Enhance response expressiveness to support agent-generated development workflows (CLI, tools, testing, etc.

[crispyberry]

Description:
As our system increasingly adopts agentic workflows for generating and executing code, we’ve observed that the current response structure is insufficient to capture the full range of actions taken by the agent. In modern development workflows, AI agents are not only generating code but also:
Calling command-line tools (e.g., npm install, pytest)
Generating and running tests
....

These actions are currently compressed into a flat or overly simplified response format, making it difficult to:

• Accurately replay or trace the sequence of operations
• Present meaningful step-by-step reasoning to users
• Integrate seamlessly with toolchains or IDE extensions
• Debug or verify what exactly the agent performed

Proposal:
We propose to enhance the response structure with a richer, composable schema that can represent:

• Intent chains (e.g., “generated a test → executed it → updated the function”)
• Action types (e.g., code_generation, cli_invocation, file_write, tool_use)
• Execution results, including stdout/stderr, tool outputs, and test results
• Dependencies between steps, forming a DAG of actions if necessary

This improvement would make it significantly easier to:

• Replay actions accurately in dev environments
• Display meaningful agent "thoughts" and execution traces in UIs
• Allow downstream automation to pick up from intermediate steps
• Improve auditing and debugging of AI-generated development flows

[crispyberry]

Input: MCP? Figma input? Not only prompt.
Output=code diffs