Capture command output and user responses in agent actions

[DylanBlakemore]

Summary

• Add Output field to AppliedActionResult to capture stdout/stderr from commands, file read results, and user responses
• Change ToolRunner.RunCommand signature from error to (string, error) to surface command output
• Feed captured output back into LLM context (truncated to 4KB) so the agent can reason about results across rounds

Details

The agent previously discarded stdout from successful commands and silently no-op'd ask_user prompts, which meant it couldn't reason about what actually happened during execution. This change threads output through every action type — run_command returns stdout on success and stderr on failure, list_files and read_files capture their results, and ask_user now prompts the user and stores their response. The formatApplyResults function in the setup command includes this output (truncated to 4KB) when building context for subsequent LLM rounds, enabling the agent to make informed decisions based on prior results.

This is Section 1 of the agent setup improvements plan (docs/AGENT_SETUP_IMPROVEMENTS.md).