Motivation

• Remove several deprecated/fallback common business actions from the action registry to avoid carrying unused tooling and special-casing in validation logic.
• Introduce a first-class "reasoning" node type to represent LLM-driven reasoning tasks with explicit params and an output schema so planners, validators, and the executor can treat LLM reasoning consistently.
• Keep param validation and binding logic in sync with the new reasoning node shape to enable safe references and downstream schema checks.

Description

• Removed deprecated actions (common.ask_ai.v1, common.classify_text.v1, common.join_list.v1, common.list_files.v1, common.split_list.v1, common.summarize.v1) from the business action registry and removed the outdated example assets under examples/search_news_summarize.
• Added a new reasoning node type and model (ReasoningNode) and extended Node.model_validate to accept/validate reasoning nodes and their optional out_params_schema.
• Plumbed reasoning support across planning and runtime: added _reasoning_params_schema and integration into param completion (planner/params_tools.py), new planner tools add_reasoning_node and update_reasoning_node, and mirrored params.expected_output_format into out_params_schema on updates and param completion (planner/structure.py, planner/workflow_builder.py).
• Executor changes to run reasoning nodes: the dynamic executor builds a composed LLM prompt, resolves a toolset, calls ask_ai, and records/serializes results into the node outputs; validation and binding checks now treat a reasoning node's expected_output_format (or mirrored out_params_schema) as its output schema (executor/dynamic_executor.py, verification/binding_checks.py, verification/node_rules.py).
• Removed the special-case logic that previously injected expected_format results into common.ask_ai.v1 outputs during workflow validation, and added visualization color for the reasoning node type (velvetflow/visualization.py).

Testing

• No automated tests were executed for these changes. The change set was prepared and validated via repository grep and local edits, and the PR was created without running CI or test suites.

Codex Task