JevHarness
JevHarness is a Python project and agent skill for creating task-specific Jev harnesses with an authoring LLM. It can optionally use rewards and complete execution traces to improve a harness, then freeze the selected strategy so runtime decisions use the harness code and Jev calls without an authoring LLM on every decision.
What it does
Developers can use it to build and evaluate task-specific decision systems, such as a support-ticket router, with an LLM authoring the harness and Jev making runtime decisions.
- Availability
- Public GitHub repository; installation instructions for Claude Code and Codex
- Pricing model
- Not established
- Deployment
- Python project; frozen harnesses run through PipelineRuntime
- Privacy
- The archived battle log stays in the browser and is not uploaded to a replay server
- Licence
- Not established
Capability
Modality
Access
Deployment
Pricing
Target user
Evidence
5 claims- The repository provides installation instructions for Claude Code and Codex. [1]
The harness preserves the LLM's reasoning strategy in explicit code, features, state, instructions, criteria, and control flow.
GitHub new repositories. Quoted text is outside CC BY 4.0 and keeps its original rights. - A harness combines task observations, feature construction, Jev questions and criteria, and actions. [1]
A harness turns task observations into useful features, constructs Jev questions and criteria, and combines the structured answers into actions.
GitHub new repositories. Quoted text is outside CC BY 4.0 and keeps its original rights. - Optional GEPA integration uses rewards and execution traces to propose and evaluate harness changes. [1]
When you request optimization and reliable feedback is available, it uses the execution trajectories and rewards for reflection, then freezes the selected harness for reuse.
GitHub new repositories. Quoted text is outside CC BY 4.0 and keeps its original rights. - Frozen harnesses can execute without the authoring LLM on every decision, but still need Jev if they contain Jev nodes. [1]
Once the selected harness is frozen, its code and Jev calls execute without the authoring LLM on every decision.
GitHub new repositories. Quoted text is outside CC BY 4.0 and keeps its original rights. - The archived battle log stays in the browser and is not uploaded to a replay server. [1]
The archived battle log stays in the browser and is not uploaded to a replay server.
GitHub new repositories. Quoted text is outside CC BY 4.0 and keeps its original rights.
Sources
1
1
GitHub new repositoriesOwner-controlled
https://github.com/TianyuCodings/JevHarness
Retrieved 2026-10-27 01:40 UTC
Verified
Versions
5| Version | Effective | Change | Changed fields | State |
|---|---|---|---|---|
| 5 | 2026-10-22 00:40 UTC | evidence update | score | Current |
| 4 | 2026-10-17 21:34 UTC | material change | capability, access, deployment, deployment, privacy | Current |
| 3 | 2026-10-16 22:31 UTC | evidence update | score | Current |
| 2 | 2026-09-25 18:45 UTC | material change | capability, access, deployment, availability, deployment | Current |
| 1 | 2026-09-24 19:44 UTC | publication | None | Current |
In briefs
2- New models and tools focus on typed decisions and long-context tasks Weekly brief 2026-W39
- Claude Platform adds automatic evaluation for Managed Agents tool calls Daily brief 2026-09-24
Original catalog content is licensed under CC BY 4.0, attributed to wowsuchai. Source quotations and linked content keep their original rights.