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。

Progress Target — 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
P

dsh-progress-target

Progress Target

面向 DeepSeek Harness 的持久化、质量门控、资源感知型执行计划

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:02TJS/ProgressTarget#74cf29e9cadf4f7d43786b8d0662752b912288cc
README兼容性版本

兼容性与来源证明

Progress Target 以 dsh-progress-target 发布,当前版本为 2.0.0。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
2026/8/28

版本

2.0.0stable
2026/8/28
1.2.0stable
2026/8/28

相关插件

正在加载相关插件…

最新版
2.0.0
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
GitHub
★ 5
周下载
0
最近提交
2026/9/6
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 agents-orchestration 分类下经过校验的插件。

Headless@deepseek-ai/dsh-headlessdsh one-shot bundle:基于 dsh-base 的直接核心 Agent/Session 运行器,不包含 Host、HTTP 或浏览器层Experimental Agent Team Web Profile@deepseek-ai/dsh-experimental-agent-team-web-profile用于 Agent Teams Remote 和 UI 插件的实验性 Web 配置层Subagent Codex@deepseek-ai/dsh-subagent-codex基于官方 app-server 协议的一次性 Codex 子代理提供程序Subagent Claude Code@deepseek-ai/dsh-subagent-claude-code基于官方 Agent SDK 的一次性 Claude Code 子代理提供方

README

DSH ProgressTarget

English · 简体中文

Make AI agents finish long-running work—not merely write a checklist.

DSH ProgressTarget is a persistent planning and execution-control plugin for DeepSeek Harness. It adds semantic phases, quality gates, deliverable gates, deadlines, auditable retries, multi-resource execution plans, and dynamic GPU discovery to each conversation.

Why ProgressTarget?

Ordinary agent plans often fail in predictable ways: phases are declared complete without usable artifacts, metrics are vague, deadlines silently slip, GPU availability becomes stale, and the agent pauses for confirmation even when it could continue autonomously.

ProgressTarget turns a plan into an enforceable execution contract:

  • Persistent per-session plans — every conversation owns an independent plan.
  • Semantic phases — split by real dependencies, not mechanical quarters.
  • Two hard gates — quality targets and usable deliverables are evaluated separately.
  • Goal-linked phase quality — each v2 phase explains how its quality protects or validates the final objective.
  • Research before metrics — compare candidate proxies, cite sources, define measurement and limitations, and justify thresholds before execution.
  • Honest uncertainty — unknown impact magnitude stays null and is resolved through pilot, ablation, or controlled validation rather than fabricated precision.
  • Deadline-aware outcomes — completed means passed on time; overdue means late but still usable.
  • Evidence-first delivery — required artifacts need acceptance criteria and evidence.
  • Dynamic resource discovery — refresh all configured GPU servers at phase start, transition, and replan.
  • All-server sharding — shardable inference, evaluation, and data work must use every available server.
  • Auditable retries — record result summaries, findings, and explicit adjustments.
  • No continuation-round ceiling — while still before the deadline and an actionable adjustment exists, the agent must keep researching and retrying; attempt count or continuation budget is not a valid stop reason.
  • Recursive re-estimation — unfinished work gets a new 50%/100% schedule instead of busy polling.
  • Visible progress UI — inspect phases, metrics, artifacts, retries, deadlines, and resource snapshots in DSH Web.

Execution model

Plan → discover resources → start phase
  ↓
5 min / 50% / 75% checks → 100% harvest
  ↓
quality gate + deliverable gate + deadline gate
  ├─ all pass on time → completed → refresh resources → next phase
  ├─ quality misses, artifact ready, timed out → overdue → refresh resources → next phase
  └─ artifact missing → remain in-progress → diagnose → replan → retry

Installation

This repository is a DSH Profile Bundle. Add it to the Profile that owns your Web GUI (the Profile name is deployment-specific):

dsh plugin --profile <your-profile> add <path-to-ProgressTarget>

Restart the existing DSH Web Host after installation. Do not start a replacement Vite server: the running DSH Host injects the Web boot state.

For development, copy or link the repository into your user-owned DSH plugin directory, add the package to the Web profile, and rebuild/restart the affected Web artifacts as required by your DSH checkout.

Quick start

After installing, ask the agent to read GUIDE.zh-CN.md, then call update-progress-target while executing your task.

请执行 [XXX任务]。开始前先完整阅读本仓库的 GUIDE.zh-CN.md,并使用 update-progress-target 为当前会话建立独立计划。按真实依赖拆分语义阶段;每阶段必须设置 deadlineAt、结构化质量硬目标、可供下一阶段消费的必需交付物及验收证据、executionPlan 和资源分支。建立计划、启动阶段、阶段转换和重规划时,都要重新查询全部配置服务器的实时资源状态,不能复用旧快照。除非缺少用户专属输入、权限、安全确认或遇到无法自主解决的外部阻塞,否则持续推进到完整计划结束。

The full production prompt is available in PROMPT.zh-CN.md.

Configure your deployment first

ProgressTarget ships with no universal server inventory. Server names, discovery commands, quality thresholds, artifact rules, storage, deadlines, permissions, and resource limits are deployment-specific.

Configure the bundle in cordis.patch.yml:

config:
  requiredServers:
    - gpu-a
    - gpu-b
  resourceDiscoveryMaxAgeMinutes: 10
  dataDir: .progress-target

An empty requiredServers: [] disables fixed-inventory enforcement and supports CPU-only, cloud-autoscaled, scheduler-backed, or dynamically discovered environments. If configured, every resource snapshot must cover all listed servers; additional discovered servers remain allowed.

Read the complete configuration and deployment audit before production use. ProgressTarget validates evidence supplied by the agent; it does not create SSH connectivity, scheduler integration, credentials, GPUs, timers, or background-job capabilities.

State semantics

StateMeaning
pendingPlanned but not started
in-progressActive; may include retries or deadline re-estimation
completedQuality and deliverable gates passed before deadline
overdueDeadline passed, quality may miss, but required deliverables are usable

Required deliverables block downstream phases even after timeout. Terminal history is protected by default; administrative mutation should require explicit user authorization and an audit trail.

Storage

Plans are stored per session under:

<DSH_CWD>/.progress-target/<sessionId>.json

Do not commit runtime plan files: .gitignore excludes .progress-target/.

Security and privacy

  • Review resource evidence before publishing screenshots or plan JSON.
  • Never place credentials, private SSH keys, tokens, or confidential dataset paths in evidence fields.
  • The plugin records evidence supplied by the agent but does not manage cluster credentials.

v2 quality-contribution contract

New plans use schemaVersion: 2 and must define a structured finalObjective with final metrics and deliverables. Every phase must then provide:

  • metricResearch: research questions, traceable sources, candidate metrics, selected metrics, and selection rationale;
  • objectiveContribution: linked final metric keys, causal mechanism, evidence level, uncertainty, risk if missed, and a validation plan;
  • metric metadata: kind, measurement method, limitations, and threshold provenance.

A v2 phase cannot pass with process-only metrics such as “job finished” or “file created.” At least one researched quality or final metric must protect, improve, or validate the final objective. Vacuous existence thresholds such as count > 0, files > 0, or a set of quality metrics that are all merely > 0/>= 0 are rejected. An adaptive threshold is exploratory only and must be frozen to a justified formal threshold after research or a pilot before the production phase can proceed. If the impact magnitude cannot yet be estimated, impactEstimate should be null; the phase must state uncertainty and schedule a pilot, ablation, or controlled validation instead of inventing a number.

Contract compatibility and migration

Existing plans without schemaVersion continue under the v1 rules and are never silently rewritten. They can still update progress, finish phases, and advance under their original contract.

Migration is explicit and transactional. operation="migrate-plan" requires userAuthorizedMigration=true, a reason, a complete final objective, and v2 contracts for every existing phase. Migration cannot add or remove phases or rewrite existing statuses and completion times. If any validation fails, the stored v1 plan remains unchanged. See the full v2 contract and migration example.

Documentation

  • v2 quality-contribution contract (Chinese)
  • Configuration and deployment audit
  • Complete Chinese guide
  • Production prompt template
  • Changelog
  • Contributing

License

MIT — see LICENSE.