[JenD86]

Modulo treated every {"name": ..., "arguments": ...} in the model's output as its own call. That made it hard to use inside an agent framework that has tools of its own:

• A call to the framework's tool got an "Unknown tool" reply and was removed from prose.
• The framework's tool definitions never reached the chat template.

Changes

• step(messages, tools=[...]) passes the framework's tool schemas (or functions, as for apply_chat_template) to the chat template.
  • Calls to those tools come back in the new StepResult.other_tool_calls. Modulo doesn't run them, and they stay in prose with their <tool_call> tags.
  • Modulo's own calls are still run and removed, tags included, and tool_messages() holds only their results.
• SteeringAgent(pass_through_unknown_tools=True) treats every call that isn't Modulo's as the framework's. It's for frameworks that describe their tools in their own system prompt, and it's needed on OLMo-2, whose template ignores tools=.
• Unknown tools: without either option, an unknown call still gets "Unknown tool", and the message now lists the framework's tool names too. That keeps the standalone demo's behaviour.
• "parameters" is read as "arguments", because Llama-3.1's template asks for that format. Objects with a "description" are echoed tool definitions, not calls.
• AGENTS.md explains how to integrate a framework with its own tools, adds the new fields, and notes the template quirks in the model table.

Testing

• CPU tests: 130 pass (test_agent_logic, test_cli, test_chat).
• Real chat templates (Qwen2.5, Llama-3.1, OLMo-2, Mistral v0.3), rendered on CPU:
  • All four render with tools=, keep the system prompt, and show tool results.
  • Qwen, Llama and Mistral list the tools; OLMo doesn't.
• Not run on a GPU. Generation isn't affected unless tools= is given.

🤖 Generated with Claude Code