Refactored RoutineDiscoveryAgent from a hardcoded sequential pipeline to an LLM-driven agentic loop with tools

[rayruizhiliao]

Files Created/Modified

File	Action	Description
bluebox/data_models/routine_discovery/state.py	Created	DiscoveryPhase enum and RoutineDiscoveryState class for BFS queue management
bluebox/llms/tools/routine_discovery_tools.py	Created	11 tool definitions with JSON schemas
bluebox/agents/routine_discovery_agent.py	Rewritten	Agentic loop replacing hardcoded pipeline

Architecture Changes

Before: Hardcoded sequential pipeline

run() → identify_transaction() → confirm_identified_transaction()
    → extract_variables() → resolve_variables() → construct_routine()
    → productionize_routine()

After: LLM-driven loop with tools

run() → _run_agent_loop():
    while phase != COMPLETE:
        response = llm_client.call_sync(messages, tools)
            for tool_call in response.tool_calls:
                result = _execute_tool(tool_call)

Tools Available to the LLM

Category	Tools
Data Access	list_transactions, get_transaction, scan_for_value
Queue Management	add_transaction_to_queue, get_queue_status, mark_transaction_complete
State Recording	record_identified_transaction, record_extracted_variables, record_resolved_variable
Construction	construct_routine, finalize_routine

Key Design Decisions

1. BFS via tools: The transaction queue is managed through add_transaction_to_queue and mark_transaction_complete - when a variable's source is another transaction, it's auto-added to the queue
2. Phase-based guidance: System prompt updates with current phase and queue status to guide the LLM
3. Same interface: The run() method signature and return type are preserved for compatibility
4. State persistence: RoutineDiscoveryState tracks all intermediate data for routine construction

Test Plan

• Run existing routine discovery tests
• Verify BFS dependency resolution with multi-transaction routines
• Compare output Routine structure with original pipeline

[dimavrem22]

Few thoughts:

I would probably yank these for now then add back in when agent works: record_identified_transaction, record_extracted_variables, record_resolved_variable

--

Edited at 6:24p, 1/30

Dima: Potentially remove queue management tools and manually write them into the agent run flow? Alternatively be very specific about the queue conditions when the agent can exit the loop

[dimavrem22]

could likely consolidate these construct_routine, finalize_routine since agent has access to docs and will get access to execution