Fix: don't skip user messages to extract identity

[ea-rus]

self._cortex.maybe_update_identity is called on every 5th message. and only the last message is sent to it
Fix: accumulate previous messages and send all of them on 5th step

[entelligence-ai-pr-reviews]

EntelligenceAI PR Summary

Refactors identity update logic in anton/core/session.py to use a buffered, batched approach for richer context across conversation turns.

• Introduced _identity_buffer list to accumulate user inputs across turns
• Every 5 turns, all buffered inputs are joined into a single concatenated string and passed to maybe_update_identity
• Buffer is cleared after each batch dispatch
• Removed the prior isinstance(user_input, str) type guard

---

Confidence Score: 2/5 - Changes Needed

Not safe to merge — the refactor in anton/core/session.py introduces a runtime TypeError by unconditionally appending user_input (typed str | list[dict]) to _identity_buffer: list[str], then calling "\n\n".join(self._identity_buffer) which will crash when user_input is a list[dict]. The PR's goal of buffering and batching identity updates across turns is a reasonable improvement, but removing the isinstance(user_input, str) guard without replacing it with equivalent type-safe handling breaks the multi-modal input path that the guard was explicitly protecting.

Key Findings:

• In _identity_buffer accumulation logic, user_input of type list[dict] is appended without type-checking, causing "\n\n".join(self._identity_buffer) to raise TypeError at batch dispatch time — a direct crash on any non-string user input.
• The original isinstance(user_input, str) guard was not merely a style choice; it was load-bearing logic that filtered out structured (multi-modal) inputs before string operations. Removing it without a replacement or serialization strategy is a correctness regression.
• The batching logic itself (every 5 turns, clear buffer) is otherwise sound in concept, but the buffer should either enforce list[str] at insertion time with a type guard, or serialize list[dict] inputs to a string representation before appending.

Files requiring special attention

• anton/core/session.py

[entelligence-ai-pr-reviews]

Refactors identity update logic in anton/core/session.py to use a buffered, batched approach for richer context across conversation turns.

• Introduced _identity_buffer list to accumulate user inputs across turns
• Every 5 turns, all buffered inputs are joined into a single concatenated string and passed to maybe_update_identity
• Buffer is cleared after each batch dispatch
• Removed the prior isinstance(user_input, str) type guard

[entelligence-ai-pr-reviews]

Correctness: user_input is typed str | list[dict], and the old code guarded with isinstance(user_input, str) before using it; the new code unconditionally appends it to _identity_buffer: list[str], so when user_input is a list[dict], the subsequent "\n\n".join(self._identity_buffer) raises TypeError: sequence item N: expected str instance, list found.

🤖 AI Agent Prompt for Cursor/Windsurf

📋 Copy this prompt to your AI coding assistant (Cursor, Windsurf, etc.) to get help fixing this issue

In anton/core/session.py around line 784, the diff removed the `isinstance(user_input, str)` guard before appending to `self._identity_buffer`. Since `user_input` is typed `str | list[dict]`, appending a `list[dict]` value to `_identity_buffer: list[str]` and then calling `'\n\n'.join(self._identity_buffer)` at line 787 will raise a TypeError at runtime. Fix: wrap the append in `if isinstance(user_input, str):` to restore the original type guard, while keeping the new buffering/flush logic intact.

[alecantu7]

Nice catch on the missing messages. One concern on the approach: _identity_buffer lives only on the live session, so if Anton restarts mid-window the buffered turns are lost and never scanned. Reading the last 5 user messages from self._history at flush time would be restart-safe and avoid the extra state.

Minor: this diff references user_msg_str, which I don't see defined in main — looks like it depends on a sibling change that hasn't landed yet.