ワンクリックでManusで任意のスキルを実行

deflake-e2e

スター7

フォーク5

更新日2026年3月27日 20:56

Triage and fix targeted E2E failures or flakes using CI artifacts, traces, local repros, and the smallest real product, fixture, or harness change.

インストール

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

Manusで実行

ソース

keppoai

keppoai/keppo-old

GitHub リポジトリを開く Creator のリポジトリを見る

ダウンロード

Manusで実行

Deflake E2E

Use this skill when a specific Playwright E2E failure or flake needs investigation and a real fix.

Scope

Prefer targeted inputs:
- spec paths
- failing test titles
- GitHub Actions run or job URLs
- Playwright traces, screenshots, or report artifacts
If the user gives no scope, ask for a specific failing area. Do not default to the full suite locally.

Workflow

Confirm the repo is on a working branch, not main, before making changes.
Read the local repo guidance first:
- AGENTS.md
- docs/rules/e2e_testing.md
Start from evidence, not reruns:
- inspect failing CI output, traces, screenshots, and merged Playwright reports when available
- cluster repeated failures by stable symptom before changing code
- separate deterministic regressions from intermittent flakes
Reproduce only the targeted scope locally:
- never run the full E2E suite locally
- use pnpm run test:e2e:base -- tests/e2e/specs/foo.spec.ts
- add --retries=0 when measuring flakiness so Playwright retries do not hide the problem
- only use --repeat-each after an initial targeted run proves the area is worth repeating
If application code changes, rerun pnpm e2e:prepare before the next browser verification. Test-only edits do not require a rebuild unless the harness says otherwise.
Prefer root-cause fixes over test hacks. Valid fix locations include:
- product code
- route logic
- fixtures and seeded backend state
- shared E2E helpers and teardown
- deterministic waiting strategy
- environment and runtime contract
Do not paper over flakes with:
- blanket retries
- arbitrary sleeps
- looser assertions that stop checking the operator-visible behavior
- document-level page.route() mocks for TanStack Start pages when the real failure is in a server function or backend state
For flaky behavior, use the smallest repeat loop that proves the signature:
- start with one targeted repro
- if needed, escalate to --repeat-each=3 or --repeat-each=5
- only push to --repeat-each=10 when the failure is rare or the user explicitly wants stronger confidence
Common failure patterns to look for:
- remount or StrictMode races losing in-flight async work
- TanStack Start server-function requests replacing older /api/... browser calls
- Better Auth same-site flows where browser cookies are present but an explicit cookie header is optional
- stale prebuilt artifacts after app code changes
- backend persistence completing before the page refreshes its table or query state
- transient Convex server errors or teardown timeouts that should be retried within a bounded budget
- selector expectations that assert too early on the wrong surface
Validate at the right level:

rerun the targeted spec after the fix
rerun nearby specs only when they cover the same helper, route, or fixture boundary
use CI for the full suite or broad shard confirmation instead of a local full run

When the fix exposes a reusable testing lesson, update the relevant rule in docs/rules/ in the same change.
Summarize:

what failed
whether it was deterministic or flaky
what evidence proved the root cause
what changed
what local validation ran
what still needs CI confirmation, if anything

Useful Commands

# Targeted local repro
pnpm run test:e2e:base -- tests/e2e/specs/foo.spec.ts --retries=0

# Repeat a targeted spec to measure flakiness
pnpm run test:e2e:base -- tests/e2e/specs/foo.spec.ts --repeat-each=5 --retries=0

# Refresh prebuilt artifacts after app-code changes
pnpm e2e:prepare

# Open a Playwright trace
pnpm exec playwright show-trace path/to/trace.zip

# Download GitHub Actions artifacts when CI evidence matters
gh run view <run-id>
gh run download <run-id> --dir /tmp/e2e-<run-id>

Output Expectations

Report the investigation like an engineering triage, not a changelog:

failure signature
root cause
chosen fix and why
validation completed
residual risk or CI follow-up

このリポジトリの他の Skills

同じリポジトリ

security-review-recent

keppoai/keppo-old

Review recent code changes and broad codebase areas for critical and high severity security vulnerabilities. Use when Codex needs to clear `./out-security-review`, select 25 non-doc, non-test files split across past-day commits, past-week commits, and broad codebase coverage, group them into 6-8 review buckets, spawn one security-review sub-agent per review bucket, persist each candidate finding to `./out-security-review/<starting_filename>_<timestamp>.md` using a filename-safe UTC ISO 8601 basic timestamp like `20260331T214512Z`, re-verify every candidate with fresh sub-agents, and emit final confirmed findings for responsible disclosure.

2026-04-097

create-video-demo

keppoai/keppo-old

Record a short Playwright-based product demo video for significant UI or user-visible functionality changes. Use when Codex needs to show a workflow in motion for a PR, either by recording an existing targeted E2E spec or by creating a disposable one-off Playwright spec that captures the changed behavior in under a minute, then uploading the final video to Vercel Blob and preparing a PR comment body that says "Demo at commit {hash}" with a 1-2 sentence summary and durable hosted link.

2026-04-047

code-architect-recent

keppoai/keppo-old

Review the last 7 days of recent commits for large architecture improvements that dramatically improve maintainability. Use when the agent needs to clear `./out-code-architect`, select 25 non-doc, non-test files from recent commits across different parts of the repo, spawn one architecture-review sub-agent per starting file, persist each candidate finding to `./out-code-architect/<starting_filename>_<timestamp>.md` using a filename-safe UTC ISO 8601 basic timestamp like `20260331T214512Z`, re-verify every candidate with fresh sub-agents, and emit final confirmed findings for issue filing.

2026-04-037

bug-finder-recent

keppoai/keppo-old

Review the last 7 days of recent commits for critical and high severity bugs that break user experience. Use when the agent needs to clear `./out-bug-finder`, select 25 non-doc, non-test files from recent commits across different parts of the repo, spawn one bug-review sub-agent per starting file, persist each candidate finding to `./out-bug-finder/<starting_filename>_<timestamp>.md` using a filename-safe UTC ISO 8601 basic timestamp like `20260331T214512Z`, re-verify every candidate with fresh sub-agents, and emit final confirmed findings for issue filing.

2026-04-037

keppo-check-workflows

keppoai/keppo-old

Review recent GitHub Actions workflow runs, separate expected noise from actionable failures, and write a machine-readable workflow health report for deterministic post-agent issue handling.

2026-04-017

setup-pr-push

keppoai/keppo-old

Print the full host-side `scripts/pr.sh` command needed to create or switch to a feature branch, commit current changes, push, and open or update a PR outside the Docker sandbox using explicit flag-style inputs.

2026-03-307

Deflake E2E

Use this skill when a specific Playwright E2E failure or flake needs investigation and a real fix.

Scope

Prefer targeted inputs:

spec paths
failing test titles
GitHub Actions run or job URLs
Playwright traces, screenshots, or report artifacts

If the user gives no scope, ask for a specific failing area. Do not default to the full suite locally.

Workflow

Confirm the repo is on a working branch, not main, before making changes.

Read the local repo guidance first:

Start from evidence, not reruns:

inspect failing CI output, traces, screenshots, and merged Playwright reports when available
cluster repeated failures by stable symptom before changing code
separate deterministic regressions from intermittent flakes

Reproduce only the targeted scope locally:

never run the full E2E suite locally
use pnpm run test:e2e:base -- tests/e2e/specs/foo.spec.ts
add --retries=0 when measuring flakiness so Playwright retries do not hide the problem
only use --repeat-each after an initial targeted run proves the area is worth repeating

If application code changes, rerun pnpm e2e:prepare before the next browser verification. Test-only edits do not require a rebuild unless the harness says otherwise.

Prefer root-cause fixes over test hacks. Valid fix locations include:

product code
route logic
fixtures and seeded backend state
shared E2E helpers and teardown
deterministic waiting strategy
environment and runtime contract

Do not paper over flakes with:

blanket retries
arbitrary sleeps
looser assertions that stop checking the operator-visible behavior
document-level page.route() mocks for TanStack Start pages when the real failure is in a server function or backend state

For flaky behavior, use the smallest repeat loop that proves the signature:

start with one targeted repro
if needed, escalate to --repeat-each=3 or --repeat-each=5
only push to --repeat-each=10 when the failure is rare or the user explicitly wants stronger confidence

Common failure patterns to look for:

remount or StrictMode races losing in-flight async work
TanStack Start server-function requests replacing older /api/... browser calls
Better Auth same-site flows where browser cookies are present but an explicit cookie header is optional
stale prebuilt artifacts after app code changes
backend persistence completing before the page refreshes its table or query state
transient Convex server errors or teardown timeouts that should be retried within a bounded budget
selector expectations that assert too early on the wrong surface

Validate at the right level:

rerun the targeted spec after the fix

rerun nearby specs only when they cover the same helper, route, or fixture boundary

use CI for the full suite or broad shard confirmation instead of a local full run

When the fix exposes a reusable testing lesson, update the relevant rule in docs/rules/ in the same change.

Summarize:

what failed

whether it was deterministic or flaky

what evidence proved the root cause

what changed

what local validation ran

what still needs CI confirmation, if anything

Useful Commands

# Targeted local repro pnpm run test:e2e:base -- tests/e2e/specs/foo.spec.ts --retries=0 # Repeat a targeted spec to measure flakiness pnpm run test:e2e:base -- tests/e2e/specs/foo.spec.ts --repeat-each=5 --retries=0 # Refresh prebuilt artifacts after app-code changes pnpm e2e:prepare # Open a Playwright trace pnpm exec playwright show-trace path/to/trace.zip # Download GitHub Actions artifacts when CI evidence matters gh run view <run-id> gh run download <run-id> --dir /tmp/e2e-<run-id>

Output Expectations

Report the investigation like an engineering triage, not a changelog:

failure signature

root cause

chosen fix and why

validation completed

residual risk or CI follow-up

name	deflake-e2e
description	Triage and fix targeted E2E failures or flakes using CI artifacts, traces, local repros, and the smallest real product, fixture, or harness change.