I have the complete plan ready. Here it is:

Create two new test artifacts: (1) a bash E2E test script that runs the real pipeline with mocked Claude/GitHub (smoke tests) and supports an optional "live" mode against real Claude for integration tests, and (2) a GitHub Actions workflow for integration tests on PRs to main.

The design follows the existing test harness pattern exactly — PASS/FAIL counters, mock binaries in temp dirs, assertion helpers, colored output. No new frameworks or dependencies.

Follows the exact pattern from sw-pipeline-test.sh — same boilerplate (header, colors, counters), same assertion helpers (assert_exit_code, assert_output_contains, assert_file_exists, assert_file_contains, assert_branch_exists, assert_state_contains), same invoke_pipeline() → captures output + exit code, same run_test() runner, same setup_env() / reset_test() / cleanup_env() lifecycle.

Two modes: --smoke (default, mocked, fast) and --live (real Claude, budget-capped).

10 Smoke Tests (mocked — always run in CI):

1. test_full_pipeline_stage_order — Run fast template (intake→build→test→PR) with mocks. Verify exit 0, state file status: complete, all stage artifacts exist.
2. test_stage_order_preserved — Parse output for stage markers, verify correct ordering.
3. test_state_file_updated_per_stage — Verify state file has all required YAML fields: pipeline, goal, status, branch, stage_progress, timestamps.
4. test_no_unhandled_errors — Run full pipeline, verify no unbound variable or command not found in output.
5. test_resume_from_interrupted — Run intake-only, manually edit state, resume. Verify continuation.
6. test_artifacts_have_valid_json — Verify all .json in pipeline-artifacts are valid via jq.
7. test_branch_created_cleanly — Feature branch exists, has commits, clean working tree.
8. test_dry_run_produces_no_artifacts — --dry-run creates no artifacts directory.
9. test_pipeline_with_custom_template — Custom template with only intake+build, verify only those stages run.
10. test_error_recovery_no_crash — Mock claude returns exit 1 during plan, verify graceful failure.

Same script with --live flag. One test: trivial goal ("Add a comment to README.md"), intake→build with real Claude. Budget protection: --max-turns 5, fast template, 120s timeout, 1ドル.00 cap. Skips gracefully if claude CLI unavailable.

.github/workflows/integration.yml — smoke tests on macOS + Ubuntu matrix for every PR. Conditional live integration job when CLAUDE_API_KEY secret exists. Upload test result artifacts.

Add "test:e2e" and "test:integration" to package.json scripts.

Add new test suite to the test suites table.

• Task 1: Create scripts/sw-e2e-test.sh boilerplate (header, colors, counters, assertions, setup/cleanup, main)
• Task 2: Implement mock environment setup (real pipeline + templates, mock binaries, mock git project)
• Task 3: Implement 10 smoke tests
• Task 4: Implement live integration mode with budget protection and skip logic
• Task 5: Create .github/workflows/integration.yml
• Task 6: Update package.json with new scripts
• Task 7: Update .claude/CLAUDE.md test suites table
• Task 8: Run smoke tests locally to verify they pass
• Task 9: Run ShellCheck and fix warnings

• npm run test:e2e runs the smoke test suite and passes
• npm run test:integration runs live tests or skips gracefully
• CI runs smoke tests on every PR to main
• Budget-capped to 1ドル.00 per live test run
• Pass/fail output matches existing test suite style
• Bash 3.2 compatible, no ShellCheck errors