fix: preserve tool-call context in tokenization (#527)

* fix: preserve tool-call context in tokenization

Only splice trainable assistant spans and keep tool_calls
in the template; error if tool_calls would be dropped.

* Fix type errors in tool-call tokenization

- Use .get() instead of direct [] access for tool_calls to handle
  message types that don't have this key
- Cast message to dict[str, Any] when appending to token_template_messages

---------

Co-authored-by: Cursor Bot <bot@cursor.com>

Author: corbt

src/art/preprocessing/tokenize.py

@@ -201,32 +201,46 @@ def tokenize_trajectory(
         set(range(cast(int, tokenizer.vocab_size))) - set(original_token_ids)
     )
     sentinal_token = tokenizer.decode(sentinal_token_id)
+    token_template_messages: list[dict[str, Any]] = []
+    for original, message in zip(messages_and_choices, messages):
+        trainable_assistant = (
+            not isinstance(original, dict) and original.logprobs is not None
+        ) or (
+            allow_training_without_logprobs
+            and isinstance(original, dict)
+            and original.get("role") == "assistant"
+        )
+        if trainable_assistant:
+            token_template_messages.append(
+                {
+                    "role": "assistant",
+                    "content": sentinal_token,
+                    **(
+                        {"tool_calls": message.get("tool_calls")}  # type: ignore[call-overload]
+                        if message.get("tool_calls")  # type: ignore[call-overload]
+                        else {}
+                    ),
+                }
+            )
+        else:
+            token_template_messages.append(cast(dict[str, Any], message))
     token_ids = cast(
         list[int],
         tokenizer.apply_chat_template(
-            cast(
-                list[dict],
-                [
-                    (
-                        message_or_choice
-                        if isinstance(message_or_choice, dict)
-                        and not message_or_choice["role"] == "assistant"
-                        else {
-                            "role": "assistant",
-                            "content": sentinal_token,
-                        }
-                    )
-                    for message_or_choice in messages_and_choices
-                ],
-            ),
+            cast(list[dict], token_template_messages),
             tools=tools,  # type: ignore
             continue_final_message=True,
         ),
     )
     assistant_mask: list[int] = [0] * len(token_ids)
     logprobs = [float("nan")] * len(token_ids)
     for message in messages_and_choices:
-        if isinstance(message, dict) and not message["role"] == "assistant":
+        if isinstance(message, dict):
+            if message["role"] != "assistant":
+                continue
+            if not allow_training_without_logprobs:
+                continue
+        elif message.logprobs is None and not allow_training_without_logprobs:
             continue
         start = token_ids.index(sentinal_token_id)
         end = start + 1
@@ -235,6 +249,11 @@ def tokenize_trajectory(
         except IndexError:
             end_token_id = None
         if isinstance(message, dict):
+            if message.get("tool_calls"):
+                raise ValueError(
+                    "Assistant message has tool_calls but is being tokenized "
+                    "via tokenizer.encode(content). This path ignores tool calls."
+                )
             content = message.get("content")
             assert isinstance(content, str)
             content_token_ids = tokenizer.encode(