Keep Leo follow-up retrieval on the active subject (#184)

hermes-agent/leoclean-bin/kb_tool.py

@@ -253,6 +253,16 @@ def parse_args() -> argparse.Namespace:
 
     context = sub.add_parser("context", help="Build an agent-ready KB context bundle for a question.")
     context.add_argument("query", nargs="+")
+    context.add_argument(
+        "--retrieval-query",
+        help="Optional read-only semantic query; operational contracts remain bound to the exact operator query.",
+    )
+    context.add_argument(
+        "--retrieval-anchor",
+        action="append",
+        default=[],
+        help="Bounded topic anchor used only to rank read-only claim retrieval; may be repeated.",
+    )
     context.add_argument("--limit", type=int, default=4)
     context.add_argument("--context-limit", type=int, default=6)
     add_output_flag(context)
@@ -1951,13 +1961,24 @@ def markdown_claim_link(claim_id: str | None, label: str | None = None) -> str:
     return f"[`{text}`]({claim_url(claim_id)})"
 
 
-def find_claims(args: argparse.Namespace, query: str, limit: int) -> list[dict[str, Any]]:
+def find_claims(
+    args: argparse.Namespace,
+    query: str,
+    limit: int,
+    retrieval_anchors: list[str] | None = None,
+) -> list[dict[str, Any]]:
     patterns = [f"%{term}%" for term in query_terms(query)]
+    anchor_terms = query_terms(" ".join(retrieval_anchors or [])) if retrieval_anchors else []
+    anchor_patterns = [f"%{term}%" for term in anchor_terms]
     min_score = 2 if len(patterns) >= 3 else 1
+    anchor_limit = max(1, min(2, limit))
     sql = f"""
     with params(patterns) as (
       values ({sql_array(patterns)})
     ),
+    anchors(patterns) as (
+      values ({sql_array(anchor_patterns)})
+    ),
     ranked as (
       select c.id,
              c.type,
@@ -1974,6 +1995,15 @@ def find_claims(args: argparse.Namespace, query: str, limit: int) -> list[dict[s
                          where lower(tag) like pattern
                       )
              ) as score,
+             (
+               select count(*)
+                 from anchors a, unnest(a.patterns) pattern
+                where lower(c.text) like pattern
+                   or exists (
+                        select 1 from unnest(coalesce(c.tags, '{{}}'::text[])) tag
+                         where lower(tag) like pattern
+                      )
+             ) as anchor_score,
              (select count(*) from claim_evidence ce where ce.claim_id = c.id) as evidence_count,
              (select count(*) from claim_edges e where e.from_claim = c.id or e.to_claim = c.id) as edge_count
         from claims c
@@ -1983,6 +2013,29 @@ def find_claims(args: argparse.Namespace, query: str, limit: int) -> list[dict[s
       select ranked.*,
              max(score) over () as max_score
         from ranked
+    ),
+    raw_selected as (
+      select eligible.*
+        from eligible
+       where score >= case when max_score >= {min_score} then {min_score} else 1 end
+       order by score desc, evidence_count desc, edge_count desc,
+                coalesce(confidence, 0) desc, length(text), text, id
+       limit {limit}
+    ),
+    anchor_selected as (
+      select eligible.*
+        from eligible
+       where anchor_score >= 1
+       order by anchor_score desc, score desc, evidence_count desc, edge_count desc,
+                coalesce(confidence, 0) desc, length(text), text, id
+       limit {anchor_limit}
+    ),
+    lanes as (
+      select raw_selected.*, 'raw'::text as selection_lane
+        from raw_selected
+      union all
+      select anchor_selected.*, 'anchor'::text as selection_lane
+        from anchor_selected
     )
     select jsonb_build_object(
              'id', id::text,
@@ -1991,16 +2044,31 @@ def find_claims(args: argparse.Namespace, query: str, limit: int) -> list[dict[s
              'confidence', confidence,
              'tags', coalesce(to_jsonb(tags), '[]'::jsonb),
              'score', score,
+             'anchor_score', anchor_score,
              'evidence_count', evidence_count,
-             'edge_count', edge_count
+             'edge_count', edge_count,
+             '_selection_lane', selection_lane
            )::text
-      from eligible
-     where score >= 1
-       and score >= case when max_score >= {min_score} then {min_score} else 1 end
-     order by score desc, evidence_count desc, edge_count desc, coalesce(confidence, 0) desc, length(text), text, id
-     limit {limit};
+      from lanes
+     order by case selection_lane when 'raw' then 0 else 1 end,
+              score desc, anchor_score desc, evidence_count desc, edge_count desc,
+              coalesce(confidence, 0) desc, length(text), text, id;
     """
-    return psql_json(args, sql)
+    rows = psql_json(args, sql)
+    raw_rows: list[dict[str, Any]] = []
+    anchor_rows: list[dict[str, Any]] = []
+    for row in rows:
+        lane = row.pop("_selection_lane", None)
+        if lane == "raw":
+            raw_rows.append(row)
+        elif lane == "anchor":
+            anchor_rows.append(row)
+    if not anchor_patterns or any(int(row.get("anchor_score") or 0) > 0 for row in raw_rows):
+        return raw_rows[:limit]
+
+    raw_ids = {str(row.get("id")) for row in raw_rows}
+    backfill = [row for row in anchor_rows if str(row.get("id")) not in raw_ids][:anchor_limit]
+    return backfill + raw_rows[: max(0, limit - len(backfill))]
 
 
 def find_context_rows(args: argparse.Namespace, query: str, limit: int) -> list[dict[str, Any]]:
@@ -3339,6 +3407,8 @@ def build_retrieval_receipt(
     return {
         "schema": RETRIEVAL_RECEIPT_SCHEMA,
         "query_sha256": hashlib.sha256(str(data.get("query") or "").encode("utf-8")).hexdigest(),
+        "retrieval_query_sha256": data.get("retrieval_query_sha256"),
+        "retrieval_anchors_sha256": data.get("retrieval_anchors_sha256"),
         "semantic_context_sha256": canonical_json_sha256(semantic_context),
         "artifact_state_sha256": canonical_json_sha256(artifact_states),
         "claim_ids": [str(claim.get("id")) for claim in data.get("claims", []) if claim.get("id")],
@@ -3362,16 +3432,26 @@ def build_retrieval_receipt(
     }
 
 
-def _bundle_once(args: argparse.Namespace, query: str, claim_limit: int, context_limit: int) -> dict[str, Any]:
+def _bundle_once(
+    args: argparse.Namespace,
+    query: str,
+    claim_limit: int,
+    context_limit: int,
+    retrieval_query: str | None = None,
+    retrieval_anchors: list[str] | None = None,
+) -> dict[str, Any]:
+    semantic_query = retrieval_query or query
     contracts = context_operational_contracts(args, query)
-    claims = find_claims(args, query, claim_limit)
+    claims = find_claims(args, semantic_query, claim_limit, retrieval_anchors)
     claim_ids = [claim["id"] for claim in claims]
     evidence = load_evidence(args, claim_ids, 4)
     edges = load_edges(args, claim_ids, 6)
     bind_claim_evidence_challenge(contracts, claims, evidence)
     return {
         "query": query,
-        "terms": query_terms(query),
+        "retrieval_query_sha256": hashlib.sha256(semantic_query.encode("utf-8")).hexdigest(),
+        "retrieval_anchors_sha256": canonical_json_sha256(retrieval_anchors or []),
+        "terms": query_terms(semantic_query),
         "operational_contracts": contracts,
         "compiled_response": compile_operational_response(contracts),
         "claims": [
@@ -3383,11 +3463,18 @@ def _bundle_once(args: argparse.Namespace, query: str, claim_limit: int, context
             }
             for claim in claims
         ],
-        "context_rows": find_context_rows(args, query, context_limit),
+        "context_rows": find_context_rows(args, semantic_query, context_limit),
     }
 
 
-def bundle(args: argparse.Namespace, query: str, claim_limit: int, context_limit: int) -> dict[str, Any]:
+def bundle(
+    args: argparse.Namespace,
+    query: str,
+    claim_limit: int,
+    context_limit: int,
+    retrieval_query: str | None = None,
+    retrieval_anchors: list[str] | None = None,
+) -> dict[str, Any]:
     try:
         max_attempts = max(1, min(int(os.environ.get("TELEO_KB_READ_ATTEMPTS", "3")), 5))
     except ValueError as exc:
@@ -3396,7 +3483,7 @@ def bundle(args: argparse.Namespace, query: str, claim_limit: int, context_limit
     last_markers: tuple[dict[str, Any], dict[str, Any]] | None = None
     for attempt in range(1, max_attempts + 1):
         before = database_read_marker(args)
-        data = _bundle_once(args, query, claim_limit, context_limit)
+        data = _bundle_once(args, query, claim_limit, context_limit, retrieval_query, retrieval_anchors)
         after = database_read_marker(args)
         last_markers = (before, after)
         consistency_status = "stable_wal_marker" if _read_marker_stable(before, after) else "wal_changed_during_read"
@@ -3607,7 +3694,14 @@ def main() -> int:
     if args.command == "status":
         data = status(args)
     elif args.command in {"search", "context"}:
-        data = bundle(args, args.query, args.limit, args.context_limit)
+        data = bundle(
+            args,
+            args.query,
+            args.limit,
+            args.context_limit,
+            getattr(args, "retrieval_query", None),
+            getattr(args, "retrieval_anchor", None),
+        )
     elif args.command == "show":
         claim = get_claim(args, args.claim_id)
         if not claim:

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

@@ -9,7 +9,7 @@ import re
 import subprocess
 import sys
 import threading
-from collections import OrderedDict
+from collections import Counter, OrderedDict
 from collections.abc import Callable
 from datetime import datetime, timezone
 from pathlib import Path
@@ -23,6 +23,10 @@ CONTEXT_ROW_LIMIT = 4
 MAX_CLAIM_TEXT_CHARS = 1_000
 MAX_CONTEXT_BODY_CHARS = 800
 MAX_EVIDENCE_EXCERPT_CHARS = 600
+MAX_RETRIEVAL_ANCHORS = 8
+MAX_RETRIEVAL_SUBJECT_ANCHORS = 2
+MAX_RETRIEVAL_FALLBACK_ANCHORS = 3
+RETRIEVAL_HISTORY_USER_TURNS = 6
 DATABASE_REASONING_PROTOCOL_VERSION = "leo-db-reasoning-v2"
 WORD_RE = re.compile(r"\b\w+(?:[-']\w+)*\b")
 LIST_PREFIX_RE = re.compile(r"^(\s*(?:[-*]|\d+[.)])\s+)(.*)$")
@@ -31,6 +35,120 @@ REQUESTED_SUBJECT_RE = re.compile(
     r"name the subject exactly once as\s+`(?P<subject>[^`\r\n]{1,120})`\s+in your answer",
     re.I,
 )
+RETRIEVAL_TOKEN_RE = re.compile(r"[A-Za-z][A-Za-z0-9.+#'-]{2,}")
+RETRIEVAL_ANCHOR_STOPWORDS = frozenset(
+    {
+        "about",
+        "across",
+        "after",
+        "all",
+        "also",
+        "and",
+        "answer",
+        "actually",
+        "any",
+        "anything",
+        "are",
+        "argue",
+        "around",
+        "before",
+        "being",
+        "been",
+        "better",
+        "both",
+        "but",
+        "can",
+        "could",
+        "did",
+        "does",
+        "each",
+        "for",
+        "from",
+        "had",
+        "has",
+        "have",
+        "how",
+        "into",
+        "less",
+        "more",
+        "not",
+        "now",
+        "only",
+        "our",
+        "over",
+        "same",
+        "say",
+        "should",
+        "still",
+        "than",
+        "that",
+        "the",
+        "their",
+        "them",
+        "then",
+        "they",
+        "this",
+        "through",
+        "use",
+        "was",
+        "were",
+        "what",
+        "when",
+        "where",
+        "which",
+        "who",
+        "why",
+        "will",
+        "with",
+        "without",
+        "would",
+        "you",
+        "your",
+        "asserting",
+        "based",
+        "body",
+        "candidate",
+        "canonical",
+        "challenge",
+        "claim",
+        "concrete",
+        "confidence",
+        "context",
+        "database",
+        "direct",
+        "distinguish",
+        "discussion",
+        "evidence",
+        "exact",
+        "explain",
+        "explicit",
+        "existing",
+        "falsifier",
+        "give",
+        "identify",
+        "initial",
+        "inference",
+        "minimum",
+        "missing",
+        "most",
+        "nothing",
+        "opposite",
+        "outcome",
+        "requirements",
+        "return",
+        "revised",
+        "reviewable",
+        "source",
+        "stage",
+        "state",
+        "strongest",
+        "support",
+        "test",
+        "tool",
+        "while",
+        "write",
+    }
+)
 COMPILED_CONTRACT_IDS = frozenset(
     {
         "mixed_packet_composition",
@@ -96,6 +214,8 @@ def _retrieval_receipt_trace(
     safe = {
         "schema": value.get("schema"),
         "query_sha256": value.get("query_sha256"),
+        "retrieval_query_sha256": value.get("retrieval_query_sha256"),
+        "retrieval_anchors_sha256": value.get("retrieval_anchors_sha256"),
         "semantic_context_sha256": value.get("semantic_context_sha256"),
         "artifact_state_sha256": value.get("artifact_state_sha256"),
         "claim_ids": sorted(str(item) for item in value.get("claim_ids") or []),
@@ -134,6 +254,100 @@ def _failure_context(reason: str) -> str:
     )
 
 
+def _history_message_text(message: Any) -> str:
+    if not isinstance(message, dict) or message.get("role") != "user":
+        return ""
+    content = message.get("content")
+    if isinstance(content, str):
+        return content
+    if not isinstance(content, list):
+        return ""
+    parts: list[str] = []
+    for item in content:
+        if isinstance(item, str):
+            parts.append(item)
+        elif isinstance(item, dict) and isinstance(item.get("text"), str):
+            parts.append(str(item["text"]))
+    return " ".join(parts)
+
+
+def _is_subject_token(token: str) -> bool:
+    return token.isupper() or any(char.isupper() for char in token[1:])
+
+
+def _contextual_retrieval_query(user_message: str, conversation_history: Any) -> tuple[str, list[str]]:
+    """Return the exact query plus bounded user-only topic anchors for separate ranking."""
+
+    query = str(user_message or "")[:MAX_QUERY_CHARS]
+    if not isinstance(conversation_history, list):
+        return query, []
+
+    prior_user_turns: list[str] = []
+    skipped_current = False
+    for message in reversed(conversation_history):
+        text = _history_message_text(message).strip()
+        if not text:
+            continue
+        if not skipped_current and text == query.strip():
+            skipped_current = True
+            continue
+        prior_user_turns.append(text)
+        if len(prior_user_turns) >= RETRIEVAL_HISTORY_USER_TURNS:
+            break
+
+    counts: Counter[str] = Counter()
+    recency: dict[str, int] = {}
+    surfaces: dict[str, str] = {}
+    subject_tokens: set[str] = set()
+    for turn_index, text in enumerate(prior_user_turns):
+        seen_in_turn: set[str] = set()
+        for token in RETRIEVAL_TOKEN_RE.findall(text):
+            surface = token.strip("-'.+")
+            if surface.casefold().endswith("'s"):
+                surface = surface[:-2]
+            normalized = surface.casefold()
+            if len(normalized) < 3 or normalized in RETRIEVAL_ANCHOR_STOPWORDS or normalized in seen_in_turn:
+                continue
+            seen_in_turn.add(normalized)
+            counts[normalized] += 1
+            recency.setdefault(normalized, RETRIEVAL_HISTORY_USER_TURNS - turn_index)
+            surfaces.setdefault(normalized, surface)
+            if _is_subject_token(surface):
+                subject_tokens.add(normalized)
+
+    current_terms: set[str] = set()
+    for token in RETRIEVAL_TOKEN_RE.findall(query):
+        surface = token.strip("-'.+")
+        if surface.casefold().endswith("'s"):
+            surface = surface[:-2]
+        normalized = surface.casefold()
+        if len(normalized) < 3 or normalized in RETRIEVAL_ANCHOR_STOPWORDS or normalized in current_terms:
+            continue
+        current_terms.add(normalized)
+        counts[normalized] += 1
+        recency[normalized] = RETRIEVAL_HISTORY_USER_TURNS + 1
+        surfaces[normalized] = surface
+        if _is_subject_token(surface):
+            subject_tokens.add(normalized)
+
+    subject_anchors = sorted(
+        subject_tokens,
+        key=lambda term: (
+            -counts[term],
+            -recency[term],
+            term,
+        ),
+    )
+    if subject_anchors:
+        selected = subject_anchors[:MAX_RETRIEVAL_SUBJECT_ANCHORS]
+    else:
+        repeated = [term for term, count in counts.items() if count >= 2]
+        repeated.sort(key=lambda term: (-counts[term], -recency[term], term))
+        selected = repeated[:MAX_RETRIEVAL_FALLBACK_ANCHORS]
+    anchors = [surfaces[term] for term in selected[:MAX_RETRIEVAL_ANCHORS]]
+    return query, anchors
+
+
 def _requires_database_truth(query: str) -> bool:
     lowered = query.lower()
     named_packet_question = "helmer" in lowered and any(
@@ -345,12 +559,14 @@ def _take_snapshot(session_id: str, query_sha256: str) -> dict[str, Any] | None:
 def _load_database_snapshot(
     user_message: str,
     *,
+    conversation_history: Any = None,
     hermes_home: Path | None = None,
     runner: Callable[..., subprocess.CompletedProcess[str]] = subprocess.run,
 ) -> dict[str, Any]:
     """Load one immutable DB-contract snapshot for generation and validation."""
 
     query = str(user_message or "")[:MAX_QUERY_CHARS]
+    retrieval_query, retrieval_anchors = _contextual_retrieval_query(query, conversation_history)
     home = hermes_home or Path(os.getenv("HERMES_HOME", str(Path.home() / ".hermes")))
     kb_tool = home / "bin" / "kb_tool.py"
     query_sha256 = hashlib.sha256(query.encode("utf-8")).hexdigest()
@@ -358,6 +574,11 @@ def _load_database_snapshot(
         "generated_at_utc": datetime.now(timezone.utc).isoformat(),
         "event": "pre_llm_call",
         "query_sha256": query_sha256,
+        "retrieval_query_sha256": hashlib.sha256(retrieval_query.encode("utf-8")).hexdigest(),
+        "retrieval_anchor_count": len(retrieval_anchors),
+        "retrieval_anchors_sha256": hashlib.sha256(
+            json.dumps(retrieval_anchors, sort_keys=True, separators=(",", ":")).encode("utf-8")
+        ).hexdigest(),
         "source": "kb_tool.py --local context",
     }
 
@@ -373,13 +594,21 @@ def _load_database_snapshot(
         "--local",
         "context",
         query,
-        "--limit",
-        str(CONTEXT_CLAIM_LIMIT),
-        "--context-limit",
-        str(CONTEXT_ROW_LIMIT),
-        "--format",
-        "json",
+        "--retrieval-query",
+        retrieval_query,
     ]
+    for anchor in retrieval_anchors:
+        command.extend(("--retrieval-anchor", anchor))
+    command.extend(
+        (
+            "--limit",
+            str(CONTEXT_CLAIM_LIMIT),
+            "--context-limit",
+            str(CONTEXT_ROW_LIMIT),
+            "--format",
+            "json",
+        )
+    )
     try:
         completed = runner(command, capture_output=True, text=True, timeout=timeout, check=False)
     except subprocess.TimeoutExpired:
@@ -454,12 +683,20 @@ def _load_database_snapshot(
 def build_database_context(
     user_message: str,
     *,
+    conversation_history: Any = None,
     hermes_home: Path | None = None,
     runner: Callable[..., subprocess.CompletedProcess[str]] = subprocess.run,
 ) -> str:
     """Return current operational contracts for one exact operator question."""
 
-    return str(_load_database_snapshot(user_message, hermes_home=hermes_home, runner=runner)["context"])
+    return str(
+        _load_database_snapshot(
+            user_message,
+            conversation_history=conversation_history,
+            hermes_home=hermes_home,
+            runner=runner,
+        )["context"]
+    )
 
 
 def _word_count(value: str) -> int:
@@ -1082,7 +1319,7 @@ def _requires_compiled_fallback(issues: list[str]) -> bool:
 
 def _pre_llm_call(**kwargs: Any) -> dict[str, str]:
     user_message = str(kwargs.get("user_message") or "")
-    snapshot = _load_database_snapshot(user_message)
+    snapshot = _load_database_snapshot(user_message, conversation_history=kwargs.get("conversation_history"))
     _store_snapshot(str(kwargs.get("session_id") or ""), snapshot)
     return {"context": str(snapshot["context"])}
 

tests/test_hermes_leoclean_db_context_plugin.py

@@ -336,6 +336,8 @@ def test_plugin_injects_bounded_retrieval_rows_and_writes_body_redacted_trace(tm
         "retrieval_receipt": {
             "schema": "livingip.teleoKbRetrievalReceipt.v1",
             "query_sha256": "1" * 64,
+            "retrieval_query_sha256": "4" * 64,
+            "retrieval_anchors_sha256": "5" * 64,
             "semantic_context_sha256": "2" * 64,
             "artifact_state_sha256": "3" * 64,
             "claim_ids": ["claim-1"],
@@ -402,12 +404,15 @@ def test_plugin_injects_bounded_retrieval_rows_and_writes_body_redacted_trace(tm
     assert record["contract_ids"] == ["reply_budget", "source_intake"]
     assert record["database_reasoning_protocol_version"] == "leo-db-reasoning-v2"
     assert record["retrieval_receipt"]["semantic_context_sha256"] == "2" * 64
+    assert record["retrieval_receipt"]["retrieval_query_sha256"] == "4" * 64
+    assert record["retrieval_receipt"]["retrieval_anchors_sha256"] == "5" * 64
     assert record["retrieval_receipt"]["read_consistency"]["system_identifier"] == "system-1"
     assert record["retrieval_receipt"]["counts"] == {"claims": 1, "context_rows": 1, "evidence_rows": 1}
     assert record["retrieval_receipt"]["injected_rows_sha256"] == injected_rows_sha256
     assert len(record["retrieval_receipt"]["receipt_sha256"]) == 64
     command, kwargs = calls[0]
     assert command[2:5] == ["--local", "context", "operator secret-shaped but nonsecret message"]
+    assert command[command.index("--retrieval-query") + 1] == "operator secret-shaped but nonsecret message"
     claim_limit = int(command[command.index("--limit") + 1])
     context_limit = int(command[command.index("--context-limit") + 1])
     assert 0 < claim_limit <= 8
@@ -417,6 +422,85 @@ def test_plugin_injects_bounded_retrieval_rows_and_writes_body_redacted_trace(tm
     assert kwargs["timeout"] == module.DEFAULT_TIMEOUT_SECONDS
 
 
+def test_plugin_anchors_followup_retrieval_to_recent_user_topics_without_assistant_text() -> None:
+    module = load_plugin()
+    query = "Narrow that candidate and return the revised falsifier."
+    history = [
+        {
+            "role": "user",
+            "content": "Does LivingIP token ownership create governance over narrative and licensing?",
+        },
+        {"role": "assistant", "content": "UNRELATED_ASSISTANT_SECRET about GLP-1."},
+        {
+            "role": "user",
+            "content": "Compare LivingIP ownership with governance rights and treasury control.",
+        },
+        {"role": "user", "content": query},
+    ]
+
+    retrieval_query, anchors = module._contextual_retrieval_query(query, history)
+
+    assert retrieval_query == query
+    assert anchors == ["LivingIP"]
+    assert "UNRELATED_ASSISTANT_SECRET" not in retrieval_query
+    assert len(anchors) <= module.MAX_RETRIEVAL_ANCHORS
+
+
+def test_plugin_keeps_operational_query_separate_from_contextual_anchors(tmp_path: Path) -> None:
+    module = load_plugin()
+    home = tmp_path / "profile"
+    kb_tool = home / "bin" / "kb_tool.py"
+    kb_tool.parent.mkdir(parents=True)
+    kb_tool.write_text("# fixture\n", encoding="utf-8")
+    calls = []
+    payload = {
+        "claims": [],
+        "context_rows": [],
+        "operational_contracts": [{"id": "reply_budget", "hard_max_words": 220}],
+        "compiled_response": None,
+    }
+
+    def fake_runner(command, **kwargs):
+        calls.append(command)
+        return subprocess.CompletedProcess(command, 0, stdout=json.dumps(payload), stderr="")
+
+    query = "Revise that candidate."
+    module.build_database_context(
+        query,
+        conversation_history=[
+            {"role": "user", "content": "LivingIP ownership needs governance and treasury rights."},
+            {"role": "user", "content": "LivingIP token ownership is not narrative control."},
+            {"role": "user", "content": query},
+        ],
+        hermes_home=home,
+        runner=fake_runner,
+    )
+
+    command = calls[0]
+    assert command[4] == query
+    retrieval_query = command[command.index("--retrieval-query") + 1]
+    assert retrieval_query == query
+    anchors = [command[index + 1] for index, item in enumerate(command) if item == "--retrieval-anchor"]
+    assert anchors == ["LivingIP"]
+    assert all("treasury rights" not in item for item in anchors)
+
+
+def test_plugin_uses_repeated_user_topics_only_when_no_explicit_subject_anchor() -> None:
+    module = load_plugin()
+    query = "Narrow that position."
+    history = [
+        {"role": "user", "content": "Compare ownership governance with licensing constraints."},
+        {"role": "assistant", "content": "UNRELATED_ASSISTANT_SECRET governance ownership."},
+        {"role": "user", "content": "Ownership governance may not imply narrative control."},
+        {"role": "user", "content": query},
+    ]
+
+    retrieval_query, anchors = module._contextual_retrieval_query(query, history)
+
+    assert retrieval_query == query
+    assert {item.casefold() for item in anchors} == {"ownership", "governance"}
+
+
 def test_plugin_replaces_contract_violating_draft_with_compiled_db_response(tmp_path: Path, monkeypatch) -> None:
     module = load_plugin()
     home = tmp_path / "profile"

tests/test_hermes_leoclean_kb_bridge_source.py

@@ -1411,7 +1411,7 @@ def test_vps_bridge_context_receipt_is_deterministic_and_retries_moving_wal(monk
     monkeypatch.setattr(module, "database_read_marker", lambda _args: next(markers))
     monkeypatch.setattr(module, "context_operational_contracts", contracts)
     monkeypatch.setattr(module, "compile_operational_response", lambda _contracts: None)
-    monkeypatch.setattr(module, "find_claims", lambda _args, _query, _limit: [])
+    monkeypatch.setattr(module, "find_claims", lambda _args, _query, _limit, _anchors=None: [])
     monkeypatch.setattr(module, "load_evidence", lambda _args, _ids, _limit: {})
     monkeypatch.setattr(module, "load_edges", lambda _args, _ids, _limit: {})
     monkeypatch.setattr(module, "find_context_rows", lambda _args, _query, _limit: [])
@@ -2211,7 +2211,7 @@ def test_vps_bridge_context_corpus_does_not_expose_persona_source_ref(monkeypatc
     assert "p.voice, p.role" in captured["sql"]
 
 
-def test_vps_bridge_ranked_retrieval_keeps_one_match_recall_floor(monkeypatch) -> None:
+def test_vps_bridge_ranked_retrieval_boosts_bounded_anchors_and_keeps_one_match_recall_floor(monkeypatch) -> None:
     module = _load_module(BRIDGE_DIR / "kb_tool.py")
     sql_statements = []
 
@@ -2220,14 +2220,75 @@ def test_vps_bridge_ranked_retrieval_keeps_one_match_recall_floor(monkeypatch) -
         return []
 
     monkeypatch.setattr(module, "psql_json", capture_sql)
-    module.find_claims(SimpleNamespace(), "singular ceramic topic with several wrapper terms", 4)
+    module.find_claims(
+        SimpleNamespace(),
+        "singular ceramic topic with several wrapper terms",
+        4,
+        ["LivingIP", "ownership"],
+    )
     module.find_context_rows(SimpleNamespace(), "singular ceramic topic with several wrapper terms", 4)
 
     assert len(sql_statements) == 2
-    for sql in sql_statements:
-        assert "max(score) over () as max_score" in sql
-        assert "score >= 1" in sql
-        assert "case when max_score >= 2 then 2 else 1 end" in sql
+    claim_sql, context_sql = sql_statements
+    assert "anchors(patterns) as" in claim_sql
+    assert "%livingip%" in claim_sql
+    assert "%ownership%" in claim_sql
+    assert "anchor_score" in claim_sql
+    assert "raw_selected as" in claim_sql
+    assert "anchor_selected as" in claim_sql
+    assert "'_selection_lane', selection_lane" in claim_sql
+    assert "where anchor_score >= 1" in claim_sql
+    assert "max(score) over () as max_score" in claim_sql
+    assert "max(score) over () as max_score" in context_sql
+    assert "case when max_score >= 2 then 2 else 1 end" in context_sql
+
+
+def test_vps_bridge_ranked_retrieval_has_no_match_all_anchor_when_anchors_absent(monkeypatch) -> None:
+    module = _load_module(BRIDGE_DIR / "kb_tool.py")
+    captured = {}
+
+    def capture_sql(_args, sql):
+        captured["sql"] = sql
+        return []
+
+    monkeypatch.setattr(module, "psql_json", capture_sql)
+    module.find_claims(SimpleNamespace(), "specific topic", 4)
+
+    assert "array[]::text[]" in captured["sql"]
+    assert "'%%'" not in captured["sql"]
+
+
+def test_vps_bridge_ranked_retrieval_preserves_raw_results_when_subject_already_present(monkeypatch) -> None:
+    module = _load_module(BRIDGE_DIR / "kb_tool.py")
+    rows = [
+        {"id": "raw-subject", "score": 4, "anchor_score": 1, "_selection_lane": "raw"},
+        {"id": "raw-cross-domain", "score": 3, "anchor_score": 0, "_selection_lane": "raw"},
+        {"id": "anchor-duplicate", "score": 4, "anchor_score": 1, "_selection_lane": "anchor"},
+    ]
+    monkeypatch.setattr(module, "psql_json", lambda _args, _sql: rows)
+
+    result = module.find_claims(SimpleNamespace(), "LivingIP ownership", 4, ["LivingIP"])
+
+    assert [row["id"] for row in result] == ["raw-subject", "raw-cross-domain"]
+    assert all("_selection_lane" not in row for row in result)
+
+
+def test_vps_bridge_ranked_retrieval_backfills_subject_without_discarding_top_raw_rows(monkeypatch) -> None:
+    module = _load_module(BRIDGE_DIR / "kb_tool.py")
+    rows = [
+        *(
+            {"id": f"raw-{index}", "score": 5 - index, "anchor_score": 0, "_selection_lane": "raw"}
+            for index in range(4)
+        ),
+        {"id": "subject-1", "score": 1, "anchor_score": 1, "_selection_lane": "anchor"},
+        {"id": "subject-2", "score": 0, "anchor_score": 1, "_selection_lane": "anchor"},
+    ]
+    monkeypatch.setattr(module, "psql_json", lambda _args, _sql: rows)
+
+    result = module.find_claims(SimpleNamespace(), "revise that conclusion", 4, ["LivingIP"])
+
+    assert [row["id"] for row in result] == ["subject-1", "subject-2", "raw-0", "raw-1"]
+    assert all("_selection_lane" not in row for row in result)
 
 
 def test_vps_bridge_combines_identity_and_restart_contracts_for_independent_paraphrase() -> None: