feat: Add state-based configuration support to MCPToolset

[vblagoje]

Why

Agents need tools that seamlessly share data through state. MCPTool already supports inputs_from_state and outputs_to_state for automatic parameter injection and output handling. MCPToolset couldn't use these features - breaking Agent workflows that rely on state propagation across multiple MCP tools.

This PR completes MCP's Agent integration story. MCPToolset now dynamically configures state mappings for each discovered tool, enabling the same seamless state-based coordination that MCPTool provides.

• fixes Add inputs_from_state and outputs_to_state support to MCPToolset #2677

What

Added per-tool state configuration to MCPToolset:

• Added inputs_from_state, outputs_to_state, outputs_to_string parameters (dict mapping tool names to configs)
• _serialize_state_config() / _deserialize_state_config() helpers handle both nested (outputs_to_state) and flat (outputs_to_string) formats
• _validate_state_configs() warns on unknown tool names
• ValueError from Tool parameter validation bubbles up (Haystack >=2.22.0)
• Updated dependency to haystack-ai>=2.19.0 (supports state params)

How can it be used

from haystack_integrations.tools.mcp import MCPToolset, StdioServerInfo

# Create toolset with per-tool state configuration
toolset = MCPToolset(
    server_info=StdioServerInfo(command="uvx", args=["mcp-server-git"]),
    tool_names=["git_status", "git_diff", "git_log"],
    
    # Map state keys to tool parameters
    inputs_from_state={
        "git_status": {"repository": "repo_path"},
        "git_diff": {"repository": "repo_path"},
        "git_log": {"repository": "repo_path"},
    },
    
    # Map tool outputs to state keys
    outputs_to_state={
        "git_status": {"status_result": {"source": "status"}},
        "git_diff": {"diff_result": {}},
    },
    
    # Convert outputs to strings for LLM
    outputs_to_string={
        "git_diff": {"source": "diff", "handler": format_diff},
    },
)

How did you test it

• 3 unit tests: state config creation, serialization roundtrip, unknown tool warning
• 1 validation test: ValueError on invalid parameter (skipped on Haystack <2.22.0)
• 6 helper tests: serialize/deserialize for both outputs_to_state and outputs_to_string formats

Notes for the reviewer

Parameter validation behavior:

• Haystack >=2.22.0: ValueError raised at tool creation time for invalid parameter names
• Haystack 2.19.0-2.21.x: Invalid parameters fail at runtime during tool invocation

[sjrl]

Could we add a dev comment here explaining why one might provide an empty dict here?

[sjrl]

@vblagoje looking good! I wanted to ask if you could add an integration test using these configs to see if everything works as expected?

[mpangrazzi]

Looks already good! I've added a few comments. +1 on adding an integration test as @sjrl mentioned.

[mpangrazzi]

More a type checker thing: the function accepts and handles None correctly, but the type hint says it only accepts dict.

[mpangrazzi]

Edge case: what if someone use source or handler as keys in state? May worth documenting?

[mpangrazzi]

I may miss something, but why not doing a full roundtrip of to_dict() / from_dict() here (so without mocking)?