fix(strategies): guard foresight sender scan against tool-call items

[dani1005]

Problem

extract_foresight opens by listing the user senders in the memcell:

sender_ids = sorted({m.sender_id for m in memcell.items if m.role == "user"})

memcell.items is a union. Only ChatMessage carries a role; ToolCallRequest has a sender but no role, and ToolCallResult has neither. So a memcell cut from an agent trajectory raises AttributeError: 'ToolCallRequest' object has no attribute 'role' on the first tool item, and the strategy dies before it extracts anything. Every session containing a tool call loses foresights for all of its users.

Found while recording a Langfuse trace fixture against 1.2.1: a four-round agent trajectory produced three everos.ome.extract_foresight error spans, one per retry.

Change

Guard the attribute read with getattr, matching how the rest of the codebase already walks this union:

call site	existing guard
extract/pipeline/user_memory.py:230	getattr(item, "role", None) != "user" — same job, collecting user senders
extract_agent_case.py:131	getattr(item, "sender_id"/"role"/"kind", None)
service/_boundary.py:467	getattr(item, "sender_id", None)

extract_foresight.py:54 was the only unguarded read left, so this completes the pattern rather than introducing a new one.

Semantics are unchanged: a tool call is issued by an agent, so its sender_id could never be a user sender, and ToolCallResult has no sender at all. The set of user senders is exactly the set of role == "user" chat messages either way.

Why the extractor needs no change

ForesightExtractor is already designed to accept an agent-shaped memcell — its docstring states that non-ChatMessage items in memcell.items are silently skipped (agent → user-memory contract), and it filters internally through everalgo's own chat_messages() helper. The crash was purely in our pre-scan, so the fix does not just move the failure one line down.

That helper is deliberately not imported here: it lives in everalgo/user_memory/_render.py, whose module docstring marks it internal and outside the public API.

Testing

New regression test builds a memcell with a ChatMessage plus ToolCallRequest / ToolCallResult and asserts the strategy still finds the one user sender. It fails with AttributeError against the unpatched line and passes with the guard.

tests/unit 1862 passed; ruff check/format, import-linter and the repo gates clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01UyKinsWs1MgoARPoB9R4NW