The problem I kept hitting in production systems: AI pipelines are observable, but not reproducible. After an incident, models, routing logic, retries, or fallbacks may have changed — logs alone don’t let you replay what actually happened.

v1.3.0 introduces a runtime determinism core:

Write-ahead log (append-only JSONL) written before side effects

Crash-safe recovery and deterministic replay (fails loud on divergence)

Runtime execution contracts (timeouts, retries, cost ceilings)

Side-effect classification to prevent unsafe retries or fallback

CLI-first inspection (list / show / trace / replay / diff)

It’s not a planner or agent framework, and not a replacement for MCP — it focuses purely on execution semantics around tools (including MCP-style tools).

Quick try (run from repo root):

git clone https://github.com/Balchandar/intentusnet

cd intentusnet pip install -e . python -m examples.deterministic_routing_demo.demo --mode with python -m examples.deterministic_routing_demo.demo --mode mcp

Docs (architecture, guarantees, demos): https://intentusnet.com

MIT licensed, open source: https://github.com/Balchandar/intentusnet

I’d really value feedback from people building real systems:

What guarantees do you expect from deterministic replay in practice?

How do you handle retries and side effects safely in AI pipelines?