Show HN: Decide – Jev decisions in the shell

Straight from the project description, with no methodology attached. The thread carries one reply that does not test either claim, and the entry copy records only that the topic is shell-native decisions instead of parsing JSON from LLM replies. Treat the multipliers as marketing until you measure your own calls.

Daily personal tasks. The single commenter says they would not reach for a decision model for their own work because an LLM does everything better outside the cost advantage, while using it to sort things cheaply and fast inside their own app, Greer, and expecting vendors to ship equivalents.

The body documents a scriptable scheme: exit 0 is decided or yes, 1 is no, 2 is unsure, 10 is your mistake, 11 is a network or model problem, with --min-confidence and --fallback for gating on confidence. That is enough to branch on uncertainty, though no example of a tuned threshold is given.

From the body: a TypeSafe or OpenRouter key is required, pre-built binaries cover macOS 26 and up while macOS 15 and Linux build from source needing Swift 6.2, and Jev is the only supported decision model so far. The entry copy adds that this was a small launch with low point totals.

Both are flags: --min-confidence sets the gate and --fallback supplies what to do when it is not met. The exit-code scheme is the other half of the story: 0 is decided or yes, 1 is no, 2 is unsure, 10 is your mistake, and 11 is a network or model problem, so a shell script can branch on uncertainty rather than parsing text.

Context comes from --context @file with a 32k context window, and JSON is supported on both sides: context can be read as JSON objects and decisions written back as JSON objects. The --each flag decides once per line from stdin, which is what makes JSONL pipelines work, and named questions with --name keep the output easier to parse when you ask several at once.

An example agent skill for getting cursory information about files cheaply before or instead of reading them, plus a Swift library called DecisionModels that decide is built on, which is modeled on Apple's FoundationModels library and is Apache 2.0. The CLI itself is Swift 6.2, and it claims token efficiency with no MCP needed.