Python: Feature/fix context provider lifecycle agentic mode

[farzad528]

Motivation and Context

When using AzureAISearchContextProvider in agentic mode (Knowledge Base retrieval), the second and subsequent queries fail with:

RuntimeError: Retrieval client not initialized. Ensure Knowledge Base is set up correctly.

This bug prevents users from having multi-turn conversations with agents that use Knowledge Base-powered RAG. The first query works, but any follow-up questions fail because the KnowledgeBaseRetrievalClient connection is prematurely closed.

Root Cause: The ChatAgent._prepare_thread_and_messages method wraps the context provider invocation with async with self.context_provider:, which calls __aexit__ after each query. This closes stateful resources (like the retrieval client) after the first query completes.

Why semantic mode was unaffected: The SearchClient used in semantic mode is stateless - each search is independent. The agentic mode's KnowledgeBaseRetrievalClient maintains a persistent connection that gets closed by the erroneous async with wrapper.

Description

Changes in python/packages/core/agent_framework/_agents.py:

Removed the async with self.context_provider: wrapper in _prepare_thread_and_messages. The context provider's lifecycle should be managed by the user (via async with ChatAgent(...) as agent) or persist across multiple invocations - not be opened/closed on every query.

Before (incorrect):

if self.context_provider:
    async with self.context_provider:  # ❌ Closes provider after each query!
        context = await self.context_provider.invoking(...)

After (correct):

if self.context_provider:
    # Note: We don't use 'async with' here because the context provider's lifecycle
    # should be managed by the user (via async with) or persist across multiple invocations.
    # Using async with here would close resources (like retrieval clients) after each query.
    context = await self.context_provider.invoking(...)

This aligns with how AggregateContextProvider works - it enters providers once in __aenter__ and exits once in __aexit__, while invoking() does not use async with.

Additional fix in python/samples/getting_started/context_providers/azure_ai_search/README.md:

Corrected the Azure AI Foundry project endpoint format in documentation from https://myproject.api.azureml.ms to https://<resource-name>.services.ai.azure.com/api/projects/<project-name>.

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the Contribution Guidelines
• All unit tests pass, and I have added new tests where possible
• Is this a breaking change? If yes, add "[BREAKING]" prefix to the title of the PR.

[markwallace-microsoft]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/core/agent_framework
_agents.py	289	52	82%	329, 390–392, 438, 492, 510, 672, 853, 856–858, 986–989, 991, 994–996, 1084, 1125, 1127, 1136–1141, 1147, 1149, 1159–1160, 1167, 1169–1170, 1178–1182, 1190–1191, 1193, 1198, 1200, 1234, 1279–1280, 1282, 1284, 1295
TOTAL	16039	2519	84%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
2296	130 💤	0 ❌	0 🔥	56.070s ⏱️

[Copilot]

Pull request overview

This PR fixes a critical bug in the context provider lifecycle management that prevented multi-turn conversations when using AzureAISearchContextProvider in agentic mode (Knowledge Base retrieval). The bug caused the second and subsequent queries to fail because the retrieval client connection was being closed after each query.

Key Changes:

• Removed the async with self.context_provider wrapper in _prepare_thread_and_messages that was incorrectly closing context provider resources after each query
• Added a clarifying comment explaining that context provider lifecycle should be managed by the user
• Corrected the Azure AI Foundry project endpoint format in documentation

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
python/packages/core/agent_framework/_agents.py	Removed the async with wrapper around context_provider.invoking() that was causing premature resource cleanup after each query, and added a comment explaining the lifecycle management approach
python/samples/getting_started/context_providers/azure_ai_search/README.md	Fixed the Azure AI Foundry project endpoint URL format from incorrect AzureML format to correct AI Foundry format

[Copilot]

Missing test coverage for the multi-turn conversation bug that was fixed. While the fix correctly removes the async with wrapper that was closing resources after each query, there's no test verifying that multiple consecutive calls to invoking() work correctly with agentic mode context providers.

Consider adding a test that:

1. Creates an agent with a context provider that tracks __aenter__/__aexit__ calls
2. Makes multiple consecutive queries via agent.run()
3. Verifies the context provider's __aexit__ is only called when the agent's context exits, not after each query