DeepSeek Harness Plugin Hub

发布与管理完整 Harness Profiles,发现适合你的插件。

探索

插件目录环境预设文档中心动态

社区

发布插件联系我们报告问题

相关链接

Plugin Hub GitHubDeepSeek Harness 官方项目系统状态隐私说明
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

独立、非官方社区项目,与 DeepSeek 官方无隶属、授权或背书关系。

Slice Agent Loop — DeepSeek Harness 插件(DSH Plugin)
← Plugins
S

@dsh-external/dsh-slice-agent-loop

Slice Agent Loop

用于内置 DeepSeek Harness 代理循环的持久化有界上下文和召回策略

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:TT-Wang/dsh-slice-agent-loop#c6de4e189b0c187354d717df4bbf6e5688be8c64
README兼容性版本

兼容性与来源证明

Slice Agent Loop 以 @dsh-external/dsh-slice-agent-loop 发布,当前版本为 0.1.0。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。

DSH 兼容范围
*
运行环境
any
发布来源
github
Registry 更新时间
2026/9/8

版本

0.1.0stable
2026/9/8
0.0.1stable
2026/8/25

相关插件

正在加载相关插件…

最新版
0.1.0
DSH
*
HMR
重启进程
Tree shaking
未声明可安全裁剪
解包体积
未提供
文件数
未提供
Surface
any
许可证
BSD-3-Clause
发布源
github
GitHub
★ 7
周下载
0
最近提交
2026/9/17
查看源码 ↗
README Badge

点击下方 Badge 复制 Markdown,粘贴到 README 即可。

这是你的 Plugin?认领权益 · 优先安全扫描

验证 package.json 声明的 GitHub 仓库,即可管理这个公开页面。认领后,Hub 会优先安排当前版本的安全扫描,并在通过后公开展示结果。

认领这个 Plugin →
报告问题
DeepSeek Harness Plugin Hub
ProfilesPlugins分类动态文档登录管理 Profiles
ProfilesPlugins分类动态文档登录

相关插件

继续浏览 memory-context 分类下经过校验的插件。

Memory Plugin@openviking/dsh-memory-plugin适用于 DeepSeek Harness 的 OpenViking 记忆与上下文套件Weknora@wxg-prc-cpg/dsh-weknora适用于 DeepSeek Harness (dsh) 的 WeKnora 知识检索工具:通过自有知识库进行语义搜索、文档阅读以及 RAG/代理回答。Memsearch Dsh@zilliz/memsearch-dsh适用于 DeepSeek Harness 的 MemSearch 插件:在多个代理之间共享 Markdown 记忆,支持捕获、步骤前上下文注入、记忆召回技能和技能候选审核面板。Reme@agentscope-ai/reme面向 TypeScript 代理的 ReMe 客户端和记忆集成

README

dsh-slice-agent-loop

中文

A bounded conversational-context policy for DeepSeek Harness 0.1.3-alpha.2. It runs alongside the stock agent loop, retaining its lifecycle, scheduler, inbox, persistence, request-series handling, and full request-reconstruction invariant.

Completed conversational spans become durable user/message surface replacements. Runtime snapshots, instruction messages, current user input, and multimodal user messages keep their original sources and positions. Raw events remain in the session log; recall retrieves them after folding or resume.

Install and compose

Install this repository with DSH's plugin installer and apply its cordis.patch.yml bundle. The patch only adds the plugin. Keep agent-loop, agent-loop-invariant, and the stock session projections enabled. The Git package includes generated lib/ artifacts.

- id: slice-agent-loop
  name: '@dsh-external/dsh-slice-agent-loop'
  config:
    maxHistoryChars: 120000
    maxRequestChars: 400000
    maxStepsPerTurn: 50
    defaultReasoningEffort: low
SettingMeaning
maxHistoryCharsHard bound on the combined rendered conversational history, including headers and recall markers.
maxRequestCharsHard bound on serialized model messages, including protected context and current input. It is a character bound, not a token estimate; system prompt/tool schemas and model capacity remain host-owned.
maxStepsPerTurnStop before dispatching beyond this many model steps; default 50.
defaultReasoningEffortoff, low, high, max, or inherit; an explicit host/model choice wins.
digestContent-routing options from src/slice/result-digest.ts.
foldTool-result folding options, including enabled, pinSteps, pinMaxChars, spillPreviewMinBytes, and backoffAfterExpansions.

If recall markers cannot fit, or protected input exceeds the message budget, request construction fails visibly. The plugin does not quietly return an oversized view. Current input is logged before budget refusal. Historical spans are planned before any replacement is appended.

recall_search searches original human, assistant and tool records; recall_turn returns a turn, including original records and tool metadata; recall_step retrieves a step; expand_result retrieves an exact result ordinal, optionally filtered by lines or a regex. Tool results with a spill locator use the locator shown in their preview.

Migration from 0.0.1

  • Remove the old override that disabled the stock loop and invariant. Apply the new additive bundle.
  • Move scheduler configuration such as maxParallelToolCalls to the stock agent-loop row.
  • mode: state, mode: stream, state, tape, and inTurnSeal are retired and fail at load. The unsafe host-file snapshots and write rollback implementation have been removed.
  • The private sliceContext.contribute registry is retired. Use the host's system-prompt/runtime-context contributions so source identity and persistence are owned by DSH.
  • Mount either the stock invariant or this package's compatibility ./invariant export, never both. Both install the same complete reconstruction check.
  • Existing logs containing required custom slice/* events still require their original reader or an explicit migration. This release does not mutate DSH's known-event vocabulary or rewrite old session files.

File reads are recorded windows; write/edit metadata contains diff hunks. They are historical observations, not proof of a complete current file or backend identity. Exact base/pointer optimizations remain disabled until the host offers a durable observation channel with complete provider text, target identity, and version. See recorded memory.

Development and verification

Use Node ^22.19.0 || >=24.0.0 and pnpm 11.7.0:

pnpm install --frozen-lockfile
npm run typecheck
npm test
npm run build

Public alpha.2 dependencies are pinned in pnpm-lock.yaml; no maintainer checkout or absolute dependency path is required. CI runs the full suite against those published packages, including stock-loop invariants and real JSONL close/resume. The build removes stale generated files before emitting Git-install artifacts.

Run npm run verify:packed for a keyless standard-installer/Loader/JSONL smoke, or npm run verify:master -- /path/to/deepseek-harness against one prepared upstream source checkout. Recorded upgrade verification separates published-release, source-master and packed-artifact evidence.

The native regression suite covers runtime retention/update/removal, opaque instruction source ownership, multimodal input, retries and steering, request-series transitions, admission failures, unload, and resume without the plugin. A separate test checks that changing a later message causes the stock invariant to reject dispatch.

The cache prefix and per-turn paid text depend on which spans change; there is no universal cache-hit or cost guarantee. Earlier custom-loop measurements are historical. This migration changes the policy and needs new model-quality/cost experiments before those numbers can be reused.