[Bug] flush crashes with TypeError: DetectionResult.__new__() missing 1 required positional argument on fresh pip 1.2.3 install — recall always empty

[Yunkuboom]

Summary

POST /api/v1/memory/flush (and the memorize extract path in general) crashes with TypeError: DetectionResult.__new__() missing 1 required positional argument on a fresh pip-installed EverOS 1.2.3 deployment, so buffered messages never become episodes and recall always returns empty.

Environment

• EverOS 1.2.3 via pip install everos==1.2.3 (Docker, python:3.12-slim)
• Installed dependency versions (from pip list):
  • everalgo-boundary 0.3.0
  • everalgo-core 0.5.0
  • everalgo-agent-memory 0.4.0
  • everalgo-user-memory 0.4.0
• Config: [llm] model=MiniMax-M3, base_url=https://api.minimaxi.com/v1; embedding/rerank/multimodal not configured (keyword-only mode)
• Server: everos server start --host 0.0.0.0 --port 8000

Reproduction

1. Start server with a bare [llm] config (no embedding).
2. Add a message:

curl -s -X POST localhost:8000/api/v1/memory/add -H 'Content-Type: application/json' \
  -d '{"session_id":"e2e-probe","app_id":"hermes","project_id":"default",
       "messages":[{"sender_id":"hermes-default","sender_name":"大拿","role":"assistant",
                    "timestamp":1787298814000,"content":"云哥的 NAS 是 HPStudioBJ，IP 192.168.3.144，Tailscale 100.81.224.74。"}]}'

Returns 200 {"data":{"message_count":1,"status":"accumulated"}} — so far so good.

3. Flush to trigger extraction:

curl -s -X POST localhost:8000/api/v1/memory/flush -H 'Content-Type: application/json' \
  -d '{"session_id":"e2e-probe","app_id":"hermes","project_id":"default"}'

Returns 500 {"error":{"code":"INTERNAL_ERROR",...}}. Server log:

FileNotFoundError? no —
TypeError: DetectionResult.__new__() missing 1 required positional argument:
│ ze.py:195 in flush_memory

(Full traceback below.)

4. Search for the content:

curl -s -X POST localhost:8000/api/v1/memory/search -H 'Content-Type: application/json' \
  -d '{"agent_id":"hermes-default","app_id":"hermes","query":"HPStudioBJ","method":"keyword","top_k":10}'

Returns {"episodes":[],"profiles":[],"agent_cases":[],"agent_skills":[],"unprocessed_messages":[]} — nothing recallable, even though the message IS durable in unprocessed_buffer (SQLite).

Expected behavior

• Flush should not 500 on the default pip install; buffered messages should become memcells → episodes/atomic_facts so search can recall them.
• If keyword-only mode is unsupported for extraction, the server should say so at startup or in /health, not crash on flush.

Suspected cause

DetectionResult is a NamedTuple(cells, tail) (per everalgo-boundary README). Something in the boundary-detection call path (everos.service.memorize.flush_memory → everalgo) constructs it with the wrong arity or the installed everalgo-boundary 0.3.0 / everalgo-core 0.5.0 pair disagree on its shape. This is a fresh pip install — no upgrade path, no legacy state — so it reproduces out of the box.

Extra context

This makes the OSS server's core memorize→search loop unusable for a brand-new user (write works, recall never does). It also affects the proposed Hermes integration (PR #329) which relies on on_session_end → /memory/flush. Happy to provide full stacktrace logs or run any diagnostic commands if helpful.

Full server traceback (truncated):

TypeError: DetectionResult.__new__() missing 1 required positional argument:
│ /usr/local/lib/python3.12/site-packages/.../ze.py:195 in flush_memory