fix(hindsight): complete recall tool methods

agent/model_metadata.py

@@ -1454,10 +1454,11 @@ def get_model_context_length(
        d. GMI /models endpoint
        e. Ollama native /api/show probe (any base_url, provider-agnostic)
        f. models.dev registry lookup (with :cloud/-cloud suffix fallback)
-    6. OpenRouter live API metadata (Kimi-family 32k guard)
-    7. Hardcoded defaults (broad family patterns, longest-key-first)
-    8. Local server query (last resort)
-    9. Default fallback (256K)"""
+    6. Exact curated defaults
+    7. OpenRouter live API metadata (Kimi-family 32k guard)
+    8. Fuzzy hardcoded defaults (broad family patterns, longest-key-first)
+    9. Local server query (last resort)
+    10. Default fallback (256K)"""
     # 0. Explicit config override — user knows best
     if config_context_length is not None and isinstance(config_context_length, int) and config_context_length > 0:
         return config_context_length
@@ -1672,15 +1673,21 @@ def get_model_context_length(
         if ctx:
             return ctx
 
-    # 6. OpenRouter live API metadata — provider-unaware fallback.
+    # 6. Exact curated defaults. Provider-aware branches above still win, but
+    # an exact local entry is more specific than provider-unaware OpenRouter
+    # metadata for bare model IDs like "hy3-preview".
+    exact_default = DEFAULT_CONTEXT_LENGTHS.get(model.lower())
+    if exact_default:
+        return exact_default
+
+    # 7. OpenRouter live API metadata — provider-unaware fallback.
     # Only consulted when the provider is unknown (no effective_provider),
     # because OpenRouter data is community-maintained and can be incorrect
     # for models that belong to known providers with curated defaults.
     if not effective_provider:
         metadata = fetch_model_metadata()
         if model in metadata:
             or_ctx = metadata[model].get("context_length", DEFAULT_FALLBACK_CONTEXT)
-            # Guard against stale OpenRouter metadata for Kimi-family models.
             if or_ctx == 32768 and _model_name_suggests_kimi(model):
                 logger.info(
                     "Rejecting OpenRouter metadata context=%s for %r "
@@ -1690,9 +1697,7 @@ def get_model_context_length(
             else:
                 return or_ctx
 
-    # 7. (reserved)
-
-    # 8. Hardcoded defaults (fuzzy match — longest key first for specificity)
+    # 8. Fuzzy hardcoded defaults (longest key first for specificity)
     # Only check `default_model in model` (is the key a substring of the input).
     # The reverse (`model in default_model`) causes shorter names like
     # "claude-sonnet-4" to incorrectly match "claude-sonnet-4-6" and return 1M.

gateway/platforms/base.py

@@ -65,7 +65,11 @@ def _thread_metadata_for_source(source, reply_to_message_id: str | None = None)
     if thread_id is None:
         return None
     metadata = {"thread_id": thread_id}
-    if _platform_name(getattr(source, "platform", None)) == "telegram" and getattr(source, "chat_type", None) == "dm":
+    if (
+        _platform_name(getattr(source, "platform", None)) == "telegram"
+        and getattr(source, "chat_type", None) == "dm"
+        and str(thread_id).isdigit()
+    ):
         metadata["telegram_dm_topic_reply_fallback"] = True
         tid = str(thread_id)
         if tid and tid not in {"", "1"}:
@@ -79,11 +83,12 @@ def _thread_metadata_for_source(source, reply_to_message_id: str | None = None)
 def _reply_anchor_for_event(event) -> str | None:
     """Return reply_to id for platforms that need reply semantics.
 
-    Telegram forum/supergroup topics should be routed by topic metadata, not by
-    replying to the triggering message. Hermes-created Telegram private-chat
-    topic lanes prefer replying to the triggering user message so the answer
-    stays attached to the active lane; synthetic/resumed sends fall back to
-    ``direct_messages_topic_id`` metadata when no message id is available.
+    Telegram forum/supergroup topics are routed by topic metadata, but keeping
+    the triggering message as the reply anchor preserves threaded reply
+    behavior in adapters that support both. Hermes-created Telegram
+    private-chat topic lanes prefer replying to the triggering user message;
+    synthetic/resumed sends fall back to ``direct_messages_topic_id`` metadata
+    when no message id is available.
     """
     source = getattr(event, "source", None)
     platform = _platform_name(getattr(source, "platform", None))
@@ -93,7 +98,7 @@ def _reply_anchor_for_event(event) -> str | None:
         # topic seed/anchor can render the bot response outside the active lane.
         return getattr(event, "message_id", None) or getattr(event, "reply_to_message_id", None)
     if platform == "telegram" and thread_id:
-        return None
+        return getattr(event, "message_id", None) or getattr(event, "reply_to_message_id", None)
     if platform == "feishu" and thread_id and getattr(event, "reply_to_message_id", None):
         return getattr(event, "reply_to_message_id", None)
     return getattr(event, "message_id", None)

gateway/run.py

@@ -1035,6 +1035,7 @@ def _reload_runtime_env_preserving_config_authority() -> None:
     MessageEvent,
     MessageType,
     _reply_anchor_for_event,
+    _thread_metadata_for_source,
     merge_pending_message_event,
 )
 from gateway.restart import (
@@ -11500,7 +11501,10 @@ async def _deliver_media_from_response(
             local_files, _ = adapter.extract_local_files(cleaned)
             local_files = BasePlatformAdapter.filter_local_delivery_paths(local_files)
 
-            _thread_meta = self._thread_metadata_for_source(event.source, self._reply_anchor_for_event(event))
+            if hasattr(self, "_thread_metadata_for_source") and hasattr(self, "_reply_anchor_for_event"):
+                _thread_meta = self._thread_metadata_for_source(event.source, self._reply_anchor_for_event(event))
+            else:
+                _thread_meta = _thread_metadata_for_source(event.source, _reply_anchor_for_event(event))
 
             _VIDEO_EXTS = {'.mp4', '.mov', '.avi', '.mkv', '.webm', '.3gp'}
             _IMAGE_EXTS = {'.jpg', '.jpeg', '.png', '.webp', '.gif'}
@@ -13767,6 +13771,7 @@ def _thread_metadata_for_source(
         if (
             getattr(source, "platform", None) == Platform.TELEGRAM
             and getattr(source, "chat_type", None) == "dm"
+            and str(thread_id).isdigit()
         ):
             metadata["telegram_dm_topic_reply_fallback"] = True
             # Telegram DM topic lanes need direct_messages_topic_id in metadata

plugins/memory/hindsight/README.md

@@ -118,7 +118,7 @@ Available in `hybrid` and `tools` memory modes:
 | Tool | Description |
 |------|-------------|
 | `hindsight_retain` | Store information with auto entity extraction; supports optional per-call `tags` |
-| `hindsight_recall` | Multi-strategy search (semantic + entity graph) |
+| `hindsight_recall` | Retrieve memories. Defaults to semantic/entity-graph recall; supports optional `method` (`recall`, `list`, `entity`) plus per-call `budget`, `max_tokens`, `types`, `tags`, `tags_match`, and exact-match `metadata` filters. `tag_groups` is passed through for `recall`/`entity`; `method="list"` uses the public list API and supports `limit`/`offset`; `method="entity"` supports `max_entity_tokens`. |
 | `hindsight_reflect` | Cross-memory synthesis (LLM-powered) |
 
 ## Environment Variables