Self-hosted workflows for coding-agent delivery. Orchestrate CodeBuddy, Codex and Claude Code with scripts and human approval gates. Includes Local workflows for coding assistants.
E2E tests
bash scripts/e2e.sh runs the end-to-end suite. It needs Python 3.8+, Node.js 20+ and npm, and no network access, API key or real coding Agent.
What it does:
Builds the npm tarball once (
npm pack), installs it into a temporary prefix, and uses thatloopforgebinary, which is the same path users take withnpx loopforge-cli.Installs every supported host and edition into throw-away git projects, then checks the entrypoint files,
status,doctor,updateand a cleanuninstall(test_e2e_install.py).Drives the installed Portable workflow end to end with
fake_agent.py, a deterministic stand-in for the host Agent. It follows the Skill'sresume-driven protocol (init,prepare --emit-prompt,assign,start,finish,approve,decide-knowledge), but writes the stage outputs recorded infixtures/scenario-slugify.jsoninstead of calling an LLM. The developer stage really changesfixtures/sample_project, and the test stage really runs its tests. The suite covers every host, resume after a crash and after an interruption, the requirement confirmation gate, review rejection, unfilled artifacts and repeated failures (test_e2e_portable_workflow.py).Covers the other Portable routes and modes (
test_e2e_workflow_modes.py): small/SOLO, SOLO overflow escalating to medium, large with a knowledge stage,manualapproval gates, and design-implementation tasks that need design context and visual evidence.Runs the Classic CodeBuddy hooks from an installed project: auto-dispatch routing between roles, the fallbacks to the main Agent, and a recorded session flowing through the observability hooks into the local dashboard, which is served over HTTP (
test_e2e_classic_runtime.py).Renders the dashboard in Chromium with Playwright (
test_e2e_dashboard_browser.py). It checks the data badge, KPIs, session table, Devflow run card and auto-dispatch health, plus the empty and load-failure states, and blocks every request outside 127.0.0.1. The test is skipped whenplaywrightis not installed. CI setsLOOPFORGE_E2E_BROWSER=1, so there a missing browser fails the job instead, and the screenshots are uploaded as thedashboard-screenshotsartifact.
Useful options:
LOOPFORGE_E2E_SOURCE=1 bash scripts/e2e.sh # run the CLI from src/ and skip npm pack
bash scripts/e2e.sh -k resume # run only tests matching a pattern
python3 tests/e2e/fake_agent.py --help # drive a project by hand
pip install playwright && python3 -m playwright install chromium
LOOPFORGE_E2E_BROWSER=1 bash scripts/e2e.sh -k DashboardBrowserE2E
When a test fails, the assertion message includes the fake Agent transcript (.fake-agent/transcript.jsonl in the project) that shows each action and the command that was rejected.
Real CodeBuddy host tests
bash scripts/test-live.sh is a separate opt-in suite. Set LOOPFORGE_LIVE_TESTS=1, LOOPFORGE_TEST_MODEL, LOOPFORGE_CODEBUDDY_ENVIRONMENT, and either CODEBUDDY_API_KEY or LOOPFORGE_TEST_KEY_FILE. It consumes model credits. Both Classic and Portable are installed from the npm package into temporary projects; a real CodeBuddy process fixes a small Python fixture, runs the acceptance tests, and saves workflow completion and artifacts. The test checks those files and model usage, then removes the temporary project. It does not publish raw model transcripts or credentials.
Package notices have one maintained source at the repository root. npm prepack, Python builds and release-archive creation generate the local-package notice, preserving the complete upstream copyright and permission text. Generated notices are ignored by Git.
Classic live acceptance keeps a real POSIX terminal session open while CodeBuddy's asynchronous team executes. A one-shot -p response saying that a team was dispatched is not a completed delivery. The live test waits for the persisted workflow summary, then verifies the actual fixture tests and token usage. Portable live acceptance uses the CLI's streamed print mode. Set LOOPFORGE_TEST_DIAGNOSTICS to an ignored, protected local directory to retain redacted replies and saved state; terminal output, raw transcripts and environment values are not published.