run-examples: 28_ask_agent_example.py fails with Anthropic 400 (orphaned tool_use without tool_result)

The run-examples workflow is failing on examples/01_standalone_sdk/28_ask_agent_example.py.

Symptom

The example fails with an Anthropic 400 invalid_request_error surfaced through LiteLLM/OpenAI as openai.BadRequestError:

messages.6: `tool_use` ids were found without `tool_result` blocks immediately after: toolu_01DqrY27EqjnoefvSxz3uwz6, ... Each `tool_use` block must have a corresponding `tool_result` block in the next message.

...mapped to openhands.sdk.llm.exceptions.types.LLMMalformedConversationHistoryError, and reported by the harness as ConversationRunError.

Root cause

Conversation.ask_agent() builds its LLM context from the cached, incrementally-maintained self.state.view. That view skips enforce_properties on linear appends, so it can transiently contain an ActionEvent whose matching ObservationEvent has not landed yet — i.e. a tool call that the main agent is still executing on another thread (this example deliberately calls ask_agent() mid-run, on a background thread while the "Explore the current directory" task is in flight).

When that orphaned tool_use reaches Anthropic with no following tool_result, it is rejected with a 400. A fresh View.from_events(active_branch) re-runs property enforcement and drops the orphaned action before serialization.

Fix

Already available (open, unmerged): #4630 — fix(sdk): drop in-flight tool calls from ask_agent context, which builds the ask-agent context from View.from_events(self._state.active_branch()) instead of the cached view.

Evidence

• Failing run (release branch rel-1.44.0, no fix): https://github.com/OpenHands/software-agent-sdk/actions/runs/32863616276
• Passing run on the fix branch (fix/ask-agent-inflight-tool-call): https://github.com/OpenHands/software-agent-sdk/actions/runs/32871786638

This issue was created by an AI agent (OpenHands) on behalf of a user.

Start with examples/01_standalone_sdk/28_ask_agent_example.py and inspect Conversation.ask_agent(), then compare the context construction with the approach described in pull request #4630. Run the example or the run-examples workflow and confirm that the in-flight tool call no longer produces an Anthropic 400 or ConversationRunError.