Expand description
§jevdev
A coding-agent harness built around Jev, TypeSafe’s System One decision model.
The harness keeps all agent state as explicit, typed, content-addressed chunks
(state) and asks Jev the per-turn questions a coding agent normally leaves
to defaults (jev::questions):
| Decision point | Question | Typed answer |
|---|---|---|
| Context | How visible should this chunk be for this query? | hide / short / long / full |
| Cache | Reuse the cached prefix or rebuild? | noul + probability |
| Routing | Can this subtask leave the frontier model? | choice + cost estimate |
| Tools | Which tool fits this intent? | ranked choice, top-k |
| Permissions | Should this command run? | allow / ask / deny |
| Security | Which files will this task touch? | sensitivity score |
Frontier models (llm), tools (tools), and deterministic code do the
actual work. Context is assembled per query (context), never appended;
routing is priced per context rebuild (router); commands pass a Cedar
policy plus Jev deep inspection (policy); and the turn loop lives in
runtime, with a terminal UI in tui.
Modules§
- config
- Harness configuration, loaded from
jevdev.tomlin the repository root. - context
- Meta-attention: context as a per-query decision.
- jev
- The Jev client.
- llm
- Frontier and helper models.
- policy
- Programmable permissions.
- router
- Cost-priced, security-aware model routing.
- runtime
- The turn loop.
- state
- Explicit, typed, content-addressed state.
- tools
- Tools with tiered disclosure.
- tui
- Terminal UI.
Constants§
- VERSION
- Crate version, as compiled.