fix(summarizing_conversation_manager): use model stream to generate summary

[mkmeral]

Description

When SummarizingConversationManager is used without a separate summarization_agent and a ContextWindowOverflowException occurs, the agent crashes with:

ConcurrencyException: Agent is already processing a request. Concurrent invocations are not supported.

Root cause: _generate_summary() calls agent("Please summarize this conversation."), which re-enters stream_async() on the same agent instance. The _invocation_lock (threading.Lock, non-reentrant) is already held by the outer stream_async() call, so the inner call fails immediately.

The fix: Split _generate_summary() into two code paths:

1. When a separate summarization_agent IS provided — keep existing behavior, call summarization_agent(...). This is safe because it's a different agent instance with its own lock.

2. When NO separate agent is provided (the bug case) — call agent.model.stream() directly instead of agent(...). Summarization only needs the model to produce text; it doesn't need tools, callback handlers, tracing, or any other agent machinery. The response stream is processed via process_stream() (the same function the event loop uses internally).

This approach:

• Zero changes to agent.py — the core stays simple
• No concurrency mechanism changes (no RLock, no lock release/reacquire hacks)
• No agent state corruption (metrics, traces, interrupt state are untouched)
• Fix is scoped exactly where the problem is

Related Issues

Fixes the re-entrant lock deadlock in SummarizingConversationManager when no dedicated summarization_agent is configured.

Documentation PR

N/A — no public API changes; behavior is now correct where it previously crashed.

Type of Change

Bug fix

Testing

How have you tested the change?

37 tests total (28 existing updated + 9 new), all passing:

• Unit tests cover both code paths (model-direct vs. dedicated agent), system prompt handling, message construction, error propagation, tool registry behavior, and agent state isolation.
• Integration test (test_full_agent_pipeline_no_reentrant_lock_on_context_overflow) reproduces the exact bug scenario end-to-end: a real Agent instance calls agent("prompt") → model raises ContextWindowOverflowException → reduce_context() fires while the lock is held → summary is generated via model.stream() → event loop retries → agent returns successfully. This test would have caught the original bug.

$ python -m pytest tests/strands/agent/test_summarizing_conversation_manager.py -v
37 passed in 0.31s

$ python -m pytest tests/strands/agent/test_agent.py -v -k "concurren"
2 passed in 0.18s

$ python -m pytest tests/strands/agent/test_agent.py -v
100 passed in 3.28s

$ python -m pytest tests/strands/agent/test_conversation_manager.py -v
26 passed in 0.89s

• I ran hatch run prepare

Checklist

• I have read the CONTRIBUTING document
• I have added any necessary tests that prove my fix is effective or my feature works
• I have updated the documentation accordingly
• I have added an appropriate example to the documentation to outline the feature, or no new docs are needed
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[mkmeral]

/strands review

[codecov]

Codecov Report

❌ Patch coverage is 90.90909% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...sation_manager/summarizing_conversation_manager.py	90.90%	1 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!

[github-actions]

Code Review Summary

Assessment: ✅ Approve

Key Themes:

• Clean separation of summarization paths that correctly addresses the re-entrant lock deadlock
• Excellent test coverage with 37 unit tests covering both code paths, edge cases, and the core bug fix scenario
• Minimal, targeted changes that don't touch agent.py core logic
• Good use of existing infrastructure (process_stream from event_loop)

Strengths:

1. Well-scoped fix: The solution addresses exactly the problem without over-engineering
2. Comprehensive tests: The test_reduce_context_succeeds_with_held_lock test directly validates the bug fix by acquiring the lock before calling reduce_context
3. Clear documentation: Docstrings explain the rationale for each code path
4. Integration testing: The test_context_overflow_triggers_summarization_and_recovery validates the end-to-end scenario

Minor Notes:

• Codecov reports 94.73% patch coverage with 1 partial line - this is acceptable for branch coverage

This is a solid bug fix that follows the project's patterns and maintains code quality. Nice work! 🎉

[afarntrog]

Just noting that we need to remember to add tracing/metrics to this.

[afarntrog]

I would really like to see an integration test for this as well. Testing both cases - agent and direct call.