fix(tutorials): repair five tutorials that fail with a real model

[danielmillerp]

What this fixes

Five tutorials fail or flake against a real model. Three have MCP servers that die at startup, and two have agents that sometimes answer without calling their tool:

Tutorial	Cause	Fix
10_async/10_temporal/020_state_machine	pydantic 2.14 break (below)	add --with pydantic<2.14 to the existing uvx pin
10_async/10_temporal/010_agent_chat	same	same pin on its uvx openai-websearch-mcp
10_async/00_base/040_other_sdks	same, plus its image has no Node, so npx for the sequential-thinking server raises FileNotFoundError: 'npx'	same pin, and install nodejs npm in its Dockerfile
10_async/10_temporal/110_pydantic_ai, 10_async/00_base/110_pydantic_ai	the prompt says to "use tools when they would help", so gpt-4o-mini sometimes answers a weather question without calling get_weather (seen failing all 3 attempts in one run)	add a guideline: when asked for the current weather in a place, call get_weather and never answer from memory. The test is unchanged.

The pydantic break. openai-websearch-mcp pins mcp==1.13.1, which imports eval_type_backport from pydantic's internals. pydantic 2.14.0, released 2026-10-08, removed it. uvx resolves each server unpinned, so it picks up 2.14.0 and the server dies at import:

ImportError: cannot import name 'eval_type_backport' from 'pydantic._internal._typing_extra'

The agent's MCP context then fails with Connection closed before the model runs, and the tests report "Did not use tool requests" or time out.

Verification

• uvx --with "mcp<2" openai-websearch-mcp: exits 1 with the ImportError above (pydantic 2.14.0).
• With --with "pydantic<2.14" added: starts and exits cleanly (exit 0). mcp-server-time and mcp-server-fetch with the same pin also exit 0.
• npx -y @modelcontextprotocol/server-sequential-thinking starts ("running on stdio") once Node is installed.
• All three tutorials failed the same way on agentex main with a model key, so this isn't a server regression.

The tutorial images need to be republished after merge for the agentex integration tests to pick them up.

🤖 Generated with Claude Code

Confidence Score: 5/5

The PR appears safe to merge; the earlier weather-tool concern has been addressed.

Summary

The PR updates three tutorials so their MCP servers use a compatible Pydantic version, adds Node and npm to an image that runs an npx server, and tells two weather agents to call get_weather.

• MCP servers stay on a Pydantic version their web-search tool supports.
• The 040_other_sdks tutorial image can run its npx server.
• Weather questions tell both Pydantic AI agents to call get_weather.

_{Reviews (5) · Last reviewed commit: "fix(tutorials): limit the get_weather in..." · Reviewed by Greptile}