Make broad Leo replies deterministic (#121)

hermes-agent/leoclean-bin/kb_tool.py

@@ -456,6 +456,18 @@ def operational_contracts(
     identity_question = any(term in lowered for term in ("soul.md", "soul file")) and any(
         term in lowered for term in ("canonical identity", "canonical", "identity", "source of truth", "database")
     )
+    visible_sender_match = re.search(
+        r"(?:current visible telegram sender|visible telegram sender|current telegram sender)\s+is\s+@?([a-z0-9_]+)",
+        query,
+        re.I,
+    )
+    participant_identity_question = bool(
+        visible_sender_match
+        and any(
+            term in lowered
+            for term in ("call this participant", "identity sources", "mixing identities", "mix identities")
+        )
+    )
     primary_direct_intent = next(
         (
             intent
@@ -585,6 +597,27 @@ def operational_contracts(
             }
         )
 
+    if participant_identity_question and visible_sender_match:
+        contracts.append(
+            {
+                "id": "telegram_participant_identity",
+                "current_visible_handle": visible_sender_match.group(1),
+                "allowed_sources": [
+                    "the current message sender handle",
+                    "that participant's explicit in-chat correction",
+                ],
+                "disallowed_sources": [
+                    "stale session context",
+                    "environment labels",
+                    "inferred personal names",
+                    "another participant's profile or messages",
+                ],
+                "cross_participant_boundary": (
+                    "re-anchor every reply to that message's visible sender; never transfer or mix identities"
+                ),
+            }
+        )
+
     mixed_markers = (
         "packet",
         "factual observation",
@@ -782,6 +815,19 @@ def compile_operational_response(contracts: list[dict[str, Any]]) -> str | None:
     """Compile a safe response when a model draft violates live DB contracts."""
 
     by_id = {str(contract.get("id") or ""): contract for contract in contracts}
+    participant = by_id.get("telegram_participant_identity")
+    if participant:
+        handle = str(participant.get("current_visible_handle") or "").lstrip("@")
+        return (
+            f"Call the current Telegram sender {handle}, exactly as their visible handle identifies them.\n"
+            "Allowed identity sources: the current message sender handle and that participant's explicit in-chat "
+            "correction.\n"
+            "Do not use stale session context, environment labels, inferred personal names, or another "
+            "participant's profile.\n"
+            "When another user replies, never transfer or mix identities; re-anchor to that message's visible "
+            "sender handle."
+        )
+
     mixed = by_id.get("mixed_packet_composition")
     if mixed:
         mapping = mixed.get("map") or {}
@@ -953,13 +999,14 @@ def compile_operational_response(contracts: list[dict[str, Any]]) -> str | None:
             return (
                 f"Not proven.\nFresh live readback.\n{format_database_count_readback(status_data)}\nProposal states are "
                 f"applied: {states.get('applied', 0)}, approved: {states.get('approved', 0)}, pending_review: "
-                f"{states.get('pending_review', 0)}. This current snapshot does not establish what changed since "
-                f"yesterday. Existing receipts identify applied types {applied_types or 'none'}, but not a time-window "
-                f"delta. Approved is not applied; {approved_needs_payload}/{len(approved_rows)} approved rows are "
+                f"{states.get('pending_review', 0)}. This current snapshot does not prove what changed since "
+                f"yesterday, so the demo claim is limited to current readback. Existing receipt artifacts identify "
+                f"applied types {applied_types or 'none'}, but not a time-window delta. Approved is not applied; "
+                f"{approved_needs_payload}/{len(approved_rows)} approved rows are "
                 "approved_needs_apply_payload and not directly applyable.\n\n"
                 "Next proof-changing follow-up: build and review a strict apply_payload for the highest-impact approved "
-                "proposal, then seek explicit apply authorization. Until its applied_at and public.* before/after row "
-                "IDs or hashes exist, do not claim the canonical database changed."
+                "proposal, then seek explicit apply authorization. Until its applied_at, public.* before/after row "
+                "IDs or hashes, and postflight readback exist, do not claim the canonical database changed."
             )
         return (
             f"Partly.\nFresh live readback.\n{format_database_count_readback(status_data)}\nProposal states are applied: "

hermes-agent/leoclean-plugins/vps/leo-db-context/__init__.py

@@ -28,6 +28,7 @@ COMPILED_CONTRACT_IDS = frozenset(
         "source_link_audit",
         "demo_capability_readback",
         "identity_canonicality_readback",
+        "telegram_participant_identity",
         "runtime_persistence",
         "shared_claims_agent_positions",
         "forecast_resolution_schema",

tests/test_hermes_leoclean_kb_bridge_source.py

@@ -251,6 +251,10 @@ def test_vps_bridge_broad_db_questions_select_query_specific_live_contracts() ->
         "Is the proposal backlog stuck on missing document-to-source links?": "source_link_audit",
         "Can we demo a real database write from Leo?": "demo_capability_readback",
         "Does editing the SOUL file alter the source of truth for identity?": "identity_canonicality_readback",
+        (
+            "The current visible Telegram sender is @m3taversal. What should Leo call this participant, which "
+            "identity sources are allowed, and how should Leo avoid mixing identities?"
+        ): "telegram_participant_identity",
     }
 
     for query, expected_id in cases.items():
@@ -316,6 +320,11 @@ def test_vps_bridge_oos_intents_preserve_specific_contracts_and_negated_actions(
         (
             "A temporary-profile GatewayRunner answered, but posted nothing to Telegram. Is delivery proven live?"
         ): {"telegram_delivery_proof"},
+        (
+            "The current visible Telegram sender is @m3taversal. An earlier answer shortened that handle. What "
+            "should Leo call this participant, which identity sources are allowed, and how should Leo avoid mixing "
+            "identities when another user replies? Do not write to memory or the KB."
+        ): {"telegram_participant_identity"},
     }
 
     for query, expected_ids in cases.items():

tests/test_working_leo_m3taversal_oos_benchmark.py

@@ -573,6 +573,7 @@ def test_db_compiled_responses_pass_all_database_backed_oos_cases() -> None:
         "OOS-04",
         "OOS-05",
         "OOS-06",
+        "OOS-09",
         "OOS-10",
         "OOS-11",
         "OOS-12",