fix: handle consecutive system messages with plain string content

[giulio-leone]

Summary

Fixes #1972

get_clean_message_list crashes with AssertionError when consecutive messages with the same role have plain string content instead of the structured list format [{"type": "text", "text": "..."}].

Root Cause

When ChatMessage.from_dict() creates a message from {"role": "system", "content": "text"}, the content remains a plain string. The merging logic at line 376 asserts isinstance(message.content, list), which fails for string content.

Fix

Normalize string content to the structured list format [{"type": "text", "text": content}] early in the processing loop — right after role conversion and before image encoding. This ensures all downstream code (merging, output, flatten) can assume content is always a list of dicts.

Changes

• src/smolagents/models.py: Added early normalization of string content to list format in get_clean_message_list
• tests/test_models.py: Added 2 regression tests (structured and flatten modes)

Tests

pytest tests/test_models.py -k 'clean_message' -v
# 9 passed, 113 deselected

[Copilot]

Pull request overview

Fixes a crash in get_clean_message_list when consecutive messages share the same role but use plain string content (instead of the structured list-of-dicts format), by normalizing string content early so merging logic is consistent.

Changes:

• Normalize ChatMessage.content from str → [{"type": "text", "text": ...}] within get_clean_message_list before downstream processing/merging.
• Add regression tests covering consecutive system messages with string content in both structured and flatten_messages_as_text modes.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/smolagents/models.py	Normalizes string message content early to prevent merge-time AssertionError for consecutive same-role messages.
tests/test_models.py	Adds regression tests for consecutive system string content in both structured and flatten modes.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The extra normalization of output_message_list[-1]["content"] inside the merge branch appears unreachable given the invariants in this function: when flatten_messages_as_text=True the "else" branch always stores a string (so it can't be a list here), and when flatten_messages_as_text=False it always stores a list (so it can't be a string here). Keeping these branches adds complexity and can mask invariant violations; consider removing them or replacing them with assertions that the stored type matches the selected mode.

Suggested change

	# Normalize it if needed before concatenating.
	if isinstance(output_message_list[-1]["content"], list):
	output_message_list[-1]["content"] = output_message_list[-1]["content"][0]["text"]
	output_message_list[-1]["content"] += "\n" + message.content[0]["text"]
	else:
	# In structured mode, output content is stored as a list of dicts.
	# Normalize it if needed before merging.
	if isinstance(output_message_list[-1]["content"], str):
	output_message_list[-1]["content"] = [
	{"type": "text", "text": output_message_list[-1]["content"]}
	]
	assert isinstance(
	output_message_list[-1]["content"], str
	), "Error: expected string content in flatten mode"
	output_message_list[-1]["content"] += "\n" + message.content[0]["text"]
	else:
	# In structured mode, output content is stored as a list of dicts.
	assert isinstance(
	output_message_list[-1]["content"], list
	), "Error: expected list content in structured mode"

[giulio-leone]

The defensive normalization is intentional — while the invariants should hold under normal operation, the function receives data from LLM tool call results that may not always conform. Keeping the normalization as a safety net prevents runtime crashes from unexpected content types. Replacing with assertions would be less resilient in production.