Rivulet agent lab

Rivulet agent lab is the shared, dependency-light Python layer behind the Rivulet operating wiki. It gives employees and agents small, testable building blocks for work that must remain reviewable before any live system is touched.

The package supports:

• task contracts with explicit authority, sources, scope, reviewer, and stop conditions;
• approval classification for externally visible or irreversible actions;
• state records that distinguish prepared, approved, executed, verified, blocked, stale, and superseded work;
• Markdown briefings and handoffs for human review; and
• safe workflow fixtures for local tests and demonstrations.

It deliberately does not authenticate, send messages, modify access, deploy, spend money, or call a customer system. Project repositories own any approved live integration and must retain their own readback behavior.

The package is available under the MIT License. For the operating model and adaptation guidance, see the companion agent-wiki-reference repository. Read CONTRIBUTING.md and SECURITY.md before opening a change or security report.

Start here

PYTHONPATH=src python3 -m rivulet_agents.cli catalog
PYTHONPATH=src python3 -m rivulet_agents.cli gate publish
PYTHONPATH=src python3 -m rivulet_agents.cli brief \
  --title "Review a source-quality trend" \
  --outcome "A source-labeled draft briefing" \
  --owner "Data delivery owner"
python3 -m unittest discover -s tests -v
python3 scripts/safety_scan.py
python3 scripts/validate_package.py

The gate command never executes an action. It makes the required approval boundary explicit. See the code-to-wiki map for ownership and integration routes.