fix Leo broad source intake truth (#124)

hermes-agent/leoclean-bin/kb_tool.py

@@ -504,7 +504,27 @@ def operational_contracts(
         None,
     )
     source_terms = ("document", "pdf", "tweet", "attachment", "ingest", "intake", "absorb", "extract")
-    if any(term in lowered for term in source_terms) and primary_direct_intent is None:
+    broad_source_object = any(term in lowered for term in ("report", "article", "file", "url", "link"))
+    broad_source_action = any(
+        term in lowered
+        for term in (
+            "send",
+            "hand",
+            "drop",
+            "upload",
+            "learn",
+            "absorb",
+            "ingest",
+            "extract",
+            "stage",
+            "add",
+            "make it part",
+        )
+    )
+    source_intake_question = any(term in lowered for term in source_terms) or (
+        broad_source_object and broad_source_action
+    )
+    if source_intake_question and primary_direct_intent is None:
         contracts.append(
             {
                 "id": "source_intake",
@@ -514,8 +534,9 @@ def operational_contracts(
                 ),
                 "capability_tier": "VPS-local prepared-input compiler and pending-review staging",
                 "safe_without_apply_authorization": [
-                    "retain artifact plus URL or storage_path plus content hash",
-                    "extract candidates into a pending_review proposal packet",
+                    "prepare and retain the artifact, UTF-8 extracted text, extraction manifest, URL or storage_path "
+                    "locator, and content hash",
+                    "run teleo-kb propose-source to validate and stage a pending_review proposal",
                 ],
                 "approval_begins": "before guarded canonical public.* apply, not before mechanical capture/staging",
                 "current_public_sources_columns": list(schema.get("sources", ())),
@@ -878,17 +899,20 @@ def compile_operational_response(contracts: list[dict[str, Any]]) -> str | None:
     source = by_id.get("source_intake")
     if source:
         safe_steps = [str(item) for item in source.get("safe_without_apply_authorization") or []]
-        retain_step = safe_steps[0] if safe_steps else "retain the artifact plus its locator and content hash"
-        stage_step = safe_steps[1] if len(safe_steps) > 1 else "stage candidates in a pending_review proposal"
+        retain_step = safe_steps[0] if safe_steps else "prepare the artifact, extracted text, and manifest"
+        stage_step = safe_steps[1] if len(safe_steps) > 1 else "stage a pending_review proposal"
         return (
-            f"- {source.get('required_lead')} Immediately {retain_step}; retain provenance for each PDF, Telegram "
-            "message, and tweet. A filename, chat label, attachment reference, or source_ref is not source identity.\n\n"
-            f"- {stage_step}. Deduplicate overlapping candidate claims, preserve evidence excerpts, source-specific "
-            "caveats, uncertainty, and contradictions. Candidates are not canonical until reviewed and applied: create "
-            "no public.sources or other public.* rows during capture or staging.\n\n"
+            f"- {source.get('required_lead')} Sending a report in chat does not by itself create a proposal or "
+            "persistent knowledge. I can analyze content available in this conversation, but chat attachment "
+            "capture and extraction are not connected to the staging command.\n\n"
+            f"- The live bounded path is to {retain_step}, then {stage_step}. It writes only "
+            "kb_stage.kb_proposals. A chat label, filename, attachment reference, or source_ref is not provenance. "
+            "Before review, deduplicate overlaps and retain exact excerpts, source caveats, uncertainty, and "
+            "contradictory interpretations. Candidates are not canonical until reviewed and applied, and no "
+            "public.sources or other public.* rows are created during staging.\n\n"
             f"- Explicit approval begins {source.get('approval_begins')}. Receipt: locator and content hash, proposal ID/"
             "status, reviewed payload, applied_at, created row IDs, and postflight readback. Next proof-changing "
-            "follow-up: run the retained artifact through the staging compiler, then review its typed candidates."
+            "follow-up: prepare the three inputs, run teleo-kb propose-source, then review its typed candidates."
         )
 
     named_packet = by_id.get("named_packet_readback")

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

@@ -83,8 +83,27 @@ def _requires_database_truth(query: str) -> bool:
     named_packet_question = "helmer" in lowered and any(
         term in lowered for term in ("7 powers", "seven powers", "powers framework", "powers packet")
     )
+    broad_source_intake_question = any(
+        term in lowered for term in ("report", "article", "file", "url", "link")
+    ) and any(
+        term in lowered
+        for term in (
+            "send",
+            "hand",
+            "drop",
+            "upload",
+            "learn",
+            "absorb",
+            "ingest",
+            "extract",
+            "stage",
+            "add",
+            "make it part",
+        )
+    )
     return bool(
         named_packet_question
+        or broad_source_intake_question
         or ("telegram" in lowered and "gatewayrunner" in lowered)
         or re.search(
             r"\b(?:databases?|db|knowledge bases?|kb|proposals?|canonical|source_ref|source rows?|"

tests/test_hermes_leoclean_db_context_plugin.py

@@ -170,6 +170,16 @@ def test_source_validator_accepts_explicit_schema_and_canonical_write_prohibitio
     assert module.response_contract_issues(response, contracts) == []
 
 
+def test_broad_report_learning_question_requires_database_truth() -> None:
+    module = load_plugin()
+    query = (
+        "If I send you a report here, can you learn it into your knowledge base and make it part of what you know? "
+        "What actually happens right now?"
+    )
+
+    assert module._requires_database_truth(query) is True
+
+
 def test_plugin_replaces_memory_only_status_answer_with_exact_live_db_receipt(tmp_path: Path, monkeypatch) -> None:
     module = load_plugin()
     home = tmp_path / "profile"

tests/test_hermes_leoclean_kb_bridge_source.py

@@ -220,6 +220,21 @@ def test_vps_bridge_operational_contracts_are_query_specific_and_schema_exact()
     assert source["current_public_sources_columns"] == list(module.CURRENT_PUBLIC_SCHEMA["sources"])
     assert "author/title/publisher/publication date are current public.sources columns" in source["must_not_claim"]
 
+    broad_source_query = (
+        "If I send you a report here, can you learn it into your knowledge base and make it part of what you know? "
+        "What actually happens right now?"
+    )
+    broad_source_contracts = {
+        item["id"]: item for item in module.operational_contracts(broad_source_query)
+    }
+    assert "source_intake" in broad_source_contracts
+    broad_source_response = module.compile_operational_response(list(broad_source_contracts.values()))
+    assert broad_source_response is not None
+    assert "Sending a report in chat does not by itself create a proposal or persistent knowledge" in broad_source_response
+    assert "chat attachment capture and extraction are not connected" in broad_source_response
+    assert "teleo-kb propose-source" in broad_source_response
+    assert "writes only kb_stage.kb_proposals" in broad_source_response
+
     mixed_contracts = {
         item["id"]: item
         for item in module.operational_contracts(