fix(mcp): close out the #952 manual verification findings

src/basic_memory/api/v2/utils.py

@@ -147,6 +147,7 @@ def to_summary(
                         from_entity_id=item.from_id,
                         from_entity_external_id=from_ext_id,
                         to_entity=to_title,
+                        to_name=item.to_name,
                         to_entity_id=item.to_id,
                         to_entity_external_id=to_ext_id,
                         created_at=item.created_at,

src/basic_memory/cli/app.py

@@ -83,8 +83,11 @@ def _post_command_messages() -> None:
     # Skip for 'mcp' command - it has its own lifespan that handles initialization
     # Skip for API-using commands (status, sync, etc.) - they handle initialization via deps.py
     # Skip for 'reset' command - it manages its own database lifecycle
+    # Skip for 'man' - it only copies packaged files; a broken local database
+    # must not block installing the offline docs
     skip_init_commands = {
         "doctor",
+        "man",
         "mcp",
         "status",
         "sync",

src/basic_memory/cli/commands/status.py

@@ -190,9 +190,16 @@ async def run_status(
             if sync_report.total == 0:
                 return project_item.name, sync_report
             if time.monotonic() >= deadline:
+                # Why the hint: indexing is done by the sync coordinator, which
+                # only runs inside a live server (bm mcp / hosted API). In a
+                # CLI-only session nothing will ever drain the pending count,
+                # so this wait cannot succeed — point at the command that
+                # actually indexes (#959).
                 raise StatusTimeout(
                     f"Timed out after {timeout:g}s waiting for '{project_item.name}' "
-                    f"to finish indexing ({sync_report.total} pending change(s) remaining)."
+                    f"to finish indexing ({sync_report.total} pending change(s) remaining). "
+                    f"If no Basic Memory server is running, pending changes are never "
+                    f"indexed — run 'bm reindex --project {project_item.name}' instead."
                 )
             await asyncio.sleep(poll_interval)
 

src/basic_memory/mcp/async_client.py

@@ -10,7 +10,7 @@
 from loguru import logger
 
 import logfire
-from basic_memory.config import ConfigManager, ProjectMode
+from basic_memory.config import ConfigManager, ProjectMode, has_cloud_credentials
 
 if TYPE_CHECKING:
     from sqlalchemy.ext.asyncio import AsyncEngine, AsyncSession, async_sessionmaker
@@ -368,7 +368,8 @@ async def get_client(
     1. Factory injection.
     2. Explicit routing flags (--local/--cloud).
     3. Per-project mode routing when project_name is provided.
-    4. Local ASGI transport by default.
+    4. Cloud routing when a workspace selector is provided.
+    5. Local ASGI transport by default.
     """
     if _client_factory:
         async with _client_factory(workspace=workspace) as client:
@@ -428,6 +429,26 @@ async def get_client(
             yield client
         return
 
+    # --- Workspace-selector routing ---
+    # Trigger: caller passed a cloud workspace selector and nothing above routed.
+    # Why: a workspace names a cloud tenant — silently serving the request from
+    #   the local ASGI app sent writes to the wrong destination (#954: a cloud
+    #   project create either failed on the cloud-style path or landed locally).
+    # Outcome: route to the cloud proxy when credentials exist; without
+    #   credentials fail fast instead of pretending the operation succeeded.
+    if workspace is not None:
+        if not has_cloud_credentials(config):
+            raise RuntimeError(
+                f"A cloud workspace was requested ('{workspace}') but no cloud "
+                "credentials were found. Run 'bm cloud login' or "
+                "'bm cloud set-key <key>' first, or omit the workspace selector "
+                "for a local operation."
+            )
+        logger.debug(f"Workspace selector '{workspace}' provided - using cloud proxy client")
+        async with _cloud_client(config, timeout, workspace=workspace) as client:
+            yield client
+        return
+
     # --- Default fallback ---
     logger.debug("Default routing - using ASGI client for local Basic Memory API")
     async with _asgi_client(timeout) as client:

src/basic_memory/mcp/project_context.py

@@ -54,6 +54,14 @@
 _WORKSPACE_PROJECT_INDEX_STATE_KEY = "workspace_project_index"
 
 
+class WorkspaceProjectLookupMiss(ValueError):
+    """A project was absent from the workspace index (as opposed to ambiguous).
+
+    Misses are retried once against a freshly rebuilt index, because the
+    session cache may simply predate an out-of-band project creation (#956).
+    """
+
+
 @dataclass(frozen=True)
 class WorkspaceProjectEntry:
     """A cloud project resolved together with the workspace that owns it."""
@@ -460,6 +468,17 @@ def _canonical_memory_path_for_workspace(
     # Outcome: lookups preserve the complete workspace/project canonical permalink.
     if not normalized_remainder:
         normalized_remainder = project_permalink
+
+    # Same index-form rule as _canonical_memory_path_for_active_route (#957):
+    # without an active workspace permalink context, stored permalinks are
+    # project-qualified and a workspace-prefixed pattern cannot match.
+    if "*" in normalized_remainder and current_workspace_permalink_context() is None:
+        return build_qualified_permalink_reference(
+            project_permalink,
+            normalized_remainder,
+            include_project=True,
+        )
+
     return build_qualified_permalink_reference(
         project_permalink,
         normalized_remainder,
@@ -477,6 +496,24 @@ def _canonical_memory_path_for_active_route(
 ) -> str:
     """Return the canonical permalink path for the currently routed project/workspace."""
     project_prefix = active_project.permalink
+
+    # Trigger: the path contains a glob wildcard (folder/*) and no server-side
+    #   workspace permalink context is active.
+    # Why: patterns match raw against the search index, so they must mirror the
+    #   stored permalink form. The contextvar is what qualified permalinks at
+    #   write time — when it is absent, stored rows are project-qualified and a
+    #   workspace prefix (from the client's cached_workspace display state)
+    #   guarantees zero matches (#957). Direct lookups keep full qualification
+    #   because the link resolver understands it; patterns have no fallback.
+    # Outcome: without the contextvar, qualify patterns with the project prefix
+    #   only; with it, fall through to normal workspace canonicalization.
+    if "*" in path and current_workspace_permalink_context() is None:
+        if not include_project:
+            return path
+        if path == project_prefix or path.startswith(f"{project_prefix}/"):
+            return path
+        return f"{project_prefix}/{path}"
+
     workspace_remainder = path
     if include_project and (path == project_prefix or path.startswith(f"{project_prefix}/")):
         # Trigger: the memory URL already names the active project root/prefix
@@ -722,9 +759,15 @@ async def _fetch_workspace_project_entries(
 
 async def _ensure_workspace_project_index(
     context: Optional[Context] = None,
+    *,
+    force_refresh: bool = False,
 ) -> WorkspaceProjectIndex:
-    """Build or load the session-local workspace/project lookup index."""
-    if context:
+    """Build or load the session-local workspace/project lookup index.
+
+    force_refresh bypasses the cached index and rebuilds from discovery —
+    used by resolve_workspace_project_identifier when a lookup misses (#956).
+    """
+    if context and not force_refresh:
         cached_raw = await context.get_state(_WORKSPACE_PROJECT_INDEX_STATE_KEY)
         cached_index = _workspace_project_index_from_state(cached_raw)
         if cached_index is not None:
@@ -865,7 +908,32 @@ async def resolve_workspace_project_identifier(
 ) -> WorkspaceProjectEntry:
     """Resolve a project by external_id (UUID), qualified name, or unqualified name."""
     index = await _ensure_workspace_project_index(context=context)
+    try:
+        return await _resolve_workspace_project_from_index(index, project, context)
+    except WorkspaceProjectLookupMiss:
+        # Trigger: the lookup missed the session-cached index.
+        # Why: a miss is exactly the signal the cache may be stale — projects
+        #   created out-of-band (CLI, a teammate in a shared workspace) post-date
+        #   the index built at session start (#956).
+        # Outcome: rebuild the index once and retry; a second miss is authoritative
+        #   and its error (with the refreshed project list) propagates.
+        logger.info(
+            f"Workspace project lookup missed for '{project}'; refreshing index and retrying"
+        )
+        refreshed = await _ensure_workspace_project_index(context=context, force_refresh=True)
+        return await _resolve_workspace_project_from_index(refreshed, project, context)
+
+
+async def _resolve_workspace_project_from_index(
+    index: WorkspaceProjectIndex,
+    project: str,
+    context: Optional[Context] = None,
+) -> WorkspaceProjectEntry:
+    """Resolve a project against one concrete index snapshot.
 
+    Raises WorkspaceProjectLookupMiss for absent projects (retryable via index
+    refresh) and plain ValueError for ambiguity, which a refresh cannot fix.
+    """
     # Fast path: direct lookup by external_id when the identifier is a UUID
     # Canonicalize via str(UUID(...)) so uppercase, brace-wrapped, or urn:uuid forms
     # all hash to the same lowercase-hyphenated key as the stored external_ids.
@@ -895,7 +963,7 @@ async def resolve_workspace_project_identifier(
                 failed_workspace.tenant_id == workspace.tenant_id
                 for failed_workspace in index.failed_workspaces
             ):
-                raise ValueError(
+                raise WorkspaceProjectLookupMiss(
                     f"Projects for workspace '{workspace.name}' ({workspace.slug}) "
                     "could not be loaded. Retry after workspace discovery recovers."
                 )
@@ -904,7 +972,7 @@ async def resolve_workspace_project_identifier(
                 for entry in index.entries
                 if entry.workspace.tenant_id == workspace.tenant_id
             )
-            raise ValueError(
+            raise WorkspaceProjectLookupMiss(
                 f"Project '{project_identifier}' was not found in workspace "
                 f"'{workspace.name}' ({workspace.slug}). Available projects: {available}"
             )
@@ -929,7 +997,7 @@ async def resolve_workspace_project_identifier(
                 "retry or use a qualified project from an indexed workspace."
             )
         available = ", ".join(entry.qualified_name for entry in index.entries)
-        raise ValueError(
+        raise WorkspaceProjectLookupMiss(
             f"Project '{project}' was not found in indexed cloud workspaces. "
             f"Available projects: {available}.{failed_note}"
         )

src/basic_memory/mcp/tools/build_context.py

@@ -54,7 +54,10 @@ def _format_entity_block(result: ContextResult) -> str:
         lines.append("")
         lines.append("### Relations")
         for rel in relation_items:
-            lines.append(f"- {rel.relation_type} [[{rel.to_entity}]]")
+            # Unresolved forward references have no resolved entity yet; fall back
+            # to the literal target text instead of rendering [[None]] (#955)
+            target = rel.to_entity or rel.to_name
+            lines.append(f"- {rel.relation_type} [[{target}]]")
 
     # --- Related entities (non-relation related results) ---
     related_entities: list[EntitySummary | ObservationSummary] = [

src/basic_memory/mcp/tools/project_management.py

@@ -477,10 +477,14 @@ async def _resolve_workspace_routing(
     if workspace is None:
         return None
 
-    explicit_cloud_routing = _explicit_routing() and not _force_local_mode()
+    forced_local = _explicit_routing() and _force_local_mode()
     config = ConfigManager().config
+    # Resolve whenever credentials make workspace discovery possible — not only
+    # under explicit --cloud. A workspace selector implies cloud routing
+    # (get_client routes it to the cloud proxy, #954), and the transport needs
+    # the tenant id in X-Workspace-ID, not a slug or display name.
     should_resolve_workspace = is_factory_mode() or (
-        explicit_cloud_routing and has_cloud_credentials(config)
+        has_cloud_credentials(config) and not forced_local
     )
     if not should_resolve_workspace:
         return workspace
@@ -521,8 +525,9 @@ async def create_memory_project(
         workspace: Optional cloud workspace selector to create the project in. Slug is
             preferred for AI callers, but tenant_id and unique name are also accepted.
             When omitted, the connection's default workspace is used. Discover values
-            via `list_workspaces`. In local mode the selector is passed through
-            without slug resolution.
+            via `list_workspaces`. A workspace selector implies cloud routing:
+            without cloud credentials the call fails fast instead of silently
+            creating a local project (#954).
         output_format: "text" returns the existing human-readable result text.
             "json" returns structured project creation metadata.
         context: Optional FastMCP context for progress/status logging.
@@ -660,8 +665,9 @@ async def delete_project(
         workspace: Optional cloud workspace selector to delete the project from.
             Slug is preferred for AI callers, but tenant_id and unique name are
             also accepted. When omitted, the connection's default workspace is
-            used. In local mode the selector is passed through without slug
-            resolution, matching create_memory_project behavior.
+            used. A workspace selector implies cloud routing: without cloud
+            credentials the call fails fast, matching create_memory_project
+            behavior (#954).
 
     Returns:
         Confirmation message about project deletion

src/basic_memory/repository/search_index_row.py

@@ -36,6 +36,7 @@ class SearchIndexRow:
     category: Optional[str] = None  # observations
     from_id: Optional[int] = None  # relations
     to_id: Optional[int] = None  # relations
+    to_name: Optional[str] = None  # relations: literal target text, set even when unresolved
     relation_type: Optional[str] = None  # relations
 
     # Matched chunk text from vector search (the actual content that matched the query)
@@ -94,6 +95,7 @@ def to_insert(self, serialize_json: bool = True):
             else self.metadata,
             "from_id": self.from_id,
             "to_id": self.to_id,
+            "to_name": self.to_name,
             "relation_type": self.relation_type,
             "entity_id": self.entity_id,
             "category": self.category,

src/basic_memory/schemas/memory.py

@@ -155,6 +155,9 @@ class RelationSummary(BaseModel):
     from_entity_id: Optional[int] = None
     from_entity_external_id: Optional[str] = None
     to_entity: Optional[str] = None
+    # Literal target text from the markdown; present even when the relation is
+    # an unresolved forward reference (to_entity is None until the target exists)
+    to_name: Optional[str] = None
     to_entity_id: Optional[int] = None
     to_entity_external_id: Optional[str] = None
     created_at: Annotated[