RecHarness v0.2-alpha

RecHarness v0.2-alpha Release Notes

RecHarness v0.2-alpha focuses on making the harness callable by general-purpose agents.

Highlights

• Agent-facing parse, assist, and verify schemas
• Multi-catalog AgentHarnessRouter
• MCP tools for catalog listing, parsing, assist, and verification
• Framework-neutral tool callables via make_recharness_tool_functions()
• Deterministic MCP-style and tool-calling agent-loop demos
• Headphones domain adapter and local dogfooding utilities
• Safer local-catalog-based answer drafting
• Local evaluation and diagnostics

What RecHarness Is

RecHarness is an agent-agnostic recommendation quality layer. It helps agents ground product recommendations in local catalogs, verify constraints and claims, and inspect recommendation traces.

What RecHarness Is Not

• Not a full shopping agent
• Not a checkout system
• Not a real-time ecommerce crawler
• Not dependent on an external LLM API
• Not a benchmark report generator for research papers

Quick Checks

uv run ruff check .
uv run pytest
uv build
python examples/integrations/tool_calling_agent_demo.py
python examples/integrations/run_headphones_dogfood.py

RecHarness v0.1.0

Initial v0.1 release of RecHarness: an agent-agnostic harness for reliable product recommendation by general-purpose agents.

Highlights:

• Pydantic schemas for products, constraints, preferences, bundles, reports, and traces
• JSONL catalog loading, validation, and coverage stats
• Rule-based preference parser
• Constraint verifier and pattern-based claim verifier
• Keyword, attribute-filter, and hybrid retrieval
• Simple transparent ranker
• RecHarness.assist() and verify_agent_recommendation() SDK flows
• CLI commands: catalog validate, assist, verify, eval, mcp serve
• JSONL trace logging
• Batch evaluation runner
• Optional MCP integration
• Example backpack catalog, missions, and agent outputs
• CI, docs, changelog, and contribution guide