Context overflow when compacting history

[tino097]

Fixes #106

Add three-tier ContextOverflowError recovery to session helpers

[entelligence-ai-pr-reviews]

EntelligenceAI PR Summary

Refactors hard_truncate_history in anton/core/session.py and adds comprehensive unit tests in tests/test_chat.py to cover history truncation edge cases.

• Replaced single-pass truncation logic with an iterative while loop in hard_truncate_history
• Added handling to strip assistant-first entries and orphaned tool_result-only user messages at the head of retained history
• Preserves non-tool_result content blocks in mixed-content user messages instead of dropping the entire message
• Added loop break condition once a valid head message is found
• Introduced four unit tests covering: no-op behavior, pair boundary preservation, orphaned tool-result regression, and mixed-content block filtering

---

Confidence Score: 5/5 - Safe to Merge

Safe to merge — this PR refactors hard_truncate_history in anton/core/session.py with an iterative while loop and adds proper handling for assistant-first entries and orphaned tool_result-only user messages at the head of retained history. The change also correctly preserves non-tool_result content blocks in mixed-content user messages rather than silently dropping entire messages, which is a meaningful correctness improvement. Comprehensive unit tests in tests/test_chat.py cover the edge cases introduced by this logic, and no review comments — new or pre-existing — flag any issues with this implementation.

Key Findings:

• The iterative while loop in hard_truncate_history correctly handles the multi-pass nature of stripping assistant-first and orphaned tool_result entries, which the previous single-pass approach would have missed in chained edge cases.
• Preserving non-tool_result content blocks in mixed-content user messages is a correctness fix over wholesale message removal, reducing potential silent data loss in conversation history.
• New unit tests in tests/test_chat.py directly exercise the edge cases (orphaned tool_result messages, assistant-first head, mixed-content preservation) introduced by the refactor, providing meaningful regression coverage.
• No review comments were generated and no pre-existing unresolved issues are flagged, giving high confidence that the implementation is clean and well-reasoned.

Files requiring special attention

• anton/core/session.py
• tests/test_chat.py

[entelligence-ai-pr-reviews]

Refactors ChatSession in anton/core/session.py to centralize and improve context overflow recovery logic.

• Introduced plan_with_recovery: non-streaming helper with three-tier recovery (normal → summarize+compact → hard-truncate)
• Introduced plan_stream_with_recovery: streaming analogue that yields StreamContextCompacted events on compaction
• Added hard_truncate_history: last-resort fallback that preserves only the tail of history while enforcing API message-ordering rules
• Promoted _compacted_this_turn from a local variable to an instance variable, reset at the start of each turn, preventing double-summarization across both paths
• Replaced all direct self._llm.plan and self._llm.plan_stream call sites in turn and _stream_and_handle_tools with the new helpers
• Refactored task-completion verifier to use a messages_factory callable, ensuring post-compaction history is reflected on retry

[tino097]

@Entelligence review

[entelligence-ai-pr-reviews]

Starting fresh pr review. This may take a few minutes

[entelligence-ai-pr-reviews]

New pr review completed! Check the pr for comments.

[torrmal]

looks good to me