fix(session_search): scope recall by chat origin

[lkz-de]

Summary

• Add chat-origin provenance to session_search results so recall hits identify the conversation they came from.
• Default gateway-origin discovery and browse calls to the current chat/session scope, including DMs, while preserving explicit cross-conversation recall via scope="all".
• Carry the scope predicates through every search path, including stale-FTS LIKE fallback, and expose scoped-result provenance through same_origin, origin, and recall_scope counts.
• Pass gateway_session_key through background turns and update English and zh-Hans session-search docs.

Why

A fresh or compacted gateway session can receive an ambiguous follow-up after its local context has expired. Before this change, session_search searched the whole profile by default and returned matching sessions without enough visible room provenance. That made a hit from another conversation easy to treat as if it came from the active chat.

The session table already stores routing provenance such as source, chat_id, chat_type, display_name, and session_key. This PR carries that metadata through the search layer and uses it to make the safer default enforceable in SQL instead of relying only on prompt guidance.

When same-chat origin cannot be established, scope="chat" fails closed. Legacy rows without chat_id or session_key are excluded from scope="chat", counted in recall_scope, and remain reachable through explicit scope="all".

Testing

• python3 -m pytest tests/tools/test_session_search.py -q
  • 53 passed in 2.33s
• uv run --with pytest --with pytest-asyncio --with python-dotenv python -m pytest tests/tools/test_session_search.py tests/test_hermes_state.py tests/agent/test_prompt_builder.py tests/gateway/test_background_command.py -q --maxfail=1
  • 345 passed in 13.65s
• python3 -m py_compile tools/session_search_tool.py hermes_state.py hermes_state_search.py agent/tool_executor.py agent/agent_runtime_helpers.py gateway/run.py tests/tools/test_session_search.py tests/gateway/test_background_command.py
• git diff --check

Overlap check

This refresh keeps this PR as the implementation vehicle and folds in runtime pieces that were missing relative to current main. #59295 covers a narrower default-scoping approach and is superseded by this PR's broader search-path coverage, provenance labels, docs, and tests. The schema-backed scope-key design in #50030 remains the main open alternative; if maintainers prefer that direction, this PR's provenance labeling could be rebased onto it. Related prompt-only guidance (#30502) remains useful, but it does not replace code-level SQL scoping and visible provenance in session_search itself.

Notes

• Gateway contexts with a chat/session origin now default discovery and browse to scope="chat"; CLI and no-origin contexts continue to default to global recall.
• profile= recall deliberately defaults to global scope rather than applying the current gateway room to a different profile database.
• The model can still request scope="all"; the safety improvement is scoped defaults plus visible provenance, not a hard access-control boundary.
• Background-turn behavior is covered at the gateway_session_key handoff seam rather than by a full end-to-end gateway test.

[alt-glitch]

This was generated by AI during triage.

Related to #50030 and #59295. This patch uses chat-origin metadata and fail-closed shared-chat scoping; the open alternatives use a schema-backed scope-key design or a narrower tool-only filter. Please select the intended recall-scope policy.

[lkz-de]

On the scope-policy question: I opened both #59295 and this PR, and this PR is the successor. #59295 narrows only the default-discovery path with a tool-local post-search filter; this PR scopes recall by stored chat-origin metadata, fails closed in group-like contexts, and keeps explicit cross-conversation recall via scope="all" with same_origin labeling. I'd suggest evaluating this PR as the fail-closed candidate and treating #59295 as superseded — I'll close it once a direction is chosen. If the alternative design in #50030 is preferred instead, the provenance-labeling portion of this PR could be rebased onto that.

[lkz-de]

Refreshed this branch onto current main and resolved the tools/session_search_tool.py conflict by preserving both the current session-link/profile handling and this PR's chat-origin scoping behaviour.

Local verification:

• scripts/run_tests.sh tests/tools/test_session_search.py tests/test_hermes_state.py tests/agent/test_prompt_builder.py -- --tb=short -q

[teknium1]

Thanks for carrying origin metadata through the result shape and pushing the initial filter into the database query.

Problems

• tools/session_search_tool.py:272-276 defines same_origin as only source + chat_id. That crosses distinct forum topics and the default per-user group boundary: gateway/config.py:918 defaults group_sessions_per_user=True, and gateway/session.py:1053-1133 includes thread and participant identity in live routing.
• The same incomplete predicate is used for browse at tools/session_search_tool.py:520-521 and discovery at :816-830, so affected foreign sessions can be returned as local history.
• Current main moved the search implementation to hermes_state_search.py in 21c7ae8563; the required predicate must cover its FTS, CJK, trigram, LIKE, and unindexed-gap paths. website/docs/user-guide/sessions.md:549 also still says search spans all past conversations.

Suggested changes

• Scope against the persisted live routing identity (or an exactly equivalent source/chat/thread/user predicate) across every retrieval path, and add topic and per-user isolation regressions.
• Update the public session-search documentation with the scoped default and explicit global escape hatch.

Automated hermes-sweeper review.

[teknium1]

same_origin only compares chat_id and source. Gateway routing distinguishes thread_id and, for ordinary groups by default, the participant (gateway/session.py:1053-1133); this labels those foreign sessions as local and bypasses the intended fail-closed boundary.

[teknium1]

Filtering only on chat_id and source leaves same-chat forum topics and per-user group sessions in scope. Please use the persisted routing identity, or an equivalent predicate including thread and the applicable participant boundary, across every search backend.

[lkz-de]

Rebased this onto current main at 470cf66b0 and pushed head fffa5bc55aec55d6bbfbc440e3a22c1510a9ffd6.

The refresh keeps main's current list_sessions_rich source / sources / session_key behaviour and layers the chat-origin scope on top with an AND-composed chat_id filter. I also re-ported the original origin-scope regression tests, covering the cross-chat search/browse filters, scope="all" foreign-result labelling, read/scroll foreign-origin warnings and the registered handler path.

Verification run on the refreshed head:

uv run --with pytest --with python-dotenv python -m pytest tests/tools/test_session_search.py tests/test_hermes_state.py tests/agent/test_prompt_builder.py -q --maxfail=1
244 passed in 10.03s

[lkz-de]

Refreshed this branch onto current main and tightened the default chat-scoped predicate.

The chat-scoped search path now prefers the stored session_key when available, so per-user group sessions and thread-scoped sessions do not collapse merely because they share the same platform chat id. It falls back to chat_id only when the current session has no stored session key; rows without a session key are excluded from key-scoped recall and remain reachable via scope='all'. The default scope='chat' behaviour, the scope='all' escape hatch and the cross-context warnings are still covered.

Focused verification passed locally:

scripts/run_tests.sh tests/tools/test_session_search.py tests/test_hermes_state.py tests/agent/test_prompt_builder.py -- --tb=short -q
# 283 passed, 0 failed