: PENDING|RUNNING|DONE|FAILED|SKIPPED. Worker calls this for each ' +
+        'task transition. Token must own the parent SPRINT_IMPLEMENTATION ClaudeJob. ' +
+        'Idempotent — safe to retry. Schrijft started_at (RUNNING) en finished_at ' +
+        '(DONE/FAILED/SKIPPED). Forbidden for demo accounts.',
+      inputSchema,
+    },
+    async ({ execution_id, status, base_sha, head_sha, skip_reason }) =>
+      withToolErrors(async () => {
+        const auth = await requireWriteAccess()
+
+        const execution = await prisma.sprintTaskExecution.findUnique({
+          where: { id: execution_id },
+          select: {
+            id: true,
+            sprint_job_id: true,
+            sprint_job: {
+              select: { claimed_by_token_id: true, status: true, kind: true },
+            },
+          },
+        })
+        if (!execution) {
+          return toolError(`SprintTaskExecution ${execution_id} not found`)
+        }
+        if (execution.sprint_job.kind !== 'SPRINT_IMPLEMENTATION') {
+          return toolError(
+            `Execution ${execution_id} hangs at job kind ${execution.sprint_job.kind}, expected SPRINT_IMPLEMENTATION`,
+          )
+        }
+        if (execution.sprint_job.claimed_by_token_id !== auth.tokenId) {
+          return toolError(
+            `Forbidden: token does not own SPRINT_IMPLEMENTATION job for execution ${execution_id}`,
+          )
+        }
+        if (
+          execution.sprint_job.status !== 'CLAIMED' &&
+          execution.sprint_job.status !== 'RUNNING'
+        ) {
+          return toolError(
+            `Sprint job is in terminal state ${execution.sprint_job.status}`,
+          )
+        }
+
+        const now = new Date()
+        const updated = await prisma.sprintTaskExecution.update({
+          where: { id: execution_id },
+          data: {
+            status,
+            ...(base_sha !== undefined ? { base_sha } : {}),
+            ...(head_sha !== undefined ? { head_sha } : {}),
+            ...(skip_reason !== undefined ? { skip_reason } : {}),
+            ...(status === 'RUNNING' ? { started_at: now } : {}),
+            ...(status === 'DONE' || status === 'FAILED' || status === 'SKIPPED'
+              ? { finished_at: now }
+              : {}),
+          },
+          select: {
+            id: true,
+            status: true,
+            base_sha: true,
+            head_sha: true,
+            verify_result: true,
+            verify_summary: true,
+            skip_reason: true,
+            started_at: true,
+            finished_at: true,
+          },
+        })
+
+        return toolJson({
+          execution_id: updated.id,
+          status: updated.status,
+          base_sha: updated.base_sha,
+          head_sha: updated.head_sha,
+          verify_result: updated.verify_result,
+          verify_summary: updated.verify_summary,
+          skip_reason: updated.skip_reason,
+          started_at: updated.started_at?.toISOString() ?? null,
+          finished_at: updated.finished_at?.toISOString() ?? null,
+        })
+      }),
+  )
+}

src/tools/update-task-status.ts

@@ -1,5 +1,6 @@
 import { z } from 'zod'
 import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
+import { prisma } from '../prisma.js'
 import { requireWriteAccess } from '../auth.js'
 import { userCanAccessTask } from '../access.js'
 import { toolError, toolJson, withToolErrors } from '../errors.js'
@@ -9,6 +10,10 @@ import { updateTaskStatusWithStoryPromotion } from '../lib/tasks-status-update.j
 const inputSchema = z.object({
   task_id: z.string().min(1),
   status: z.enum(TASK_STATUS_API_VALUES as [string, ...string[]]),
+  // PBI-50: optionele sprint_run_id voor SPRINT_IMPLEMENTATION-flow.
+  // Wanneer aanwezig: server valideert dat task in deze sprint zit, run
+  // actief is, en de huidige token een ClaudeJob in deze run heeft geclaimt.
+  sprint_run_id: z.string().min(1).optional(),
 })
 
 export function registerUpdateTaskStatusTool(server: McpServer) {
@@ -17,11 +22,14 @@ export function registerUpdateTaskStatusTool(server: McpServer) {
     {
       title: 'Update task status',
       description:
-        'Set the status of a task. Allowed values: todo, in_progress, review, done. ' +
+        'Set the status of a task. Allowed values: todo, in_progress, review, done, failed. ' +
+        'Optional sprint_run_id binds the update to a SPRINT_IMPLEMENTATION run for ' +
+        'cascade-propagation; the server validates that the task belongs to the sprint ' +
+        'and that the calling token has claimed a job in that run. ' +
         'Forbidden for demo accounts.',
       inputSchema,
     },
-    async ({ task_id, status }) =>
+    async ({ task_id, status, sprint_run_id }) =>
       withToolErrors(async () => {
         const auth = await requireWriteAccess()
         const dbStatus = taskStatusFromApi(status)
@@ -31,15 +39,74 @@ export function registerUpdateTaskStatusTool(server: McpServer) {
         if (!(await userCanAccessTask(task_id, auth.userId))) {
           return toolError(`Task ${task_id} not found or not accessible`)
         }
-        const { task, storyStatusChange } = await updateTaskStatusWithStoryPromotion(
+
-          task_id,
+        // PBI-50: validate explicit sprint_run_id binding.
-          dbStatus,
+        if (sprint_run_id) {
-        )
+          const sprintRun = await prisma.sprintRun.findUnique({
+            where: { id: sprint_run_id },
+            select: { id: true, status: true, sprint_id: true },
+          })
+          if (!sprintRun) {
+            return toolError(`SprintRun ${sprint_run_id} not found`)
+          }
+          if (
+            sprintRun.status !== 'QUEUED' &&
+            sprintRun.status !== 'RUNNING' &&
+            sprintRun.status !== 'PAUSED'
+          ) {
+            return toolError(
+              `SprintRun ${sprint_run_id} is in terminal state ${sprintRun.status}; cannot update task status against it`,
+            )
+          }
+
+          // Task moet in deze sprint zitten
+          const task = await prisma.task.findUnique({
+            where: { id: task_id },
+            select: { story: { select: { sprint_id: true } } },
+          })
+          if (!task || task.story.sprint_id !== sprintRun.sprint_id) {
+            return toolError(
+              `Task ${task_id} is not in sprint ${sprintRun.sprint_id} (sprint_run ${sprint_run_id})`,
+            )
+          }
+
+          // Token-coupling: huidige token moet een actieve ClaudeJob in deze
+          // SprintRun hebben geclaimt (typisch de SPRINT_IMPLEMENTATION-job).
+          const tokenJob = await prisma.claudeJob.findFirst({
+            where: {
+              sprint_run_id,
+              claimed_by_token_id: auth.tokenId,
+              status: { in: ['CLAIMED', 'RUNNING'] },
+            },
+            select: { id: true },
+          })
+          if (!tokenJob) {
+            return toolError(
+              `Forbidden: current token has no active claim in sprint_run ${sprint_run_id}`,
+            )
+          }
+        }
+
+        const { task, storyStatusChange, sprintRunChanged } =
+          await updateTaskStatusWithStoryPromotion(task_id, dbStatus, undefined, sprint_run_id)
+
+        // Voor SPRINT-flow: stuur expliciete sprint_run_status mee zodat
+        // worker zijn loop kan breken bij FAILED/PAUSED zonder extra query.
+        let sprintRunStatusChange: string | null = null
+        if (sprintRunChanged && sprint_run_id) {
+          const updated = await prisma.sprintRun.findUnique({
+            where: { id: sprint_run_id },
+            select: { status: true },
+          })
+          sprintRunStatusChange = updated?.status ?? null
+        }
+
         return toolJson({
           id: task.id,
           status: taskStatusToApi(task.status),
           implementation_plan: task.implementation_plan,
           story_status_change: storyStatusChange,
+          sprint_run_status_change: sprintRunStatusChange,
         })
       }),
   )

src/tools/verify-sprint-task.ts

@@ -0,0 +1,151 @@
+// PBI-50 F3-T1: verify_sprint_task
+//
+// Execution-aware verify-tool voor SPRINT_IMPLEMENTATION-flow.
+// Verschilt van verify_task_against_plan in:
+//   - input via execution_id (niet task_id)
+//   - base_sha komt uit SprintTaskExecution.base_sha; voor task[1..N] zonder
+//     base_sha vult de tool dynamisch met head_sha van vorige DONE-execution
+//   - plan_snapshot komt uit execution.plan_snapshot (frozen op claim-tijd)
+//   - resultaat opgeslagen op execution-row, niet op ClaudeJob.verify_result
+//   - response geeft allowed_for_done direct mee
+
+import { execFile } from 'node:child_process'
+import { promisify } from 'node:util'
+import { z } from 'zod'
+import type { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js'
+import { prisma } from '../prisma.js'
+import { requireWriteAccess } from '../auth.js'
+import { toolError, toolJson, withToolErrors } from '../errors.js'
+import { classifyDiffAgainstPlan } from '../verify/classify.js'
+import { checkVerifyGate } from './update-job-status.js'
+
+const exec = promisify(execFile)
+
+const inputSchema = z.object({
+  execution_id: z.string().min(1),
+  worktree_path: z.string().min(1),
+  summary: z.string().max(2000).optional(),
+})
+
+export function registerVerifySprintTaskTool(server: McpServer) {
+  server.registerTool(
+    'verify_sprint_task',
+    {
+      title: 'Verify SprintTaskExecution against frozen plan',
+      description:
+        'Run `git diff <base_sha>...HEAD` in the worktree and classify against the ' +
+        'frozen plan_snapshot of this SprintTaskExecution. Returns ALIGNED|PARTIAL|EMPTY|' +
+        'DIVERGENT plus reasoning + allowed_for_done (computed via the standard verify-gate ' +
+        'with the execution\'s frozen verify_required/verify_only). ' +
+        'For task[1..N] zonder base_sha vult de tool die in op basis van de head_sha van de ' +
+        'vorige DONE-execution. Optional summary is opgeslagen voor PARTIAL/DIVERGENT-rationale ' +
+        'en gebruikt door de gate. ' +
+        'Call this BEFORE update_task_execution(DONE) for each task in the sprint batch. ' +
+        'Forbidden for demo accounts.',
+      inputSchema,
+      annotations: { readOnlyHint: false },
+    },
+    async ({ execution_id, worktree_path, summary }) =>
+      withToolErrors(async () => {
+        const auth = await requireWriteAccess()
+
+        const execution = await prisma.sprintTaskExecution.findUnique({
+          where: { id: execution_id },
+          select: {
+            id: true,
+            sprint_job_id: true,
+            order: true,
+            base_sha: true,
+            plan_snapshot: true,
+            verify_required_snapshot: true,
+            verify_only_snapshot: true,
+            sprint_job: {
+              select: { claimed_by_token_id: true, status: true, kind: true },
+            },
+          },
+        })
+        if (!execution) {
+          return toolError(`SprintTaskExecution ${execution_id} not found`)
+        }
+        if (execution.sprint_job.kind !== 'SPRINT_IMPLEMENTATION') {
+          return toolError(
+            `Execution ${execution_id} hangs at job kind ${execution.sprint_job.kind}, expected SPRINT_IMPLEMENTATION`,
+          )
+        }
+        if (execution.sprint_job.claimed_by_token_id !== auth.tokenId) {
+          return toolError(
+            `Forbidden: token does not own SPRINT_IMPLEMENTATION job for execution ${execution_id}`,
+          )
+        }
+
+        // Resolve base_sha. Voor task[0] is dit gevuld bij claim. Voor
+        // task[1..N] wordt dit dynamisch gevuld op basis van de vorige
+        // DONE-execution's head_sha. Persist na fill zodat herhaalde calls
+        // dezelfde base gebruiken.
+        let baseSha = execution.base_sha
+        if (!baseSha) {
+          const previousDone = await prisma.sprintTaskExecution.findFirst({
+            where: {
+              sprint_job_id: execution.sprint_job_id,
+              order: { lt: execution.order },
+              status: 'DONE',
+              head_sha: { not: null },
+            },
+            orderBy: { order: 'desc' },
+            select: { head_sha: true },
+          })
+          if (!previousDone?.head_sha) {
+            return toolError(
+              `MISSING_BASE_SHA: execution ${execution_id} has no base_sha and no previous DONE-execution with head_sha. Did you skip update_task_execution(DONE) on a prior task?`,
+            )
+          }
+          baseSha = previousDone.head_sha
+          await prisma.sprintTaskExecution.update({
+            where: { id: execution_id },
+            data: { base_sha: baseSha },
+          })
+        }
+
+        let diff: string
+        try {
+          const { stdout } = await exec('git', ['diff', `${baseSha}...HEAD`], {
+            cwd: worktree_path,
+          })
+          diff = stdout
+        } catch (err) {
+          return toolError(
+            `git diff failed in worktree (${worktree_path}): ${(err as Error).message ?? 'unknown error'}`,
+          )
+        }
+
+        const { result, reasoning } = classifyDiffAgainstPlan({
+          diff,
+          plan: execution.plan_snapshot,
+        })
+
+        await prisma.sprintTaskExecution.update({
+          where: { id: execution_id },
+          data: {
+            verify_result: result,
+            ...(summary !== undefined ? { verify_summary: summary } : {}),
+          },
+        })
+
+        const gate = checkVerifyGate(
+          result,
+          execution.verify_only_snapshot,
+          execution.verify_required_snapshot,
+          summary,
+        )
+
+        return toolJson({
+          execution_id: execution.id,
+          result: result.toLowerCase() as 'aligned' | 'partial' | 'empty' | 'divergent',
+          reasoning,
+          base_sha: baseSha,
+          allowed_for_done: gate.allowed,
+          reason: gate.allowed ? null : gate.error,
+        })
+      }),
+  )
+}

src/tools/wait-for-job.ts

@@ -15,7 +15,9 @@ const execFileP = promisify(execFile)
 import { requireWriteAccess } from '../auth.js'
 import { toolJson, toolError, withToolErrors } from '../errors.js'
 import { createWorktreeForJob } from '../git/worktree.js'
+import { getWorktreeRoot } from '../git/worktree-paths.js'
 import { setupProductWorktrees, releaseLocksOnTerminal } from '../git/job-locks.js'
+import { pushBranchForJob } from '../git/push.js'
 
 /** Parse `https://github.com/<owner>/<name>(.git)?` → `<name>`. */
 export function repoNameFromUrl(repoUrl: string | null | undefined): string | null {
@@ -225,45 +227,96 @@ const inputSchema = z.object({
 const STALE_ERROR_MSG = 'agent did not complete job within 2 attempts'
 
 export async function resetStaleClaimedJobs(userId: string): Promise<void> {
-  // Jobs that exceeded the retry limit → FAILED
+  // PBI-50: lease-driven stale-detection. Jobs in CLAIMED of RUNNING met
-  const failedRows = await prisma.$queryRaw<
+  // verlopen lease_until (default 5 min, verlengd door job_heartbeat) worden
-    Array<{ id: string; task_id: string; product_id: string }>
+  // gereset. Legacy jobs zonder lease_until vallen terug op de oude
-  >`
+  // claimed_at + 30-min-regel.
+  type StaleRow = {
+    id: string
+    task_id: string | null
+    product_id: string
+    kind: string
+    sprint_run_id: string | null
+    branch: string | null
+  }
+
+  const failedRows = await prisma.$queryRaw<StaleRow[]>`
     UPDATE claude_jobs
     SET status = 'FAILED',
         finished_at = NOW(),
         error = ${STALE_ERROR_MSG}
     WHERE user_id = ${userId}
-      AND status = 'CLAIMED'
+      AND status IN ('CLAIMED', 'RUNNING')
-      AND claimed_at < NOW() - INTERVAL '30 minutes'
       AND retry_count >= 2
-    RETURNING id, task_id, product_id
+      AND (
+        lease_until < NOW()
+        OR (lease_until IS NULL AND claimed_at < NOW() - INTERVAL '30 minutes')
+      )
+    RETURNING id, task_id, product_id, kind::text AS kind, sprint_run_id, branch
   `
 
-  // Jobs under the retry limit → back to QUEUED, increment retry_count
   const requeuedRows = await prisma.$queryRaw<
-    Array<{ id: string; task_id: string; product_id: string; retry_count: number }>
+    (StaleRow & { retry_count: number })[]
   >`
     UPDATE claude_jobs
     SET status = 'QUEUED',
         claimed_by_token_id = NULL,
         claimed_at = NULL,
         plan_snapshot = NULL,
+        lease_until = NULL,
         retry_count = retry_count + 1
     WHERE user_id = ${userId}
-      AND status = 'CLAIMED'
+      AND status IN ('CLAIMED', 'RUNNING')
-      AND claimed_at < NOW() - INTERVAL '30 minutes'
       AND retry_count < 2
-    RETURNING id, task_id, product_id, retry_count
+      AND (
+        lease_until < NOW()
+        OR (lease_until IS NULL AND claimed_at < NOW() - INTERVAL '30 minutes')
+      )
+    RETURNING id, task_id, product_id, kind::text AS kind, sprint_run_id, branch, retry_count
   `
 
   if (failedRows.length === 0 && requeuedRows.length === 0) return
 
   // PBI-9: release any product-worktree locks held by these stale jobs.
-  // No-op for jobs without registered locks (TASK_IMPLEMENTATION).
   for (const j of failedRows) await releaseLocksOnTerminal(j.id)
   for (const j of requeuedRows) await releaseLocksOnTerminal(j.id)
 
+  // PBI-50: voor stale FAILED SPRINT_IMPLEMENTATION jobs — push de branch
+  // zodat het werk niet verloren gaat (geen mark-ready / PR-promotie),
+  // en zet SprintRun.failure_reason met een verwijzing naar de laatst
+  // RUNNING execution voor diagnose.
+  for (const j of failedRows.filter((r) => r.kind === 'SPRINT_IMPLEMENTATION')) {
+    if (j.branch && j.product_id) {
+      const repoRoot = await resolveRepoRoot(j.product_id).catch(() => null)
+      if (repoRoot) {
+        const worktreeDir = getWorktreeRoot()
+        const worktreePath = path.join(worktreeDir, j.id)
+        try {
+          await pushBranchForJob({ worktreePath, branchName: j.branch })
+        } catch (err) {
+          console.warn(`[stale-reset] push failed for stale sprint-job ${j.id}:`, err)
+        }
+      }
+    }
+    if (j.sprint_run_id) {
+      const lastRunning = await prisma.sprintTaskExecution.findFirst({
+        where: { sprint_job_id: j.id, status: 'RUNNING' },
+        orderBy: { order: 'desc' },
+        select: { order: true, task_id: true },
+      })
+      const reasonSuffix = lastRunning
+        ? `, last execution: order ${lastRunning.order} task ${lastRunning.task_id}`
+        : ''
+      await prisma.sprintRun.update({
+        where: { id: j.sprint_run_id },
+        data: {
+          status: 'FAILED',
+          failure_reason: `stale: lease verlopen${reasonSuffix}`,
+        },
+      })
+    }
+  }
+
   // Notify UI via SSE for each transition (best-effort)
   try {
     const pg = new Client({ connectionString: process.env.DATABASE_URL })
@@ -308,12 +361,15 @@ export async function tryClaimJob(
 ): Promise<string | null> {
   // Atomic claim in a single transaction — also captures plan_snapshot from task.
   //
-  // Sprint-flow filter (PBI-46):
+  // PBI-50: claim-filter discrimineert via cj.kind:
-  //   Idea-jobs (task_id IS NULL) blijven onafhankelijk claimable.
+  //   - IDEA_GRILL/IDEA_MAKE_PLAN/PLAN_CHAT: standalone idea-jobs.
-  //   Task-jobs zijn alleen claimable wanneer ze aan een actieve SprintRun
+  //   - TASK_IMPLEMENTATION/SPRINT_IMPLEMENTATION: alleen via actieve SprintRun
-  //   hangen (status QUEUED of RUNNING). Legacy task-jobs zonder sprint_run_id
+  //     (status QUEUED of RUNNING). Legacy task-jobs zonder sprint_run_id en
-  //   en jobs in PAUSED/FAILED/CANCELLED/DONE SprintRuns worden overgeslagen.
+  //     jobs in PAUSED/FAILED/CANCELLED/DONE SprintRuns worden overgeslagen.
   //   Bij eerste claim van een nog QUEUED SprintRun → status RUNNING.
+  //
+  // PBI-50 lease: lease_until = NOW() + 5min op claim. resetStaleClaimedJobs
+  // reset bij verlopen lease.
   const rows = await prisma.$transaction(async (tx) => {
     const found = productId
       ? await tx.$queryRaw<
@@ -327,8 +383,10 @@ export async function tryClaimJob(
             AND cj.product_id = ${productId}
             AND cj.status = 'QUEUED'
             AND (
-              cj.task_id IS NULL
+              cj.kind IN ('IDEA_GRILL', 'IDEA_MAKE_PLAN', 'PLAN_CHAT')
-              OR (cj.sprint_run_id IS NOT NULL AND sr.status IN ('QUEUED', 'RUNNING'))
+              OR (cj.kind IN ('TASK_IMPLEMENTATION', 'SPRINT_IMPLEMENTATION')
+                  AND cj.sprint_run_id IS NOT NULL
+                  AND sr.status IN ('QUEUED', 'RUNNING'))
             )
           ORDER BY cj.created_at ASC
           LIMIT 1
@@ -344,8 +402,10 @@ export async function tryClaimJob(
           WHERE cj.user_id = ${userId}
             AND cj.status = 'QUEUED'
             AND (
-              cj.task_id IS NULL
+              cj.kind IN ('IDEA_GRILL', 'IDEA_MAKE_PLAN', 'PLAN_CHAT')
-              OR (cj.sprint_run_id IS NOT NULL AND sr.status IN ('QUEUED', 'RUNNING'))
+              OR (cj.kind IN ('TASK_IMPLEMENTATION', 'SPRINT_IMPLEMENTATION')
+                  AND cj.sprint_run_id IS NOT NULL
+                  AND sr.status IN ('QUEUED', 'RUNNING'))
             )
           ORDER BY cj.created_at ASC
           LIMIT 1
@@ -362,7 +422,8 @@ export async function tryClaimJob(
       SET status = 'CLAIMED',
           claimed_by_token_id = ${tokenId},
           claimed_at = NOW(),
-          plan_snapshot = ${snapshot}
+          plan_snapshot = ${snapshot},
+          lease_until = NOW() + INTERVAL '5 minutes'
       WHERE id = ${jobId}
     `
 
@@ -482,6 +543,177 @@ async function getFullJobContext(jobId: string) {
     }
   }
 
+  // PBI-50: SPRINT_IMPLEMENTATION — single-session sprint runner.
+  // Eén ClaudeJob per SprintRun handelt sequentieel alle TO_DO-tasks af.
+  // Bij claim: maak frozen scope-snapshot via SprintTaskExecution-rows,
+  // resolve worktree (verse branch of hergebruikt via previous_run_id),
+  // capture base_sha. Worker werkt uitsluitend op deze frozen snapshot.
+  if (job.kind === 'SPRINT_IMPLEMENTATION') {
+    if (!job.sprint_run_id) {
+      await rollbackClaim(job.id)
+      return null
+    }
+    const sprintRun = await prisma.sprintRun.findUnique({
+      where: { id: job.sprint_run_id },
+      include: {
+        sprint: {
+          include: {
+            product: true,
+            stories: {
+              where: { status: { not: 'DONE' } },
+              include: {
+                pbi: {
+                  select: { id: true, code: true, title: true, priority: true, sort_order: true, status: true },
+                },
+                tasks: {
+                  where: { status: 'TO_DO' },
+                  orderBy: [{ priority: 'asc' }, { sort_order: 'asc' }],
+                },
+              },
+              orderBy: [{ priority: 'asc' }, { sort_order: 'asc' }],
+            },
+          },
+        },
+      },
+    })
+    if (!sprintRun) {
+      await rollbackClaim(job.id)
+      return null
+    }
+
+    const repoRoot = await resolveRepoRoot(sprintRun.sprint.product_id)
+    if (!repoRoot) {
+      await rollbackClaim(job.id)
+      return null
+    }
+
+    // Branch resolution: previous_run_id + branch → reuse; anders verse.
+    const isResume = !!(sprintRun.previous_run_id && sprintRun.branch)
+    const branchName = isResume
+      ? sprintRun.branch!
+      : `feat/sprint-${job.sprint_run_id.slice(-8)}`
+
+    let worktreePath: string
+    let baseSha: string
+    try {
+      const wt = await createWorktreeForJob({
+        repoRoot,
+        jobId: job.id,
+        branchName,
+        reuseBranch: isResume,
+      })
+      worktreePath = wt.worktreePath
+
+      const { stdout: headSha } = await execFileP('git', ['rev-parse', 'HEAD'], {
+        cwd: worktreePath,
+      })
+      baseSha = headSha.trim()
+    } catch (err) {
+      console.warn(`[wait-for-job] sprint-worktree setup failed for ${job.id}:`, err)
+      await rollbackClaim(job.id)
+      return null
+    }
+
+    // Verzamel ordered tasks in flat list, behoud volgorde
+    const orderedTasks = sprintRun.sprint.stories.flatMap((s) =>
+      s.tasks.map((t) => ({ ...t, story_pbi_id: s.pbi.id })),
+    )
+
+    // Persist branch + base_sha + scope-snapshot in één transactie
+    await prisma.$transaction([
+      prisma.claudeJob.update({
+        where: { id: job.id },
+        data: { branch: branchName, base_sha: baseSha },
+      }),
+      prisma.sprintTaskExecution.createMany({
+        data: orderedTasks.map((t, idx) => ({
+          sprint_job_id: job.id,
+          task_id: t.id,
+          order: idx,
+          plan_snapshot: t.implementation_plan ?? '',
+          verify_required_snapshot: t.verify_required,
+          verify_only_snapshot: t.verify_only,
+          base_sha: idx === 0 ? baseSha : null,
+          status: 'PENDING' as const,
+        })),
+      }),
+      prisma.sprintRun.update({
+        where: { id: job.sprint_run_id },
+        data: { branch: branchName },
+      }),
+    ])
+
+    // Lookup execution_ids in volgorde voor de response
+    const executions = await prisma.sprintTaskExecution.findMany({
+      where: { sprint_job_id: job.id },
+      orderBy: { order: 'asc' },
+      select: { id: true, task_id: true, order: true, base_sha: true },
+    })
+    const execIdByTaskId = new Map(executions.map((e) => [e.task_id, e.id]))
+
+    // Dedupe PBIs uit de stories (één PBI kan meerdere stories hebben)
+    const pbiMap = new Map<string, typeof sprintRun.sprint.stories[number]['pbi']>()
+    for (const s of sprintRun.sprint.stories) pbiMap.set(s.pbi.id, s.pbi)
+
+    return {
+      job_id: job.id,
+      kind: job.kind,
+      status: 'claimed',
+      sprint: {
+        id: sprintRun.sprint.id,
+        sprint_goal: sprintRun.sprint.sprint_goal,
+        status: sprintRun.sprint.status,
+      },
+      sprint_run: {
+        id: sprintRun.id,
+        pr_strategy: sprintRun.pr_strategy,
+        branch: branchName,
+        previous_run_id: sprintRun.previous_run_id,
+      },
+      product: {
+        id: sprintRun.sprint.product.id,
+        name: sprintRun.sprint.product.name,
+        repo_url: sprintRun.sprint.product.repo_url,
+        definition_of_done: sprintRun.sprint.product.definition_of_done,
+        auto_pr: sprintRun.sprint.product.auto_pr,
+      },
+      pbis: Array.from(pbiMap.values()).map((p) => ({
+        id: p.id,
+        code: p.code,
+        title: p.title,
+        priority: p.priority,
+        sort_order: p.sort_order,
+        status: p.status,
+      })),
+      stories: sprintRun.sprint.stories.map((s) => ({
+        id: s.id,
+        code: s.code,
+        title: s.title,
+        pbi_id: s.pbi_id,
+        priority: s.priority,
+        sort_order: s.sort_order,
+        status: s.status,
+      })),
+      task_executions: orderedTasks.map((t, idx) => ({
+        execution_id: execIdByTaskId.get(t.id)!,
+        task_id: t.id,
+        code: t.code,
+        title: t.title,
+        story_id: t.story_id,
+        order: idx,
+        plan_snapshot: t.implementation_plan ?? '',
+        verify_required: t.verify_required,
+        verify_only: t.verify_only,
+        base_sha: idx === 0 ? baseSha : null,
+      })),
+      worktree_path: worktreePath,
+      branch_name: branchName,
+      repo_url: sprintRun.sprint.product.repo_url,
+      base_sha: baseSha,
+      heartbeat_interval_seconds: 60,
+    }
+  }
+
   // TASK_IMPLEMENTATION (default) — bestaande gedrag onaangetast.
   const { task } = job
   if (!task) return null