Skip to content

Dsh Context Curator

A DSH compaction backend that swaps only the summarize() hook: instead of asking a model to rewrite history into a summary, it asks Jev two yes/no questions per tool call and drops only the stale ones, keeping user and assistant prose verbatim.

Stars
2
Language
JavaScript
License
MIT
Published by
Community
Author
LXBWOW
Last verified
Oct 2026

Highlights

  • Drop-in for @deepseek-ai/dsh-compaction-basic: same trigger timing, retention policy, log transactions and surface replacement.
  • Ships in shadow mode by default, so the whole pipeline runs and logs decisions while DSH still uses its own summary.
  • First message in the span, the last N messages, and any result carrying an image or attachment are never deleted.
  • Eleven named fallback reasons hand the compaction back to DSH's native summarizer instead of degrading the session.

Watch out

The install script copies the shipped `standard` preset and edits one line, so a DSH upgrade will not update that copy and you have to re-run the script. The preset references the plugin by absolute path, so moving the directory means reinstalling. Reasoning blocks are outside the curation range entirely, so they are neither counted nor preserved. Replacing earlier history can invalidate the provider's prompt cache from the first changed token onward; the plugin records cache read and write tokens so you can check whether the token savings survive the cache rewrite, but it does not prevent the rewrite. The algorithm is ported from tamaratran/fast-jev-compaction, so treat upstream as the source of truth for the core logic. Verification is deliberately staged: in shadow mode the plugin's own checklist explicitly marks context size, context survival and continued agent work as not exercised, because DSH's own summary is what gets submitted. 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.

More Jev resources like this

2GitHub stars
An MCP server packaged as both a Codex and a Claude Code plugin, exposing eight read-only tools that turn Jev probabilities into proceed/review recommendations for routing, ranking, extraction and claim checks.
Jev Agents & MCP#community#awesomejev#catalog-import
5GitHub stars
An MCP server exposing a single judge tool that answers noul, choice and score questions about any JSON application state in one fast request, backed by TypeSafe's System One model with an opt-in Cloudflare Workers AI provider.
Jev Agents & MCP#community#awesomejev#catalog-import
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
72GitHub stars
MCP server over stdio that gives coding agents fourteen Jev judgment tools, turning evidence plus an enumerable question into auto, review, or escalate actions from calibrated probabilities.
Jev Agents & MCP#community#jevbest#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
6GitHub 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
Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai