coze_lab 0.1.53 → 0.1.55

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "coze_lab",
-  "version": "0.1.53",
+  "version": "0.1.55",
   "description": "Configure local AI agents (Claude Code, Codex, OpenClaw) to report traces to CozeLoop",
   "keywords": [
     "cozeloop",

package/scripts/claude-code/cozeloop_hook.py

@@ -230,6 +230,7 @@ def _make_finish_event_processor(upload_events: Optional[List[str]] = None):
     def _processor(info):
         try:
             if not getattr(info, "is_event_fail", False):
+                hook_log("upload success")
                 return
             detail = getattr(info, "detail_msg", "") or ""
             if upload_events is not None:
@@ -393,6 +394,73 @@ def format_content(content: Any, truncate: int = 4096) -> str:
     return str(content)[:truncate]
 
 
+def _normalize_assistant_content(content: Any) -> Dict[str, Any]:
+    """Split supported Claude content without retaining protected payloads."""
+    result = {
+        "text": [],
+        "reasoning": [],
+        "tool_calls": [],
+        "parts": [],
+        "protected": {"source": "", "count": 0, "bytes": 0},
+        "unknown_types": {},
+    }
+    if isinstance(content, str):
+        if content:
+            result["text"].append(content)
+        return result
+    if not isinstance(content, list):
+        return result
+    for item in content:
+        if not isinstance(item, dict):
+            continue
+        item_type = item.get("type", "")
+        if item_type == "text":
+            text = item.get("text", "")
+            if text:
+                result["text"].append(text)
+                result["parts"].append(
+                    ModelMessagePart(type=ModelMessagePartType.TEXT, text=text)
+                )
+        elif item_type == "thinking":
+            thinking = item.get("thinking", "")
+            if thinking:
+                result["reasoning"].append(thinking)
+        elif item_type == "redacted_thinking":
+            data = item.get("data", "")
+            if isinstance(data, str) and data:
+                result["protected"]["source"] = "redacted_thinking"
+                result["protected"]["count"] += 1
+                result["protected"]["bytes"] += len(data.encode("utf-8"))
+        elif item_type == "tool_use":
+            result["tool_calls"].append(ModelToolCall(
+                id=item.get("id", ""),
+                type="function",
+                function=ModelToolCallFunction(
+                    name=item.get("name", ""),
+                    arguments=(
+                        json.dumps(item.get("input", {}), ensure_ascii=False)
+                        if isinstance(item.get("input"), dict)
+                        else str(item.get("input", ""))
+                    ),
+                ),
+            ))
+        else:
+            key = item_type or "unknown"
+            result["unknown_types"][key] = result["unknown_types"].get(key, 0) + 1
+    return result
+
+
+def _set_protected_reasoning_tags(span, normalized: Dict[str, Any]):
+    protected = normalized.get("protected") or {}
+    if not protected.get("count"):
+        return
+    span.set_tags({
+        "protected_reasoning_source": protected.get("source", ""),
+        "protected_reasoning_count": protected.get("count", 0),
+        "protected_reasoning_bytes": protected.get("bytes", 0),
+    })
+
+
 # --- Message Parsing and Grouping ---
 
 def is_tool_result_message(msg: Dict[str, Any]) -> bool:
@@ -1080,44 +1148,11 @@ def send_turns_to_cozeloop(turns: List[Dict[str, Any]], session_id: str, history
                                     tool_choice=ModelToolChoice(type="", function=None)
                                 ))
 
-                                # Build output: text -> parts, tool_use -> tool_calls, thinking -> reasoning_content
-                                text_parts = []
-                                tool_call_list = []
-                                parts_list = []
-                                thinking_parts = []
-                                if isinstance(raw_content, list):
-                                    for item in raw_content:
-                                        if not isinstance(item, dict):
-                                            continue
-                                        item_type = item.get("type", "")
-                                        if item_type == "text":
-                                            text = item.get("text", "")
-                                            if text:
-                                                text_parts.append(text)
-                                                parts_list.append(ModelMessagePart(type=ModelMessagePartType.TEXT, text=text))
-                                        elif item_type == "thinking":
-                                            thinking = item.get("thinking", "")
-                                            if thinking:
-                                                thinking_parts.append(thinking)
-                                        elif item_type == "redacted_thinking":
-                                            pass  # encrypted, cannot extract
-                                        elif item_type == "tool_use":
-                                            tool_call_list.append(ModelToolCall(
-                                                id=item.get("id", ""),
-                                                type="function",
-                                                function=ModelToolCallFunction(
-                                                    name=item.get("name", ""),
-                                                    arguments=json.dumps(item.get("input", {}), ensure_ascii=False) if isinstance(item.get("input"), dict) else str(item.get("input", ""))
-                                                )
-                                            ))
-                                        else:
-                                            parts_list.append(ModelMessagePart(
-                                                type=ModelMessagePartType.TEXT,
-                                                text=json.dumps(item, ensure_ascii=False)[:4096]
-                                            ))
-                                elif isinstance(raw_content, str) and raw_content:
-                                    text_parts.append(raw_content)
-
+                                normalized = _normalize_assistant_content(raw_content)
+                                text_parts = normalized["text"]
+                                thinking_parts = normalized["reasoning"]
+                                tool_call_list = normalized["tool_calls"]
+                                parts_list = normalized["parts"]
                                 content_text = "\n".join(text_parts) if text_parts else ""
                                 reasoning_text = "\n".join(thinking_parts) if thinking_parts else ""
                                 finish_reason = "tool_calls" if tool_call_list else "stop"
@@ -1138,6 +1173,7 @@ def send_turns_to_cozeloop(turns: List[Dict[str, Any]], session_id: str, history
                                 )
 
                                 model_span.set_output(ModelOutput(choices=[output_choice]))
+                                _set_protected_reasoning_tags(model_span, normalized)
 
                                 # Set token usage for this specific model call
                                 usage = assistant_message_obj.get("usage", {})
@@ -1243,42 +1279,11 @@ def send_turns_to_cozeloop(turns: List[Dict[str, Any]], session_id: str, history
                                                     tool_choice=ModelToolChoice(type="", function=None)
                                                 ))
 
-                                                # Build output for sub-agent model call
-                                                sub_text_parts = []
-                                                sub_tc_list = []
-                                                sub_parts_list = []
-                                                sub_thinking_parts = []
-                                                if isinstance(sub_content, list):
-                                                    for item in sub_content:
-                                                        if not isinstance(item, dict):
-                                                            continue
-                                                        item_type = item.get("type", "")
-                                                        if item_type == "text":
-                                                            t = item.get("text", "")
-                                                            if t:
-                                                                sub_text_parts.append(t)
-                                                                sub_parts_list.append(ModelMessagePart(type=ModelMessagePartType.TEXT, text=t))
-                                                        elif item_type == "thinking":
-                                                            t = item.get("thinking", "")
-                                                            if t:
-                                                                sub_thinking_parts.append(t)
-                                                        elif item_type == "redacted_thinking":
-                                                            pass
-                                                        elif item_type == "tool_use":
-                                                            sub_tc_list.append(ModelToolCall(
-                                                                id=item.get("id", ""),
-                                                                type="function",
-                                                                function=ModelToolCallFunction(
-                                                                    name=item.get("name", ""),
-                                                                    arguments=json.dumps(item.get("input", {}), ensure_ascii=False) if isinstance(item.get("input"), dict) else str(item.get("input", ""))
-                                                                )
-                                                            ))
-                                                        else:
-                                                            sub_parts_list.append(ModelMessagePart(
-                                                                type=ModelMessagePartType.TEXT,
-                                                                text=json.dumps(item, ensure_ascii=False)[:4096]
-                                                            ))
-
+                                                sub_normalized = _normalize_assistant_content(sub_content)
+                                                sub_text_parts = sub_normalized["text"]
+                                                sub_tc_list = sub_normalized["tool_calls"]
+                                                sub_parts_list = sub_normalized["parts"]
+                                                sub_thinking_parts = sub_normalized["reasoning"]
                                                 sub_content_text = "\n".join(sub_text_parts) if sub_text_parts else ""
                                                 sub_reasoning_text = "\n".join(sub_thinking_parts) if sub_thinking_parts else ""
                                                 sub_finish = "tool_calls" if sub_tc_list else "stop"
@@ -1296,6 +1301,7 @@ def send_turns_to_cozeloop(turns: List[Dict[str, Any]], session_id: str, history
                                                         metadata={}
                                                     )
                                                 )]))
+                                                _set_protected_reasoning_tags(sub_model_span, sub_normalized)
 
                                                 # Distribute tokens evenly; remainder goes to last step
                                                 step_in = per_step_in + (remainder_in if sk == n_model_steps - 1 else 0)
@@ -1435,7 +1441,7 @@ def send_turns_to_cozeloop(turns: List[Dict[str, Any]], session_id: str, history
 def _rt_state_key_header():
     return "rt_root_header"
 
-def _flatten_completed_steps(turns: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+def _flatten_completed_steps(turns: List[Dict[str, Any]], is_terminal: bool = False) -> List[Dict[str, Any]]:
     """把 turns 线性展开成 step 列表；只取【已完成】的 step。
 
     一个 step（assistant 一次模型调用 + 其触发的 tool_results）视为已完成的判定：
@@ -1454,7 +1460,7 @@ def _flatten_completed_steps(turns: List[Dict[str, Any]]) -> List[Dict[str, Any]
     for idx, item in enumerate(flat):
         is_global_last = (idx == len(flat) - 1)
         has_results = bool(item["step"].get("tool_results"))
-        if (not is_global_last) or has_results:
+        if is_terminal or (not is_global_last) or has_results:
             completed.append(item)
     return completed, len(flat)
 
@@ -1546,7 +1552,7 @@ def send_steps_realtime(turns, session_id, history_turns, state, coze_tags_overr
             debug_log(f"[rt] root created+finished trace_id={getattr(root_ctx,'trace_id','?')}")
 
         # ---- 发新完成的 step ----
-        completed, total_steps = _flatten_completed_steps(turns)
+        completed, total_steps = _flatten_completed_steps(turns, is_terminal=is_terminal)
         last_global = state.get("rt_last_global_step", 0)
         new_items = completed[last_global:]
         # 批量节流：非终态时，距上次上报不足 REALTIME_BATCH_INTERVAL 秒则本次不发，
@@ -1581,16 +1587,33 @@ def send_steps_realtime(turns, session_id, history_turns, state, coze_tags_overr
             if step_start_dt is not None and step_end_dt is not None:
                 _set_finish_time_safe(mspan, step_end_dt)
             raw_content = amo.get("content", [])
-            text_parts = []
-            if isinstance(raw_content, list):
-                for it in raw_content:
-                    if isinstance(it, dict) and it.get("type") == "text" and it.get("text"):
-                        text_parts.append(it["text"])
-            elif isinstance(raw_content, str):
-                text_parts.append(raw_content)
-            mspan.set_input(format_content((assistant_msg.get("message", {}) or {}).get("content")))
-            if text_parts:
-                mspan.set_output("\n".join(text_parts))
+            normalized = _normalize_assistant_content(raw_content)
+            turn_user = item["turn"].get("user_message", {}).get("message", {})
+            turn_user_content = turn_user.get("content") if turn_user else None
+            input_messages = list(_build_history_messages(history_turns or []))
+            if not is_empty_content(turn_user_content):
+                input_messages.append(_make_message("user", format_content(turn_user_content)))
+            mspan.set_input(ModelInput(
+                messages=input_messages,
+                tools=[],
+                tool_choice=ModelToolChoice(type="", function=None),
+            ))
+            output_choice = ModelChoice(
+                finish_reason="tool_calls" if normalized["tool_calls"] else "stop",
+                index=0,
+                message=ModelMessage(
+                    role="assistant",
+                    content="\n".join(normalized["text"]),
+                    reasoning_content="\n".join(normalized["reasoning"]),
+                    parts=normalized["parts"],
+                    name="",
+                    tool_calls=normalized["tool_calls"],
+                    tool_call_id="",
+                    metadata={},
+                ),
+            )
+            mspan.set_output(ModelOutput(choices=[output_choice]))
+            _set_protected_reasoning_tags(mspan, normalized)
             usage = amo.get("usage", {})
             it_tok = _usage_int(usage, "input_tokens") + _usage_int(usage, "cache_creation_input_tokens") + _usage_int(usage, "cache_read_input_tokens")
             if it_tok > 0:

package/scripts/codex/cozeloop_hook.py

@@ -575,6 +575,48 @@ def extract_assistant_text(payload: Dict[str, Any]) -> str:
     return "\n".join(parts)
 
 
+def _protected_reasoning_meta(source: str, payload: Any) -> Dict[str, Any]:
+    """Return safe metadata for protected reasoning without retaining payload."""
+    if not isinstance(payload, str) or not payload:
+        return {}
+    return {
+        "source": source,
+        "count": 1,
+        "bytes": len(payload.encode("utf-8")),
+    }
+
+
+def extract_reasoning_record(payload: Dict[str, Any]) -> Dict[str, Any]:
+    """Normalize a Codex reasoning item, excluding encrypted content itself."""
+    parts = [
+        item.get("text", "")
+        for item in payload.get("summary", [])
+        if isinstance(item, dict)
+        and item.get("type") == "summary_text"
+        and item.get("text")
+    ]
+    return {
+        "content": "",
+        "reasoning": "\n".join(parts),
+        "protected": _protected_reasoning_meta(
+            "encrypted_content", payload.get("encrypted_content")
+        ),
+        "_ts": payload.get("_ts"),
+        "_line_number": payload.get("_line_number", -1),
+        "source": "response_item.reasoning",
+    }
+
+
+def _json_text(value: Any) -> str:
+    """Render structured tool values deterministically for trace display."""
+    if isinstance(value, str):
+        return value
+    try:
+        return json.dumps(value, ensure_ascii=False)
+    except (TypeError, ValueError):
+        return str(value)
+
+
 def extract_message_content_text(payload: Dict[str, Any]) -> str:
     """Extract all text content from a message payload regardless of role."""
     parts = []
@@ -669,6 +711,10 @@ def _turn_timestamps(turn):
         dt = _parse_ts(item)
         if dt:
             values.append(dt)
+    for item in turn.get("model_records", []):
+        dt = _parse_ts(item)
+        if dt:
+            values.append(dt)
     for sc in turn.get("subagent_calls", []):
         for key in ("_start_ts", "_end_ts"):
             dt = _parse_ts_value(sc.get(key))
@@ -695,13 +741,45 @@ def _assistant_bounds(turn):
     return min(values), max(values)
 
 
-def extract_token_events(entries: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
-    """扫 rollout event_msg/token_count 事件, 与 turn 完全解耦.
+def extract_model_invocations(turns: List[Dict[str, Any]],
+                              entries: List[Dict[str, Any]],
+                              is_terminal: bool = False) -> List[Dict[str, Any]]:
+    """Group visible model output, tool calls, and usage into model invocations.
 
-    每条事件独立采集, 依赖 _line_number 做全局水位线幂等. 只收 last_token_usage 非空的事件
-    (空的是心跳/占位, 不代表真实模型调用). total_token_usage 一并保留供 tag 上报.
+    Codex emits one non-empty ``token_count.last_token_usage`` after each model
+    invocation. It is the only reliable boundary shared by tool and no-tool
+    calls, so records since the previous boundary are reported together. The
+    cumulative ``total_token_usage`` is retained as metadata only.
     """
     events: List[Dict[str, Any]] = []
+    tool_results: Dict[str, Dict[str, Any]] = {}
+
+    for turn in turns:
+        for result in turn.get("tool_results", []):
+            call_id = result.get("call_id")
+            if call_id:
+                tool_results[call_id] = result
+        for record in turn.get("model_records", []):
+            events.append({
+                "kind": "record",
+                "line": record.get("_line_number", -1),
+                "timestamp": record.get("_ts"),
+                "record": record,
+                "turn": turn,
+            })
+        for tool_call in turn.get("tool_calls", []):
+            events.append({
+                "kind": "tool",
+                "line": tool_call.get("_line_number", -1),
+                "timestamp": tool_call.get("_ts"),
+                "tool_call": tool_call,
+                "turn": turn,
+            })
+
+    turn_starts = sorted(
+        ((turn.get("start_line", -1), turn) for turn in turns),
+        key=lambda item: item[0],
+    )
     for entry in entries:
         if entry.get("type") != "event_msg":
             continue
@@ -712,13 +790,114 @@ def extract_token_events(entries: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
         last_usage = info.get("last_token_usage")
         if not isinstance(last_usage, dict) or not last_usage:
             continue
+        line = entry.get("_line_number", 0)
+        boundary_turn = None
+        for start_line, turn in turn_starts:
+            if start_line > line:
+                break
+            boundary_turn = turn
         events.append({
-            "line": entry.get("_line_number", 0),
+            "kind": "usage",
+            "line": line,
             "timestamp": entry.get("timestamp"),
             "usage": last_usage,
             "total_usage": info.get("total_token_usage") or {},
+            "turn": boundary_turn,
         })
-    return events
+
+    events.sort(key=lambda item: (item.get("line", -1),
+                                  {"record": 0, "tool": 1, "usage": 2}.get(item.get("kind"), 3)))
+    invocations: List[Dict[str, Any]] = []
+    current: Dict[str, Any] = {
+        "content_parts": [],
+        "reasoning_parts": [],
+        "protected": {"source": "", "count": 0, "bytes": 0},
+        "tool_calls": [],
+        "turn": None,
+        "start_timestamp": None,
+        "end_timestamp": None,
+        "line": -1,
+        "usage": {},
+        "total_usage": {},
+    }
+    seen = {"content": set(), "reasoning": set()}
+
+    def append_text(category: str, value: Any):
+        if not isinstance(value, str):
+            return
+        text = value.strip()
+        if not text or text in seen[category]:
+            return
+        seen[category].add(text)
+        current[f"{category}_parts"].append(text)
+
+    def finish_current(boundary_line: int, boundary_ts: Optional[str],
+                       usage: Optional[Dict[str, Any]] = None,
+                       total_usage: Optional[Dict[str, Any]] = None):
+        has_payload = (
+            current["content_parts"] or current["reasoning_parts"]
+            or current["tool_calls"] or current["protected"].get("count")
+            or usage
+        )
+        if not has_payload:
+            return
+        current["line"] = boundary_line
+        current["end_timestamp"] = boundary_ts or current["end_timestamp"]
+        current["usage"] = usage or {}
+        current["total_usage"] = total_usage or {}
+        current["content"] = "\n".join(current.pop("content_parts"))
+        current["reasoning"] = "\n".join(current.pop("reasoning_parts"))
+        for tool_call in current["tool_calls"]:
+            tool_call["result"] = tool_results.get(tool_call.get("call_id"))
+        invocations.append(dict(current))
+
+    for event in events:
+        kind = event["kind"]
+        if kind == "usage":
+            if current["turn"] is None:
+                current["turn"] = event.get("turn")
+            if current["start_timestamp"] is None:
+                current["start_timestamp"] = event.get("timestamp")
+            finish_current(event["line"], event.get("timestamp"),
+                           event.get("usage"), event.get("total_usage"))
+            current = {
+                "content_parts": [],
+                "reasoning_parts": [],
+                "protected": {"source": "", "count": 0, "bytes": 0},
+                "tool_calls": [],
+                "turn": None,
+                "start_timestamp": None,
+                "end_timestamp": None,
+                "line": -1,
+                "usage": {},
+                "total_usage": {},
+            }
+            seen = {"content": set(), "reasoning": set()}
+            continue
+
+        current["turn"] = current["turn"] or event.get("turn")
+        current["start_timestamp"] = current["start_timestamp"] or event.get("timestamp")
+        current["end_timestamp"] = event.get("timestamp") or current["end_timestamp"]
+        current["line"] = max(current["line"], event.get("line", -1))
+        if kind == "record":
+            record = event["record"]
+            append_text("content", record.get("content"))
+            append_text("reasoning", record.get("reasoning"))
+            protected = record.get("protected") or {}
+            if protected.get("count"):
+                current["protected"]["source"] = protected.get("source", "")
+                current["protected"]["count"] += protected.get("count", 0)
+                current["protected"]["bytes"] += protected.get("bytes", 0)
+        elif kind == "tool":
+            current["tool_calls"].append(dict(event["tool_call"]))
+
+    if is_terminal and (
+        current["content_parts"] or current["reasoning_parts"]
+        or current["tool_calls"] or current["protected"].get("count")
+    ):
+        finish_current(current["line"], current["end_timestamp"])
+
+    return invocations
 
 
 def group_messages_into_turns(entries: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
@@ -758,6 +937,7 @@ def group_messages_into_turns(entries: List[Dict[str, Any]]) -> List[Dict[str, A
                     "user_message": None,
                     "user_message_text": "",
                     "assistant_messages": [],
+                    "model_records": [],
                     "tool_calls": [],
                     "tool_results": [],
                     "input_messages": [],
@@ -775,6 +955,17 @@ def group_messages_into_turns(entries: List[Dict[str, Any]]) -> List[Dict[str, A
                 if current_turn is not None:
                     info = payload.get("info") or {}
                     current_turn["token_usage"] = info.get("last_token_usage", {})
+            elif msg_type in ("agent_reasoning", "agent_message") and current_turn is not None:
+                text = payload.get("text") if msg_type == "agent_reasoning" else payload.get("message")
+                if isinstance(text, str) and text:
+                    current_turn["model_records"].append({
+                        "content": text if msg_type == "agent_message" else "",
+                        "reasoning": text if msg_type == "agent_reasoning" else "",
+                        "protected": {},
+                        "_ts": payload.get("_ts"),
+                        "_line_number": entry.get("_line_number", -1),
+                        "source": f"event_msg.{msg_type}",
+                    })
             continue
 
         # --- Content items ---
@@ -795,6 +986,14 @@ def group_messages_into_turns(entries: List[Dict[str, Any]]) -> List[Dict[str, A
                 elif role == "assistant":
                     if current_turn is not None:
                         current_turn["assistant_messages"].append(payload)
+                        current_turn["model_records"].append({
+                            "content": extract_assistant_text(payload),
+                            "reasoning": "",
+                            "protected": {},
+                            "_ts": payload.get("_ts"),
+                            "_line_number": entry.get("_line_number", -1),
+                            "source": "response_item.message",
+                        })
                 elif role in ("developer", "system"):
                     if current_turn is not None:
                         current_turn["input_messages"].append({
@@ -810,16 +1009,25 @@ def group_messages_into_turns(entries: List[Dict[str, Any]]) -> List[Dict[str, A
                                 "content": text,
                             })
 
-            elif item_type == "function_call":
+            elif item_type == "reasoning":
+                if current_turn is not None:
+                    record = extract_reasoning_record(payload)
+                    record["_line_number"] = entry.get("_line_number", -1)
+                    current_turn["model_records"].append(record)
+
+            elif item_type in ("function_call", "custom_tool_call"):
                 if current_turn is None:
                     continue
                 call_id = payload.get("call_id")
                 name = payload.get("name", "")
-                args_raw = payload.get("arguments", "{}")
-                try:
-                    args = json.loads(args_raw)
-                except (json.JSONDecodeError, TypeError):
-                    args = {"_raw": args_raw}
+                if item_type == "custom_tool_call":
+                    args = payload.get("input", "")
+                else:
+                    args_raw = payload.get("arguments", "{}")
+                    try:
+                        args = json.loads(args_raw)
+                    except (json.JSONDecodeError, TypeError):
+                        args = {"_raw": args_raw}
 
                 if name == "spawn_agent":
                     subagent_call = {
@@ -846,11 +1054,13 @@ def group_messages_into_turns(entries: List[Dict[str, Any]]) -> List[Dict[str, A
                         "call_id": call_id,
                         "name": name,
                         "input": args,
+                        "custom": item_type == "custom_tool_call",
                         "_ts": payload.get("_ts"),
+                        "_line_number": entry.get("_line_number", -1),
                     })
                     pending_calls[call_id] = {"kind": "tool", "_start_ts": payload.get("_ts")}
 
-            elif item_type == "function_call_output":
+            elif item_type in ("function_call_output", "custom_tool_call_output"):
                 if current_turn is None:
                     continue
                 call_id = payload.get("call_id")
@@ -896,10 +1106,35 @@ def group_messages_into_turns(entries: List[Dict[str, Any]]) -> List[Dict[str, A
     if current_turn is not None:
         turns.append(current_turn)
 
+    # Codex may duplicate visible commentary/reasoning in response_item and
+    # event_msg. Prefer response_item, then retain event-only fallbacks.
+    for turn in turns:
+        seen = set()
+        deduped = []
+        records = sorted(
+            turn.get("model_records", []),
+            key=lambda record: (
+                1 if str(record.get("source", "")).startswith("event_msg.") else 0,
+                record.get("_line_number", -1),
+            ),
+        )
+        for record in records:
+            category = "reasoning" if record.get("reasoning") else "content"
+            text = record.get(category, "").strip()
+            key = (category, text)
+            if text and key in seen:
+                continue
+            if text:
+                seen.add(key)
+            deduped.append(record)
+        turn["model_records"] = sorted(
+            deduped, key=lambda record: record.get("_line_number", -1)
+        )
+
     # Drop turns with no user input and no assistant response
     turns = [
         t for t in turns
-        if t["user_message"] is not None or t["assistant_messages"]
+        if t["user_message"] is not None or t["assistant_messages"] or t.get("model_records")
     ]
 
     return turns
@@ -1380,14 +1615,15 @@ def send_turns_to_cozeloop(turns: List[Dict[str, Any]], session_id: str, model_n
     return ctx
 
 
-# --- 实时 step 级上报（方案 B：每个 tool_call 结束即可见）---------------------
+# --- 实时 model invocation 上报 -------------------------------------------
 #
 # 与 claude-code 同架构：一个会话 = 一条稳定 trace。首次建 root(codex_request)立即 finish 并存
-# to_header() 到 state["rt_root_header"]；之后每次触发把【新完成的 tool_call】(有 result 的)作为
-# tool span 挂到固定 root、当场 flush。codex 的实时单元是 tool_call（turn 内平铺，按 call_id 关联
-# result）。全局水位线 state["rt_last_tool"] 跟踪已发的 tool_call 数。
+# to_header() 到 state["rt_root_header"]；之后按 token_count 边界把模型可见内容、推理、工具调用和
+# last_token_usage 合成 model_call span，工具执行作为其子 span。全局行号水位线
+# state["rt_last_invocation_line"] 保证重复 hook 不重发。
 
-def send_steps_realtime(turns, session_id, state, model_name="codex", coze_tags_override=None, is_terminal=False, token_events=None):
+def send_steps_realtime(turns, session_id, state, model_name="codex", coze_tags_override=None,
+                        is_terminal=False, model_invocations=None):
     import cozeloop
     from cozeloop.spec.tracespec import Runtime
     token = get_fresh_token()
@@ -1439,7 +1675,8 @@ def send_steps_realtime(turns, session_id, state, model_name="codex", coze_tags_
                 if turn.get("user_message_text"):
                     first_user_text = turn["user_message_text"]
                     break
-            root_start_dt = _parse_ts_value(turns[0].get("tool_calls", [{}])[0].get("_ts")) if turns and turns[0].get("tool_calls") else None
+            all_timestamps = [dt for turn in turns for dt in _turn_timestamps(turn)]
+            root_start_dt = min(all_timestamps) if all_timestamps else None
             root_span = client.start_span(name="codex_request", span_type="main",
                                           start_time=root_start_dt, start_new_trace=True)
             root_span.set_runtime(Runtime(library="codex-cli"))
@@ -1460,120 +1697,162 @@ def send_steps_realtime(turns, session_id, state, model_name="codex", coze_tags_
                 return None
             hook_log(f"[rt] root created trace_id={getattr(root_ctx,'trace_id','?')}")
 
-        # 线性展开所有 tool_call（带 result 的才算完成），按全局序增量发
-        flat = []
-        for ti, turn in enumerate(turns):
-            results = {r.get("call_id"): r for r in turn.get("tool_results", [])}
-            for tc in turn.get("tool_calls", []):
-                cid = tc.get("call_id")
-                flat.append({"tc": tc, "result": results.get(cid), "turn_index": ti})
-        # 只发已完成（有 result）的，且全局最后一个若无 result 则留到下次
-        completed = []
-        for idx, item in enumerate(flat):
-            is_last = (idx == len(flat) - 1)
-            if item["result"] is not None or (not is_last):
-                completed.append(item)
-        last_tool = state.get("rt_last_tool", 0)
-        new_items = completed[last_tool:]
-        # 批量节流：非终态时距上次上报<5s 则本次不发，攒到下次一起发。终态不节流。
-        if new_items and not is_terminal:
+        # One atomic model span per Codex invocation. Upgrade existing sessions
+        # without replaying events already emitted by the pre-0.1.55 watermarks.
+        legacy_line = max(state.get("rt_last_model_line", -1),
+                          state.get("rt_last_token_line", -1))
+        last_invocation_line = state.get("rt_last_invocation_line", legacy_line)
+        pending_invocations = [
+            invocation for invocation in (model_invocations or [])
+            if invocation.get("line", -1) > last_invocation_line
+        ]
+        if pending_invocations and not is_terminal:
             since = time.time() - state.get("rt_last_upload_ts", 0)
             if since < REALTIME_BATCH_INTERVAL:
-                hook_log(f"[rt] batch throttle {since:.1f}s<{REALTIME_BATCH_INTERVAL}s defer {len(new_items)} tool(s)")
-                new_items = []
-        sent = 0
-        for item in new_items:
-            tc = item["tc"]; result = item["result"]
-            gidx = last_tool + sent
-            tname = tc.get("name", "unknown")
-            start_dt = _parse_ts_value(tc.get("_ts"))
-            end_dt = _parse_ts_value(result.get("_ts")) if result else None
-            # 每个 tool_call 用一个 model_call_+tool 子树（与 claude-code 视觉一致）
-            mspan = client.start_span(name=f"model_call_{gidx}", span_type="model",
-                                      start_time=start_dt, child_of=root_ctx)
+                hook_log(
+                    f"[rt] batch throttle {since:.1f}s<{REALTIME_BATCH_INTERVAL}s "
+                    f"defer {len(pending_invocations)} invocation(s)"
+                )
+                pending_invocations = []
+
+        invocation_sent = 0
+        tool_sent = 0
+        model_call_base = state.get(
+            "rt_model_call_count", state.get("rt_last_tool", 0)
+        )
+        for invocation in pending_invocations:
+            line = invocation.get("line", -1)
+            turn = invocation.get("turn") or {}
+            tool_calls = invocation.get("tool_calls") or []
+            model_tool_calls = []
+            for tool_call in tool_calls:
+                model_tool_calls.append(ModelToolCall(
+                    id=tool_call.get("call_id", ""),
+                    type="function",
+                    function=ModelToolCallFunction(
+                        name=tool_call.get("name", ""),
+                        arguments=_json_text(tool_call.get("input", {})),
+                    ),
+                ))
+
+            start_dt = _parse_ts_value(invocation.get("start_timestamp"))
+            end_dt = _parse_ts_value(invocation.get("end_timestamp")) or start_dt
+            mspan = client.start_span(
+                name=f"model_call_{invocation_sent + model_call_base}",
+                span_type="model", start_time=start_dt, child_of=root_ctx,
+            )
             mspan.set_runtime(Runtime(library="codex-cli"))
-            try: mspan.set_model_name(model_name)
-            except Exception: pass
-            if start_dt and end_dt:
-                _set_finish_time_safe(mspan, end_dt)
-            mspan.set_input(truncate_text(json.dumps(tc.get("input", {}), ensure_ascii=False)))
-            mspan_ctx = client.get_span_from_header(mspan.to_header())
-            mspan.finish()
-
-            tspan = client.start_span(name=f"tool_{tname}", span_type="tool",
-                                      start_time=start_dt, child_of=mspan_ctx)
-            tspan.set_runtime(Runtime(library="codex-cli"))
-            if end_dt:
-                _set_finish_time_safe(tspan, end_dt)
-            tspan.set_tags({"tool_name": tname, "tool_call_id": tc.get("call_id"), "step_index": gidx})
-            tspan.set_input(truncate_text(json.dumps(tc.get("input", {}), ensure_ascii=False)))
-            if result is not None:
-                tspan.set_output(truncate_text(str(result.get("output", ""))))
-            tspan.finish()
-            sent += 1
-            client.flush()  # 每个 tool_call 立即 flush → 结束即可见
-            if upload_events:
-                hook_log(f"[rt] upload failed after tool flush detail={upload_events[-1][:500]}")
-                return None
-
-        state["rt_last_tool"] = last_tool + sent
-        if sent > 0:
-            state["rt_last_upload_ts"] = time.time()
-
-        # --- token_count 独立增量, 与 tool span 同一 root 下, 按行号水位线幂等 ---
-        # 每条 token_count 挂一个 model_usage_<line> span:
-        #   last_token_usage → set_input_tokens/set_output_tokens (SDK 官方求和字段)
-        #   total_token_usage → tags (只观察, 不参与后端聚合, 否则会重复累加)
-        # 只在 flush 成功后推进 rt_last_token_line, 保证上报失败下次能重发.
-        token_sent = 0
-        last_token_line = state.get("rt_last_token_line", -1)
-        pending_tokens = [ev for ev in (token_events or []) if ev["line"] > last_token_line]
-        max_flushed_line = last_token_line
-        for ev in pending_tokens:
-            line = ev["line"]
-            last = ev.get("usage") or {}
-            total = ev.get("total_usage") or {}
-            in_tk = _usage_int(last, "input_tokens")
-            out_tk = _usage_int(last, "output_tokens")
-            start_dt = _parse_ts_value(ev.get("timestamp"))
-            tk_span = client.start_span(name=f"model_usage_{line}", span_type="model",
-                                        start_time=start_dt, child_of=root_ctx)
-            tk_span.set_runtime(Runtime(library="codex-cli"))
-            try: tk_span.set_model_name(model_name)
-            except Exception: pass
-            if start_dt is not None:
-                _set_finish_time_safe(tk_span, start_dt)
+            try:
+                mspan.set_model_name(model_name)
+            except Exception:
+                pass
+            if start_dt is not None and end_dt is not None:
+                _set_finish_time_safe(mspan, _max_dt(start_dt, end_dt))
+
+            input_messages = []
+            for message in turn.get("input_messages", []):
+                input_messages.append(_make_model_message(
+                    message.get("role", "user"), message.get("content", "")
+                ))
+            if not input_messages and turn.get("user_message_text"):
+                input_messages.append(_make_model_message("user", turn["user_message_text"]))
+            mspan.set_input(ModelInput(
+                messages=input_messages,
+                tools=[],
+                tool_choice=ModelToolChoice(type="", function=None),
+            ))
+            mspan.set_output(ModelOutput(choices=[ModelChoice(
+                finish_reason="tool_calls" if model_tool_calls else "stop",
+                index=0,
+                message=ModelMessage(
+                    role="assistant",
+                    content=invocation.get("content", ""),
+                    reasoning_content=invocation.get("reasoning", ""),
+                    parts=[],
+                    name="",
+                    tool_calls=model_tool_calls,
+                    tool_call_id="",
+                    metadata={},
+                ),
+            )]))
+
+            usage = invocation.get("usage") or {}
+            in_tk = _usage_int(usage, "input_tokens")
+            out_tk = _usage_int(usage, "output_tokens")
             if in_tk > 0:
-                try: tk_span.set_input_tokens(in_tk)
-                except Exception: pass
+                try:
+                    mspan.set_input_tokens(in_tk)
+                except Exception:
+                    pass
             if out_tk > 0:
-                try: tk_span.set_output_tokens(out_tk)
-                except Exception: pass
-            tk_tags = {
-                "token_event_line": line,
-                "token_source": "codex_token_count",
-                "cached_input_tokens": _usage_int(last, "cached_input_tokens"),
-                "reasoning_output_tokens": _usage_int(last, "reasoning_output_tokens"),
-                "total_tokens": _usage_int(last, "total_tokens"),
+                try:
+                    mspan.set_output_tokens(out_tk)
+                except Exception:
+                    pass
+
+            tags = {
+                "invocation_boundary_line": line,
+                "token_source": "codex_token_count" if usage else "unavailable",
+                "cached_input_tokens": _usage_int(usage, "cached_input_tokens"),
+                "reasoning_output_tokens": _usage_int(usage, "reasoning_output_tokens"),
+                "total_tokens": _usage_int(usage, "total_tokens"),
             }
+            total = invocation.get("total_usage") or {}
             if total:
-                tk_tags["total_input_tokens"] = _usage_int(total, "input_tokens")
-                tk_tags["total_output_tokens"] = _usage_int(total, "output_tokens")
-                tk_tags["total_cached_input_tokens"] = _usage_int(total, "cached_input_tokens")
-                tk_tags["total_reasoning_output_tokens"] = _usage_int(total, "reasoning_output_tokens")
-                tk_tags["total_all_tokens"] = _usage_int(total, "total_tokens")
-            tk_span.set_tags(tk_tags)
-            tk_span.finish()
+                tags.update({
+                    "total_input_tokens": _usage_int(total, "input_tokens"),
+                    "total_output_tokens": _usage_int(total, "output_tokens"),
+                    "total_cached_input_tokens": _usage_int(total, "cached_input_tokens"),
+                    "total_reasoning_output_tokens": _usage_int(total, "reasoning_output_tokens"),
+                    "total_all_tokens": _usage_int(total, "total_tokens"),
+                })
+            protected = invocation.get("protected") or {}
+            if protected.get("count"):
+                tags.update({
+                    "protected_reasoning_source": protected.get("source", ""),
+                    "protected_reasoning_count": protected.get("count", 0),
+                    "protected_reasoning_bytes": protected.get("bytes", 0),
+                })
+            mspan.set_tags(tags)
+            mspan_ctx = client.get_span_from_header(mspan.to_header())
+            mspan.finish()
+
+            for tool_call in tool_calls:
+                result = tool_call.get("result")
+                if result is None:
+                    continue
+                tool_start_dt = _parse_ts_value(tool_call.get("_ts")) or start_dt
+                tool_end_dt = _parse_ts_value(result.get("_ts")) or tool_start_dt
+                tname = tool_call.get("name", "unknown")
+                tspan = client.start_span(
+                    name=f"tool_{tname}", span_type="tool",
+                    start_time=tool_start_dt, child_of=mspan_ctx,
+                )
+                tspan.set_runtime(Runtime(library="codex-cli"))
+                if tool_end_dt is not None:
+                    _set_finish_time_safe(tspan, _max_dt(tool_start_dt, tool_end_dt))
+                tspan.set_tags({
+                    "tool_name": tname,
+                    "tool_call_id": tool_call.get("call_id"),
+                    "model_invocation_line": line,
+                })
+                tspan.set_input(truncate_text(_json_text(tool_call.get("input", {}))))
+                tspan.set_output(truncate_text(_json_text(result.get("output", ""))))
+                tspan.finish()
+                tool_sent += 1
+
             client.flush()
             if upload_events:
-                hook_log(f"[rt] upload failed after token flush line={line} detail={upload_events[-1][:500]}")
-                if max_flushed_line > last_token_line:
-                    state["rt_last_token_line"] = max_flushed_line
+                hook_log(
+                    f"[rt] upload failed after invocation flush line={line} "
+                    f"detail={upload_events[-1][:500]}"
+                )
                 return None
-            max_flushed_line = line
-            token_sent += 1
-        if token_sent > 0:
-            state["rt_last_token_line"] = max_flushed_line
+            state["rt_last_invocation_line"] = line
+            invocation_sent += 1
+
+        if invocation_sent > 0:
+            state["rt_model_call_count"] = model_call_base + invocation_sent
             state["rt_last_upload_ts"] = time.time()
 
         if is_terminal:
@@ -1595,13 +1874,17 @@ def send_steps_realtime(turns, session_id, state, model_name="codex", coze_tags_
             if upload_events:
                 hook_log(f"[rt] upload failed after final flush detail={upload_events[-1][:500]}")
                 return None
-            hook_log(f"[rt] finalized total_tools={state['rt_last_tool']}")
+            hook_log(f"[rt] finalized total_model_calls={state.get('rt_model_call_count', 0)}")
 
         if upload_events:
             hook_log(f"[rt] upload failed state not advanced detail={upload_events[-1][:500]}")
             return None
 
-        hook_log(f"[rt] sent {sent} tool span(s), {token_sent} token span(s), last_tool={state['rt_last_tool']}, last_token_line={state.get('rt_last_token_line', -1)}, terminal={is_terminal}")
+        hook_log(
+            f"[rt] sent {invocation_sent} model invocation(s), {tool_sent} tool span(s), "
+            f"last_invocation_line={state.get('rt_last_invocation_line', -1)}, "
+            f"terminal={is_terminal}"
+        )
         return True
     except Exception as e:
         hook_log(f"[rt] error: {e}")
@@ -1676,8 +1959,7 @@ def main():
     state_file = get_state_file_path(transcript_path)
     state = load_state(state_file)
 
-    # 方案 B（实时）：读【全量】entries 重建完整 turn/tool 序列。实时路径用全局 tool 水位线
-    # state["rt_last_tool"] 增量，每次触发把新完成的 tool_call 发出，结束即可见，不节流。
+    # 实时路径读取全量 entries 重建 invocation 序列，再用 token 边界行号水位线增量上报。
     entries = read_rollout_messages(transcript_path, 0)
 
     if not entries:
@@ -1719,9 +2001,13 @@ def main():
 
     # Group entries into turns
     turns = group_messages_into_turns(entries)
-    # 独立扫 token_count 事件, 与 turn 解耦, 走独立行号水位线上报.
-    token_events = extract_token_events(entries)
-    debug_log(f"Grouped into {len(turns)} turns, {len(token_events)} token events")
+    model_invocations = extract_model_invocations(
+        turns, entries, is_terminal=is_terminal_event
+    )
+    debug_log(
+        f"Grouped into {len(turns)} turns, "
+        f"{len(model_invocations)} model invocations"
+    )
 
     # If this is a subagent, save data for parent to include later
     if is_subagent:
@@ -1758,18 +2044,25 @@ def main():
                     break
         coze_tags_override = state.get("coze_tags") or None
 
-        # 方案 B 实时：把新完成的 tool_call 作为 span 挂稳定 root，结束即可见。不节流。
+        # 方案 B 实时：按 token_count 边界把内容、推理、工具调用和 usage 合成 model_call。
         ok = send_steps_realtime(turns, session_id, state, model_name=model_name,
                                  coze_tags_override=coze_tags_override,
                                  is_terminal=is_terminal_event,
-                                 token_events=token_events)
+                                 model_invocations=model_invocations)
         if ok is None:
             hook_log(f"realtime send failed session_id={session_id}")
             debug_log("Realtime send failed, state not advanced")
             return
         save_state(state_file, state)
-        hook_log(f"rt state saved event={hook_event} last_tool={state.get('rt_last_tool')} session_id={session_id}")
-        debug_log(f"State saved. rt_last_tool={state.get('rt_last_tool')}")
+        hook_log(
+            f"rt state saved event={hook_event} "
+            f"last_invocation_line={state.get('rt_last_invocation_line')} "
+            f"session_id={session_id}"
+        )
+        debug_log(
+            f"State saved. rt_last_invocation_line="
+            f"{state.get('rt_last_invocation_line')}"
+        )
     else:
         hook_log(f"skip no turns session_id={session_id}")
         debug_log("No turns to send")