Agent-side Athena Decision Service integration

[vredchenko]

Summary

Implement Athena Decision Service integration in the Agent (not backend). This follows the separation of concerns principle where:

• Backend handles recommendations (ML predictions stored and routed via SSE)
• Agent handles instructions (user confirmation, conversion to microscope format, execution via AthenaAPI)

Background

PR #190 originally implemented Athena integration in the backend (consumer.py), which violated the architectural boundary. That approach was removed. This issue captures the correct agent-side approach.

Implementation Notes

Decision Type Mapping

• Grid square predictions → DecisionType.GRID_SQUARE_SELECTION
• Foil hole predictions → DecisionType.FOIL_HOLE_SELECTION
• Decision value: "true" if prediction > 0.5, else "false"
• Details field: include prediction value and model name

Bulk vs Single Requests

• Single foilhole: Use POST /api/v1/Decision
• Multiple foilholes: Use POST /api/v1/Decisions (bulk endpoint)
• Reason: Prevents microscope blocking - sequential decisions would cause the microscope to process the first and block subsequent ones

Plugin Configuration

• Use PluginType.CUSTOM
• decidedBy: "smartem_ml_model"

User Confirmation Modes

The agent should support three modes:

1. Auto-accept: Automatically forward all recommendations as instructions
2. Manual confirm: Present each recommendation to user for approval
3. Observe only: Display recommendations but user takes manual action outside system

Tasks

• Add recommendation instruction handlers to agent (fs_watcher.py)
• Create Athena handler module for decision conversion and submission
• Implement user confirmation flow (3 modes)
• Connect to AthenaAPI for actual instruction submission
• Add configuration for user confirmation mode
• Test with Athena mock server

Related

• PR Athena Decision Service integration #190 (cleaned up backend-side integration)
• src/athena_api/client.py - record_decisions() bulk method already available