dsh-jev-judge
面向编码代理的判断内核：由小型 Jev 判定器回答范围明确的问题，让主模型保留上下文和令牌，用于真正的推理。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:HunterXing/jev-judge#97df1228952ac6dd670726ca2d45dcad262e5678Jev Judge 以 dsh-jev-judge 发布，当前版本为 0.1.4。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 简体中文
A judgment kernel for coding agents. The small, frequent calls an agent makes — does this tool output matter, did anything actually verify that claim, does this page carry instructions aimed at the model, did the user really ask for this command — go to a small judge instead of the main model. The agent keeps its context and its tokens for the work that needs reasoning.
Read the scope before installing. This layer treats a context budget, not code quality or execution speed: it takes no part in compiling, caching or optimising, and changes only what the model sees. Measured over 55 days of real sessions on the machine it was written on — context windows of 262k–1M, zero pressure-triggered compactions, zero injection hits, and 0.4/day of the one decision point that maps onto an outcome — it was uninstalled there. See R10 of
docs/verification.md. It earns its place when the host's context really is scarce, or when the output is mostly scanned rather than worked on: shell logs, introspection dumps, browser snapshots.
It runs in DeepSeek Harness as a plugin bundle, in Claude Code, Codex and OpenCode through command hooks or MCP, and anywhere else that can start an MCP server or run a command.
An agent loop makes dozens of small judgments per run. Handing every one to the main model is slow and expensive; encoding every one as a fixed rule breaks on real input. This package is the layer in between: one bounded question over a small state, answered by a small judge, consumed by code that owns the thresholds, the fallbacks and the final action.
The design follows mu, which loads a judgment
kernel into the pi coding agent. What this project adds is reach: the kernel is
host-free, so the same decision points run in agents that mu does not cover.
docs/mu-analysis.md is the study, including the wording
discipline and the measurements behind it.
dsh plugin --profile <profile> add dsh-jev-judge
dsh --profile <profile> --dump-config # shows the layer
Or install it from Settings → Plugins in the Web GUI, by registry name, git address or local path.
settings.json / hooks.json.claude mcp add jev-judge -- npx -y dsh-jev-judge mcp
docs/agents.md has the exact configuration for each host and
what a verdict is able to do in each.
| Decision point | The question | Default |
|---|---|---|
judge.items | One yes/no question asked of many items — hundreds of log lines, files or findings, without reading them all | on when called |
tool.admission | Which chunks of a long tool result are needed for this task, so the rest never enters the context | active |
tool.injection | Which passages of fetched content read as instructions aimed at the model rather than material for a reader | active |
turn.completion | Does this "it's done" claim have anything that verified it | active |
turn.continue | Did the turn stop in front of a step it announced, or ask for a go-ahead it already had | active |
memory.capture | Is this message a correction worth keeping, or just the next task | active |
tool.risk | For a command the rules already flagged: did the user actually ask for this | shadow |
Every point carries a deterministic fallback equal to the behaviour without this
package, and every verdict is written to a ledger — which is how a point earns
active from real sessions instead of from optimism.
tool.risk starts in shadow on purpose: its verdict can only loosen policy, so
it has to earn that from your own ledger before it is allowed to.
The provider is a private record you own, in the format published by the
typesafe-ai-jev-skill
Agent Skill:
$JEV_SKILL_CONFIG → ~/.config/typesafe-ai-jev-skill.json
{
"providerName": "your-provider",
"baseUrl": "https://api.your-provider.example",
"model": "jev",
"apiKeyEnv": "JEV_API_KEY",
"protocol": "systemone",
"endpointPath": "v1/systemone",
"authHeader": "Authorization",
"authScheme": "Bearer",
"timeoutSeconds": 30
}
jev-judge verify # offline; reports the contract, never the key
jev-judge doctor # what will happen: judges, modes, ledger
Kernel settings, all optional, live in ~/.config/jev-judge/kernel.json:
{
"modes": { "default": "shadow", "tool.admission": "active" },
"routes": { "turn.continue": ["jev"] },
"tiers": ["jev"],
"ledger": { "path": "~/.config/jev-judge/ledger.ndjson", "recordState": false },
"options": { "judge.items": { "maxItems": 20 } },
"timeoutMs": 8000
}
jev-judge ledger reads the ledger; --json prints the records as stored.
tool.admission acts by default only on output whose size nobody chose: a log, a
dump, a listing, a page. The default list:
["bash", "cordis_inspect_*", "mcp__*", "web_fetch", "webfetch", "web_search", "websearch", "fetch"]
read, grep, glob, skill, write, edit — the tools whose shape the
agent asked for — pass through whole. Thinning one of those saves tokens and
buys a re-read that may come back thinned differently. Names match without case
(bash and Claude Code's Bash are the same tool), and a trailing * matches a
prefix.
{ "options": { "admission": { "toolNames": ["bash", "mcp__*"], "minChars": 4000 } } }
In DeepSeek Harness this goes in the plugin row's config.options, i.e. the
profile's cordis.patch.yml.
choice question always carries an escape answer.docs/verification.md).From docs/verification.md, on a real 81-line test log
with two failures, judged by the live provider 20 items per call:
No runtime dependencies and no build step, on purpose: the package loads straight
from a checkout, so dsh plugin add <path> needs no build-script approval.
node --test # 217 tests, no network
node src/cli.js --help
Releases are cut by pushing a version tag: .github/workflows/release.yml
publishes to npm through trusted publishing (no token in the repository) after
the same suite passes. See docs/market.md.
MIT. TypeSafe, System One and Jev belong to their respective rights
holders; this project is not affiliated with them. The judgment-kernel design is
adapted from mu (MIT) — see
ATTRIBUTION.md.