feat: playwrite mcp に変更

[yu-sugimoto]

No description provided.

[Copilot]

Pull request overview

Worker のスクリーンショット取得を従来の take-screenshot.mjs 実行から Playwright MCP 経由に切り替え、分析結果から推定した device_type（desktop/mobile）をワーカー実行まで伝搬できるようにする変更です。

Changes:

• Worker に @playwright/mcp を導入し、MCP ツール経由で before/after スクリーンショットを取得するよう変更
• analyzer が device_type を推定して proposals.json に含め、backend がそれを読み取り implementation worker まで引き回すよう変更
• K8s worker 実行環境に CLAUDE_CODE_MAX_OUTPUT_TOKENS を追加し、ログ取得のデフォルト tail を増加

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
docker/worker.Dockerfile	@playwright/mcp のインストール追加、take-screenshot.mjs のコピー削除
docker/worker-analyze.py	MCP サーバ/ツール定義追加、device_type 推定・出力形式変更、MCP でスクリーンショット取得
docker/worker-implement.py	MCP サーバ/ツール定義追加、device_type をもとに MCP でスクリーンショット/デバッグ、plan.json から device_type 読み取り
docker/take-screenshot.mjs	旧スクリーンショットスクリプト削除
backend/app/workflow/state.py	workflow state に device_type を追加
backend/app/workflow/session_implementation_graph.py	worker に渡す plan を S3 上で {plan, device_type} に拡張
backend/app/workflow/session_analyzer_graph.py	S3 の proposals.json から device_type を読み取るよう変更
backend/app/usecase/session_usecase.py	analyzer 結果の device_type をログ/implementation 起動に反映
backend/app/infra/s3_client.py	get_proposals アクセサ削除（呼び出し側が直接 download_json へ）
backend/app/infra/k8s_client.py	worker env 追加、ログ tail デフォルト変更
backend/README.md	デプロイ手順の追記

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The MCP server is started via npx @playwright/mcp@latest, which is non-deterministic and may trigger network fetches at runtime. Consider pinning the MCP package version (and/or invoking the already-installed global binary) so worker executions don’t depend on latest resolution.

Suggested change

	PLAYWRIGHT_MCP_SERVERS = {
	"playwright": {
	"command": "npx",
	"args": ["@playwright/mcp@latest", "--headless", "--browser", "chromium"],
	},
	"playwright_mobile": {
	"command": "npx",
	"args": ["@playwright/mcp@latest", "--headless", "--browser", "chromium", "--device", "iPhone 15"],
	PLAYWRIGHT_MCP_PACKAGE_SPEC = os.environ.get("PLAYWRIGHT_MCP_PACKAGE_SPEC", "@playwright/mcp@0.5.0")
	PLAYWRIGHT_MCP_SERVERS = {
	"playwright": {
	"command": "npx",
	"args": [PLAYWRIGHT_MCP_PACKAGE_SPEC, "--headless", "--browser", "chromium"],
	},
	"playwright_mobile": {
	"command": "npx",
	"args": [PLAYWRIGHT_MCP_PACKAGE_SPEC, "--headless", "--browser", "chromium", "--device", "iPhone 15"],

[Copilot]

The MCP server configuration uses npx @playwright/mcp@latest, which is non-deterministic and can require network access at runtime. Pin the MCP package version (and keep it compatible with the pinned Playwright version) to avoid unexpected breakage.

Suggested change

	PLAYWRIGHT_MCP_SERVERS = {
	"playwright": {
	"command": "npx",
	"args": ["@playwright/mcp@latest", "--headless", "--browser", "chromium"],
	},
	"playwright_mobile": {
	"command": "npx",
	"args": ["@playwright/mcp@latest", "--headless", "--browser", "chromium", "--device", "iPhone 15"],
	PLAYWRIGHT_MCP_PACKAGE = "@playwright/mcp@0.4.4"
	PLAYWRIGHT_MCP_SERVERS = {
	"playwright": {
	"command": "npx",
	"args": [PLAYWRIGHT_MCP_PACKAGE, "--headless", "--browser", "chromium"],
	},
	"playwright_mobile": {
	"command": "npx",
	"args": [PLAYWRIGHT_MCP_PACKAGE, "--headless", "--browser", "chromium", "--device", "iPhone 15"],

[Copilot]

implement_changes() exceptions are logged but then ignored, and the workflow continues to launch/screenshot/commit. This can produce “successful” artifacts even when implementation failed. Consider re-raising (or exiting non-zero) after logging so failures don’t get silently converted into partial results.

Suggested change

	emit_log("implementing", f"Implementation interrupted: {e}", detail=str(e))
	emit_log("implementing", f"Implementation interrupted: {e}", detail=str(e))
	raise

[Copilot]

S3Client.download_json() calls json.loads() and will raise on invalid JSON; extract_results() doesn’t catch that, so a malformed/partial proposals.json will crash the graph instead of returning a clean failure. Wrap the download/parse in a try/except (e.g., json.JSONDecodeError) and return a failed status with a helpful error.

[Copilot]

The npm install -g playwright@1.49.0 @playwright/mcp@latest line pulls @playwright/mcp from npm using the mutable latest dist-tag, which makes worker image builds non-reproducible and exposes the build/runtime to supply-chain risk if that package or its dist-tag is compromised. Because this global dependency runs inside the worker container (likely with access to repo data and possibly credentials), a malicious update to @playwright/mcp would be automatically trusted and executed on the next image build without explicit review. Pin @playwright/mcp to a specific version and update it intentionally as needed to reduce this attack surface.