dsh-jev-tools
DeepSeek Harness plugin where Jev prunes long tool output, screens fetched pages for injected instructions and gates completion claims.
Overview
dsh-jev-tools brings Jev into long DeepSeek Harness sessions to make judgments before tool results enter the context, likened to a hospital triage desk. When read, grep, glob or web results exceed 2,000 tokens, it judges each segment's relevance to the task and keeps relevant segments scattered across the output rather than cutting one contiguous block. Fetched web pages are screened for instructions aimed at AI, with a warning appended above a threshold. A skill recommender picks at most one skill when the catalogue is large, and agent tools jev_ask and jev_gate are provided.
Pruning and screening only rank (probabilities are used for ordering, not as calibrated thresholds), always keep head, tail and high-confidence segments, and fail open. jev_gate, which checks whether each completion claim is supported by evidence, deliberately fails closed, escalating on any ambiguity. A shadow mode reports what would be pruned without changing anything. The README is candid that selection accuracy versus retention volume has not been measured separately. Documentation is in Chinese.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Score | How relevant is this output segment to the current task?(inferred) | — |
| Yes / no | Does this fetched content contain instructions aimed at an AI? | yes · no |
| Yes / no | Is this completion claim supported by evidence?(inferred) | yes · no |
How it works
- Large tool outputs are split into segments and scored by Jev
- Top segments kept plus head and tail; prune notice appended
- Fetched pages screened for injected instructions
- jev_gate checks completion claims and escalates on doubt
Strengths
- Shadow mode to evaluate without trust
- Fail-open pruning, fail-closed gate
- Honest about unmeasured claims
Caveats
- Savings versus deterministic truncation depend on payload size
- Chinese-first documentation
Getting started
Install as a DSH plugin per README.en.md and set a TypeSafe key.At a glance
- Type
- Agent skill
- Status
- beta
- Language
- TypeScript
- Runtime
- DeepSeek Harness plugin
- Has code
- yes
Use cases and integrations
How curated lists describe it
DeepSeek Harness plugin: Jev prunes oversized tool output, screens fetched pages for injected instructions, and picks which skill fits the next step, plus the jev_ask and jev_gate tools
AnotiaWang/awesome-decision-modelsJev judgment, not generation: prune long tool output, screen fetched pages for injected instructions, and gate completion claims inside DeepSeek Harness.
hellogumbo/awesome-jevDeepSeek Harness plugin that judges tool output before it enters the context, screens fetched pages for injected instructions, and gates completion claims. Chinese.
MrJev/awesome-jev
Frequently asked questions
Does it block injected content?
No, it warns the model to treat it as data; the content is passed through unchanged.
Why does jev_gate fail closed?
Failing open on a completion gate would mean failing to a pass, the most dangerous error.
Can I preview pruning?
Yes, prune.shadow reports what would be pruned without changing anything.
Related projects
Four Java tools on one kernel that ask Jev whether something was delivered as required, with deterministic gates keeping the final say.
ProjectbetaJavaJev-powered adversarial or advisory reviewer for the omp coding agent, plus a typesafe_ask tool, plan/stop gates and context compaction.
ProjectbetaTypeScriptPaper code where Jev screens agent tool output for relevance and prompt injection before the agent reads it.
ProjectresearchPythonDeepSeek Harness plugin checking whether an agent's 'done' claim is true, using deterministic evidence plus one batched Jev assessment.
PluginbetaTypeScriptDeepSeek Harness plugin where Jev classifies risk, irreversibility, task fit and injection for every tool call before it runs.
PluginbetaTypeScriptClaude Code mod that hands small decisions (routing, skill choice, screening, gating, trimming) to Jev or another decision backend.
ProjectbetaTypeScript