feat: プロンプト修正

[yu-sugimoto]

No description provided.

[Copilot]

Pull request overview

This PR refactors the “analyze / implement / create PR” worker pipeline by centralizing prompt text and shared worker utilities, and adjusts the plan JSON shape passed through S3 so downstream steps have richer context.

Changes:

• Added shared modules for worker prompts (docker/prompts.py) and common worker utilities (docker/worker_common.py), and updated workers to use them.
• Updated analysis/implementation flow to include richer plan context (title/concept/files) and improved patch/PR creation handling (e.g., empty patch early-exit).
• Removed React StrictMode wrapper in the frontend entrypoint.

Reviewed changes

Copilot reviewed 9 out of 9 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
frontend/src/main.tsx	Removes StrictMode from app root render.
docker/worker_common.py	New shared logging/S3/tool-stream helpers for workers.
docker/worker.Dockerfile	Copies new shared worker modules into the worker image.
docker/worker-implement.py	Switches to shared prompts/common utils; adjusts plan handling and patch creation flow.
docker/worker-createpr.py	Uses shared prompts/common utils; adds plan context to PR prompt; skips PR creation on empty patch.
docker/worker-analyze.py	Uses shared prompts/common utils; validates proposals; includes instruction in proposals.json.
docker/prompts.py	New centralized prompt definitions and prompt-building helpers.
backend/app/workflow/session_implementation_graph.py	Changes how plan.json is constructed before uploading to S3.
backend/app/usecase/session_usecase.py	Passes richer per-proposal plan JSON into implementation workflow.

Comments suppressed due to low confidence (1)

frontend/src/main.tsx:25

• React StrictMode was removed entirely. This disables important dev-only checks (including double-invocation to surface side effects) and may hide issues that StrictMode would catch. If the goal is to avoid StrictMode behavior only for the worker/screenshot flow, consider keeping StrictMode enabled for normal dev builds and gating it behind an env flag instead of removing it globally.

createRoot(document.getElementById('root')!).render(<RouterProvider router={router} />)

---

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

[Copilot]

create_k8s_job now unconditionally json.loads() a string proposal_plan. If proposal_plan is ever a non-JSON string (e.g., legacy plain-text plans / already-formatted prompts), this will raise JSONDecodeError and fail the whole job creation path. To keep backward compatibility and avoid hard failures, wrap the loads in try/except JSONDecodeError and fall back to treating raw_plan as the plan value (e.g., {"plan": raw_plan}).

[Copilot]

After centralizing message processing, tool-use logging now relies on _emit_tool_detail, but it only handles Read/Write/Edit. Bash tool invocations (and potentially other tools) will no longer produce the helpful "Running: ..." logs that existed before, reducing observability when debugging worker failures. Consider extending _emit_tool_detail to handle Bash (and any other relevant tools) similarly to the previous per-worker implementations.

Suggested change

	elif tool_name in ("Bash", "bash"):
	# Log Bash tool invocations with a concise "Running: ..." message.
	cmd = (
	params.get("command")
	or params.get("cmd")
	or params.get("script")
	)
	if cmd is None:
	return
	if not isinstance(cmd, str):
	try:
	cmd_str = json.dumps(cmd, ensure_ascii=False)
	except TypeError:
	cmd_str = str(cmd)
	else:
	cmd_str = cmd
	# Use only the first line for the summary to keep logs compact.
	summary = cmd_str.strip().splitlines()[0][:200]
	emit_log(phase, f"Running: {summary}", detail=cmd_str)
	elif tool_name in PLAYWRIGHT_MCP_TOOLS:
	# Emit a concise log for Playwright MCP tools to aid debugging.
	emit_log(phase, f"Running Playwright tool: {tool_name}", detail=json.dumps(params, ensure_ascii=False))

[Copilot]

The implementation worker drops "instruction" from the plan object before passing it into implement_changes (it filters out "instruction"), but implement_changes formats the plan using plan_data.get("instruction") for the "User Request" header. This will usually render as "N/A", removing critical user context from the implementation prompt. Keep "instruction" in the plan passed to implement_changes, or explicitly inject the separate instruction variable when building formatted_plan.

Suggested change

	# Pass the full plan object (title, concept, plan, files) to implement_changes
	proposal_plan = {
	k: v
	for k, v in plan_data.items()
	if k not in ("device_type", "instruction")
	# Pass the full plan object (title, concept, plan, files, instruction) to implement_changes
	proposal_plan = {
	k: v
	for k, v in plan_data.items()
	if k != "device_type"