docs/bugs/PHASE_00_IMPLEMENTATION_ORDER.md

Phase 00: Implementation Order & Summary

Total Effort: 5-8 hours Parallelizable: Yes (all 3 phases are independent)

---

Executive Summary

The DeepCritical drug repurposing agent produces garbage results because the search tools are broken:

Tool	Problem	Fix
BioRxiv	API doesn't support search	Replace with Europe PMC
PubMed	Raw queries, no preprocessing	Add query cleaner
ClinicalTrials	No filtering	Add status/type filters

The Microsoft Agent Framework (Magentic) is working correctly. The orchestration layer is fine. The data layer is broken.

---

Phase Specs

Phase	Title	Effort	Priority	Dependencies
01	Replace BioRxiv with Europe PMC	2-3 hrs	P0	None
02	PubMed Query Preprocessing	2-3 hrs	P0	None
03	ClinicalTrials Filtering	1-2 hrs	P1	None

---

Recommended Execution Order

Since all phases are independent, they can be done in parallel by different developers.

If doing sequentially, order by impact:

1. Phase 01 - BioRxiv is completely broken (returns random papers)
2. Phase 02 - PubMed is partially broken (returns suboptimal results)
3. Phase 03 - ClinicalTrials returns too much noise

---

TDD Workflow (Per Phase)

1. Write failing tests
2. Run tests (confirm they fail)
3. Implement fix
4. Run tests (confirm they pass)
5. Run ALL tests (confirm no regressions)
6. Manual verification
7. Commit

---

Verification After All Phases

After completing all 3 phases, run this integration test:

# Full system test
uv run python -c "
import asyncio
from src.tools.europepmc import EuropePMCTool
from src.tools.pubmed import PubMedTool
from src.tools.clinicaltrials import ClinicalTrialsTool

async def test_all():
    query = 'long covid treatment'

    print('=== Europe PMC (Preprints) ===')
    epmc = EuropePMCTool()
    results = await epmc.search(query, 2)
    for r in results:
        print(f'  - {r.citation.title[:60]}...')

    print()
    print('=== PubMed ===')
    pm = PubMedTool()
    results = await pm.search(query, 2)
    for r in results:
        print(f'  - {r.citation.title[:60]}...')

    print()
    print('=== ClinicalTrials.gov ===')
    ct = ClinicalTrialsTool()
    results = await ct.search(query, 2)
    for r in results:
        print(f'  - {r.citation.title[:60]}...')

asyncio.run(test_all())
"

Expected: All results should be relevant to "long covid treatment"

---

Test Magentic Integration

After all phases are complete, test the full Magentic workflow:

# Test Magentic mode (requires OPENAI_API_KEY)
uv run python -c "
import asyncio
from src.orchestrator_magentic import MagenticOrchestrator

async def test_magentic():
    orchestrator = MagenticOrchestrator(max_rounds=3)

    print('Running Magentic workflow...')
    async for event in orchestrator.run('What drugs show promise for Long COVID?'):
        print(f'[{event.type}] {event.message[:100]}...')

asyncio.run(test_magentic())
"

---

Files Changed (All Phases)

File	Phase	Action
src/tools/europepmc.py	01	CREATE
tests/unit/tools/test_europepmc.py	01	CREATE
src/agents/tools.py	01	MODIFY
src/tools/search_handler.py	01	MODIFY
src/tools/biorxiv.py	01	DELETE
tests/unit/tools/test_biorxiv.py	01	DELETE
src/tools/query_utils.py	02	CREATE
tests/unit/tools/test_query_utils.py	02	CREATE
src/tools/pubmed.py	02	MODIFY
src/tools/clinicaltrials.py	03	MODIFY
tests/unit/tools/test_clinicaltrials.py	03	MODIFY

---

Success Criteria (Overall)

• All unit tests pass
• All integration tests pass (real APIs)
• Query "What drugs show promise for Long COVID?" returns relevant results from all 3 sources
• Magentic workflow produces a coherent research report
• No regressions in existing functionality

---

Related Documentation

• P0 Critical Bugs - Root cause analysis
• P0 Magentic Audit - Framework verification
• P0 Actionable Fixes - Fix summaries