dsh-jev
Discontinued DeepSeek Harness decision layer whose policies let Jev only narrow permissions and fail closed; successor is jevcore.
Overview
This dsh-jev repository was a decision layer for DeepSeek Harness that used Jev to restrict visible tools, assess calls before they ran and route between models. Development was discontinued in October 2026; version 0.1.4 is the last, tested only against an early alpha of DeepSeek Harness. The README directs users to jevcore by PerryLink, which provides a harness-independent core, a DSH plugin and an MCP server.
The repository remains available under MIT and documents ideas others are invited to adopt: monotone permission narrowing, where Jev can only restrict and call assessments compose as deny over hold over ask over allow, with errors and timeouts failing closed rather than allowing; a failure policy that rejects allow at configuration time; shadow mode and an offline mock as defaults; and versioned benchmarks, such as a protected audit trail blocked in 10 of 10 runs.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Should this tool call be allowed, asked, held or denied? | allow · ask · hold · deny |
| Choice | Which tools should be visible for this step?(inferred) | — |
How it works
- Before a step, Jev narrows the visible tool set
- Before a call, Jev assesses it
- Assessments compose with existing policies (deny > hold > ask > allow)
- Errors or timeouts fail closed
Strengths
- Permissions can only narrow, never widen
- Shadow mode and mock provider by default
- Documented benchmarks
Caveats
- Discontinued; no further releases
- Tested only on an old DSH alpha
At a glance
- Type
- Project
- Status
- prototype
- Language
- TypeScript
- Runtime
- DeepSeek Harness plugin
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
*(archived)* - Decision layer for DeepSeek Harness whose failure policy rejects the value `allow` at configuration time, from plain JavaScript as well as TypeScript. Its verify script installs the built tarballs into a fresh consumer before testing them.
MrJev/awesome-jev
Frequently asked questions
Is it maintained?
No, it was discontinued in October 2026; the README recommends jevcore.
What is monotone permission narrowing?
Jev may only restrict permissions, and failures fail closed rather than allowing.
Can I still use the code?
Yes, it stays available under MIT.
Related projects
Plugin suite adding Jev judgments to DeepSeek Harness: tool pruning, dead-loop blocking, a safety guard and skill routing.
PluginbetaTypeScriptDeepSeek Harness plugin where Jev classifies risk, irreversibility, task fit and injection for every tool call before it runs.
PluginbetaTypeScriptStrict-allowlist shell harness where Jev walks a TOML decision tree to pick a bash command, flags and paths, confirming writes.
ProjectprototypePythonAuditable Zig terminal agent pairing Codex for coding with Jev typed preflight and postflight checks under deterministic policy gates.
CLIproductionZigGateway that checks agent tool calls with policy plus Jev evidence, then binds approval to a single-use permit for that exact action.
MCP serverbetaTypeScriptMCP proxy and hooks that screen agent tool calls before they run and their results before the agent reads them, using Jev judgments.
MCP serverbetaTypeScript