open-swe/tests/e2e/README.md

Playwright E2E — the full Slack → implement → PR → reply flow

This drives the whole happy path through two mock UIs:

1. A user asks Open SWE to implement something in a mock Slack thread.
2. The real agent runs (via langgraph dev): it implements the change in a local temp-dir sandbox, pushes a branch, and opens a PR on a fake GitHub.
3. It posts the PR link back to the same Slack thread — visible in the mock UI.

What is faked vs. real

Only the LLM and the external SaaS HTTP boundaries are faked. All agent code runs for real.

Piece	Real or fake
Slack webhook → process_slack_mention → run dispatch	real (agent.webapp)
get_agent, deepagents loop, tools, middleware, prompt	real
open_pull_request, slack_thread_reply tools	real
Sandbox	real local provider, rooted in a throwaway temp dir
Git remote ("GitHub")	real git, a local bare repo the agent clones/pushes
The LLM	fake — a scripted model (fake_llm.py) emitting a fixed tool sequence
api.github.com REST (PR create) + dashboard GitHub OAuth login	fake (/fake-gh/...), state rendered at /mock/github
slack.com/api (post message, etc.)	fake (/fake-slack/...), thread rendered at /mock/slack
GitHub App token mint, api.github.com/user identity	stubbed (offline)

The fake GitHub/Slack stores are the single source of truth the mock UIs render, so what Playwright asserts on is exactly what the real agent produced.

Files

• e2e_env.py — env + constants set before any agent.* import (sandbox=local, fake API URLs, isolated GIT_CONFIG_GLOBAL, bot-token-only mode).
• fake_llm.py — the scripted BaseChatModel (the only faked agent piece).
• patches.py — monkeypatches the boundaries (LLM, GitHub/Slack URLs, token mint).
• agent_entrypoint.py — langgraph agent graph: applies patches, re-exports the real traced_agent.
• harness.py — langgraph http.app: the real agent.webapp plus the fake GitHub/Slack APIs, the mock UIs, and the control/compose endpoints.
• fakes.py — in-memory PR/Slack stores + git seeding of the bare remote.
• langgraph.e2e.json — dev-server config pointing at the two entrypoints above.
• static/{slack,github}.html — the mock Slack/GitHub UIs (external SaaS we can't run locally). The dashboard is not mocked — it's the real ui/ app.
• global-setup.ts — builds the real ui/ SPA (once) so the harness can serve it.

The dashboard — the real ui/ app

The dashboard is not mocked. The bot's "Open in Web" link (DASHBOARD_BASE_URL/agents/{thread_id}) loads the actual built ui/ React app — served same-origin from the harness so the session cookie and /dashboard/api/* calls work without CORS. The signed session cookie is real (minted via /control/login), so per-user authorization is genuine; the only extra fake is the OAuth-token store (an external credential).

The UI is built by global-setup.ts with VITE_DASHBOARD_API_BASE_URL pointed at the harness. It builds once; set E2E_FORCE_UI_BUILD=1 to rebuild (e.g. after a UI change or port change). Requires Corepack with pnpm enabled.

Run

cd tests/e2e
npm install
npx playwright install chromium
npx playwright test          # boots langgraph dev automatically, then runs

Watch it in human time:

SLOW_MO=700 npx playwright test --headed

Artifacts (replay a run)

Every test records a trace (DOM-snapshot timeline + network + console + source) and a video; failures also get a screenshot. Locally they land in test-results/<test>/ and are embedded in playwright-report/:

npx playwright show-report                       # browse runs; each has a Trace tab
npx playwright show-trace test-results/<test>/trace.zip   # open one trace directly

In CI the Playwright E2E job uploads both playwright-report/ and test-results/ as the playwright-report artifact on the run. Download it, then npx playwright show-report <unzipped-dir> (or drag a trace.zip onto https://trace.playwright.dev) to replay.

Poke at it by hand (from the repo root):

uv run langgraph dev --config tests/e2e/langgraph.e2e.json --port 2024 \
  --no-browser --allow-blocking --no-reload
# open http://127.0.0.1:2024/mock/slack  and  /mock/github