User-side extraction strategies crash on ToolCallRequest because they assume every memcell item has role

[rendigua2025-gif]

Summary

Some user-side extraction strategies assume every item in memcell.items has a role attribute. Tool-call trajectories can include ToolCallRequest items, which do not have role, causing downstream strategies to crash and eventually enter dead_letter.

Observed affected strategies:

extract_atomic_facts
extract_foresight

Why this matters

Tool-use trajectories are common for coding agents and assistant workflows. A single tool-call item should not break user memory extraction for the whole memcell.

In the observed run, the agent case itself was extracted, but user-side extraction strategies failed independently because they iterated over mixed item types as if every item were a normal conversational message.

Observed evidence

OME records:

extract_atomic_facts: dead_letter
extract_foresight: dead_letter

Failure message:

AttributeError: 'ToolCallRequest' object has no attribute 'role'

Relevant code locations:

src/everos/memory/strategies/extract_atomic_facts.py:66
src/everos/memory/strategies/extract_foresight.py:54
src/everos/service/_boundary.py

The extraction strategies contain logic shaped like:

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

But _boundary.py can map assistant tool calls into ToolCallRequest, which does not expose role.

Reproduction shape

Use a synthetic agent/coding trajectory with assistant tool calls, for example:

1. User reports a failing focused test.
2. Assistant emits a tool call to run the test.
3. Tool returns trace output.
4. Assistant identifies the root cause.
5. Assistant emits another tool call to apply or validate a fix.
6. Tool returns passing output.
7. Assistant summarizes the result.

Then flush/process the session. The memcell contains mixed item types, including tool-call request objects.

Expected behavior

Tool-call items should not crash extraction strategies that only need user messages.

Possible fixes:

• Filter by item type before reading role.
• Use getattr(m, "role", None) == "user" where appropriate.
• Centralize sender/user-message extraction in a helper that understands mixed memcell item types.
• Ignore non-text/tool-call items in user-side strategies unless they are explicitly supported.

Environment

EverOS: 1.0.0 source checkout
Runtime: Docker Linux runtime
Python: 3.12
Data: synthetic agent trajectory only

No real user memory or secrets were used.