| name | clash-real-e2e-tester |
| description | Test Clash desktop Copilot/ACP flows with subagents using agent-browser. Use this skill whenever the user asks for real desktop E2E, ACP agent verification, Codex/Claude/Gemini harness testing, session history persistence checks, or screenshots from a user-perspective Electron run. It prevents mock-only tests from being mistaken for real E2E. |
Clash Real E2E Tester
Use this skill to hand a Clash desktop/Copilot test task to a subagent or to run it directly. The goal is to verify the product from a user's point of view, not to prove that mocks can pass.
Test Levels
Always name the level being run:
-
Static startup
- Command:
pnpm --filter @clash/desktop test:startup:static
- Verifies bundled/local harness setup and registry invariants.
- No UI claim is allowed from this level.
-
Stub UI smoke
- Command:
pnpm --filter @clash/desktop test:startup:ui
- Starts Electron and a stub ACP agent.
- Useful for deterministic UI regressions.
- Must be reported as stub/mock, never as real agent validation.
-
Real Codex desktop E2E
- Command:
pnpm test:startup:real-codex
- Starts desktop dev, uses the bundled
codex-acp, and drives Electron with agent-browser.
- This is the minimum acceptable test when the user asks whether real Codex/Copilot works.
Subagent Prompt
When delegating, use this prompt shape:
You are testing Clash desktop Copilot as a real user.
Use the current Clash repository root as the repo.
Rules:
- Use agent-browser against the Electron dev window, not the previously installed app.
- Do not use a stub agent when the task asks for real E2E. If you use a stub smoke test, label it clearly as stub.
- Interact through the UI: click buttons, type into the composer with keyboard input, and take screenshots. Use DOM eval only for waiting/assertions, not to fake user actions.
- Do not claim success without screenshot paths and concrete visible evidence.
- Wait for the agent turn to finish before clicking New session.
- After clicking New session, open Session history and verify the previous session is still listed.
- Capture failures too: screenshot plus relevant stdout/stderr tails.
- Clean up spawned Electron/Vite/agent child processes at the end.
Run:
1. `pnpm --filter @clash/desktop test:startup:static` for deterministic startup/package coverage.
2. `pnpm --filter @clash/desktop test:startup:ui` only if a stub UI smoke test is explicitly useful, and label it as stub.
3. `pnpm test:startup:real-codex` for the real Codex path.
For the real run, verify:
- The prompt is typed into the composer: `Run \`pwd\` with your shell tool, then answer with only the path.`
- A tool-call row appears, such as `Ran pwd`.
- The final answer includes a `.clash/projects/` working directory path.
- The stop/running state clears before a fresh session is created.
- The old session remains visible in Session history after `+`.
Return:
- Commands run and pass/fail status.
- Screenshot paths as Markdown image links.
- Any observed transport diagnostics, such as WebSocket timeout/fallback, separated from assistant-message rendering issues.
- A short list of product regressions found, with repro steps.
Pass/Fail Bar
A real desktop E2E passes only if all are true:
- It uses the development Electron window launched from the current repo.
- It uses the real Codex ACP path, not a fake/stub provider.
- The UI is driven through
agent-browser user actions.
- Screenshots show the final state.
- Tool call, final answer, idle state, fresh session, and history persistence are all verified.
Common Failure Modes
- Old installed app opened: kill it and relaunch desktop dev from the repo.
- Submit click opens the model menu: scope submit clicks to
.clash-chat-input-surface and the right-bottom primary button.
- New session wipes history: session creation must persist a local row immediately, before ACP ready.
- Tool output duplicated: tool updates must be merged by tool call id.
- Mock mistaken for E2E: rerun with
pnpm test:startup:real-codex.
- Transport warning rendered as chat text: treat WebSocket fallback warnings as diagnostics, not assistant messages.
