jev-agent-kit
Typed judgment layer for Claude Code, Codex and Hermes Agent: deterministic facts in code, language judgments in Jev, with receipts.
Overview
Jev Agent Kit is a Python toolkit for the narrow, frequent decisions inside an agent loop that do not justify another LLM turn: is this task done, which handler fits, does the final message claim the tests passed, does this tool call look destructive. Its principle is to keep deterministic facts in code and use Jev only for judgments that depend on language. For example, code counts passed tests from JUnit reports while Jev checks whether the summary claims they passed; code parses a shell command for recursive deletes and force flags while Jev judges destructiveness given those facts and the stated purpose.
On top of TypeSafe's API it adds question sets with declared, minimized input fields, deterministic facts computed before each call, a pinned model, thresholds tied to one question and model version, a receipt for every decision, calibration tooling, an optional routing policy, and hooks plus a skill for Claude Code, Codex and Hermes Agent. Calibration is available but not required.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does the summary claim the tests passed? | yes · no |
| Yes / no | Given the parsed facts and stated purpose, is this command destructive? | yes · no |
| Choice | Which handler fits this request: code, a model, or a person? | code · model · person |
How it works
- Compute deterministic facts in code
- Send minimal declared fields and the question to a pinned Jev model
- Apply per-question thresholds or the built-in policy
- Record a receipt; act, escalate or ask a person
Strengths
- Clear split between code facts and model judgment
- Receipts and pinned versions make decisions auditable
- Hooks and skill for three agent hosts
Caveats
- Requires a TypeSafe key
- Thresholds need tuning per question and model version
Getting started
`pip install -e .`, `export TYPESAFE_API_KEY=...`, then `python examples/verify_claim.py`.At a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- library, agent hooks and skill
- Has code
- yes
Use cases and integrations
How curated lists describe it
A typed judgment layer for Claude Code, Codex and Hermes Agent using TypeSafe.
wh000wh000/awesome-jev-live
Frequently asked questions
What does the kit decide?
Language-dependent judgments like whether a claim matches evidence or a command is destructive.
What stays in code?
Deterministic facts such as test counts and parsed command flags.
Is calibration required?
No, it is optional tooling.
Related projects
MCP server giving coding agents fast typed decisions (ship gates, risk checks, file triage, page checks) as accept, review or escalate.
MCP serverbetaTypeScriptOpen-source Jev skills for Claude Code, Codex, Cursor and Grok: one client to gate, route, guard, verify, triage, review and act.
ProjectbetaTypeScriptProject memory for Claude Code that asks Jev whether each chat statement is worth keeping and checks tool calls against saved rules.
ProjectbetaTypeScriptClaude Code plugin adding typed Jev decisions to hooks: task kind, command risk, failure class, completion; never grants permissions.
PluginbetaJavaScriptMIT plugin and local MCP server that hands an agent's closed questions to Jev or local Laya, with policy gates and receipts.
Agent skillproductionPythonIntent compiler skill that turns vague coding requests into a typed IntentSpec, with Jev or LAYA as optional decision backends.
Projectbeta