Add guardrails hard rules recipe

[webchick]

Summary

• Adds agents/guardrails_hard_rules_python/ — a new recipe demonstrating the post-LLM guardrail pattern
• The LLM classifies user-submitted content as safe / review / block; deterministic hard rules can override to block regardless of what the LLM said
• When a hard rule fires, the LLM's original reasoning is preserved in the returned verdict so every decision is auditable

What it teaches

The core insight: for any decision with real consequences, you shouldn't rely on the LLM alone. Hard rules catch unambiguous violations deterministically; the LLM handles the grey zone. This recipe shows how to layer the two cleanly — and how to preserve auditability when the deterministic layer wins.

Fills the Guardrails gap from the cookbook wishlist. Sourced from the classifiers/_helpers.py pattern in dependency-scout, simplified to a generic content moderation scenario.

Recipe structure

agents/guardrails_hard_rules_python/
├── models/signals.py          # ContentSignals input model
├── models/verdict.py          # LLMVerdict + Verdict (with overridden_by_hard_rule)
├── guardrails/hard_rules.py   # _hard_block() + apply_hard_rules() — pure functions
├── activities/classify.py     # Calls Claude via tool_choice, then apply_hard_rules
├── workflows/classify_workflow.py
├── worker.py
├── start_workflow.py          # Two examples: override fires, then LLM verdict stands
└── tests/test_guardrails.py   # 10 tests: unit, activity (mocked Anthropic), workflow

Test plan

• uv run pytest tests/ --timeout=30 → 10/10 passing, no API key needed
• ANTHROPIC_API_KEY=... uv run python -m worker + uv run python -m start_workflow — verify output matches README

🤖 Generated with Claude Code