jev-harness
Pi agent harness where Jev routes the model tier, picks runbook sections, gates tool calls and verifies reports on Neon Postgres.
Overview
jev-harness by promtengineer is an agent harness built around a System One model. Pi runs the agent loop, a Gemini model does the work, and Jev makes the small decisions at four hooks: a router choosing a fast or powerful tier once per run (keeping the prompt cache warm), a context picker asking one yes/no per runbook section so only relevant sections join the system prompt (22 questions in one call), a gate rating each tool call as read, additive, modify or destroy and whether it touches another branch, and a verifier asking how complete the work is and whether every claim is backed by a tool result.
The gate returns allow, checkpoint (create a Neon branch first), ask a human, or block, and the verifier can send the agent back once. The agent works on a real Postgres database with a Neon branch per run, and every decision is logged with its numbers. It accompanies a video asking whether System 1 can make agents better, with a benchmark, dashboard and results. The README notes that clipping long tool output was the biggest token driver, though published results were measured without it.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Fast or powerful tier for this request? | fast · powerful |
| Yes / no | Does this runbook section apply to the ticket? | — |
| Choice | What is the worst this tool call does? | read · additive · modify · destroy |
| Yes / no | Is every claim in the report backed by a tool result? | — |
How it works
- Before start, route the tier and select runbook sections
- On each tool call, gate to allow, checkpoint, ask or block
- At the end, verify completeness and claims; retry at most once
- Log every decision
Strengths
- Neon branch checkpoints before risky calls
- Selective context loading
- Benchmark and dashboard included
Caveats
- Requires Neon, Gemini and TypeSafe access
- Results tied to one benchmark
At a glance
- Type
- Library
- Status
- research
- Language
- TypeScript
- Runtime
- Pi agent extension
- Has code
- yes
- License
- MIT
- Stars
- 9
Use cases and integrations
How curated lists describe it
A Pi agent harness built around TypeSafe's Jev (System One model): router, context picker, gate, verifier. Tested on Neon Postgres branches.
logicrw/awesome-jev-projectsA Pi agent harness built around TypeSafe's Jev (System One model): router, context picker, gate, verifier. Tested on Neon Postgres branches.
wh000wh000/awesome-jev-live
Frequently asked questions
What does Jev decide?
Model tier, which runbook sections apply, tool-call risk and whether the final report is supported.
What is a checkpoint?
Creating a Neon database branch before a risky call.
Who owns thresholds?
Code; Jev only supplies numbers.
Related projects
DeepSeek Harness plugin with 12 opt-in Jev or Luna Decisions judgments: skill and file selection, supervision, filtering and approval help.
PluginbetaTypeScriptVerified computer-use agent: a Jev guard checks claimed completions and gates irreversible actions; Jev also takes simple steps.
ProjectbetaPythonClaude Code plugin adding typed Jev decisions to hooks: task kind, command risk, failure class, completion; never grants permissions.
PluginbetaJavaScriptPi-based coding agent where Jev gates every tool call, picks the model tier, checks 'done' claims and trims tool output.
ProjectbetaTypeScriptAuditable Zig terminal agent pairing Codex for coding with Jev typed preflight and postflight checks under deterministic policy gates.
CLIproductionZigMCP server giving coding agents fast typed decisions (ship gates, risk checks, file triage, page checks) as accept, review or escalate.
MCP serverbetaTypeScript