diff --git a/examples/01_standalone_sdk/27_compare_content_vs_telemetry.py b/examples/01_standalone_sdk/27_compare_content_vs_telemetry.py
new file mode 100644
index 0000000000..502b0789da
--- /dev/null
+++ b/examples/01_standalone_sdk/27_compare_content_vs_telemetry.py
@@ -0,0 +1,288 @@
+"""
+Helper: Compare assistant content between telemetry logs and persisted events.
+
+- Runs a short conversation using the configured LLM (LLM_MODEL/LLM_API_KEY)
+- Ensures telemetry logging is enabled and a persistence_dir is used
+- After the run, scans:
+  - Telemetry logs under LOG_DIR/completions for entries written during this run
+  - Conversation events under <persistence_dir>/<conv_id_hex>/events
+- For each case where assistant content is empty either in telemetry or in events,
+  appends a section into a Markdown report with the JSON for:
+  - The telemetry LLM response (response object)
+  - The matching (or nearest) assistant MessageEvent (if any)
+
+Usage:
+  LLM_MODEL=gemini/gemini-2.5-pro LLM_API_KEY=... uv run python \
+    examples/01_standalone_sdk/27_compare_content_vs_telemetry.py
+
+Optional env:
+  LOG_DIR defaults to ./logs; the report is stored under LOG_DIR/comparisons
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import time
+from pathlib import Path
+from typing import Any
+
+from pydantic import SecretStr
+
+from openhands.sdk import LLM, Agent, Conversation
+
+
+def ensure_dir(p: str) -> None:
+    Path(p).mkdir(parents=True, exist_ok=True)
+
+
+def read_json(p: Path) -> Any:
+    with p.open("r", encoding="utf-8") as f:
+        return json.load(f)
+
+
+def content_text_from_message_content(payload: Any) -> str:
+    """Heuristic extraction of text from message.content payloads.
+
+    Supports several shapes:
+    - None -> ""
+    - str -> itself
+    - list[dict or str] -> concatenate any 'text' fields from dict items or str items
+    """
+    if payload is None:
+        return ""
+    if isinstance(payload, str):
+        return payload
+    if isinstance(payload, list):
+        texts: list[str] = []
+        for item in payload:
+            if isinstance(item, str):
+                texts.append(item)
+            elif isinstance(item, dict):
+                # Common shapes: {"type": "text", "text": "..."}
+                t = item.get("text")
+                if isinstance(t, str):
+                    texts.append(t)
+        return "\n".join(texts)
+    return ""
+
+
+def content_text_from_event_llm_message_content(payload: Any) -> str:
+    """Extract text from persisted event llm_message.content list.
+
+    Events typically store a list of TextContent dicts with 'text' keys.
+    """
+    if not payload:
+        return ""
+    if isinstance(payload, list):
+        texts: list[str] = []
+        for item in payload:
+            if isinstance(item, dict):
+                t = item.get("text")
+                if isinstance(t, str):
+                    texts.append(t)
+        return "\n".join(texts)
+    return ""
+
+
+def main() -> None:
+    api_key = os.getenv("LLM_API_KEY")
+    assert api_key, "LLM_API_KEY is required"
+    model = os.getenv("LLM_MODEL", "gemini/gemini-2.5-pro")
+
+    log_dir = os.getenv("LOG_DIR", "logs")
+    completions_dir = os.path.join(log_dir, "completions")
+    comparisons_dir = os.path.join(log_dir, "comparisons")
+    ensure_dir(completions_dir)
+    ensure_dir(comparisons_dir)
+
+    # Build LLM with telemetry logging enabled
+    llm = LLM(
+        model=model,
+        api_key=SecretStr(api_key),
+        log_completions=True,
+        log_completions_folder=completions_dir,
+        usage_id="agent",
+    )
+
+    agent = Agent(llm=llm)
+
+    # Use .conversations as base; conversation id folder appended automatically
+    persistence_base = ".conversations"
+    ensure_dir(persistence_base)
+
+    # Mark start time to filter telemetry written during this run
+    t0 = time.time() - 1.0  # small slack
+
+    conversation = Conversation(
+        agent=agent,
+        workspace=os.getcwd(),
+        persistence_dir=persistence_base,
+    )
+
+    # Generate only text turns (no tool calls) to force MessageEvent
+    conversation.send_message("Reply with only OK (no tools)")
+    conversation.run()
+    conversation.send_message("Now reply with: This is a second text-only message.")
+    conversation.run()
+
+    # Locate events dir
+    # persistence_dir is non-None since we passed a base dir
+    assert conversation.state.persistence_dir is not None
+    conv_dir = Path(conversation.state.persistence_dir)
+    events_dir = conv_dir / "events"
+
+    # Collect telemetry records written after t0 and map by response id
+    telemetry_files = sorted(Path(completions_dir).glob("*.json"))
+    telemetry_by_id: dict[str, dict[str, Any]] = {}
+    for tf in telemetry_files:
+        try:
+            data = read_json(tf)
+        except Exception:
+            continue
+        ts = float(data.get("timestamp", 0.0))
+        if ts < t0:
+            continue
+        resp = data.get("response", {})
+        rid = resp.get("id")
+        if isinstance(rid, str) and rid:
+            telemetry_by_id[rid] = {"path": str(tf), "data": data}
+
+    # Collect LLM-convertible events grouped by llm_response_id
+    events_by_id: dict[str, dict[str, list[dict[str, Any]]]] = {}
+    if events_dir.is_dir():
+        for ef in sorted(events_dir.glob("*.json")):
+            try:
+                ev = read_json(ef)
+            except Exception:
+                continue
+            kind = ev.get("kind")
+            # MessageEvent (assistant) – use llm_response_id when present
+            if kind == "MessageEvent":
+                if ev.get("source") == "agent":
+                    rid = ev.get("llm_response_id")
+                    if isinstance(rid, str) and rid:
+                        events_by_id.setdefault(rid, {"message": [], "action": []})
+                        events_by_id[rid]["message"].append(
+                            {"path": str(ef), "data": ev}
+                        )
+            # ActionEvent – always has llm_response_id
+            elif kind == "ActionEvent":
+                rid = ev.get("llm_response_id")
+                if isinstance(rid, str) and rid:
+                    events_by_id.setdefault(rid, {"message": [], "action": []})
+                    events_by_id[rid]["action"].append({"path": str(ef), "data": ev})
+
+    # Prepare markdown report
+    out_path = Path(comparisons_dir) / (f"content_vs_telemetry_{int(time.time())}.md")
+
+    def dump_json(obj: Any) -> str:
+        try:
+            return json.dumps(obj, indent=2, ensure_ascii=False, sort_keys=True)
+        except Exception:
+            return str(obj)
+
+    lines: list[str] = []
+    lines.append("# Content vs Telemetry Comparison\n")
+    lines.append(f"- Model: **{model}**\n")
+    lines.append(f"- Conversation dir: `{conv_dir}`\n")
+    lines.append(f"- Events dir: `{events_dir}`\n")
+    lines.append(f"- Telemetry scan start: `{t0}`\n")
+    lines.append("")
+
+    # Helper to check emptiness for telemetry message content
+    def telemetry_content_text(rec: dict[str, Any]) -> str:
+        resp = rec["data"].get("response", {})
+        # Chat completions shape
+        choices = resp.get("choices") or []
+        if choices:
+            msg = (choices[0] or {}).get("message", {})
+            content = msg.get("content")
+            return content_text_from_message_content(content).strip()
+        # Responses API shape
+        out_text = resp.get("output_text")
+        if isinstance(out_text, str):
+            return out_text.strip()
+        return ""
+
+    # Helpers to extract MessageEvent-side content and reasoning
+    def message_event_text(ev: dict[str, Any]) -> str:
+        lm = ev.get("llm_message", {})
+        return content_text_from_event_llm_message_content(lm.get("content")).strip()
+
+    def message_event_reasoning(ev: dict[str, Any]) -> str:
+        lm = ev.get("llm_message", {})
+        rc = lm.get("reasoning_content")
+        if isinstance(rc, str):
+            return rc.strip()
+        return ""
+
+    # Telemetry-side reasoning (chat-completions shape)
+    def telemetry_reasoning_text(rec: dict[str, Any]) -> str:
+        resp = rec["data"].get("response", {})
+        choices = resp.get("choices") or []
+        if choices:
+            msg = (choices[0] or {}).get("message", {})
+            rc = msg.get("reasoning_content")
+            if isinstance(rc, str):
+                return rc.strip()
+        return ""
+
+    lines.append(
+        "## MessageEvent-focused cases (empty content and/or reasoning present)\n"
+    )
+
+    # Compare telemetry entries with events by response id (MessageEvents only)
+    for rid, tele in telemetry_by_id.items():
+        ev_bucket = events_by_id.get(rid, {"message": [], "action": []})
+        if not ev_bucket.get("message"):
+            continue  # focus only on responses that resulted in MessageEvent
+
+        tele_txt = telemetry_content_text(tele)
+        tele_empty = len(tele_txt) == 0
+        tele_reason = telemetry_reasoning_text(tele)
+
+        # Consider the first MessageEvent for this response id
+        msg_event = ev_bucket["message"][0]
+        etxt = message_event_text(msg_event["data"]) or ""
+        e_reason = message_event_reasoning(msg_event["data"]) or ""
+        evt_empty = len(etxt) == 0
+
+        # Report only when we see the phenomenon of interest
+        if not ((tele_empty and tele_reason) or (evt_empty and e_reason)):
+            continue
+
+        lines.append(f"### Response `{rid}`\n")
+        # Telemetry block
+        lines.append(f"- Telemetry file: `{tele['path']}`\n")
+        lines.append(f"- Telemetry content empty: **{tele_empty}**\n")
+        lines.append(f"- Telemetry reasoning present: **{bool(tele_reason)}**\n")
+        if tele_reason:
+            preview = tele_reason[:200].replace("\n", " ")
+            lines.append(f"  - Reasoning preview: {preview}...\n")
+        lines.append("- Telemetry response JSON:\n")
+        lines.append("```json")
+        lines.append(dump_json(tele["data"].get("response", {})))
+        lines.append("```")
+
+        # Event block (MessageEvent)
+        lines.append(f"- Event (MessageEvent) file: `{msg_event['path']}`\n")
+        lines.append(f"  - Event content empty: **{evt_empty}**\n")
+        lines.append(f"  - Event reasoning present: **{bool(e_reason)}**\n")
+        if e_reason:
+            eprev = e_reason[:200].replace("\n", " ")
+            lines.append(f"    - Reasoning preview: {eprev}...\n")
+        lines.append("  - Event JSON:\n")
+        lines.append("```json")
+        lines.append(dump_json(msg_event["data"]))
+        lines.append("```")
+        lines.append("")
+
+    with out_path.open("w", encoding="utf-8") as f:
+        f.write("\n".join(lines) + "\n")
+
+    print(f"Wrote report to: {out_path}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/openhands-sdk/openhands/sdk/llm/utils/telemetry.py b/openhands-sdk/openhands/sdk/llm/utils/telemetry.py
index 6d4174ce1f..ddbb117a5e 100644
--- a/openhands-sdk/openhands/sdk/llm/utils/telemetry.py
+++ b/openhands-sdk/openhands/sdk/llm/utils/telemetry.py
@@ -303,15 +303,26 @@ def log_llm_call(
 
 def _safe_json(obj: Any) -> Any:
     # Centralized serializer for telemetry logs.
-    # Today, responses are Pydantic ModelResponse or ResponsesAPIResponse; rely on it.
-    if isinstance(obj, ModelResponse) or isinstance(obj, ResponsesAPIResponse):
-        # Use mode='json' to ensure proper serialization of nested Pydantic models
-        # and avoid PydanticSerializationUnexpectedValue warnings.
-        # exclude_none=True reduces payload size by omitting None fields.
+    # Prefer Pydantic-safe dumps to avoid circular refs and exclude runtime-only fields.
+    if isinstance(obj, (ModelResponse, ResponsesAPIResponse)):
         return obj.model_dump(mode="json", exclude_none=True)
 
-    # Fallbacks for non-Pydantic objects used elsewhere in the log payload
+    # Generic Pydantic models (e.g., ToolDefinition) – use model_dump to honor
+    # field exclusions (like executor) and custom field serializers.
+    if isinstance(obj, BaseModel):
+        return obj.model_dump(mode="json", exclude_none=True)
+
+    # Common non-JSON-native types
+    if isinstance(obj, set):
+        return list(obj)
+    if isinstance(obj, bytes):
+        try:
+            return obj.decode("utf-8", errors="ignore")
+        except Exception:
+            return str(obj)
+
+    # Fallbacks for arbitrary objects
     try:
-        return obj.__dict__
+        return getattr(obj, "__dict__", str(obj))
     except Exception:
         return str(obj)