refactor: keep anthropic_content_blocks in-memory only (no state.db column)

Drop the hermes_state.py column + persistence plumbing from the salvaged
interleaved-thinking fix. The ordered-block channel covers the failure
window in-memory (turn replayed within the live conversation loop). A
session reloaded from disk after a crash falls back to reconstruction;
if that replay 400s, the thinking-signature recovery (#43667) strips
reasoning_details and retries — one degraded call in a rare resume path
instead of a schema column. Replaces the DB-roundtrip test with a
fallback-shape test.

hermes_state.py

@@ -488,7 +488,6 @@ CREATE TABLE IF NOT EXISTS messages (
     reasoning TEXT,
     reasoning_content TEXT,
     reasoning_details TEXT,
-    anthropic_content_blocks TEXT,
     codex_reasoning_items TEXT,
     codex_message_items TEXT,
     platform_message_id TEXT,
@@ -2241,7 +2240,6 @@ class SessionDB:
         reasoning: str = None,
         reasoning_content: str = None,
         reasoning_details: Any = None,
-        anthropic_content_blocks: Any = None,
         codex_reasoning_items: Any = None,
         codex_message_items: Any = None,
         platform_message_id: str = None,
@@ -2264,10 +2262,6 @@ class SessionDB:
             json.dumps(reasoning_details)
             if reasoning_details else None
         )
-        anthropic_content_blocks_json = (
-            json.dumps(anthropic_content_blocks)
-            if anthropic_content_blocks else None
-        )
         codex_items_json = (
             json.dumps(codex_reasoning_items)
             if codex_reasoning_items else None
@@ -2290,10 +2284,9 @@ class SessionDB:
             cursor = conn.execute(
                 """INSERT INTO messages (session_id, role, content, tool_call_id,
                    tool_calls, tool_name, timestamp, token_count, finish_reason,
-                   reasoning, reasoning_content, reasoning_details, anthropic_content_blocks,
+                   reasoning, reasoning_content, reasoning_details, codex_reasoning_items,
-                   codex_reasoning_items,
                    codex_message_items, platform_message_id, observed)
-                   VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
+                   VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
                 (
                     session_id,
                     role,
@@ -2307,7 +2300,6 @@ class SessionDB:
                     reasoning,
                     reasoning_content,
                     reasoning_details_json,
-                    anthropic_content_blocks_json,
                     codex_items_json,
                     codex_message_items_json,
                     platform_message_id,
@@ -2356,9 +2348,6 @@ class SessionDB:
                 role = msg.get("role", "unknown")
                 tool_calls = msg.get("tool_calls")
                 reasoning_details = msg.get("reasoning_details") if role == "assistant" else None
-                anthropic_content_blocks = (
-                    msg.get("anthropic_content_blocks") if role == "assistant" else None
-                )
                 codex_reasoning_items = (
                     msg.get("codex_reasoning_items") if role == "assistant" else None
                 )
@@ -2369,9 +2358,6 @@ class SessionDB:
                 reasoning_details_json = (
                     json.dumps(reasoning_details) if reasoning_details else None
                 )
-                anthropic_content_blocks_json = (
-                    json.dumps(anthropic_content_blocks) if anthropic_content_blocks else None
-                )
                 codex_items_json = (
                     json.dumps(codex_reasoning_items) if codex_reasoning_items else None
                 )
@@ -2388,10 +2374,9 @@ class SessionDB:
                 conn.execute(
                     """INSERT INTO messages (session_id, role, content, tool_call_id,
                        tool_calls, tool_name, timestamp, token_count, finish_reason,
-                       reasoning, reasoning_content, reasoning_details, anthropic_content_blocks,
+                       reasoning, reasoning_content, reasoning_details, codex_reasoning_items,
-                       codex_reasoning_items,
                        codex_message_items, platform_message_id, observed)
-                       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
+                       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)""",
                     (
                         session_id,
                         role,
@@ -2405,7 +2390,6 @@ class SessionDB:
                         msg.get("reasoning") if role == "assistant" else None,
                         msg.get("reasoning_content") if role == "assistant" else None,
                         reasoning_details_json,
-                        anthropic_content_blocks_json,
                         codex_items_json,
                         codex_message_items_json,
                         platform_msg_id,
@@ -2748,7 +2732,6 @@ class SessionDB:
             rows = self._conn.execute(
                 "SELECT role, content, tool_call_id, tool_calls, tool_name, "
                 "finish_reason, reasoning, reasoning_content, reasoning_details, "
-                "anthropic_content_blocks, "
                 "codex_reasoning_items, codex_message_items, platform_message_id, observed "
                 f"FROM messages WHERE session_id IN ({placeholders})"
                 f"{active_clause} ORDER BY id",
@@ -2796,12 +2779,6 @@ class SessionDB:
                     except (json.JSONDecodeError, TypeError):
                         logger.warning("Failed to deserialize reasoning_details, falling back to None")
                         msg["reasoning_details"] = None
-                if row["anthropic_content_blocks"]:
-                    try:
-                        msg["anthropic_content_blocks"] = json.loads(row["anthropic_content_blocks"])
-                    except (json.JSONDecodeError, TypeError):
-                        logger.warning("Failed to deserialize anthropic_content_blocks, falling back to None")
-                        msg["anthropic_content_blocks"] = None
                 if row["codex_reasoning_items"]:
                     try:
                         msg["codex_reasoning_items"] = json.loads(row["codex_reasoning_items"])

run_agent.py

@@ -1597,7 +1597,6 @@ class AIAgent:
                     reasoning=msg.get("reasoning") if role == "assistant" else None,
                     reasoning_content=msg.get("reasoning_content") if role == "assistant" else None,
                     reasoning_details=msg.get("reasoning_details") if role == "assistant" else None,
-                    anthropic_content_blocks=msg.get("anthropic_content_blocks") if role == "assistant" else None,
                     codex_reasoning_items=msg.get("codex_reasoning_items") if role == "assistant" else None,
                     codex_message_items=msg.get("codex_message_items") if role == "assistant" else None,
                 )

tests/agent/test_anthropic_thinking_block_order.py

@@ -176,59 +176,39 @@ class TestInterleavedThinkingBlockOrder:
             "in the latest assistant message cannot be modified'."
         )
 
-    def test_interleaved_order_survives_db_roundtrip(self, tmp_path):
+    def test_replay_falls_back_gracefully_without_ordered_blocks(self):
-        """The ordered-block channel must survive SQLite persistence + reload.
+        """Without the ordered-block channel, conversion must not crash.
 
-        This is the exact path that fails after a gateway crash: the session
+        The channel is intentionally NOT persisted to state.db (in-memory
-        is reloaded from state.db via get_messages_as_conversation, then
+        only): a session reloaded from disk after a crash loses the field
-        replayed. If the verbatim block list is dropped or not deserialized,
+        and falls back to reconstruction. That replay may take one HTTP 400,
-        the reconstruction reorders signed thinking blocks -> HTTP 400.
+        which the thinking-signature recovery (#43667) absorbs by stripping
+        reasoning_details and retrying. This test pins the fallback shape:
+        conversion still produces a valid assistant message from the
+        parallel reasoning_details + tool_calls fields.
         """
-        import hermes_state
-
         response = _interleaved_response()
-        original_order = _original_block_order(response)
-
         transport = get_transport("anthropic_messages")
         normalized = transport.normalize_response(response)
         assistant_msg = _stored_assistant_message(normalized)
+        # Simulate a disk reload: the in-memory-only channel is gone.
+        assistant_msg.pop("anthropic_content_blocks", None)
 
-        db = hermes_state.SessionDB(tmp_path / "state.db")
+        messages = [
-        sid = "sess_roundtrip"
+            assistant_msg,
-        db.create_session(sid, source="test")
+            {"role": "tool", "tool_call_id": "toolu_1", "content": "a ok"},
-        db.append_message(
+            {"role": "tool", "tool_call_id": "toolu_2", "content": "b ok"},
-            session_id=sid,
+        ]
-            role="assistant",
-            content=assistant_msg["content"],
-            tool_calls=assistant_msg["tool_calls"],
-            reasoning_details=assistant_msg.get("reasoning_details"),
-            anthropic_content_blocks=assistant_msg.get("anthropic_content_blocks"),
-        )
-        db.append_message(session_id=sid, role="tool", tool_call_id="toolu_1", content="a ok")
-        db.append_message(session_id=sid, role="tool", tool_call_id="toolu_2", content="b ok")
-
-        # Reload via the conversation-restore path used on resume / crash recovery.
-        loaded = db.get_messages_as_conversation(sid)
-        reloaded_assistant = [m for m in loaded if m.get("role") == "assistant"]
-        assert reloaded_assistant, "no assistant message after DB reload"
-        # The ordered-block channel must come back as a deserialized list.
-        blocks = reloaded_assistant[0].get("anthropic_content_blocks")
-        assert isinstance(blocks, list) and len(blocks) == 4, (
-            "anthropic_content_blocks was not persisted/deserialized correctly"
-        )
-
         _system, anthropic_messages = convert_messages_to_anthropic(
-            loaded, base_url=None, model="claude-opus-4-8",
+            messages, base_url=None, model="claude-opus-4-8",
         )
         assistant_out = [m for m in anthropic_messages if m.get("role") == "assistant"]
         assert assistant_out, "no assistant message in converted output"
-        replayed_order = _replayed_block_order(assistant_out[-1]["content"])
+        content = assistant_out[-1]["content"]
-
+        assert isinstance(content, list) and content, "fallback produced empty content"
-        assert replayed_order == original_order, (
+        # Reconstruction keeps both tool_use blocks (answered by results).
-            "Interleaved block order was lost across the SQLite round-trip.\n"
+        tool_ids = [b.get("id") for b in content if isinstance(b, dict) and b.get("type") == "tool_use"]
-            f"  original: {original_order}\n"
+        assert set(tool_ids) == {"toolu_1", "toolu_2"}
-            f"  replayed: {replayed_order}"
-        )
 
 
 class TestInterleavedReplayCredentialRedaction: