AgentSkillsCN

self-test

启动一个测试版的Clawd实例，并对其运行端到端测试。

SKILL.md
--- frontmatter
name: self-test
description: Spin up a test Clawd instance and run E2E tests against it
user-invocable: true

Self-Test Skill

IMPORTANT: Do NOT explore the codebase. Do NOT read source files, package.json, or project structure. Go directly to Step 1. Simple E2E tests should complete under $0.50.

Spin up a separate test Clawd instance and run E2E tests against it using Playwright.

CRITICAL: Do NOT navigate to clawd:4000 or any parent instance URL. You must spin up your OWN test instance using the script below and test against THAT.

Step 1: Spin Up a Test Instance

Detect the current branch and launch the test instance:

bash
BRANCH=$(git -C /workspace branch --show-current 2>/dev/null || echo "main")
echo "Testing branch: $BRANCH"
bash /workspace/scripts/test-clawd.sh --branch "$BRANCH"
  • •The script outputs the test instance URL (e.g. http://test-clawd-1234567890:5000) and container name
  • •Login credentials are test / test
  • •Wait for the script to confirm the instance is ready before proceeding

Step 2: Run E2E Tests with Playwright MCP

Use the Playwright MCP browser tools to test the instance from Step 1:

  1. •Navigate to the test instance URL (from Step 1 output) with browser_navigate
  2. •Log in with test / test
  3. •Use browser_snapshot to inspect page state (preferred over screenshots)
  4. •Interact with elements using browser_click, browser_type, browser_fill_form
  5. •Assert conditions by inspecting snapshots

Tips:

  • •The test instance has no project-repos.json, so the New Session dialog shows raw URL + branch inputs instead of the repository dropdown
  • •PLAYWRIGHT_BROWSERS_PATH=/opt/playwright-browsers is already set
  • •Test sessions can use dangerous permission mode to avoid approval dialog friction during E2E tests

Step 3: Clean Up

When testing is complete, always clean up:

bash
bash /workspace/scripts/cleanup-test-clawd.sh <container-name>

Step 4: Report Results

Summarize which tests passed/failed, any errors encountered, and snapshots of failures.

Test Recipes

Recipe: Verify Plan Markdown View

  1. •Create a new session (any repo, auto_edits mode)
  2. •Send: "Enter plan mode and write a short 3-bullet plan for adding a README"
  3. •Wait for the plan card to appear in the chat
  4. •Click the plan card to open the full-screen overlay
  5. •Assert: overlay is visible, contains markdown content, has a close button
  6. •Close the overlay and assert it's dismissed

Recipe: Verify Session Creation

  1. •Click "New Session" from the session list
  2. •Fill in a repo URL (e.g. https://github.com/octocat/Hello-World), branch master, permission mode normal
  3. •Submit the form
  4. •Assert: session appears in the list, status transitions to "running"
  5. •Send a simple message like "What files are in this repo?" and verify a response appears

Recipe: Verify Skill Invocation

  1. •Create a new session (any repo, auto_edits mode)
  2. •Type / in the input to trigger the skill picker
  3. •Assert: skill options appear (e.g. "wrapup")
  4. •Select a skill and verify it loads into the input

Troubleshooting

  • •If the test instance takes too long to start, check logs: docker logs <container-name>
  • •Test instances run on port 5000 within the Docker network — not exposed to the host
  • •Each test instance gets a unique name based on timestamp, so multiple can coexist