fix: cap openai <2.45 for openai-agents 0.14.x compatibility

openai 2.45.0 made `InputTokensDetails.cache_write_tokens` a required
field, but openai-agents 0.14.x (the latest release) still constructs
`InputTokensDetails(cached_tokens=0)` in agents/usage.py. As a result
every `Runner.run_streamed` call raises a pydantic ValidationError at
context setup, before any LLM request.

The repo's own uv.lock pins openai 2.40.0 so unit tests pass, but the
tutorial agents resolve their dependencies fresh (no lockfile) and pick
up openai 2.45.0, which is why the "Test Tutorial Agents" workflow fails
on every branch for all openai-agents based tutorials (010_multiturn,
020_streaming, temporal openai-agents tools/HIL, state_machine, and the
other-SDKs base tutorial).

Capping openai <2.45 in the agentex-sdk deps makes fresh resolves land
on openai 2.44.0, which is compatible. Drop the ceiling once openai-agents
ships a fix.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: declan-scale

adk/pyproject.toml

@@ -41,7 +41,14 @@ dependencies = [
     # LLM provider integrations
     "litellm>=1.83.7,<2",
     "openai-agents>=0.14.3,<0.15",
-    "openai>=2.2,<3", # Required by openai-agents; litellm now supports openai 2.x (issue #13711 resolved: https://github.com/BerriAI/litellm/issues/13711)
+    # Cap <2.45: openai 2.45.0 makes InputTokensDetails.cache_write_tokens a
+    # required field, but openai-agents 0.14.x still builds
+    # InputTokensDetails(cached_tokens=0) (agents/usage.py), so every
+    # Runner.run_streamed raises a pydantic ValidationError at context setup.
+    # openai-agents 0.14.8 is the latest release, so there is no newer version
+    # to bump to; drop this ceiling once openai-agents ships a fix.
+    # litellm now supports openai 2.x (issue #13711 resolved: https://github.com/BerriAI/litellm/issues/13711)
+    "openai>=2.2,<2.45",
     "claude-agent-sdk>=0.1.0",
     "pydantic-ai-slim>=1.0,<2",
     "langgraph-checkpoint>=2.0.0",