WebUI non-default profile sessions written to default state.db (DEFAULT_DB_PATH module-level constant)

[yeeyuny]

Bug Description

When using WebUI with a non-default profile (e.g. switching to zhangtingban via the profile dropdown), sessionsare incorrectly written to the default profile's state.db instead of the correct profile's state.db.

This means:

• In WebUI UI: sessions appear correctly under the switched profile (the _index.json metadata has the correctprofile field)
• In Hermes Agent TUI: sessions are missing from the expected profile and appear under the default profileinstead
• session_search in TUI for the target profile returns no results for WebUI-created sessions

Root Cause

In api/streaming.py line ~3556:

─ python
_session_db = SessionDB()

SessionDB() with no arguments uses the module-level constant DEFAULT_DB_PATH:

─ python

hermes_state.py line 34

DEFAULT_DB_PATH = get_hermes_home() / "state.db"

This constant is computed at import time and never changes, even though streaming.py correctly updatesos.environ['HERMES_HOME'] to the target profile home before the agent runs (lines 3201-3209). The module-levelconstant is already frozen to the default profile's path.

Fix

Pass db_path explicitly based on the already-resolved _profile_home:

─ python

streaming.py ~line 3556

_state_db_path = Path(_profile_home) / "state.db" if _profile_home else None
_session_db = SessionDB(db_path=_state_db_path)

This ensures SessionDB connects to the correct profile's state.db.

Reproduction

1. Start WebUI with multiple profiles configured
2. Switch to a non-default profile (e.g. zhangtingban)
3. Create a session and send messages
4. In TUI, switch to the same non-default profile
5. Run session_search — the WebUI session is not found
6. Switch TUI to default profile — the WebUI session appears there instead

Environment

• Hermes Agent: v0.14.0
• WebUI: v0.51.92
• OS: WSL2 (Ubuntu)
• Profiles: default, zhangtingban, us-stock

[nesquena-hermes]

Summary

The diagnosis here is correct. DEFAULT_DB_PATH in the agent is computed at module-import time:

# hermes-agent/hermes_state.py:34
DEFAULT_DB_PATH = get_hermes_home() / "state.db"

get_hermes_home() reads os.environ['HERMES_HOME'], so whichever profile the WebUI server was launched under wins, and the constant is then frozen for the rest of the process. SessionDB.__init__ at hermes_state.py:335 reads db_path or DEFAULT_DB_PATH, so calling SessionDB() with no arguments always lands on that frozen default — even when the per-request code at api/streaming.py:3696-3704 has already swapped os.environ['HERMES_HOME'] to the active session's profile home.

Code reference

In api/streaming.py:4140-4146, well after the env mutation has run:

_session_db = None
try:
    from hermes_state import SessionDB
    _session_db = SessionDB()
except Exception as _db_err:
    print(f"[webui] WARNING: SessionDB init failed ...", flush=True)

This _session_db then gets attached to the agent at api/streaming.py:4448 (agent._session_db = _session_db) and is what session_search (hermes-agent/tools/session_search_tool.py:378-405) reads through. So WebUI-only sessions created under a non-default profile end up indexed in the default profile's state.db, and CLI/TUI session_search against the named profile cannot see them. The reporter's reproduction matches this.

Notably this is a separate, parallel bug to #2762 / PR #2827, which fixed the state_sync writer path. api/state_sync.py:_get_state_db(profile=...) at lines 49-79 already does the right thing for the sync_session_usage path — it resolves the profile-specific home and refuses to fall back. The streaming attach at api/streaming.py:4144 was missed in that pass because it constructs the DB once at agent-build time rather than per-write.

Recommended fix

The reporter's proposed patch is right, with one wrinkle: by the time line 4144 runs, _profile_home has already been resolved at api/streaming.py:3649-3651:

_profile_home_path = get_hermes_home_for_profile(getattr(s, 'profile', None))
_profile_home = str(_profile_home_path)

So the fix can read that variable directly rather than reconstructing the path:

_state_db_path = (Path(_profile_home) / "state.db") if _profile_home else None
_session_db = SessionDB(db_path=_state_db_path)

There's also a session-cache angle: the agent gets reused via SESSION_AGENT_CACHE (api/streaming.py:4439-4448), and _session_db is replaced on reuse. As long as the cached agent's _agent_sig already encodes the profile (worth double-checking — but agents are keyed by session_id and sessions are profile-stamped, so this should already be safe), the new db_path will flow through correctly on subsequent turns.

Test plan

The existing tests/test_issue2762_state_sync_profile_kwarg.py covers the state-sync writer side. A parallel test for the streaming attach would assert:

1. _run_agent_streaming for a session with profile='zhangtingban' constructs SessionDB(db_path=<zhangtingban-home>/state.db), not the default.
2. agent._session_db.db_path matches the session's profile home after attach.

Both can be exercised by monkeypatching hermes_state.SessionDB.__init__ to capture the db_path arg, similar to how test_issue2762_state_sync_profile_kwarg.py:1-50 patches the resolver. Labeling as bug + a back-reference to #2762 would also make the link explicit, since the fix lives in the same family of corrections.