Conn-Pod
A library that splits AI work across different models: one executes, another verifies. Each role has declared read and write scope, bounded retries and an evidence receipt tied to the version. Every task runs in its own Git worktree, under a sandbox.
- Context
- I already had three projects calling AI models, each with its own way of executing, reviewing and keeping evidence. What was missing was a single contract any of them could use.
- Pressure
- The model that writes code cannot be the one that approves it. And an agent with disk access is attack surface: the boundary has to be verifiable, not agreed upon.
- Decision
- Split executor and verifier by model, run each task in its own worktree and keep writes with the host, which then checks that only the declared paths changed. A subscription is the default profile; paid API usage is opt-in only.
- Evidence
- 143 green tests in the second slice. In the live smoke run one model executed and another verified with a verified status, and the verifier got EPERM when trying to read HOME, credentials and the network.
- Impact
- Jarbas already consumes the library. The next step is routing each microtask to the smallest model that can handle it, with measured calibration.