Merge branch 'fix-tests-suite-loop'

# Conflicts: # package-lock.json # packages/server/src/server/speech/speech-config-resolver.ts # packages/server/src/shared/tool-call-display.ts
2026-03-14 22:06:42 +07:00
parent 3ae1102d13 a495138374
commit e526f375c8
50 changed files with 1330 additions and 2544 deletions
@@ -15,12 +15,10 @@
 */

 import assert from 'node:assert'
-import { $ } from 'zx'
 import { mkdtemp, rm } from 'fs/promises'
 import { tmpdir } from 'os'
 import { join } from 'path'
-
-$.verbose = false
+import { runLocalPaseo } from './helpers/local-cli.ts'

 console.log('=== Daemon Commands ===\n')

@@ -28,11 +26,15 @@ console.log('=== Daemon Commands ===\n')
 const port = 10000 + Math.floor(Math.random() * 50000)
 const paseoHome = await mkdtemp(join(tmpdir(), 'paseo-test-home-'))

+function daemonCommand(args: string[]) {
+  return runLocalPaseo(['daemon', ...args], { PASEO_HOME: paseoHome })
+}
+
 try {
  // Test 1: daemon --help shows subcommands
  {
    console.log('Test 1: daemon --help shows subcommands')
-    const result = await $`npx paseo daemon --help`.nothrow()
+    const result = await runLocalPaseo(['daemon', '--help'])
    assert.strictEqual(result.exitCode, 0, 'daemon --help should exit 0')
    assert(result.stdout.includes('start'), 'help should mention start')
    assert(result.stdout.includes('status'), 'help should mention status')
@@ -45,8 +47,7 @@ try {
  // Test 2: daemon pair works without daemon process
  {
    console.log('Test 2: daemon pair prints local pairing URL')
-    const result =
-      await $`PASEO_HOME=${paseoHome} npx paseo daemon pair`.nothrow()
+    const result = await daemonCommand(['pair'])
    assert.strictEqual(result.exitCode, 0, 'daemon pair should succeed')
    assert(result.stdout.includes('Scan to pair:'), 'output should include scan header')
    assert(result.stdout.includes('#offer='), 'output should include pairing offer fragment')
@@ -56,8 +57,7 @@ try {
  // Test 3: daemon status reports stopped when daemon not running
  {
    console.log('Test 3: daemon status reports stopped when not running')
-    const result =
-      await $`PASEO_HOME=${paseoHome} npx paseo daemon status`.nothrow()
+    const result = await daemonCommand(['status'])
    assert.strictEqual(result.exitCode, 0, 'status should succeed when daemon is stopped')
    const output = result.stdout.toLowerCase()
    assert(output.includes('status'), 'status table should include Status row')
@@ -68,8 +68,7 @@ try {
  // Test 4: daemon pair --json outputs valid JSON
  {
    console.log('Test 4: daemon pair --json outputs JSON')
-    const result =
-      await $`PASEO_HOME=${paseoHome} npx paseo daemon pair --json`.nothrow()
+    const result = await daemonCommand(['pair', '--json'])
    assert.strictEqual(result.exitCode, 0, 'daemon pair --json should succeed')
    const pairing = JSON.parse(result.stdout)
    assert.strictEqual(pairing.relayEnabled, true, 'pairing should report relay enabled')
@@ -81,8 +80,7 @@ try {
  // Test 5: daemon status --json outputs valid JSON
  {
    console.log('Test 5: daemon status --json outputs JSON')
-    const result =
-      await $`PASEO_HOME=${paseoHome} npx paseo daemon status --json`.nothrow()
+    const result = await daemonCommand(['status', '--json'])
    assert.strictEqual(result.exitCode, 0, '--json status should succeed')
    const status = JSON.parse(result.stdout)
    assert.strictEqual(typeof status.serverId, 'string', 'json status should include serverId')
@@ -95,8 +93,7 @@ try {
  // Test 6: daemon stop handles daemon not running gracefully
  {
    console.log('Test 6: daemon stop handles daemon not running')
-    const result =
-      await $`PASEO_HOME=${paseoHome} npx paseo daemon stop`.nothrow()
+    const result = await daemonCommand(['stop'])
    // Stop should succeed even if daemon is not running (idempotent).
    assert.strictEqual(result.exitCode, 0, 'stop should succeed when daemon not running')
    const output = result.stdout + result.stderr
@@ -110,18 +107,17 @@ try {
  // Test 7: daemon restart starts daemon and can be stopped
  {
    console.log('Test 7: daemon restart starts daemon and can be stopped')
-    const result =
-      await $`PASEO_HOME=${paseoHome} npx paseo daemon restart --port ${String(port)}`.nothrow()
+    const result = await daemonCommand(['restart', '--port', String(port)])
    assert.strictEqual(result.exitCode, 0, 'restart should succeed even when previously stopped')
    assert(result.stdout.toLowerCase().includes('restarted'), 'output should report restart')

-    const cleanup = await $`PASEO_HOME=${paseoHome} npx paseo daemon stop --force`.nothrow()
+    const cleanup = await daemonCommand(['stop', '--force'])
    assert.strictEqual(cleanup.exitCode, 0, 'cleanup stop should succeed after restart')
    console.log('✓ daemon restart starts and stop cleanup succeeds\n')
  }
 } finally {
  // Best-effort daemon cleanup in case assertions fail before explicit stop.
-  await $`PASEO_HOME=${paseoHome} npx paseo daemon stop --force`.nothrow()
+  await daemonCommand(['stop', '--force'])
  // Clean up temp directory
  await rm(paseoHome, { recursive: true, force: true })
 }
@@ -20,12 +20,10 @@
 */

 import assert from 'node:assert'
-import { $ } from 'zx'
 import { mkdtemp, rm } from 'fs/promises'
 import { tmpdir } from 'os'
 import { join } from 'path'
-
-$.verbose = false
+import { runLocalPaseo } from './helpers/local-cli.ts'

 console.log('=== LS Command Tests ===\n')

@@ -37,7 +35,7 @@ try {
  // Test 1: paseo --help shows ls command
  {
    console.log('Test 1: paseo --help shows ls command')
-    const result = await $`npx paseo --help`.nothrow()
+    const result = await runLocalPaseo(['--help'])
    assert.strictEqual(result.exitCode, 0, 'paseo --help should exit 0')
    assert(result.stdout.includes('ls'), 'help should mention ls command')
    console.log('✓ paseo --help shows ls command\n')
@@ -46,7 +44,7 @@ try {
  // Test 2: paseo ls --help shows options
  {
    console.log('Test 2: paseo ls --help shows options')
-    const result = await $`npx paseo ls --help`.nothrow()
+    const result = await runLocalPaseo(['ls', '--help'])
    assert.strictEqual(result.exitCode, 0, 'paseo ls --help should exit 0')
    assert(result.stdout.includes('-a'), 'help should mention -a flag')
    assert(result.stdout.includes('--all'), 'help should mention --all flag')
@@ -60,8 +58,10 @@ try {
  // Test 3: paseo ls returns error when no daemon running
  {
    console.log('Test 3: paseo ls handles daemon not running')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npx paseo ls`.nothrow()
+    const result = await runLocalPaseo(['ls'], {
+      PASEO_HOST: `localhost:${port}`,
+      PASEO_HOME: paseoHome,
+    })
    // Should fail because daemon not running
    assert.notStrictEqual(result.exitCode, 0, 'should fail when daemon not running')
    const output = result.stdout + result.stderr
@@ -76,8 +76,10 @@ try {
  // Test 4: paseo ls --json returns valid JSON error
  {
    console.log('Test 4: paseo ls --json handles errors')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npx paseo ls --json`.nothrow()
+    const result = await runLocalPaseo(['ls', '--json'], {
+      PASEO_HOST: `localhost:${port}`,
+      PASEO_HOME: paseoHome,
+    })
    // Should still fail (daemon not running)
    assert.notStrictEqual(result.exitCode, 0, 'should fail when daemon not running')
    // But output should be valid JSON if present
@@ -98,8 +100,10 @@ try {
  // Test 5: paseo ls -a flag is accepted
  {
    console.log('Test 5: paseo ls -a flag is accepted')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npx paseo ls -a`.nothrow()
+    const result = await runLocalPaseo(['ls', '-a'], {
+      PASEO_HOST: `localhost:${port}`,
+      PASEO_HOME: paseoHome,
+    })
    // Will fail due to no daemon, but flag should be parsed without error
    // (no "unknown option" error)
    const output = result.stdout + result.stderr
@@ -111,8 +115,10 @@ try {
  // Test 6: paseo ls -g flag is accepted
  {
    console.log('Test 6: paseo ls -g flag is accepted')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npx paseo ls -g`.nothrow()
+    const result = await runLocalPaseo(['ls', '-g'], {
+      PASEO_HOST: `localhost:${port}`,
+      PASEO_HOME: paseoHome,
+    })
    const output = result.stdout + result.stderr
    assert(!output.includes('unknown option'), 'should accept -g flag')
    assert(!output.includes('error: option'), 'should not have option parsing error')
@@ -122,8 +128,10 @@ try {
  // Test 7: paseo ls -ag combined flags are accepted
  {
    console.log('Test 7: paseo ls -ag combined flags are accepted')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npx paseo ls -ag`.nothrow()
+    const result = await runLocalPaseo(['ls', '-ag'], {
+      PASEO_HOST: `localhost:${port}`,
+      PASEO_HOME: paseoHome,
+    })
    const output = result.stdout + result.stderr
    assert(!output.includes('unknown option'), 'should accept -ag flags')
    assert(!output.includes('error: option'), 'should not have option parsing error')
@@ -133,8 +141,10 @@ try {
  // Test 8: -q (quiet) flag is accepted globally
  {
    console.log('Test 8: -q (quiet) flag is accepted')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npx paseo -q ls`.nothrow()
+    const result = await runLocalPaseo(['-q', 'ls'], {
+      PASEO_HOST: `localhost:${port}`,
+      PASEO_HOME: paseoHome,
+    })
    const output = result.stdout + result.stderr
    assert(!output.includes('unknown option'), 'should accept -q flag')
    assert(!output.includes('error: option'), 'should not have option parsing error')
@@ -144,8 +154,10 @@ try {
  // Test 9: paseo ls --ui is rejected (flag removed)
  {
    console.log('Test 9: paseo ls --ui is rejected')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npx paseo ls --ui`.nothrow()
+    const result = await runLocalPaseo(['ls', '--ui'], {
+      PASEO_HOST: `localhost:${port}`,
+      PASEO_HOME: paseoHome,
+    })
    assert.notStrictEqual(result.exitCode, 0, 'should fail for removed --ui flag')
    const output = result.stdout + result.stderr
    assert(output.includes('unknown option'), 'should report unknown option for --ui')
@@ -30,6 +30,29 @@ type ProviderModel = {
  description?: string
 }

+const EXPECTED_CLAUDE_MODELS = [
+  {
+    id: 'claude-sonnet-4-5-20250929',
+    model: 'Sonnet 4.5',
+    descriptionFragment: 'Best for everyday tasks',
+  },
+  {
+    id: 'claude-sonnet-4-6',
+    model: 'Sonnet 4.6',
+    descriptionFragment: 'Best for everyday tasks',
+  },
+  {
+    id: 'claude-opus-4-6',
+    model: 'Opus 4.6',
+    descriptionFragment: 'Most capable',
+  },
+  {
+    id: 'claude-haiku-4-5-20251001',
+    model: 'Haiku 4.5',
+    descriptionFragment: 'Fastest',
+  },
+] as const
+
 let claudeModelIdsFromJson: string[] = []
 let claudeModelsFromJson: ProviderModel[] = []

@@ -60,6 +83,27 @@ async function runProviderModelsJson(
  assert.fail(`provider models ${provider} exhausted retries`)
 }

+function assertClaudeModels(data: ProviderModel[]): void {
+  assert.strictEqual(data.length, EXPECTED_CLAUDE_MODELS.length, 'claude output should match the current catalog size')
+
+  const byId = new Map(data.map((model) => [model.id, model]))
+  const ids = [...byId.keys()].sort()
+  const expectedIds = EXPECTED_CLAUDE_MODELS.map((model) => model.id).sort()
+
+  assert.strictEqual(byId.size, data.length, 'claude model IDs should be unique')
+  assert.deepStrictEqual(ids, expectedIds, 'claude IDs should match the current catalog')
+
+  for (const expectedModel of EXPECTED_CLAUDE_MODELS) {
+    const actualModel = byId.get(expectedModel.id)
+    assert(actualModel, `claude output should include ${expectedModel.id}`)
+    assert.strictEqual(actualModel.model, expectedModel.model, `${expectedModel.id} should keep its display name`)
+    assert(
+      (actualModel.description ?? '').includes(expectedModel.descriptionFragment),
+      `${expectedModel.id} description should mention ${expectedModel.descriptionFragment}`
+    )
+  }
+}
+
 try {
  // Test 1: provider --help shows subcommands
  {
@@ -114,25 +158,7 @@ try {
  {
    console.log('Test 5: provider models claude lists canonical model aliases')
    const data = await runProviderModelsJson('claude')
-    assert.strictEqual(data.length, 3, 'should have exactly 3 claude models')
-    const byId = new Map(data.map((model) => [model.id, model]))
-    const ids = [...byId.keys()].sort()
-    assert.deepStrictEqual(ids, ['default', 'haiku', 'opus'], 'claude IDs should be default/opus/haiku')
-    assert.strictEqual(byId.get('default')?.model, 'Sonnet 4.5', 'default claude alias should map to Sonnet 4.5')
-    assert.strictEqual(byId.get('opus')?.model, 'Opus 4.6', 'opus claude alias should map to Opus 4.6')
-    assert.strictEqual(byId.get('haiku')?.model, 'Haiku 4.5', 'haiku claude alias should map to Haiku 4.5')
-    assert(
-      (byId.get('default')?.description ?? '').includes('Best for everyday tasks'),
-      'default claude description should mention everyday tasks'
-    )
-    assert(
-      (byId.get('opus')?.description ?? '').includes('Most capable'),
-      'opus claude description should mention most capable'
-    )
-    assert(
-      (byId.get('haiku')?.description ?? '').includes('Fastest'),
-      'haiku claude description should mention fastest'
-    )
+    assertClaudeModels(data)
    console.log('✓ provider models claude lists canonical model aliases\n')
  }

@@ -184,14 +210,8 @@ try {
    console.log('Test 9: provider models --json outputs valid JSON')
    const data = await runProviderModelsJson('claude')
    assert(Array.isArray(data), 'output should be an array')
-    assert.strictEqual(data.length, 3, 'should have exactly 3 models for claude')
    assert(data.every((m) => m.model && m.id), 'each model should have name and id')
-    const byId = new Map(data.map((model) => [model.id, model]))
-    const ids = [...byId.keys()].sort()
-    assert.deepStrictEqual(ids, ['default', 'haiku', 'opus'], 'claude JSON IDs should be deterministic')
-    assert.strictEqual(byId.get('default')?.model, 'Sonnet 4.5', 'default claude alias should map to Sonnet 4.5')
-    assert.strictEqual(byId.get('opus')?.model, 'Opus 4.6', 'opus claude alias should map to Opus 4.6')
-    assert.strictEqual(byId.get('haiku')?.model, 'Haiku 4.5', 'haiku claude alias should map to Haiku 4.5')
+    assertClaudeModels(data)
    claudeModelIdsFromJson = data.map((m) => m.id)
    claudeModelsFromJson = data
    console.log('✓ provider models --json outputs valid JSON\n')
@@ -204,7 +224,7 @@ try {
    const result = await ctx.paseo(['provider', 'models', 'claude', '--quiet'])
    assert.strictEqual(result.exitCode, 0, 'should exit 0')
    const lines = result.stdout.trim().split('\n').filter(Boolean)
-    assert.strictEqual(lines.length, 3, 'should have 3 lines')
+    assert.strictEqual(lines.length, EXPECTED_CLAUDE_MODELS.length, 'should have one line per Claude catalog model')
    assert.deepStrictEqual(
      [...lines].sort(),
      [...claudeModelIdsFromJson].sort(),
@@ -212,12 +232,12 @@ try {
    )
    assert.deepStrictEqual(
      [...lines].sort(),
-      ['default', 'haiku', 'opus'],
-      '--quiet should print canonical claude IDs'
+      EXPECTED_CLAUDE_MODELS.map((model) => model.id).sort(),
+      '--quiet should print the current Claude catalog IDs'
    )
    assert(
-      claudeModelsFromJson.some((m) => m.id === 'default'),
-      'captured --json output should still include default claude model id'
+      claudeModelsFromJson.some((m) => m.id === 'claude-sonnet-4-5-20250929'),
+      'captured --json output should still include the Claude default model id'
    )
    console.log('✓ provider models --quiet outputs model IDs only\n')
  }
@@ -14,19 +14,30 @@ import assert from 'node:assert'
 import { $ } from 'zx'
 import { mkdtemp, rm } from 'fs/promises'
 import { tmpdir } from 'os'
-import { join } from 'path'
+import { dirname, join } from 'path'
+import { fileURLToPath } from 'url'

 $.verbose = false

 console.log('=== Delete Command Tests ===\n')

+const cliRoot = dirname(fileURLToPath(import.meta.url))
+const repoRoot = join(cliRoot, '..', '..', '..')
 const port = 10000 + Math.floor(Math.random() * 50000)
 const paseoHome = await mkdtemp(join(tmpdir(), 'paseo-delete-test-home-'))

+async function runCli(args: string[]) {
+  return $`npm --prefix ${repoRoot} run cli -- ${args}`.nothrow()
+}
+
+async function runDelete(args: string[]) {
+  return $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npm --prefix ${repoRoot} run cli -- delete ${args}`.nothrow()
+}
+
 try {
  {
    console.log('Test 1: delete --help shows options')
-    const result = await $`npm run cli -- delete --help`.nothrow()
+    const result = await runDelete(['--help'])
    assert.strictEqual(result.exitCode, 0, 'delete --help should exit 0')
    assert(result.stdout.includes('--all'), 'help should mention --all flag')
    assert(result.stdout.includes('--cwd'), 'help should mention --cwd option')
@@ -37,8 +48,7 @@ try {

  {
    console.log('Test 2: delete requires ID, --all, or --cwd')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npm run cli -- delete`.nothrow()
+    const result = await runDelete([])
    assert.notStrictEqual(result.exitCode, 0, 'should fail without id, --all, or --cwd')
    const output = result.stdout + result.stderr
    const hasError =
@@ -52,8 +62,7 @@ try {

  {
    console.log('Test 3: delete handles daemon not running')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npm run cli -- delete abc123`.nothrow()
+    const result = await runDelete(['abc123'])
    assert.notStrictEqual(result.exitCode, 0, 'should fail when daemon not running')
    const output = result.stdout + result.stderr
    const hasError =
@@ -66,8 +75,7 @@ try {

  {
    console.log('Test 4: delete --all flag is accepted')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npm run cli -- delete --all`.nothrow()
+    const result = await runDelete(['--all'])
    const output = result.stdout + result.stderr
    assert(!output.includes('unknown option'), 'should accept --all flag')
    assert(!output.includes('error: option'), 'should not have option parsing error')
@@ -76,8 +84,7 @@ try {

  {
    console.log('Test 5: delete --cwd flag is accepted')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npm run cli -- delete --cwd /tmp`.nothrow()
+    const result = await runDelete(['--cwd', '/tmp'])
    const output = result.stdout + result.stderr
    assert(!output.includes('unknown option'), 'should accept --cwd flag')
    assert(!output.includes('error: option'), 'should not have option parsing error')
@@ -86,8 +93,7 @@ try {

  {
    console.log('Test 6: delete with ID and --host flag is accepted')
-    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npm run cli -- delete abc123 --host localhost:${port}`.nothrow()
+    const result = await runDelete(['abc123', '--host', `localhost:${port}`])
    const output = result.stdout + result.stderr
    assert(!output.includes('unknown option'), 'should accept --host flag')
    assert(!output.includes('error: option'), 'should not have option parsing error')
@@ -96,7 +102,7 @@ try {

  {
    console.log('Test 7: paseo --help shows delete command')
-    const result = await $`npm run cli -- --help`.nothrow()
+    const result = await runCli(['--help'])
    assert.strictEqual(result.exitCode, 0, 'paseo --help should exit 0')
    assert(result.stdout.includes('delete'), 'help should mention delete command')
    console.log('✓ paseo --help shows delete command\n')
@@ -105,7 +111,7 @@ try {
  {
    console.log('Test 8: -q (quiet) flag is accepted with delete')
    const result =
-      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npm run cli -- -q delete abc123`.nothrow()
+      await $`PASEO_HOST=localhost:${port} PASEO_HOME=${paseoHome} npm --prefix ${repoRoot} run cli -- -q delete abc123`.nothrow()
    const output = result.stdout + result.stderr
    assert(!output.includes('unknown option'), 'should accept -q flag')
    assert(!output.includes('error: option'), 'should not have option parsing error')
@@ -32,6 +32,8 @@ const impossibleSchema = JSON.stringify({
  ],
 })

+const OPEN_CODE_STRUCTURED_MODEL = 'opencode/gpt-5-nano'
+
 let ctx: E2EContext

 async function setup(): Promise<void> {
@@ -94,7 +96,7 @@ async function test_all_providers_return_structured_output(): Promise<void> {
  await runProviderCase({
    provider: 'opencode',
    mode: 'default',
-    model: 'opencode/kimi-k2.5-free',
+    model: OPEN_CODE_STRUCTURED_MODEL,
  })
 }

@@ -0,0 +1,9 @@
+import { $, ProcessPromise } from 'zx'
+import { join } from 'node:path'
+
+const CLI_ENTRY = join(import.meta.dirname, '..', '..', 'dist', 'index.js')
+
+export function runLocalPaseo(args: string[], env: NodeJS.ProcessEnv = {}): ProcessPromise {
+  $.verbose = false
+  return $({ env: { ...process.env, ...env } })`${process.execPath} ${CLI_ENTRY} ${args}`.nothrow()
+}
@@ -36,102 +36,35 @@ for (let i = 0; i < args.length; i++) {

 $.verbose = false

-console.log('🧪 Paseo CLI E2E Test Runner\n')
-console.log('='.repeat(50))
-
-// Discover all test files
-const files = await readdir(__dirname)
-const testFiles = files
-  .filter(f => f.match(/^\d{2}-.*\.test\.ts$/))
-  .sort()
-
-if (testFiles.length === 0) {
-  console.log('❌ No test files found')
-  if (jsonOutputPath) {
-    await writeFile(
-      jsonOutputPath,
-      JSON.stringify(
-        {
-          suite: 'cli-local',
-          command: 'npm run test:local --workspace=@getpaseo/cli',
-          counts: {
-            passed: 0,
-            failed: 0,
-            skipped: 0,
-          },
-          skippedTests: [],
-          failures: [],
-          error: 'No test files found',
-        },
-        null,
-        2
-      ) + '\n'
-    )
-  }
-  process.exit(1)
-}
-
-console.log(`Found ${testFiles.length} test file(s):\n`)
-for (const file of testFiles) {
-  console.log(`  - ${file}`)
-}
-console.log()
-
-let passed = 0
-let failed = 0
-const failures: { test: string; error: string }[] = []
-
-for (const testFile of testFiles) {
-  const testPath = join(__dirname, testFile)
-  const testName = testFile.replace(/\.test\.ts$/, '')
+type Failure = { test: string; error: string }

+async function runCommand(label: string, command: string): Promise<void> {
  console.log(`\n${'─'.repeat(50)}`)
-  console.log(`📋 Running ${testName}...`)
+  console.log(`🔧 ${label}...`)
  console.log('─'.repeat(50))

-  try {
-    const result = await $`PASEO_LOCAL_SPEECH_AUTO_DOWNLOAD=${testEnvDefaults.PASEO_LOCAL_SPEECH_AUTO_DOWNLOAD} PASEO_DICTATION_ENABLED=${testEnvDefaults.PASEO_DICTATION_ENABLED} PASEO_VOICE_MODE_ENABLED=${testEnvDefaults.PASEO_VOICE_MODE_ENABLED} npx tsx ${testPath}`.nothrow()
-    if (result.exitCode === 0) {
-      console.log(`\n✅ ${testName} PASSED`)
-      passed++
-    } else {
-      console.log(`\n❌ ${testName} FAILED (exit code: ${result.exitCode})`)
-      if (result.stderr) {
-        console.log('stderr:', result.stderr)
-      }
-      failed++
-      failures.push({ test: testName, error: result.stderr || `Exit code: ${result.exitCode}` })
-    }
-  } catch (e) {
-    const error = e instanceof Error ? e.message : String(e)
-    console.log(`\n❌ ${testName} FAILED`)
-    console.log('Error:', error)
-    failed++
-    failures.push({ test: testName, error })
+  const result = await $`bash -lc ${command}`.nothrow()
+  if (result.exitCode !== 0) {
+    const error = result.stderr || result.stdout || `Exit code: ${result.exitCode}`
+    console.error(`\n❌ ${label} failed`)
+    console.error(error)
+    throw new Error(error)
  }
 }

-// Summary
-console.log('\n' + '='.repeat(50))
-console.log('📊 Test Results')
-console.log('='.repeat(50))
-console.log(`  ✅ Passed: ${passed}`)
-console.log(`  ❌ Failed: ${failed}`)
-console.log(`  📝 Total:  ${passed + failed}`)
-
-if (failures.length > 0) {
-  console.log('\n❌ Failed tests:')
-  for (const { test, error } of failures) {
-    console.log(`  - ${test}`)
-    if (error) {
-      console.log(`    ${error.split('\n')[0]}`)
-    }
+async function writeJsonSummary({
+  passed,
+  failed,
+  failures,
+}: {
+  passed: number
+  failed: number
+  failures: Failure[]
+}) {
+  if (!jsonOutputPath) {
+    return
  }
-}

-console.log()
-
-if (jsonOutputPath) {
  await writeFile(
    jsonOutputPath,
    JSON.stringify(
@@ -155,4 +88,87 @@ if (jsonOutputPath) {
  )
 }

+console.log('🧪 Paseo CLI E2E Test Runner\n')
+console.log('='.repeat(50))
+
+// Discover all test files
+const files = await readdir(__dirname)
+const testFiles = files
+  .filter(f => f.match(/^\d{2}-.*\.test\.ts$/))
+  .sort()
+
+if (testFiles.length === 0) {
+  console.log('❌ No test files found')
+  await writeJsonSummary({ passed: 0, failed: 0, failures: [] })
+  process.exit(1)
+}
+
+console.log(`Found ${testFiles.length} test file(s):\n`)
+for (const file of testFiles) {
+  console.log(`  - ${file}`)
+}
+console.log()
+
+let passed = 0
+let failed = 0
+const failures: Failure[] = []
+
+await runCommand('Building relay', 'npm run build --workspace=@getpaseo/relay')
+await runCommand('Building server', 'npm run build --workspace=@getpaseo/server')
+await runCommand('Building CLI', 'npm run build --workspace=@getpaseo/cli')
+
+for (const testFile of testFiles) {
+  const testPath = join(__dirname, testFile)
+  const testName = testFile.replace(/\.test\.ts$/, '')
+
+  console.log(`\n${'─'.repeat(50)}`)
+  console.log(`📋 Running ${testName}...`)
+  console.log('─'.repeat(50))
+
+  try {
+    const result = await $`PASEO_LOCAL_SPEECH_AUTO_DOWNLOAD=${testEnvDefaults.PASEO_LOCAL_SPEECH_AUTO_DOWNLOAD} PASEO_DICTATION_ENABLED=${testEnvDefaults.PASEO_DICTATION_ENABLED} PASEO_VOICE_MODE_ENABLED=${testEnvDefaults.PASEO_VOICE_MODE_ENABLED} npx tsx ${testPath}`.nothrow()
+    if (result.exitCode === 0) {
+      console.log(`\n✅ ${testName} PASSED`)
+      passed++
+    } else {
+      console.log(`\n❌ ${testName} FAILED (exit code: ${result.exitCode})`)
+      if (result.stderr) {
+        console.log('stderr:', result.stderr)
+      }
+      failed++
+      failures.push({ test: testName, error: result.stderr || `Exit code: ${result.exitCode}` })
+      break
+    }
+  } catch (e) {
+    const error = e instanceof Error ? e.message : String(e)
+    console.log(`\n❌ ${testName} FAILED`)
+    console.log('Error:', error)
+    failed++
+    failures.push({ test: testName, error })
+    break
+  }
+}
+
+// Summary
+console.log('\n' + '='.repeat(50))
+console.log('📊 Test Results')
+console.log('='.repeat(50))
+console.log(`  ✅ Passed: ${passed}`)
+console.log(`  ❌ Failed: ${failed}`)
+console.log(`  📝 Total:  ${passed + failed}`)
+
+if (failures.length > 0) {
+  console.log('\n❌ Failed tests:')
+  for (const { test, error } of failures) {
+    console.log(`  - ${test}`)
+    if (error) {
+      console.log(`    ${error.split('\n')[0]}`)
+    }
+  }
+}
+
+console.log()
+
+await writeJsonSummary({ passed, failed, failures })
+
 process.exit(failed > 0 ? 1 : 0)
@@ -0,0 +1,20 @@
+import { createE2ETestContext } from '/Users/moboudra/.paseo/worktrees/1luy0po7/beefy-parrot/packages/cli/tests/helpers/test-daemon.ts'
+
+async function main() {
+  const ctx = await createE2ETestContext({ timeout: 180000 })
+  try {
+    const run = await ctx.paseo([
+      'run',
+      '--provider','opencode',
+      '--mode','default',
+      '--model','opencode/kimi-k2.5-free',
+      'Say hello in one short sentence.'
+    ], { timeout: 180000 })
+    console.log('EXIT', run.exitCode)
+    console.log('STDOUT\n' + run.stdout)
+    console.log('STDERR\n' + run.stderr)
+  } finally {
+    await ctx.stop()
+  }
+}
+main().catch((error) => { console.error(error); process.exitCode = 1 })
@@ -0,0 +1,43 @@
+import { createE2ETestContext } from '../helpers/test-daemon.ts'
+
+const schema = JSON.stringify({
+  type: 'object',
+  properties: { summary: { type: 'string' } },
+  required: ['summary'],
+  additionalProperties: false,
+})
+
+async function main() {
+  const ctx = await createE2ETestContext({ timeout: 180000 })
+  try {
+    const run = await ctx.paseo([
+      'run',
+      '--provider','opencode',
+      '--mode','default',
+      '--model','opencode/kimi-k2.5-free',
+      '--output-schema', schema,
+      'Return valid JSON with a short summary for provider opencode.'
+    ], { timeout: 180000 })
+    console.log('RUN EXIT', run.exitCode)
+    console.log('RUN STDOUT\n' + run.stdout)
+    console.log('RUN STDERR\n' + run.stderr)
+
+    const ls = await ctx.paseo(['ls','--json'])
+    console.log('LS\n' + ls.stdout)
+    const agents = JSON.parse(ls.stdout)
+    const agentId = agents[0]?.id
+    if (agentId) {
+      const inspect = await ctx.paseo(['inspect', agentId])
+      console.log('INSPECT\n' + inspect.stdout)
+      const logs = await ctx.paseo(['logs','--tail','100', agentId], { timeout: 30000 })
+      console.log('LOGS\n' + logs.stdout)
+    }
+  } finally {
+    await ctx.stop()
+  }
+}
+
+main().catch((error) => {
+  console.error(error)
+  process.exitCode = 1
+})