docs/Memory.md

CDSS — Memory & Progress Tracker

Current Status

Date: August 30, 2026 State: Phases 1-3 implemented and reviewed. All bugs fixed. Starting Phase 4.

Bug Fixes Applied (Aug 30)

• Phase 1 Fix: evidenceScanner/requirements.txt — fixed missing newline between python-dateutil and fastapi, removed duplicate tenacity entry.
• Phase 3 Fix: orchestrator/requirements.txt — removed unnecessary heavy ML deps (torch, sentence-transformers, qdrant-client, transformers). Orchestrator only makes HTTP calls.
• Phase 3 Fix: orchestrator/tools.py — changed type[BaseModel] to Type[BaseModel] for Python <3.12 compatibility. Increased httpx timeout to 90s.
• Phase 3 Fix (CRITICAL): orchestrator/crew.py — split the single evidence-gathering task (which asked LLM to call 2 tools and return 2 JSONs in 1 string) into 2 separate tasks (Task 2: RAG, Task 3: Scanner). Added a 5th CrewAI agent role (web_scanner). Each task now has unambiguous single-tool instructions.

What Has Been Done

• Research Phase: Audited all 6 project folders (clinicalTextClassifier, vectorDbIngestion, vectorRetrieval, evidenceScanner, dataFusion, clinicalReasoning, orchestrator).
• Schema Alignment: Mapped exact Pydantic I/O models across all agents to identify integration gaps.
• Agent 3 Discovery: Confirmed the new evidenceScanner folder exists and contains valid logic, but lacks a FastAPI server.
• Agent 4 Adapters: Confirmed Agent 4 inherently supports raw JSON inputs from Agent 2/3 and can handle missing Knowledge Graph data gracefully.
• Documentation: Created PRD.md, Architecture.md, Rules.md, Phases.md, Design.md, Risk_Analysis.md, and this Memory.md.
• User Decisions Locked:
  • Keep CrewAI for Orchestration.
  • Drop Knowledge Graph (kg_evidence=[]).
  • Do not modify Agent 1's output schema (transform in Orchestrator).

What Is Pending (Phases)

• Phase 1: Evidence Scanner Server
  • Fixed tenacity retry bug in web_scraper.py
  • Added app.py FastAPI wrapper for evidenceScanner.
• Phase 2: Orchestrator Adapters
  • Create orchestrator/adapters.py.
• Phase 3: Orchestrator Rewrite
  • Update crew.py to wire all 5 agents.
  • Update app.py to handle the while loop (feedback loop) checking confidence ≥ 0.70.
• Phase 4: Full System Test
  • Created start_all.ps1 startup script.
  • Syntax-verified all new/modified files.
  • Launch all ports (8000, 8002, 8003, 8004, 8005, 9000).
  • Test POST /analyze.

Known Issues / Gotchas

• evidenceScanner currently has a bug in its web_scraper.py retry logic where non-200 HTTP codes return None instead of raising exceptions, which bypasses tenacity.
• Need to ensure top_k=5 is dynamically passed through to SearchRequest during Agent 1 → Agent 2 transformation.