triage-ci-flake（分诊 ci 片）

Triage CI Failure

Overview

Systematic workflow for triaging and fixing test failures in CI, especially flaky tests that pass locally but fail in CI. Tests that made it to main are usually flaky due to timing, bundling, or environment differences.

CRITICAL RULE: You MUST run the reproduction workflow before proposing any fixes. No exceptions.

When to Use

• CI test fails on main branch after PR was merged
• Test passes locally but fails in CI
• Test failure labeled as "flaky" or intermittent
• E2E or integration test timing out in CI only

MANDATORY First Steps

YOU MUST EXECUTE THESE COMMANDS. Reading code or analyzing logs does NOT count as reproduction.

1. Extract suite name, test name, and error from CI logs
2. EXECUTE: Kill port 3000 to avoid conflicts
3. EXECUTE: pnpm dev $SUITE_NAME (use run_in_background=true)
4. EXECUTE: Wait for server to be ready (check with curl or sleep)
5. EXECUTE: Run the specific failing test with Playwright directly (npx playwright test test/TEST_SUITE_NAME/e2e.spec.ts:31:3 --headed -g "TEST_DESCRIPTION_TARGET_GOES_HERE")
6. If test passes, EXECUTE: pnpm prepare-run-test-against-prod
7. EXECUTE: pnpm dev:prod $SUITE_NAME and run test again

Only after EXECUTING these commands and seeing their output can you proceed to analysis and fixes.

"Analysis from logs" is NOT reproduction. You must RUN the commands.

Core Workflow

digraph triage_ci {
    "CI failure reported" [shape=box];
    "Extract details from CI logs" [shape=box];
    "Identify suite and test name" [shape=box];
    "Run dev server: pnpm dev $SUITE" [shape=box];
    "Run specific test by name" [shape=box];
    "Did test fail?" [shape=diamond];
    "Debug with dev code" [shape=box];
    "Run prepare-run-test-against-prod" [shape=box];
    "Run: pnpm dev:prod $SUITE" [shape=box];
    "Run specific test again" [shape=box];
    "Did test fail now?" [shape=diamond];
    "Debug bundling issue" [shape=box];
    "Unable to reproduce - check logs" [shape=box];
    "Fix and verify" [shape=box];

    "CI failure reported" -> "Extract details from CI logs";
    "Extract details from CI logs" -> "Identify suite and test name";
    "Identify suite and test name" -> "Run dev server: pnpm dev $SUITE";
    "Run dev server: pnpm dev $SUITE" -> "Run specific test by name";
    "Run specific test by name" -> "Did test fail?";
    "Did test fail?" -> "Debug with dev code" [label="yes"];
    "Did test fail?" -> "Run prepare-run-test-against-prod" [label="no"];
    "Run prepare-run-test-against-prod" -> "Run: pnpm dev:prod $SUITE";
    "Run: pnpm dev:prod $SUITE" -> "Run specific test again";
    "Run specific test again" -> "Did test fail now?";
    "Did test fail now?" -> "Debug bundling issue" [label="yes"];
    "Did test fail now?" -> "Unable to reproduce - check logs" [label="no"];
    "Debug with dev code" -> "Fix and verify";
    "Debug bundling issue" -> "Fix and verify";
}

Step-by-Step Process

1. Extract CI Details

From CI logs or GitHub Actions URL, identify:

• Suite name: Directory name (e.g., i18n, fields, lexical)
• Test file: Full path (e.g., test/i18n/e2e.spec.ts)
• Test name: Exact test description
• Error message: Full stack trace
• Test type: E2E (Playwright) or integration (Vitest)

2. Reproduce with Dev Code

CRITICAL: Always run the specific test by name, not the full suite.

SERVER MANAGEMENT RULES:

1. ALWAYS kill all servers before starting a new one
2. NEVER assume ports are free
3. ALWAYS wait for server ready confirmation before running tests

# ========================================
# STEP 2A: STOP ALL SERVERS
# ========================================
lsof -ti:3000 | xargs kill -9 2>/dev/null || echo "Port 3000 clear"

# ========================================
# STEP 2B: START DEV SERVER
# ========================================
# Start dev server with the suite (in background with run_in_background=true)
pnpm dev $SUITE_NAME

# ========================================
# STEP 2C: WAIT FOR SERVER READY
# ========================================
# Wait for server to be ready (REQUIRED - do not skip)
until curl -s http://localhost:3000/admin > /dev/null 2>&1; do sleep 1; done && echo "Server ready"

# ========================================
# STEP 2D: RUN SPECIFIC TEST
# ========================================
# Run ONLY the specific failing test using Playwright directly
# For E2E tests (DO NOT use pnpm test:e2e as it spawns its own server):
pnpm exec playwright test test/$SUITE_NAME/e2e.spec.ts -g "exact test name"

# For integration tests:
pnpm test:int $SUITE_NAME -t "exact test name"

Did the test fail?

• ✅ YES: You reproduced it! Proceed to debug with dev code.
• ❌ NO: Continue to step 3 (bundled code test).

3. Reproduce with Bundled Code

If test passed with dev code, the issue is likely in bundled/production code.

IMPORTANT: You MUST stop the dev server before starting prod server.

# ========================================
# STEP 3A: STOP ALL SERVERS (INCLUDING DEV SERVER FROM STEP 2)
# ========================================
lsof -ti:3000 | xargs kill -9 2>/dev/null || echo "Port 3000 clear"

# ========================================
# STEP 3B: BUILD AND PACK FOR PROD
# ========================================
# Build all packages and pack them (this takes time - be patient)
pnpm prepare-run-test-against-prod

# ========================================
# STEP 3C: START PROD SERVER
# ========================================
# Start prod dev server (in background with run_in_background=true)
pnpm dev:prod $SUITE_NAME

# ========================================
# STEP 3D: WAIT FOR SERVER READY
# ========================================
# Wait for server to be ready (REQUIRED - do not skip)
until curl -s http://localhost:3000/admin > /dev/null 2>&1; do sleep 1; done && echo "Server ready"

# ========================================
# STEP 3E: RUN SPECIFIC TEST
# ========================================
# Run the specific test again using Playwright directly
pnpm exec playwright test test/$SUITE_NAME/e2e.spec.ts -g "exact test name"
# OR for integration tests:
pnpm test:int $SUITE_NAME -t "exact test name"

Did the test fail now?

• ✅ YES: Bundling or production build issue. Look for:

- Missing exports in package.json - Build configuration problems - Code that behaves differently when bundled

• ❌ NO: Unable to reproduce locally. Proceed to step 4.

4. Unable to Reproduce

If you cannot reproduce locally after both attempts:

• Review CI logs more carefully for environment differences
• Check for race conditions (run test multiple times: for i in {1..10}; do pnpm test:e2e...; done)
• Look for CI-specific constraints (memory, CPU, timing)
• Consider if it's a true race condition that's highly timing-dependent

Common Flaky Test Patterns

Race Conditions

• Page navigating while assertions run
• Network requests not settled before assertions
• State updates not completed

Fix patterns:

• Use Playwright's web-first assertions (toBeVisible(), toHaveText())
• Wait for specific conditions, not arbitrary timeouts
• Use waitForFunction() with condition checks

Test Pollution

• Tests leaving data in database
• Shared state between tests
• Missing cleanup in afterEach

Fix patterns:

• Track created IDs and clean up in afterEach
• Use isolated test data
• Don't use deleteAll that affects other tests

Timing Issues

• setTimeout/sleep instead of condition-based waiting
• Not waiting for page stability
• Animations/transitions not complete

Fix patterns:

• Use waitForPageStability() helper
• Wait for specific DOM states
• Use Playwright's built-in waiting mechanisms

Linting Considerations

When fixing e2e tests, be aware of these eslint rules:

• playwright/no-networkidle - Avoid waitForLoadState('networkidle') (use condition-based waiting instead)
• payload/no-wait-function - Avoid custom wait() functions (use Playwright's built-in waits)
• payload/no-flaky-assertions - Avoid non-retryable assertions
• playwright/prefer-web-first-assertions - Use built-in Playwright assertions

Existing code may violate these rules - when adding new code, follow the rules even if existing code doesn't.

Verification

After fixing:

# Ensure dev server is running on port 3000
# Run test multiple times to confirm stability
for i in {1..10}; do
  pnpm exec playwright test test/$SUITE_NAME/e2e.spec.ts -g "exact test name" || break
done

# Run full suite
pnpm exec playwright test test/$SUITE_NAME/e2e.spec.ts

# If you modified bundled code, test with prod build
lsof -ti:3000 | xargs kill -9 2>/dev/null
pnpm prepare-run-test-against-prod
pnpm dev:prod $SUITE_NAME
until curl -s http://localhost:3000/admin > /dev/null; do sleep 1; done
pnpm exec playwright test test/$SUITE_NAME/e2e.spec.ts

The Iron Law

NO FIX WITHOUT REPRODUCTION FIRST

If you propose a fix before completing steps 1-3 of the workflow, you've violated this skill.

This applies even when:

• The fix seems obvious from the logs
• You've seen this error before
• Time pressure from the team
• You're confident about the root cause
• The logs show clear stack traces

No exceptions. Run the reproduction workflow first.

Rationalization Table

Every excuse for skipping reproduction, and why it's wrong:

Red Flags - STOP

These thoughts mean you're about to violate the skill:

• "Let me just fix the obvious issue" → STOP. Execute reproduction first.
• "I'll increase the timeout" → STOP. Find root cause via reproduction.
• "This looks like X from the logs" → STOP. Run the test.
• "I'll run the full suite" → STOP. Run specific test first.
• "Let me check the code" → STOP. Run reproduction first.
• "I can see the problem in the stack trace" → STOP. Reproduce to confirm.
• "The logs are clear enough" → STOP. Execute the commands.
• "Let me analyze..." → STOP. Analysis comes after reproduction.

All of these mean: Stop immediately. Execute the reproduction commands. No analysis without execution.

Example Session

CORRECT approach - executes commands first:

User: CI failed on i18n suite - https://github.com/.../actions/runs/XXX

Agent:
1. Extracted details from CI logs:
   - Suite: i18n
   - Test: "ensure translations update correctly when switching language"
   - File: test/i18n/e2e.spec.ts:124
   - Error: "Error loading language: en" TypeError: network error

2. Executing reproduction workflow:
   [Uses Bash tool to kill ports]
   $ lsof -ti:3000 | xargs kill -9 2>/dev/null || echo "Port 3000 clear"

   [Uses Bash tool with run_in_background=true]
   $ pnpm dev i18n

   [Waits for server to start with curl check]
   $ until curl -s http://localhost:3000/admin > /dev/null; do sleep 1; done && echo "Server ready"

   [Uses Bash tool to run Playwright directly]
   $ pnpm exec playwright test test/i18n/e2e.spec.ts -g "ensure translations update correctly when switching language"

   Output shows: Test PASSED locally

3. Test passed with dev code, trying bundled code:
   [Kills ports again]
   $ lsof -ti:3000 | xargs kill -9 2>/dev/null

   [Uses Bash tool]
   $ pnpm prepare-run-test-against-prod

   [Uses Bash tool with run_in_background=true]
   $ pnpm dev:prod i18n

   [Waits for server]
   $ until curl -s http://localhost:3000/admin > /dev/null; do sleep 1; done

   [Uses Bash tool]
   $ pnpm exec playwright test test/i18n/e2e.spec.ts -g "ensure translations update correctly when switching language"

   Output shows: Test FAILED - reproduced!

4. Now that I've reproduced it, analyzing the failure...
   [Code analysis, root cause identification]

5. Implementing fix: [specific changes]

6. Verification:
   [Uses Bash tool]
   $ for i in {1..10}; do pnpm test:e2e i18n -g "exact test"; done
   All 10 runs passed!

INCORRECT approach - analyzes without executing:

User: CI failed on i18n suite

Agent: Looking at the error logs, I can see this is a race condition...
[Analyzes code, proposes fix without running any commands]

❌ WRONG - This violates the skill. Must execute reproduction commands first.

Common Mistakes

Key Principles

1. Reproduce before fixing: Never propose a fix without reproducing the issue
2. Test specifically: Run the exact failing test, not the full suite
3. Dev first, prod second: Check dev code before bundled code
4. Follow the workflow: No shortcuts - the steps exist to save time
5. Verify stability: Run tests multiple times to confirm fix

Completion: Creating a PR

After you have:

1. ✅ Reproduced the issue
2. ✅ Implemented a fix
3. ✅ Verified the fix passes locally (multiple runs)
4. ✅ Tested with prod build (if applicable)

You MUST prompt the user to create a PR:

The fix has been verified and is ready for review. Would you like me to create a PR with these changes?

Summary of changes:
- [List files modified]
- [Brief description of the fix]
- [Verification results]

IMPORTANT:

• DO NOT automatically create a PR - always ask the user first
• Provide a clear summary of what was changed and why
• Include verification results (number of test runs, pass rate)
• Let the user decide whether to create the PR immediately or make additional changes first

This ensures the user has visibility and control over what gets submitted for review.