Limitations

krino v0.1 is experimental. These are the limits you should know before you use it.

• TypeScript only. Agents in Python or Go cannot use krino today. The trace format is language-neutral, so a port could reuse the CLI.
• Claude Agent SDK: run-start selection only. The SDK gives hooks, not the loop. krino picks tools once per run, and its agreement metric is per run, not per step.
• Step-0 selection only. Tasks that need a new tool later in the run get no saving. This is the price of keeping the prompt cache intact.
• The risk gate is shadow-only. It records what it would do. It never blocks a call.
• Local JSONL files do not suit serverless. Traces live on one machine's disk, with no cross-machine view.
• Shadow decisions cost a little. You pay for decisions you do not use yet. The report shows that spend.
• Short processes can cut decisions off. A process that exits before krino gets an answer records the decision as cutOff. Call flushAll before exit.
• One real decision provider. Jev is the only real provider in v0.1. Its price, API, or availability can change.
• Token budgets are estimated. krino counts tokens as characters ÷ 4, with a 10% margin, so context trimming is approximate.
• APIs may change in any release.

The full list, with what each one costs and when to revisit it, is in "Known trade-offs" in the architecture document.