fix(state): read-only session search silently returned no results; add stateless memory/session_search MCP shims

hermes_cli/codex_runtime_plugin_migration.py

@@ -599,6 +599,12 @@ def _build_hermes_tools_mcp_entry() -> dict:
     }
     if env:
         out["env"] = env
+    # Session-id delivery: a NAME, never a value (burn-in rule — a literal id
+    # written at migrate time would be frozen into config.toml and name the
+    # wrong session forever). Codex snapshots the names listed here from its
+    # own process env at MCP spawn, so the shim's own-lineage exclusion
+    # follows the ACTIVE session (#26604 keep_open resolution, option (a)).
+    out["env_vars"] = ["HERMES_SESSION_ID"]
     # Generous timeouts — browser_navigate or delegate_task can take a
     # while; we don't want codex's MCP client to give up too early.
     out["startup_timeout_sec"] = 30.0

hermes_state.py

@@ -50,6 +50,7 @@
     _COMPRESSION_CHILD_SQL,
     _FTS_CJK_TRIGGERS,
     _FTS_TRIGGERS,
+    _fts_object_missing,
     _LISTABLE_CHILD_SQL,
     _PREVIEW_RAW_SELECT,
     _ephemeral_child_sql,
@@ -4205,6 +4206,19 @@ def _do(conn):
             )
         self._execute_write(_do)
 
+    def update_claude_sdk_session_id(
+        self, session_id: str, sdk_session_id: Optional[str]
+    ) -> None:
+        """Persist (or clear, with None) the claude-agent-sdk session id used
+        to resume the SDK conversation across gateway restarts and
+        agent-cache eviction (#25267 continuity)."""
+        def _do(conn):
+            conn.execute(
+                "UPDATE sessions SET claude_sdk_session_id = ? WHERE id = ?",
+                (sdk_session_id, session_id),
+            )
+        self._execute_write(_do)
+
     def update_system_prompt(
         self, session_id: str, system_prompt: Optional[str]
     ) -> None:

hermes_state_common.py

@@ -166,6 +166,13 @@ def _sql_session_last_active_by_id(session_id_expr: str) -> str:
 FTS_STORAGE_VERSION = 1
 
 
+def _fts_object_missing(exc: BaseException) -> bool:
+    """True when an FTS probe failure means the table/module is ABSENT
+    (disable search) rather than transiently unavailable (keep it on)."""
+    msg = str(exc).lower()
+    return "no such table" in msg or "no such module" in msg
+
+
 # Cap on user-controlled FTS5 query input before regex/sanitizer processing.
 # Search queries do not need to be arbitrarily large, and bounding them keeps
 # sanitizer/runtime behavior predictable under adversarial input.
@@ -206,6 +213,7 @@ def _sql_session_last_active_by_id(session_id_expr: str) -> str:
     model TEXT,
     model_config TEXT,
     system_prompt TEXT,
+    claude_sdk_session_id TEXT,
     system_prompt_hash TEXT,
     parent_session_id TEXT,
     started_at REAL NOT NULL,

hermes_state_search.py

@@ -1299,8 +1299,12 @@ def _run_trigram_search(
         with self._read_ctx() as conn:
             try:
                 tri_cursor = conn.execute(tri_sql, tri_params)
-            except sqlite3.OperationalError:
+            except sqlite3.OperationalError as exc:
                 # Query failed at runtime — let the caller fall back.
+                # A missing tokenizer is permanent: log it once instead of
+                # silently retrying on every CJK query.
+                if self._is_trigram_unavailable_error(exc):
+                    self._warn_trigram_unavailable(exc)
                 return None
             return [dict(row) for row in tri_cursor.fetchall()]