DSPy System One agent patterns

A series of six small Python files that build one coding agent, each adding a single design pattern on top of the previous one, using DSPy's RLM sandbox and TypeSafe's Jev for closed-question decisions.

The signature states the question, Jev returns the probability, and plain Python owns the policy. Jev never writes text; it answers closed questions with probabilities, and named thresholds in ordinary Python turn those probabilities into actions.

The README lists: is this command safe, which tool fits, how much of this past turn matters, and could a helper take this step. For example, IsCommandSafeToRun sits inside run_shell_command and commands below DENY_BELOW are denied by default.

Create the venv and install dspy[typesafe]==3.4.0 with dspy-monty-interpreter, pytest and python-dotenv, put OPENAI_API_KEY and TYPESAFE_API_KEY in .env, then run agent_v0_baseline.py through agent_v5_final.py in order and finally fit_thresholds.py.

Monty is the sandbox the frontier model writes code inside: it starts in microseconds and has no filesystem, network or environment access, so sandboxed code can only reach the outside world through the tools the project registers.

DSPy caches identical requests, so a second run replays the first run's answers in seconds; agent_v5_final.py switches the cache off so its numbers are real. The README gives no cost figures for the demos.

The owner is cmpnd-ai. The repo has 38 stars and 0 forks, is written in Python, and was created and pushed 2026-09-30. GitHub metadata lists no licence, and the digest excerpt is truncated inside the repository layout, so licensing is unconfirmed.