fix: track real message count since last summary

[pradneshfernandez]

Fixes #5504

Summary

LLMContextSummarizer._should_summarize() was calculating unsummarized messages using len(context.messages) - 1. The issue was that this calculation never reset after running a summary, and it blindly assumed a leading system message was always present.

Because of this, max_unsummarized_messages was effectively evaluating total context size rather than new messages. It also caused context summarization to loop continuously, as the newly injected summary message was counted as a fresh turn and immediately triggered another summary.

Fix

Following the fix suggested in #5504:

• Replaced the hardcoded - 1 with a _messages_since_summary_baseline marker to track state accurately.
• Updated _apply_summary() to reset the baseline to num_system_preserved + 1, accounting for both the preserved system message and the newly added summary message.
• Updated _should_summarize() to compute messages_since_summary against this new baseline instead of len(messages) - 1.

Note: This PR only addresses the message-count trigger. The related token-trigger issue (#5505) is intentionally left out and will be handled separately.

Test plan

• uv run pytest tests/test_llm_context_summarizer.py tests/test_context_summarization.py
• Verified manually that summaries no longer re-trigger immediately after running.

[markbackman]

Thanks for the fix. I pushed a commit on top (bc49297) that keeps your approach but changes when the count is computed.

Your version cached the number of excluded leading messages in _messages_since_summary_baseline, set once in __init__ and again in _apply_summary. That number goes stale in two cases: the app replaces the context after a summary (via set_messages or LLMMessagesUpdateFrame), or a system message is added after the summarizer was constructed.

The commit replaces the cached int with a reference to the summary message the summarizer injected (_summary_message), and a new _messages_since_summary() that computes the count from the current context each time _should_summarize runs. It excludes messages[0] if it is a system message, and the next slot if it is the same object as _summary_message. Matching by identity means a reset context naturally drops the summary from the count, and a user message that happens to look like a summary is not mistaken for one.

I also added three tests in tests/test_llm_context_summarizer.py: the threshold fires at exactly N messages with and without a leading system message, only messages added since a summary count, and replacing the context resets the count. The last one fails against the cached-int version.

[markbackman]

LGTM! Thanks for the contribution.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
...t/processors/aggregators/llm_context_summarizer.py	96.85% <100.00%> (+0.23%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.