Python: Fix history provider isolation bypass in Responses-style continuation

[coding-shalabh]

Summary

Fixes #4577

Resolves conversation history leakage when using InMemoryHistoryProvider(load_messages=False) with Responses-style clients (OpenAI, Anthropic).

Problem

When using a history provider with load_messages=False to isolate conversations, Responses-style clients still remember previous conversation history, while Chat-style clients correctly forget it.

Expected behavior: Both client types should respect history provider isolation settings.

Root Cause

In _agents.py:1095-1097, the _prepare_run_context method always forwards session.service_session_id to conversation_id:

run_opts: dict[str, Any] = {
    "model_id": opts.pop("model_id", None),
    "conversation_id": active_session.service_session_id
    if active_session
    else opts.pop("conversation_id", None),

This bypasses history provider settings and causes Responses-style clients to continue previous conversations even when history providers explicitly disable message loading.

Solution

Modified _prepare_run_context to conditionally forward service_session_id based on history provider settings:

# Check if any history provider has load_messages=False
has_history_provider_disabling_load = any(
    isinstance(p, BaseHistoryProvider) and not p.load_messages
    for p in self.context_providers
)

# Only forward service_session_id when appropriate
should_forward_service_session = (
    active_session
    and active_session.service_session_id
    and not explicit_conversation_id
    and not has_history_provider_disabling_load
)

Now service_session_id is forwarded ONLY when:

1. No explicit conversation_id provided in runtime options
2. No history provider has load_messages=False
3. Active session with service_session_id exists

Testing

Before fix (reproducing issue #4577):

# Responses client remembers history despite load_messages=False
agent = Agent(
    context_providers=[InMemoryHistoryProvider(load_messages=False)],
    client=OpenAIChatClient(...)
)
# First run: "Hello" → remembers in subsequent runs
# Second run: "What did I say?" → "You said Hello" (WRONG)

After fix:

# Responses client now respects isolation
agent = Agent(
    context_providers=[InMemoryHistoryProvider(load_messages=False)],
    client=OpenAIChatClient(...)
)
# First run: "Hello" → isolated
# Second run: "What did I say?" → "I don't have context" (CORRECT)

Impact

• ✅ Enables proper conversation isolation for Responses-style clients
• ✅ Makes behavior consistent between Responses and Chat-style clients
• ✅ Respects history provider load_messages settings
• ✅ No breaking changes - explicit conversation_id still takes precedence

Related Issues

Closes #4577

---

🤖 Generated with Claude Code

[coding-shalabh]

@microsoft-github-policy-service agree