fix: guard json.loads() in JSON tensor parsing — prevent hard crash on None or malformed tensor values

[amathxbt]

Bug Fix: Unguarded json.loads() crashes entire client when JSON tensor value is None or malformed

Summary

Two unguarded json.loads() calls in _conversions.py cause a hard crash whenever a JSON tensor's value field is missing or contains invalid JSON. Also found a missing None check on tensor names that silently corrupts the output dictionary.

---

Bug 1 — convert_to_model_output: unguarded json.loads(value)

# BEFORE (crashes if value is None or malformed JSON)
for tensor in output.get("jsons", []):
    if isinstance(tensor, (AttributeDict, dict)):
        name = tensor.get("name")
        value = tensor.get("value")
        output_dict[name] = np.array(json.loads(value))  # 💥 JSONDecodeError / TypeError

If value is None (field missing from the response) → TypeError: the JSON object must be str, bytes or bytearray, not NoneType

If value is a malformed JSON string → json.JSONDecodeError

Both crash the entire inference call with a traceback that gives no useful context.

---

Bug 2 — convert_array_to_model_output: same unguarded json.loads(value)

# BEFORE (same crash in the array-based output path)
for tensor in array_data[2]:
    name = tensor[0]
    value = tensor[1]
    json_data[name] = np.array(json.loads(value))  # 💥 same crash

---

Bug 3 — None tensor name silently corrupts output dict

In both functions, if name is None (missing from the response), the code does:

output_dict[None] = np.array(...)

This inserts a None key into the output dict without any warning, making it impossible for callers to use the result correctly.

---

Fix

Wrapped both json.loads() calls in try/except (json.JSONDecodeError, TypeError) and added a None name guard with a continue + logging.warning() in both functions.

# AFTER
if name is None:
    logging.warning("JSON tensor is missing 'name' field; skipping.")
    continue
try:
    output_dict[name] = np.array(json.loads(value))
except (json.JSONDecodeError, TypeError) as e:
    logging.warning(f"Failed to parse JSON tensor '{name}' — value={value!r}: {e}. Skipping.")

---

Impact

Any model whose output includes a JSON tensor with a missing or malformed value field will crash the entire SDK client mid-inference with no useful error message. This affects both the event-based (convert_to_model_output) and array-based (convert_array_to_model_output) output paths.

---

@adambalogh @kylexqian — both output conversion paths had the same unguarded call. Fixed in both places. Happy to add tests if needed. 🙏