pizzint-demo

Demo project for the Intelligent Control Process (ICP): runs the packaged pizzint-skill through an approved LLM provider (Anthropic, OpenAI, or local Ollama), governed by the .control/ directory.

Layout (ICP future-state pattern)

.control/          governance: session start, change tracker, decision log,
                   next-session queue, change types, review loops
skills/
  pizzint-skill/   the packaged skill (unmodified)
src/
  assistant_skills/  helper the skill imports
  skill_runtime/     SkillRequest / SkillContext / SkillResult contracts
  pizzint_demo/      providers, data provider, loader, redaction, CLI
tests/             unit tests
standards.toml     project standard: approved providers' models (DEC-002)

Run

No dependencies — Python 3.10+ stdlib only.

cd pizzint-demo
export PYTHONPATH=src

# Fully offline (no model, skill output is the briefing)
python -m pizzint_demo "what's the pentagon pizza index?" --provider none

# Local Ollama (default provider; model from standards.toml)
python -m pizzint_demo "any pizza activity spikes?"

# Anthropic / OpenAI (keys from environment only — DEC-003)
export ANTHROPIC_API_KEY=...   # never committed; .env is gitignored
python -m pizzint_demo "explain pizzint" --provider anthropic

# Options: --live (fetch public dashboard, falls back to simulation),
#          --seed N (vary the simulated snapshot), --archie (persona)

Tests

python -m unittest discover tests -v

Governance quick reference

• Approved providers only (DEC-001) — an unapproved provider name is rejected in code.
• Model per provider comes from standards.toml (DEC-002); --model warns about deviation.
• Secrets from environment; outbound logging passes a redaction filter (DEC-003).
• Skill guardrail disclaimer is enforced even if the model drops it (DEC-006).
• Current work state: .control/CHANGE_TRACKER.md · decisions: .control/DECISION_LOG.md