refactor(ai): Drive the assistant loop through the shared agents runner

[nfebe]

Closes #182.

The assistant's tool loop, step budget, and tool approval move onto the shared whilesmartgo/agents library, so that logic lives in one tested place instead of a hand-rolled loop. The MCP server (#186) already exposes the same tools; both now run on the same runner. Auto-run, per-call approval, suggestion parsing, and the hidden-context split are preserved.

Note: #186 and this branch both touch the agents dependency version, so whichever merges second needs a trivial rebase.

[sourceant]

Code Review Summary

This PR refactors the AI assistant's execution loop to use the whilesmartgo/agents library. It replaces a manual tool execution loop with a standardized runner, simplifying the API handlers and improving testability. Key features like auto-run, per-call approval, and suggestion parsing are preserved through well-designed adapters.

🚀 Key Improvements

• Adopted a shared agent runner library, reducing local maintenance of complex tool-loop logic.
• Implemented CapturingEngine to preserve model metadata across library calls.
• Simplified ApproveSession handler using the runner's Resume capability.
• Added comprehensive test coverage for the per-call tool approval workflow.

💡 Minor Suggestions

• Ensure that fromEngineToolCalls and toEngineToolCalls are consistently available in the internal/ai package to avoid compilation issues in environments with different file splits.

[sourceant]

Review complete. No specific code suggestions were generated. See the overview comment for a summary.