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。

Guardian Mode — DeepSeek Harness 插件(DSH Plugin)
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
← Plugins
G

dsh-guardian-mode

Guardian Mode

第五种 DeepSeek Harness 模式,支持可配置的 Codex、Claude Code 或 DSH 审计、用户批准的修复轮次、渐进式 Cordis/技能提升、安全暂停以及 Web/TUI 审查控制。

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:yhfgyyf/dsh-guardian-mode#da86c976d232246b9f111b668ad41c756f4b6360
README兼容性版本

兼容性与来源证明

Guardian Mode 以 dsh-guardian-mode 发布,当前版本为 0.5.0。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
web
发布来源
github
Registry 更新时间
2026/9/1

版本

0.5.0stable
2026/9/1
0.4.3stable
2026/8/31
0.4.1stable
2026/8/25
查看其余 2 个版本收起版本
0.3.0stable
2026/8/24
0.1.0stable
2026/8/23

相关插件

正在加载相关插件…

最新版
0.5.0
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
web
许可证
MIT
发布源
github
周下载
0
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

相关插件

继续浏览 security-access 分类下经过校验的插件。

Doctor@linxin666/dsh-doctorDSH 配置档案的事务性救援模式,配备受监督的启动器、隔离的恢复容器、确定性修复、健康监控以及本地 Web 恢复控制台Pocketdsh-pocket把 DeepSeek Harness 装进你的口袋:一个包、一个设置页,手机扫码即同步访问电脑上的 DSH(局域网 + 公网,实时同屏)。DSCODE@toddzheng024/dscode-bundle完整的 DeepSeek 编码代理,支持持久化 shell、Ultra 协作和自动权限审查。Auto Reviewdsh-auto-review针对 DeepSeek Harness 审批请求的第二模型 AI 自动审查:只读审查子代理在审批应答链上决定允许或拒绝,并采用故障关闭回退机制和完整的会话日志审计。

README

dsh-guardian-mode

The fifth mode of DeepSeek Harness (DSH): preset id guardian, combining PTC code presentation, independent review, and a human-approved Cordis remediation loop.

An agent on this preset keeps full standard-mode capabilities (shell, filesystem, web, skills, goals, subagents, workflows, Code Mode tool presentation). Cordis self-modification tools stay model-hidden during ordinary work and are exposed temporarily only after the user accepts a critical remediation. Separately, every session drives two isolated reviewer roles. The reviewer backend is configurable as Codex, Claude Code, or the host DSH LLM runtime. The default remains one persistent Codex app-server:

RoleDefault modelEffortJob
summarizergpt-5.6-lunamediumincremental trace summary per round
auditorgpt-5.6-solmaxindependent audit → pass / warning / critical

Codex and Claude Code keep separate persistent role sessions. The DSH backend uses direct, tool-free llm.stream() calls instead of starting another DSH Agent, so it cannot recursively enter Guardian mode. Those calls are stateless, so Guardian includes the current objective and a bounded tail of sidecar review memory in every DSH audit.

All unaccepted feedback and reviewer state is written to a sidecar (${DSH_HOME:-~/.dsh}/guardian/sidecars/<sessionId>.json). Only explicit human acceptance appends a bounded <guardian-remediation> prompt and capability lease at the context tail. The model then loads the named skills through DSH's stable skill tool; prior messages are never rewritten and raw reviewer output remains private.

Install

# in your dsh profile (profiles/web and profiles/tui use the same pattern)
cd ~/.dsh/profiles/web
pnpm add dsh-guardian-mode@github:yhfgyyf/dsh-guardian-mode
# Recommended stable tool discovery for Guardian and Auto target presets:
pnpm add dsh-progressive-tools@github:yhfgyyf/dsh-progressive-tools
# add both bundles to package.json dsh.profile.bundles (both profiles),
# then restart the profile.

The bundle patch adds one dual-face row:

- insert:
    - id: guardian-bundle
      name: dsh-guardian-mode

The node half mounts the host guardians service, registers the /guardian command, and (when a webserver is present) the Remote API. The same row's browser half (dsh.client) renders the guardian strip in the composer dock.

Using the mode

  • Start a session with --preset guardian (TUI) or pick guardian in the Web preset chip, or /preset guardian on a blank session.
  • /guardian status — round, cadence interval, last verdict, pause state.
  • /guardian now — force an audit (out of cadence).
  • /guardian history — recent audits from the sidecar.
  • /guardian accept [audit-id] — approve the latest/specified remediation.
  • /guardian resume — clear a non-critical-review failure/manual pause.

Reviewer configuration

Configure the guardian-bundle row in the profile's cordis.patch.yml. No configuration preserves the existing Codex defaults:

- id: guardian-bundle
  config:
    reviewer: codex
    binary: codex
    args: [app-server, --stdio]
    models:
      summarizer: { model: gpt-5.6-luna, effort: medium }
      auditor: { model: gpt-5.6-sol, effort: max }

summarizer and auditor are stable, responsibility-based keys; their model names remain fully configurable. Legacy luna / sol keys are still accepted and are migrated to the new names at runtime.

Claude Code uses print mode with JSON-schema output, plan permission mode, safe mode, and an empty tool set. Set Claude-supported model names explicitly:

- id: guardian-bundle
  config:
    reviewer: claude-code
    claudeBinary: claude
    claudeArgs: []
    models:
      summarizer: { model: haiku, effort: medium }
      auditor: { model: opus, effort: max }

The DSH backend routes directly through a registered provider. A per-role provider overrides dshProvider when summary and audit use different routes:

- id: guardian-bundle
  config:
    reviewer: dsh
    dshProvider: deepseek-official
    dshMaxTokens: 4096
    models:
      summarizer: { model: deepseek-v4-flash, effort: off }
      auditor: { model: deepseek-v4-flash, effort: high }

Changing reviewer does not translate model names. Guardian fails loudly if the selected backend does not support a configured model; it never silently substitutes an audit model.

Behavior

  • Cadence: the first audit requires at least two steps and 60 seconds; later audits run every three steps or three minutes, with a 60-second minimum gap. Anomalies audit at the next safe boundary.
  • Warning approval: a warning leaves the main Agent running. The user may execute the proposed repair as-is or edit it first. An accepted repair uses DSH's native next-step steering path, so the current tool call finishes before the edited instruction runs; an idle Agent runs it immediately.
  • Critical approval: critical pauses the main Agent and active Goal first. The user may execute the proposed repair as-is or edit it first. Acceptance immediately starts the repair turn, temporarily exposes Cordis tools, and appends a capability lease. The repair Agent must load editing-cordis-compositions through the stable skill loader, and loads cordis-plugin-development only for plugin or model-facing-tool work. The original task resumes only after the repair audit is no longer critical.
  • Three consecutive failures (reviewer unreachable, timeouts, malformed replies) pause the session with reason failures.
  • Every 5 rounds a full objective-alignment audit runs (objective + boundary rules + recent summaries).
  • Final audit runs when the session is disposed (or /guardian now with the Remote API final: true).
  • Fixed capability: guardian (GUARDIAN_CAPABILITY). The auto router keeps routing only standard / code / minimal / cordis.

Remote API (browser)

Third-party routes, declared by this package:

MethodPathBody / query
GET/api/guardian/snapshot?session=<id>
GET/api/guardian/watch?session=<id> (SSE event: guardian)
POST/api/guardian/request-now{ sessionId, final? }
POST/api/guardian/accept{ sessionId, auditId?, editedText? }
POST/api/guardian/resume{ sessionId }

The Web dock strip registers at conversation.input.dock order 5 — rendered between the Todo strip (order 0) and the Goal strip (order 10).

TUI

dsh-tui-app renders an independent color-coded block (pass=green, warning/critical/paused=red) beside the config row:

  • a — execute the proposed remediation unchanged
  • e — load it into the composer; Enter executes the edited text, Esc cancels
  • c — copy feedback while paused
  • r — resume a non-critical-review pause
  • Esc / Ctrl+C — stop current work

Development

npm test            # unit + integration
npm run check       # syntax, package manifest, tests
npm run pack:check  # npm pack --dry-run

scripts/build-preset.mjs regenerates presets/guardian/agent.cordis.yml from the shipped code + cordis compositions (checked-in result, so the package works standalone). Tests use test/fixtures/fake-codex.mjs and fake-claude.mjs; no real reviewer login is required. Backend, models, effort, binaries, CLI arguments, DSH provider route, timeout, and DSH output limit are configuration rather than constants.

Compatibility

  • Never calls session.delete or any session-removal API; disposal is observed via the host session/disposed event for a final audit only.
  • Auto still routes only the original four modes. When the companion auto router supports capability hints, those names are appended after routing and do not alter the original user prompt.
  • Images, ordinary skills, goals, subagents, and workflows flow unchanged. Guardian's two composition skills are progressive, critical-approval-only additions (see presets/guardian/agent.cordis.yml).
  • Persisted messages remain byte-for-byte unchanged. Acceptance only appends remediation, runtime-catalog, and continuation tail messages, so the prior message prefix remains eligible for KV-cache reuse. With dsh-progressive-tools, Cordis restriction changes affect discovery results rather than the model-visible system/tool prefix. Without that companion, DSH normally rebuilds the Code Mode SDK when visibility changes. An actual plugin/system-prompt repair still takes effect through DSH's normal restart/new-task prefix rebuild.
  • Does not modify the global node_modules; install as a profile bundle.