dsh-jev
Plugin suite adding Jev judgments to DeepSeek Harness: tool pruning, dead-loop blocking, a safety guard and skill routing.
Overview
dsh-jev (the Cordis plugin suite) adds a semantic judgment layer to DeepSeek Harness using Jev's Noul, Choice and Score primitives at about 150 ms. Modules include a batched client, a loop guard that judges whether each step makes progress and injects a corrective hint, a safety guard that reviews shell and file operations for risks such as rm -rf, privilege escalation or credential leakage and denies or asks, a tool pruner that injects only the top relevant tools, a skill router that recommends one skill, an optional result shaper, and agent tools (jev_ask, jev_rank, jev_check) that hand the primitives to the model.
The README is unusually candid: all repository gates pass and each surface works, but an A/B pilot on 20 real DeepSWE tasks found no advantage and no stable harm, no net token saving and a weak, mixed wall-time gain, with run-to-run variance larger than any measured effect. It describes the layer as fast, cheap and functional with unproven end-to-end payoff, and ships the report and a reproduction kit.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Did this step make real progress (or is the agent looping)?(inferred) | yes · no |
| Choice | Should this shell or file operation be allowed, denied or sent for approval?(inferred) | allow · deny · ask |
| Score | How relevant is each tool to the current intent?(inferred) | — |
How it works
- Hooks fire at tool pre/post-execute and prompt assembly
- Batched Jev questions judge progress, risk and relevance
- Guards deny, ask or inject hints; pruner trims tools
- Agent can call jev_ask/jev_rank/jev_check directly
Strengths
- Fast, reproducible verdicts
- Comprehensive tests and mutation checks
- Published A/B evidence and reproduction kit
Caveats
- A/B pilot showed no measurable outcome improvement
- Specific to DeepSeek Harness
Getting started
Install the dsh-jev npm package as DeepSeek Harness plugins per the README.At a glance
- Type
- Plugin
- Status
- beta
- Language
- TypeScript
- Runtime
- DeepSeek Harness plugins
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Jev (System One decision model) plugin suite for DeepSeek Harness (dsh)
logicrw/awesome-jev-projectsJev (System One decision model) plugin suite for DeepSeek Harness (dsh).
wh000wh000/awesome-jev-liveJev (System One decision model) plugin suite for DeepSeek Harness (dsh).
hellogumbo/awesome-jev
Frequently asked questions
Does it improve agent outcomes?
Its own A/B pilot on 20 tasks found no advantage and no stable harm.
What modules are included?
A client, loop guard, safety guard, tool pruner, skill router, result shaper and agent tools.
How fast are decisions?
About 150 ms each per the README.
Related projects
Python library adding a Jev-based System 1 layer to agents for intent routing, tool safety checks and stuck-loop detection.
ProjectbetaPythonDiscontinued DeepSeek Harness decision layer whose policies let Jev only narrow permissions and fail closed; successor is jevcore.
ProjectprototypeTypeScriptDeepSeek Harness plugin with jev_decision tools, an opt-in risk and loop auto-guard, and a live self-test benchmark against the Jev API.
BenchmarkbetaTypeScriptSupervisor for unattended coding agents: one batched Jev Choice over the trajectory decides continue, pause or escalate.
ProjectbetaPythonAsync Python intent and tool router: Jev picks the handler, and a code-side confidence gate stops weak or invented decisions.
ProjectbetaPythonStop hook for coding agents where Jev checks each attempted stop against plain-language rules and sends the agent back if one is violated.
HookbetaPython