User repository on GitHub
Jev-judged context compaction for DeepSeek Harness: semantic tool-result pruning + deterministic receipt compaction
- Forks
- 3
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
Dsh Jev Prune
A DeepSeek Harness plugin that replaces DSH's size-based context compaction with Jev judgments: tool results that are still needed are never trimmed however large, and spent ones get a code-written receipt instead of a model-written summary.
- Stars
- 4
- Language
- JavaScript
- License
- MIT
- Published by
- Community
- Author
- yangyu666
- Last verified
- Oct 2026
About this project
Jev-judged context compaction for DeepSeek Harness. Structured judgments from TypeSafe Jev drive DSH's two-layer context compaction. The compaction algorithms are untouched; the judgment backend is pluggable (Jev / rules / a self-hosted model).
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- JavaScript
Highlights
- Layer 1 asks Jev whether a tool result is still needed; with no judgment available it falls back to DSH's original threshold behaviour.
- Layer 2 removes eligible read-only steps and injects a receipt whose tool name, command, path, character count and seq are computed by code.
- Probabilities are read as relative quantiles within one session, never as a fixed threshold, and small populations degrade to a stricter floor.
- A fencing token per compaction stops an asynchronous DSH compaction from being overwritten by someone else's receipt while pairing stays valid.
Watch out
One task shape and one machine: the numbers come from a three-arm comparison on a 37-step read-every-file run, with layer 1 trimming 4 to 7 nodes per run and uncached input tokens falling from a 75,781 to 87,872 baseline range to 25,516 to 45,071. Cache hits were 78 to 94% of input and hit and miss prices differ by roughly 50x, so total-token comparisons mislead by design. Layer 2 only fires on long runs, DSH's own model summaries remain, receipt runtime strings are Chinese today, and it targets DSH 0.1.x on Node 22.19+. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
The README insists on reporting token classes separately for exactly that reason: cache hits were 78 to 94% of input and hit and miss prices differ by about 50x, so comparing total tokens is misleading by construction. The headline number is uncached, full-price input tokens per completed run, taken from the usage object the API actually returned in the DSH session logs rather than from any estimator.
Under one rule: what should not be generated by a model is not generated by a model. DSH's region compaction asks the model to write a summary, which invites summary hallucination, so layer 2 emits a deterministic receipt from the plugin's own JavaScript. Tool name, command, path, character count and seq are computed, and the model's own words are quoted verbatim with no inference added.
That was a real bug the README documents. Replacing a whole call-plus-result step also removed the assistant message carrying it, so in one long run the model's intermediate notes were erased step by step, 12 of 13, and it stopped issuing tool calls mid-task. Receipts now carry each step's assistant-visible text verbatim, text blocks only with reasoning drafts excluded, bounded by receiptTextChars, default 400, and 0 disables it.
Layer 2 is deliberately conservative and every gate must hold: the is-it-still-needed and did-it-change-state axes must each fall inside the session's trailing quantile, the tool must not be write-type or in neverCompactTools, results matching error, assert, fail or todo are never moved, and the span must save enough characters while the receipt stays under a set ratio of the original tokens. Too few candidates to order means nothing moves at all.
Because the judge is a small model whose output distribution is narrow. The README says a fixed threshold would not carry stable information, so values are consumed as relative quantiles within a single session, where only the ordering between candidates is meaningful. That is why the same probability can act differently across sessions.
The mode degrades to a stricter absolute floor instead of giving up, since a two- or three-item population is too small to order. If the population falls below two candidates, nothing is moved out at all, on the principle that a single sample is not a distribution. The degradation is always reported, never silent.
Version 0.1.1 for DSH 0.2.0-rc.2, which supports both root and preset-local services and passed real CLI installation and activation; the older 0.1.0 is rejected by that host. It needs Node 22.19 or newer and TYPESAFE_API_KEY set before starting DSH, with DSH 0.1.5-rc.2 as the regression baseline.
More Jev resources like this
1GitHub stars
A DeepSeek Harness plugin that exposes System One decisions as agent tools and adds an opt-in auto-guard, with a labeled benchmark it re-runs against the live API.
Jev Agents & MCP#community#jevbest#catalog-import
Community
xienda/dsh-jev-verifyJavaScript
xienda/dsh-jev-verifyJavaScript20GitHub stars
TypeScript recipe library with 248 focused functions that call Jev for routing, reranking, gating and labelling, returning structured ready-or-review decisions for Node backends.
Jev Agents & MCP#community#awesomejev#catalog-import
7GitHub stars
MCP server exposing one evaluate tool for Jev Choice, Noul, and Score questions, plus a separate eval harness that compares Jev against LLMs on shared labeled sets.
Jev Agents & MCP#community#jevbest#catalog-import
Community
arunav25/jev-mcpJavaScript
arunav25/jev-mcpJavaScript6GitHub stars
Skill and CLI that helps Codex, Claude Code, and other agents add bounded Jev judgments for routing, ranking, classification, and next-step choice.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
thehan-co/jevrielJavaScript
thehan-co/jevrielJavaScript6GitHub stars
Shell gate for coding assistants that judges commands by effect and blast radius, asking Jev twelve questions or a chat model, with fast rules first.
Jev Agents & MCP#community#awesomejev#catalog-import
4GitHub stars
OpenClaw plugin adding a jev_route tool that classifies task type, complexity and risk with TypeSafe Jev, then maps the result to a route through testable JavaScript policy.
Jev Agents & MCP#community#awesomejev#catalog-import
Community
trietphan/jev-clawJavaScript
trietphan/jev-clawJavaScriptFrom the Jev community
Posts from builders shipping with Jev right now.