AbstractRuntime 0.2.0__py3-none-any.whl → 0.4.0__py3-none-any.whl

abstractruntime/memory/active_context.py

@@ -0,0 +1,746 @@
+"""Active context policy + provenance-based recall utilities.
+
+Goal: keep a strict separation between:
+- Stored memory (durable): RunStore/LedgerStore/ArtifactStore
+- Active context (LLM-visible view): RunState.vars["context"]["messages"]
+
+This module is intentionally small and JSON-safe. It does not implement semantic
+search or "graph compression"; it establishes the contracts needed for those.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from datetime import datetime
+from typing import Any, Dict, Iterable, List, Optional, Sequence, Tuple, Union
+
+from ..core.models import RunState
+from ..core.vars import get_context, get_limits, get_runtime
+from ..storage.artifacts import ArtifactMetadata, ArtifactStore
+from ..storage.base import RunStore
+
+
+def _parse_iso(ts: str) -> datetime:
+    value = (ts or "").strip()
+    if not value:
+        raise ValueError("empty timestamp")
+    # Accept both "+00:00" and "Z"
+    if value.endswith("Z"):
+        value = value[:-1] + "+00:00"
+    return datetime.fromisoformat(value)
+
+
+@dataclass(frozen=True)
+class TimeRange:
+    """A closed interval [start, end] in ISO8601 strings.
+
+    If start or end is None, the range is unbounded on that side.
+    """
+
+    start: Optional[str] = None
+    end: Optional[str] = None
+
+    def contains(self, *, start: Optional[str], end: Optional[str]) -> bool:
+        """Return True if [start,end] intersects this range.
+
+        Spans missing timestamps are treated as non-matching when a range is used,
+        because we cannot prove they belong to the requested interval.
+        """
+
+        if self.start is None and self.end is None:
+            return True
+        if not start or not end:
+            return False
+
+        span_start = _parse_iso(start)
+        span_end = _parse_iso(end)
+        range_start = _parse_iso(self.start) if self.start else None
+        range_end = _parse_iso(self.end) if self.end else None
+
+        # Intersection test for closed ranges:
+        # span_end >= range_start AND span_start <= range_end
+        if range_start is not None and span_end < range_start:
+            return False
+        if range_end is not None and span_start > range_end:
+            return False
+        return True
+
+
+SpanId = Union[str, int]  # artifact_id or 1-based index into _runtime.memory_spans
+
+
+class ActiveContextPolicy:
+    """Runtime-owned utilities for memory spans and active context."""
+
+    def __init__(
+        self,
+        *,
+        run_store: RunStore,
+        artifact_store: ArtifactStore,
+    ) -> None:
+        self._run_store = run_store
+        self._artifact_store = artifact_store
+
+    # ------------------------------------------------------------------
+    # Spans: list + filter
+    # ------------------------------------------------------------------
+
+    def list_memory_spans(self, run_id: str) -> List[Dict[str, Any]]:
+        """Return the run's archived span index (`_runtime.memory_spans`)."""
+        run = self._require_run(run_id)
+        return self.list_memory_spans_from_run(run)
+
+    @staticmethod
+    def list_memory_spans_from_run(run: RunState) -> List[Dict[str, Any]]:
+        """Return the archived span index (`_runtime.memory_spans`) from an in-memory RunState."""
+        runtime_ns = get_runtime(run.vars)
+        spans = runtime_ns.get("memory_spans")
+        if not isinstance(spans, list):
+            return []
+        out: List[Dict[str, Any]] = []
+        for s in spans:
+            if isinstance(s, dict):
+                out.append(dict(s))
+        out.sort(key=lambda d: str(d.get("created_at") or ""), reverse=True)
+        return out
+
+    def filter_spans(
+        self,
+        run_id: str,
+        *,
+        time_range: Optional[TimeRange] = None,
+        tags: Optional[Dict[str, Any]] = None,
+        tags_mode: str = "all",
+        authors: Optional[List[str]] = None,
+        locations: Optional[List[str]] = None,
+        query: Optional[str] = None,
+        limit: int = 1000,
+    ) -> List[Dict[str, Any]]:
+        """Filter archived spans by time range, tags, and a basic keyword query.
+
+        Notes:
+        - This is a metadata filter, not semantic retrieval.
+        - `query` matches against summary messages (if present) and metadata tags.
+        """
+        run = self._require_run(run_id)
+        return self.filter_spans_from_run(
+            run,
+            artifact_store=self._artifact_store,
+            time_range=time_range,
+            tags=tags,
+            tags_mode=tags_mode,
+            authors=authors,
+            locations=locations,
+            query=query,
+            limit=limit,
+        )
+
+    @staticmethod
+    def filter_spans_from_run(
+        run: RunState,
+        *,
+        artifact_store: ArtifactStore,
+        time_range: Optional[TimeRange] = None,
+        tags: Optional[Dict[str, Any]] = None,
+        tags_mode: str = "all",
+        authors: Optional[List[str]] = None,
+        locations: Optional[List[str]] = None,
+        query: Optional[str] = None,
+        limit: int = 1000,
+    ) -> List[Dict[str, Any]]:
+        """Like `filter_spans`, but operates on an in-memory RunState."""
+        spans = ActiveContextPolicy.list_memory_spans_from_run(run)
+        if not spans:
+            return []
+
+        summary_by_artifact = ActiveContextPolicy.summary_text_by_artifact_id_from_run(run)
+        lowered_query = (query or "").strip().lower() if query else None
+
+        # Notes are small; for keyword filtering we can load their text safely.
+        # IMPORTANT: only do this when we actually have a query (avoids unnecessary I/O).
+        #
+        # Also include summary text from the note's linked conversation span(s) when available,
+        # so searching for a topic can surface both the span *and* the derived note.
+        note_by_artifact: Dict[str, str] = {}
+        if lowered_query:
+            for span in spans:
+                if not isinstance(span, dict):
+                    continue
+                if str(span.get("kind") or "") != "memory_note":
+                    continue
+                artifact_id = str(span.get("artifact_id") or "")
+                if not artifact_id:
+                    continue
+                try:
+                    payload = artifact_store.load_json(artifact_id)
+                except Exception:
+                    continue
+                if not isinstance(payload, dict):
+                    continue
+
+                parts: list[str] = []
+                note = payload.get("note")
+                if isinstance(note, str) and note.strip():
+                    parts.append(note.strip())
+
+                sources = payload.get("sources")
+                if isinstance(sources, dict):
+                    span_ids = sources.get("span_ids")
+                    if isinstance(span_ids, list):
+                        for sid in span_ids:
+                            if not isinstance(sid, str) or not sid.strip():
+                                continue
+                            summary = summary_by_artifact.get(sid.strip())
+                            if isinstance(summary, str) and summary.strip():
+                                parts.append(summary.strip())
+
+                if parts:
+                    note_by_artifact[artifact_id] = "\n".join(parts).strip()
+
+        def _artifact_meta(artifact_id: str) -> Optional[ArtifactMetadata]:
+            try:
+                return artifact_store.get_metadata(artifact_id)
+            except Exception:
+                return None
+
+        mode = str(tags_mode or "all").strip().lower() or "all"
+        if mode in {"and"}:
+            mode = "all"
+        if mode in {"or"}:
+            mode = "any"
+        if mode not in {"all", "any"}:
+            mode = "all"
+
+        authors_norm = {str(a).strip().lower() for a in (authors or []) if isinstance(a, str) and a.strip()}
+        locations_norm = {str(l).strip().lower() for l in (locations or []) if isinstance(l, str) and l.strip()}
+
+        out: List[Dict[str, Any]] = []
+        for span in spans:
+            artifact_id = str(span.get("artifact_id") or "")
+            if not artifact_id:
+                continue
+
+            if time_range is not None:
+                if not time_range.contains(
+                    start=span.get("from_timestamp"),
+                    end=span.get("to_timestamp"),
+                ):
+                    continue
+
+            meta = _artifact_meta(artifact_id)
+
+            if tags:
+                if not ActiveContextPolicy._tags_match(span=span, meta=meta, required=tags, mode=mode):
+                    continue
+
+            if authors_norm:
+                created_by = span.get("created_by")
+                author = str(created_by).strip().lower() if isinstance(created_by, str) else ""
+                if not author or author not in authors_norm:
+                    continue
+
+            if locations_norm:
+                loc = span.get("location")
+                loc_str = str(loc).strip().lower() if isinstance(loc, str) else ""
+                if not loc_str:
+                    # Fallback: allow location to be stored as a tag.
+                    span_tags = span.get("tags") if isinstance(span.get("tags"), dict) else {}
+                    tag_loc = span_tags.get("location") if isinstance(span_tags, dict) else None
+                    loc_str = str(tag_loc).strip().lower() if isinstance(tag_loc, str) else ""
+                if not loc_str or loc_str not in locations_norm:
+                    continue
+
+            if lowered_query:
+                haystack = ActiveContextPolicy._span_haystack(
+                    span=span,
+                    meta=meta,
+                    summary=summary_by_artifact.get(artifact_id),
+                    note=note_by_artifact.get(artifact_id),
+                )
+                if lowered_query not in haystack:
+                    continue
+
+            out.append(span)
+            if len(out) >= limit:
+                break
+        return out
+
+    # ------------------------------------------------------------------
+    # Rehydration: stored span -> active context
+    # ------------------------------------------------------------------
+
+    def rehydrate_into_context(
+        self,
+        run_id: str,
+        *,
+        span_ids: Sequence[SpanId],
+        placement: str = "after_summary",
+        dedup_by: str = "message_id",
+        max_messages: Optional[int] = None,
+    ) -> Dict[str, Any]:
+        """Rehydrate archived span(s) into `context.messages` and persist the run.
+
+        Args:
+            run_id: Run to mutate.
+            span_ids: Sequence of artifact_ids or 1-based indices into `_runtime.memory_spans`.
+            placement: Where to insert. Supported: "after_summary" (default), "after_system", "end".
+            dedup_by: Dedup key (default: metadata.message_id).
+            max_messages: Optional cap on inserted messages across all spans (None = unlimited).
+        """
+        run = self._require_run(run_id)
+        out = self.rehydrate_into_context_from_run(
+            run,
+            span_ids=span_ids,
+            placement=placement,
+            dedup_by=dedup_by,
+            max_messages=max_messages,
+        )
+
+        self._run_store.save(run)
+        return out
+
+    def rehydrate_into_context_from_run(
+        self,
+        run: RunState,
+        *,
+        span_ids: Sequence[SpanId],
+        placement: str = "after_summary",
+        dedup_by: str = "message_id",
+        max_messages: Optional[int] = None,
+    ) -> Dict[str, Any]:
+        """Like `rehydrate_into_context`, but operates on an in-memory RunState.
+
+        This mutates `run.vars["context"]["messages"]` (and `run.output["messages"]` when present),
+        but does NOT persist the run.
+        """
+        spans = self.list_memory_spans_from_run(run)
+        resolved_artifacts: List[str] = self.resolve_span_ids_from_spans(span_ids, spans)
+        if not resolved_artifacts:
+            return {"inserted": 0, "skipped": 0, "artifacts": []}
+
+        ctx = get_context(run.vars)
+        active = ctx.get("messages")
+        if not isinstance(active, list):
+            active = []
+
+        inserted_total = 0
+        skipped_total = 0
+        per_artifact: List[Dict[str, Any]] = []
+
+        # Build a dedup set for active context.
+        existing_keys = self._collect_message_keys(active, dedup_by=dedup_by)
+
+        # Normalize cap.
+        try:
+            max_messages_int = int(max_messages) if max_messages is not None else None
+        except Exception:
+            max_messages_int = None
+        if max_messages_int is not None and max_messages_int < 0:
+            max_messages_int = None
+
+        def _preview_inserted(messages: Sequence[Dict[str, Any]]) -> str:
+            """Build a small, human-friendly preview for UI/observability."""
+            if not messages:
+                return ""
+            lines: list[str] = []
+            for m in messages[:3]:
+                if not isinstance(m, dict):
+                    continue
+                role = str(m.get("role") or "").strip()
+                content = str(m.get("content") or "").strip()
+                if not content:
+                    continue
+                # If the synthetic memory note marker is present, keep it as-is (no "role:" prefix).
+                if content.startswith("[MEMORY NOTE]"):
+                    lines.append(content)
+                else:
+                    lines.append(f"{role}: {content}" if role else content)
+            text = "\n".join([l for l in lines if l]).strip()
+            if len(text) > 360:
+                return text[:357] + "…"
+            return text
+
+        for artifact_id in resolved_artifacts:
+            if max_messages_int is not None and inserted_total >= max_messages_int:
+                # Deterministic: stop inserting once the global cap is reached.
+                per_artifact.append(
+                    {"artifact_id": artifact_id, "inserted": 0, "skipped": 0, "error": "max_messages"}
+                )
+                continue
+
+            archived = self._artifact_store.load_json(artifact_id)
+            archived_messages = archived.get("messages") if isinstance(archived, dict) else None
+
+            # Support rehydrating memory_note spans into context as a single synthetic message.
+            # This is the most practical "make recalled memory LLM-visible" behavior for
+            # visual workflows (users expect "Recall into context" to work with notes).
+            if not isinstance(archived_messages, list):
+                note_text = None
+                if isinstance(archived, dict):
+                    raw_note = archived.get("note")
+                    if isinstance(raw_note, str) and raw_note.strip():
+                        note_text = raw_note.strip()
+                if note_text is not None:
+                    archived_messages = [
+                        {
+                            "role": "system",
+                            "content": f"[MEMORY NOTE]\n{note_text}",
+                            "timestamp": str(archived.get("created_at") or "") if isinstance(archived, dict) else "",
+                            "metadata": {
+                                "kind": "memory_note",
+                                "rehydrated": True,
+                                "source_artifact_id": artifact_id,
+                                "message_id": f"memory_note:{artifact_id}",
+                            },
+                        }
+                    ]
+                else:
+                    per_artifact.append(
+                        {"artifact_id": artifact_id, "inserted": 0, "skipped": 0, "error": "missing_messages"}
+                    )
+                    continue
+
+            to_insert: List[Dict[str, Any]] = []
+            skipped = 0
+            for m in archived_messages:
+                if not isinstance(m, dict):
+                    continue
+                m_copy = dict(m)
+                meta_copy = m_copy.get("metadata")
+                if not isinstance(meta_copy, dict):
+                    meta_copy = {}
+                    m_copy["metadata"] = meta_copy
+                # Mark as rehydrated view (do not mutate the archived artifact payload).
+                meta_copy.setdefault("rehydrated", True)
+                meta_copy.setdefault("source_artifact_id", artifact_id)
+
+                key = self._message_key(m_copy, dedup_by=dedup_by)
+                if key and key in existing_keys:
+                    skipped += 1
+                    continue
+                if key:
+                    existing_keys.add(key)
+                to_insert.append(m_copy)
+
+            if max_messages_int is not None:
+                remaining = max(0, max_messages_int - inserted_total)
+                if remaining <= 0:
+                    per_artifact.append(
+                        {"artifact_id": artifact_id, "inserted": 0, "skipped": 0, "error": "max_messages"}
+                    )
+                    continue
+                if len(to_insert) > remaining:
+                    to_insert = to_insert[:remaining]
+
+            idx = self._insertion_index(active, artifact_id=artifact_id, placement=placement)
+            active[idx:idx] = to_insert
+
+            inserted_total += len(to_insert)
+            skipped_total += skipped
+            entry: Dict[str, Any] = {"artifact_id": artifact_id, "inserted": len(to_insert), "skipped": skipped}
+            preview = _preview_inserted(to_insert)
+            if preview:
+                entry["preview"] = preview
+            per_artifact.append(entry)
+
+        ctx["messages"] = active
+        if isinstance(getattr(run, "output", None), dict):
+            run.output["messages"] = active
+
+        return {"inserted": inserted_total, "skipped": skipped_total, "artifacts": per_artifact}
+
+    # ------------------------------------------------------------------
+    # Deriving what the LLM sees
+    # ------------------------------------------------------------------
+
+    def select_active_messages_for_llm(
+        self,
+        run_id: str,
+        *,
+        max_history_messages: Optional[int] = None,
+    ) -> List[Dict[str, Any]]:
+        """Return the active-context view that should be sent to an LLM.
+
+        This does NOT mutate the run; it returns a derived view.
+
+        Rules (minimal, stable):
+        - Always preserve system messages
+        - Apply `max_history_messages` to non-system messages only
+        - If max_history_messages is None, read it from `_limits.max_history_messages`
+        """
+        run = self._require_run(run_id)
+        return self.select_active_messages_for_llm_from_run(
+            run,
+            max_history_messages=max_history_messages,
+        )
+
+    @staticmethod
+    def select_active_messages_for_llm_from_run(
+        run: RunState,
+        *,
+        max_history_messages: Optional[int] = None,
+    ) -> List[Dict[str, Any]]:
+        """Like `select_active_messages_for_llm`, but operates on an in-memory RunState."""
+        ctx = get_context(run.vars)
+        messages = ctx.get("messages")
+        if not isinstance(messages, list):
+            return []
+
+        if max_history_messages is None:
+            limits = get_limits(run.vars)
+            try:
+                max_history_messages = int(limits.get("max_history_messages", -1))
+            except Exception:
+                max_history_messages = -1
+
+        return ActiveContextPolicy.select_messages_view(messages, max_history_messages=max_history_messages)
+
+    @staticmethod
+    def select_messages_view(
+        messages: Sequence[Any],
+        *,
+        max_history_messages: int,
+    ) -> List[Dict[str, Any]]:
+        """Select an LLM-visible view from a message list under a simple history limit."""
+        system_msgs: List[Dict[str, Any]] = [m for m in messages if isinstance(m, dict) and m.get("role") == "system"]
+        convo_msgs: List[Dict[str, Any]] = [m for m in messages if isinstance(m, dict) and m.get("role") != "system"]
+
+        if max_history_messages == -1:
+            return system_msgs + convo_msgs
+        if max_history_messages < 0:
+            return system_msgs + convo_msgs
+        if max_history_messages == 0:
+            return system_msgs
+        return system_msgs + convo_msgs[-max_history_messages:]
+
+    # ------------------------------------------------------------------
+    # Internals
+    # ------------------------------------------------------------------
+
+    def _require_run(self, run_id: str) -> RunState:
+        run = self._run_store.load(run_id)
+        if run is None:
+            raise KeyError(f"Unknown run_id: {run_id}")
+        return run
+
+    @staticmethod
+    def resolve_span_ids_from_spans(span_ids: Sequence[SpanId], spans: Sequence[Dict[str, Any]]) -> List[str]:
+        resolved: List[str] = []
+        for sid in span_ids:
+            if isinstance(sid, int):
+                idx = sid - 1
+                if 0 <= idx < len(spans):
+                    artifact_id = spans[idx].get("artifact_id")
+                    if isinstance(artifact_id, str) and artifact_id:
+                        resolved.append(artifact_id)
+                continue
+            if isinstance(sid, str):
+                s = sid.strip()
+                if not s:
+                    continue
+                # If it's a digit string, treat as 1-based index.
+                if s.isdigit():
+                    idx = int(s) - 1
+                    if 0 <= idx < len(spans):
+                        artifact_id = spans[idx].get("artifact_id")
+                        if isinstance(artifact_id, str) and artifact_id:
+                            resolved.append(artifact_id)
+                    continue
+                resolved.append(s)
+        # Preserve order but dedup
+        seen = set()
+        out: List[str] = []
+        for a in resolved:
+            if a in seen:
+                continue
+            seen.add(a)
+            out.append(a)
+        return out
+
+    def _insertion_index(self, active: List[Any], *, artifact_id: str, placement: str) -> int:
+        if placement == "end":
+            return len(active)
+
+        if placement == "after_system":
+            i = 0
+            while i < len(active):
+                m = active[i]
+                if not isinstance(m, dict) or m.get("role") != "system":
+                    break
+                i += 1
+            return i
+
+        # default: after_summary
+        for i, m in enumerate(active):
+            if not isinstance(m, dict):
+                continue
+            if m.get("role") != "system":
+                continue
+            meta = m.get("metadata") if isinstance(m.get("metadata"), dict) else {}
+            if meta.get("kind") == "memory_summary" and meta.get("source_artifact_id") == artifact_id:
+                return i + 1
+
+        # Fallback: after system messages.
+        return self._insertion_index(active, artifact_id=artifact_id, placement="after_system")
+
+    def _collect_message_keys(self, messages: Iterable[Any], *, dedup_by: str) -> set[str]:
+        keys: set[str] = set()
+        for m in messages:
+            if not isinstance(m, dict):
+                continue
+            key = self._message_key(m, dedup_by=dedup_by)
+            if key:
+                keys.add(key)
+        return keys
+
+    def _message_key(self, message: Dict[str, Any], *, dedup_by: str) -> Optional[str]:
+        if dedup_by == "message_id":
+            meta = message.get("metadata")
+            if isinstance(meta, dict):
+                mid = meta.get("message_id")
+                if isinstance(mid, str) and mid:
+                    return mid
+            return None
+        # Unknown dedup key: disable dedup
+        return None
+
+    @staticmethod
+    def summary_text_by_artifact_id_from_run(run: RunState) -> Dict[str, str]:
+        ctx = get_context(run.vars)
+        messages = ctx.get("messages")
+        if not isinstance(messages, list):
+            return {}
+        out: Dict[str, str] = {}
+        for m in messages:
+            if not isinstance(m, dict):
+                continue
+            if m.get("role") != "system":
+                continue
+            meta = m.get("metadata")
+            if not isinstance(meta, dict):
+                continue
+            if meta.get("kind") != "memory_summary":
+                continue
+            artifact_id = meta.get("source_artifact_id")
+            if isinstance(artifact_id, str) and artifact_id:
+                out[artifact_id] = str(m.get("content") or "")
+        return out
+
+    @staticmethod
+    def _span_haystack(
+        *,
+        span: Dict[str, Any],
+        meta: Optional[ArtifactMetadata],
+        summary: Optional[str],
+        note: Optional[str] = None,
+    ) -> str:
+        parts: List[str] = []
+        if summary:
+            parts.append(summary)
+        if note:
+            parts.append(note)
+        for k in ("kind", "compression_mode", "focus", "from_timestamp", "to_timestamp", "created_by", "location"):
+            v = span.get(k)
+            if isinstance(v, str) and v:
+                parts.append(v)
+        # Span tags are persisted in run vars (topic/person/project, etc).
+        span_tags = span.get("tags")
+        if isinstance(span_tags, dict):
+            for k, v in span_tags.items():
+                if isinstance(v, str) and v:
+                    parts.append(str(k))
+                    parts.append(v)
+
+        if meta is not None:
+            parts.append(meta.content_type or "")
+            for k, v in (meta.tags or {}).items():
+                parts.append(k)
+                parts.append(v)
+        return " ".join(parts).lower()
+
+    @staticmethod
+    def _tags_match(
+        *,
+        span: Dict[str, Any],
+        meta: Optional[ArtifactMetadata],
+        required: Dict[str, Any],
+        mode: str = "all",
+    ) -> bool:
+        def _norm(s: str) -> str:
+            return str(s or "").strip().lower()
+
+        tags: Dict[str, str] = {}
+        if meta is not None and meta.tags:
+            for k, v in meta.tags.items():
+                if not isinstance(k, str) or not isinstance(v, str):
+                    continue
+                kk = _norm(k)
+                vv = _norm(v)
+                if kk and vv and kk not in tags:
+                    tags[kk] = vv
+
+        span_tags = span.get("tags")
+        if isinstance(span_tags, dict):
+            for k, v in span_tags.items():
+                if not isinstance(k, str):
+                    continue
+                kk = _norm(k)
+                if not kk or kk in tags:
+                    continue
+                if isinstance(v, str):
+                    vv = _norm(v)
+                    if vv:
+                        tags[kk] = vv
+
+        # Derived tags from span ref (cheap and keeps filtering usable even
+        # if artifact metadata is missing).
+        for k in ("kind", "compression_mode", "focus"):
+            v = span.get(k)
+            if isinstance(v, str) and v:
+                kk = _norm(k)
+                if kk and kk not in tags:
+                    tags[kk] = _norm(v)
+
+        required_norm: Dict[str, List[str]] = {}
+        for k, v in (required or {}).items():
+            if not isinstance(k, str):
+                continue
+            kk = _norm(k)
+            if not kk or kk == "kind":
+                continue
+            values: List[str] = []
+            if isinstance(v, str):
+                vv = _norm(v)
+                if vv:
+                    values.append(vv)
+            elif isinstance(v, (list, tuple)):
+                for it in v:
+                    if isinstance(it, str) and it.strip():
+                        values.append(_norm(it))
+            if values:
+                # preserve order but dedup
+                seen: set[str] = set()
+                deduped = []
+                for x in values:
+                    if x in seen:
+                        continue
+                    seen.add(x)
+                    deduped.append(x)
+                required_norm[kk] = deduped
+
+        if not required_norm:
+            return True
+
+        def _key_matches(key: str) -> bool:
+            cand = tags.get(key)
+            if cand is None:
+                return False
+            allowed = required_norm.get(key) or []
+            return cand in allowed
+
+        op = str(mode or "all").strip().lower() or "all"
+        if op not in {"all", "any"}:
+            op = "all"
+        if op == "any":
+            return any(_key_matches(k) for k in required_norm.keys())
+        return all(_key_matches(k) for k in required_norm.keys())