claude-harness/dsh/tests/approve.test.mjs

import assert from 'node:assert/strict'
import { createHash } from 'node:crypto'
import { existsSync, readFileSync } from 'node:fs'
import { join } from 'node:path'
import test from 'node:test'
import { registerApproveCommand } from '../bundle/command-approve.js'
import { hashPlan, parseApproval, serializeApproval } from '../bundle/lib/gates.js'
import { createApproveTool } from '../bundle/tool-approve.js'
import { cleanup, tempDir, write } from './helpers.mjs'

const PLAN_TEXT = [
  '# Plan X',
  '',
  '**Spec:** `openspec/changes/demo/`',
  '',
  '## Task 1',
  '',
  'Do the work.',
  '',
].join('\n')

const ISO = /^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3}Z$/u
const APPROVE = 'Approve'

/** A throwaway repository with the design artifacts and one demo plan. */
function repoFixture({ builderCheck = true, withPlan = true } = {}) {
  const tmp = tempDir('dsh-approve-')
  const repo = join(tmp, 'repo')
  write(join(repo, 'openspec', 'changes', 'demo', 'proposal.md'), '# Proposal\n')
  write(join(repo, 'openspec', 'changes', 'demo', 'design.md'), '# Design\n')
  if (builderCheck) write(join(repo, 'openspec', 'changes', 'demo', 'builder-check.md'), '# Check\n')
  let planPath = null
  if (withPlan) {
    planPath = join(repo, 'docs', 'superpowers', 'plans', 'plan-x.md')
    write(planPath, PLAN_TEXT)
  }
  return { tmp, repo, planPath }
}

function approvalPath(repo) {
  return join(repo, 'openspec', 'changes', 'demo', 'approval.md')
}

/** An ask() answer batch shaped like the DSH UserQuestionService result. */
function approveAnswer(kind = 'design') {
  return { answers: [{ id: `approve-${kind}`, selected: [APPROVE], custom: undefined }] }
}

/**
 * Wire the command against a fake `commands.register` that captures the
 * registered definition, and a scripted `userQuestions.ask`. The fake
 * `workingDirectory` resolves every session to the fixture repo.
 */
function setup(repo, { answer, missingService = false } = {}) {
  const captured = {}
  const commands = {
    register(definition) {
      Object.assign(captured, definition)
    },
  }
  const calls = []
  let userQuestions
  if (!missingService) {
    userQuestions = {
      async ask(request) {
        calls.push(request)
        return answer
      },
    }
  }
  registerApproveCommand({
    commands,
    userQuestions,
    workingDirectory: { get() { return repo } },
    logger: { warn() {} },
  })
  return { captured, calls, run: (rawInput) => captured.handler({ rawInput }) }
}

function shaOf(path) {
  return createHash('sha256').update(readFileSync(path)).digest('hex')
}

test('approve design refuses without builder-check.md', async () => {
  const { tmp, repo } = repoFixture({ builderCheck: false })
  try {
    const harness = setup(repo, { answer: approveAnswer('design') })
    const result = await harness.run('demo design')

    assert.match(result.text, /builder-check\.md/u)
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
    assert.equal(harness.calls.length, 0, 'the owner is not asked when artifacts are missing')
  } finally {
    cleanup(tmp)
  }
})

test('approve design writes design section (by owner, at ISO)', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const existingPlan = { by: 'aep', at: '2026-10-25T11:30:00+04:00', sha256: 'a'.repeat(64) }
    write(approvalPath(repo), serializeApproval({ plan: existingPlan }))

    const harness = setup(repo, { answer: approveAnswer('design') })
    const result = await harness.run('demo design')

    assert.equal(result.kind, 'success')
    const request = harness.calls[0].questions[0]
    assert.equal(request.question, 'Approve demo design?')

    const parsed = parseApproval(readFileSync(approvalPath(repo), 'utf8'))
    assert.equal(parsed.design.by, 'owner')
    assert.match(parsed.design.at, ISO)
    assert.equal(Number.isNaN(Date.parse(parsed.design.at)), false)
    assert.deepEqual(parsed.plan, existingPlan, 'the pre-existing plan section is untouched')
  } finally {
    cleanup(tmp)
  }
})

test('approve design refuses without confirmation', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const harness = setup(repo, {
      answer: { answers: [{ id: 'approve-design', selected: ['Cancel'] }] },
    })
    const result = await harness.run('demo design')

    assert.match(result.text, /not confirmed/u)
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
  } finally {
    cleanup(tmp)
  }
})

test('approve plan refuses without design section', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const harness = setup(repo, { answer: approveAnswer('plan') })
    const result = await harness.run('demo plan')

    assert.match(result.text, /design not approved/u)
    assert.match(result.text, /\/approve demo design/u)
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
    assert.equal(harness.calls.length, 0, 'the owner is not asked before the design gate passes')
  } finally {
    cleanup(tmp)
  }
})

test('approve plan refuses when two plans reference the change', async () => {
  const { tmp, repo } = repoFixture()
  try {
    write(join(repo, 'docs', 'superpowers', 'plans', 'plan-y.md'), PLAN_TEXT)
    write(
      approvalPath(repo),
      serializeApproval({ design: { by: 'aep', at: '2026-10-25T10:00:00+04:00' } }),
    )
    const before = readFileSync(approvalPath(repo), 'utf8')

    const harness = setup(repo, { answer: approveAnswer('plan') })
    const result = await harness.run('demo plan')

    assert.match(result.text, /plan-x\.md/u)
    assert.match(result.text, /plan-y\.md/u)
    assert.equal(readFileSync(approvalPath(repo), 'utf8'), before, 'approval.md is untouched')
    assert.equal(harness.calls.length, 0, 'the owner is not asked when the plan is ambiguous')
  } finally {
    cleanup(tmp)
  }
})

test('approve plan writes plan section with sha256 of the plan', async () => {
  const { tmp, repo, planPath } = repoFixture()
  try {
    const design = { by: 'aep', at: '2026-10-25T10:00:00+04:00' }
    write(approvalPath(repo), serializeApproval({ design }))

    const harness = setup(repo, { answer: approveAnswer('plan') })
    const result = await harness.run('demo plan')

    assert.equal(result.kind, 'success')
    const request = harness.calls[0].questions[0]
    assert.equal(request.question, 'Approve demo plan?')

    const parsed = parseApproval(readFileSync(approvalPath(repo), 'utf8'))
    assert.equal(parsed.plan.by, 'owner')
    assert.match(parsed.plan.at, ISO)
    assert.equal(parsed.plan.sha256, shaOf(planPath))
    assert.equal(parsed.plan.sha256, hashPlan(planPath))
    assert.deepEqual(parsed.design, design, 'the design section is untouched')
  } finally {
    cleanup(tmp)
  }
})

test('approve refuses when userQuestions is missing', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const harness = setup(repo, { missingService: true })
    const result = await harness.run('demo design')

    assert.match(result.text, /userQuestions service unavailable/u)
    assert.match(result.text, /\/approve requires it/u)
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
  } finally {
    cleanup(tmp)
  }
})

test('approve with unknown kind returns usage', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const harness = setup(repo, { answer: approveAnswer() })

    const unknown = await harness.run('demo review')
    assert.match(unknown.text, /Usage: \/approve <change-id> design\|plan/u)
    assert.equal(unknown.kind, 'error')

    const missing = await harness.run('demo')
    assert.match(missing.text, /Usage: \/approve <change-id> design\|plan/u)

    assert.equal(harness.calls.length, 0, 'nothing is asked')
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
  } finally {
    cleanup(tmp)
  }
})

// Pins for the Task 2 review findings (R-5): branches without a first test.

test('approve plan refuses when no plan references the change', async () => {
  const { tmp, repo } = repoFixture({ withPlan: false })
  try {
    write(
      approvalPath(repo),
      serializeApproval({ design: { by: 'aep', at: '2026-10-25T10:00:00+04:00' } }),
    )
    const before = readFileSync(approvalPath(repo), 'utf8')

    const harness = setup(repo, { answer: approveAnswer('plan') })
    const result = await harness.run('demo plan')

    assert.equal(result.kind, 'error')
    assert.match(result.text, /no plan for demo found/u)
    assert.match(result.text, /docs\/superpowers\/plans\//u)
    assert.equal(readFileSync(approvalPath(repo), 'utf8'), before, 'approval.md is untouched')
    assert.equal(harness.calls.length, 0, 'the owner is not asked without a plan')
  } finally {
    cleanup(tmp)
  }
})

test('approve refuses an unsafe change id', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const harness = setup(repo, { answer: approveAnswer('design') })

    const result = await harness.run('../x design')
    assert.match(result.text, /Invalid change id "\.\.\/x"\./u)
    assert.match(result.text, /Usage: \/approve <change-id> design\|plan/u)
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
    assert.equal(harness.calls.length, 0, 'the owner is not asked for an invalid id')
  } finally {
    cleanup(tmp)
  }
})

test('approve accepts the fallback answer shapes {approved}/{confirmed}/true', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const shapes = [{ approved: true }, { confirmed: true }, true]
    for (const answer of shapes) {
      const harness = setup(repo, { answer })
      const result = await harness.run('demo design')
      assert.equal(result.kind, 'success', `answer shape ${JSON.stringify(answer)} approves`)
      const parsed = parseApproval(readFileSync(approvalPath(repo), 'utf8'))
      assert.equal(parsed.design.by, 'owner')
      assert.match(parsed.design.at, ISO)
    }
  } finally {
    cleanup(tmp)
  }
})

// T3.1 regression: the repo root is resolved per handler invocation from the
// invocation agent's session, not fixed at registration.

test('approve resolves the session cwd per invocation (T3.1)', async () => {
  const first = repoFixture()
  const second = repoFixture()
  try {
    const roots = new Map([
      ['s1', first.repo],
      ['s2', second.repo],
    ])
    const captured = {}
    const commands = { register(definition) { Object.assign(captured, definition) } }
    registerApproveCommand({
      commands,
      userQuestions: { async ask() { return approveAnswer('design') } },
      workingDirectory: { get(session) { return roots.get(session) } },
      logger: { warn() {} },
    })

    const firstResult = await captured.handler({ agent: { session: 's1' }, rawInput: 'demo design' })
    assert.equal(firstResult.kind, 'success')
    assert.equal(existsSync(approvalPath(first.repo)), true, 'approval.md written in the s1 repo')
    assert.equal(existsSync(approvalPath(second.repo)), false, 'the s2 repo is untouched')

    const secondResult = await captured.handler({ agent: { session: 's2' }, rawInput: 'demo design' })
    assert.equal(secondResult.kind, 'success')
    assert.equal(existsSync(approvalPath(second.repo)), true, 'approval.md written in the s2 repo')
  } finally {
    cleanup(first.tmp)
    cleanup(second.tmp)
  }
})

// T3.9 pin: the process.cwd() fallback of currentDirectory is part of the
// T3.1 contract — a workingDirectory that throws must not break /approve.
// The test changes the process directory (each test file runs in its own
// process) so the write lands in an assertable place, then restores it.

test('approve falls back to process.cwd() when the session cwd lookup throws (T3.9)', async () => {
  const { tmp, repo } = repoFixture()
  const previous = process.cwd()
  try {
    process.chdir(repo)
    const captured = {}
    const commands = { register(definition) { Object.assign(captured, definition) } }
    registerApproveCommand({
      commands,
      userQuestions: { async ask() { return approveAnswer('design') } },
      workingDirectory: {
        get() {
          throw new Error('cannot get property "session" without inject')
        },
      },
      logger: { warn() {} },
    })

    const result = await captured.handler({ agent: { session: 's1' }, rawInput: 'demo design' })
    assert.equal(result.kind, 'success')
    assert.equal(existsSync(approvalPath(repo)), true, 'approval.md is written into process.cwd()')
  } finally {
    process.chdir(previous)
    cleanup(tmp)
  }
})

// `request_approval` (bounded #77): the model drives the approval gate, the
// owner only confirms in the userQuestions dialog, and the marker is still
// written by bundle code through the same lib/approval.js as `/approve`.

const FIXED_ISO = '2026-10-25T12:00:00.000Z'

/**
 * Build the tool with fakes: `getWorkingDirectory` is the DSH service factory
 * (`workingDirectory.get(session)`), `getUserQuestions` resolves lazily, and
 * `now` is pinned so the written `at` is exact.
 */
function setupTool(repo, { answer, missingService = false, at = FIXED_ISO } = {}) {
  const calls = []
  const getUserQuestions = missingService
    ? () => null
    : () => ({
        async ask(request) {
          calls.push(request)
          return answer
        },
      })
  const tool = createApproveTool({
    getWorkingDirectory: () => ({ get() { return repo } }),
    getUserQuestions,
    logger: { warn() {} },
    now: () => new Date(at),
  })
  return { tool, calls, run: (args, exec = {}) => tool.execute(args, exec) }
}

test('request_approval design: owner confirms, section is written by the harness', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const harness = setupTool(repo, { answer: approveAnswer('design') })
    assert.equal(harness.tool.name, 'request_approval')
    assert.match(harness.tool.description, /written by the harness only after the owner confirms/u)
    assert.deepEqual(harness.tool.parameters.required, ['change_id', 'kind'])
    assert.deepEqual(harness.tool.parameters.properties.kind.enum, ['design', 'plan'])
    assert.deepEqual(harness.tool.output.render({}, { approved: true, message: 'ok' }), [
      { type: 'text', text: 'ok' },
    ])

    const result = await harness.run({ change_id: 'demo', kind: 'design' }, { agent: { session: 's1' } })

    assert.equal(result.approved, true)
    assert.equal(result.message, `Approved demo design at ${FIXED_ISO}; wrote openspec/changes/demo/approval.md.`)

    const question = harness.calls[0].questions[0]
    assert.equal(question.question, 'Approve demo design?')
    assert.equal(question.header, 'Approval gate')
    assert.deepEqual(question.options.map((option) => option.label), ['Approve', 'Cancel'])
    assert.equal(harness.calls[0].agent.session, 's1', 'the receiving agent is forwarded')

    const parsed = parseApproval(readFileSync(approvalPath(repo), 'utf8'))
    assert.equal(parsed.design.by, 'owner')
    assert.equal(parsed.design.at, FIXED_ISO)
  } finally {
    cleanup(tmp)
  }
})

test('request_approval design: cancel writes nothing', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const harness = setupTool(repo, {
      answer: { answers: [{ id: 'approve', selected: ['Cancel'], custom: undefined }] },
    })
    const result = await harness.run({ change_id: 'demo', kind: 'design' })

    assert.equal(result.approved, false)
    assert.match(result.message, /not confirmed/u)
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
  } finally {
    cleanup(tmp)
  }
})

test('request_approval refuses when the userQuestions service is missing', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const harness = setupTool(repo, { missingService: true })
    const result = await harness.run({ change_id: 'demo', kind: 'design' })

    assert.equal(result.approved, false)
    assert.match(result.message, /userQuestions service unavailable/u)
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
  } finally {
    cleanup(tmp)
  }
})

test('request_approval refuses an unknown kind with usage', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const harness = setupTool(repo, { answer: approveAnswer('design') })
    const result = await harness.run({ change_id: 'demo', kind: 'review' })

    assert.equal(result.approved, false)
    assert.match(result.message, /usage: request_approval/u)
    assert.equal(harness.calls.length, 0, 'the owner is not asked')
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
  } finally {
    cleanup(tmp)
  }
})

test('request_approval refuses an unsafe change id', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const harness = setupTool(repo, { answer: approveAnswer('design') })
    const result = await harness.run({ change_id: '../x', kind: 'design' })

    assert.equal(result.approved, false)
    assert.match(result.message, /Invalid change id "\.\.\/x"/u)
    assert.equal(harness.calls.length, 0, 'the owner is not asked')
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
  } finally {
    cleanup(tmp)
  }
})

test('request_approval plan writes the sha256 of the plan', async () => {
  const { tmp, repo, planPath } = repoFixture()
  try {
    const design = { by: 'aep', at: '2026-10-25T10:00:00+04:00' }
    write(approvalPath(repo), serializeApproval({ design }))

    const harness = setupTool(repo, { answer: approveAnswer('plan') })
    const result = await harness.run({ change_id: 'demo', kind: 'plan' })

    assert.equal(result.approved, true)
    assert.equal(
      result.message,
      `Approved demo plan at ${FIXED_ISO} (sha256 ${shaOf(planPath)}); wrote openspec/changes/demo/approval.md.`,
    )

    const parsed = parseApproval(readFileSync(approvalPath(repo), 'utf8'))
    assert.equal(parsed.plan.by, 'owner')
    assert.equal(parsed.plan.at, FIXED_ISO)
    assert.equal(parsed.plan.sha256, shaOf(planPath))
    assert.deepEqual(parsed.design, design, 'the design section is untouched')
  } finally {
    cleanup(tmp)
  }
})

test('request_approval design refuses without builder-check.md (prerequisites hold)', async () => {
  const { tmp, repo } = repoFixture({ builderCheck: false })
  try {
    const harness = setupTool(repo, { answer: approveAnswer('design') })
    const result = await harness.run({ change_id: 'demo', kind: 'design' })

    assert.equal(result.approved, false)
    assert.match(result.message, /builder-check\.md/u)
    assert.equal(harness.calls.length, 0, 'the owner is not asked before the prerequisites pass')
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
  } finally {
    cleanup(tmp)
  }
})

test('request_approval falls back to process.cwd() when the session cwd lookup throws', async () => {
  const { tmp, repo } = repoFixture()
  const previous = process.cwd()
  try {
    process.chdir(repo)
    const tool = createApproveTool({
      getWorkingDirectory: () => ({
        get() {
          throw new Error('cannot get property "session" without inject')
        },
      }),
      getUserQuestions: () => ({ async ask() { return approveAnswer('design') } }),
      logger: { warn() {} },
      now: () => new Date(FIXED_ISO),
    })

    const result = await tool.execute({ change_id: 'demo', kind: 'design' }, { agent: { session: 's1' } })
    assert.equal(result.approved, true)
    assert.equal(existsSync(approvalPath(repo)), true, 'approval.md is written into process.cwd()')
  } finally {
    process.chdir(previous)
    cleanup(tmp)
  }
})

// Refactor guard (bounded #77): `/approve` still reports a failing
// userQuestions service as a refusal with the service's reason, and writes
// nothing. The shared runApproval only writes after a `true` answer.
test('approve reports a userQuestions failure and writes nothing', async () => {
  const { tmp, repo } = repoFixture()
  try {
    const captured = {}
    const commands = { register(definition) { Object.assign(captured, definition) } }
    registerApproveCommand({
      commands,
      userQuestions: {
        async ask() {
          throw new Error('dialog closed')
        },
      },
      workingDirectory: { get() { return repo } },
      logger: { warn() {} },
    })

    const result = await captured.handler({ rawInput: 'demo design' })

    assert.equal(result.kind, 'error')
    assert.equal(result.text, 'Refused: userQuestions failed: dialog closed')
    assert.equal(existsSync(approvalPath(repo)), false, 'no approval.md is written')
  } finally {
    cleanup(tmp)
  }
})