DeepSeek Harness Plugin Hub

Publish and manage complete Harness Profiles. Discover Plugins for your next setup.

Explore

PluginsPresetsDocsNews

Community

Publish a pluginContactReport an issue

Resources

Plugin Hub on GitHubDeepSeek HarnessSystem statusPrivacy notice
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

Independent and unofficial. Not affiliated with, authorized by, or endorsed by DeepSeek.

DeepSeek Harness Plugin Hub
ProfilesPluginsCategoriesNewsDocsSign inManage Profiles
ProfilesPluginsCategoriesNewsDocsSign in

COMMUNITY CHALLENGES

Find your next challenge.

Real problems from the community. Find one that sparks an idea, explore the discussion, and build something people need.

Showing 101–120 of 288 community requestsUpdated hourly from community discussions
All updatesDSH ReleasesFeature Requests
All3 days7 days14 days
TopNew
OpenShippedSolved by pluginWon't fixAll statuses
#Problems worth solvingCommunity demand
101
PLUGIN
2

Add a prompt-message budget trigger to compaction

Add a configurable maximum prompt-message count to compaction-basic, with policy inheritance and per-route overrides, so compaction triggers when message count exceeds gateway limits even if token usage remains below the threshold.

Explore challenge ↗

The problem

Add a configurable maximum prompt-message count to compaction-basic, with policy inheritance and per-route overrides, so compaction triggers when message count exceeds gateway limits even if token usage remains below the threshold.

Possibly addressed by:Context Compaction Optimizer · Compaction Threshold · Smart Compaction

Start with these discussions

  • compaction-basic: add a prompt-message budget — long sessions die on gateways that cap message countgithub · ▲1 · 1 comments
  • compaction-basic: add a prompt-message budget — long sessions die on gateways that cap message countgithub · ▲1 · 0 comments
Last active Oct 21, 2026
1 authors2 posts · 2 reactions1 comments
102
HARNESS
2hackernews1

Render images inline in conversations

Render generated, uploaded, imported, and agent-created visual content such as SVGs, Turtle drawings, and image attachments directly in chat history, with appropriate artifact handling and session persistence.

Explore challenge ↗

The problem

Render generated, uploaded, imported, and agent-created visual content such as SVGs, Turtle drawings, and image attachments directly in chat history, with appropriate artifact handling and session persistence.

Possibly addressed by:Image Inline · Chat Imagine · Chat Image Lightbox

Start with these discussions

  • Feature request: agent-emitted images inline in assistant messages (assistant-side rendering)github · ▲1 · 2 comments
  • Making simple images! (With a bug)github · ▲1 · 1 comments
  • Comment on: Pi 1.0
103
HARNESS
2

Add Hermes integration

The author wants DeepSeek Harness to integrate with Hermes, enabling the two systems to work together.

Explore challenge ↗

The problem

The author wants DeepSeek Harness to integrate with Hermes, enabling the two systems to work together.

Possibly addressed by:Hermes Link · Hermes Dsh Collab · Hermes Dsh Bridge

Start with these discussions

  • Deepseek harness 好定hermes 好?x · ▲0 · 0 comments
  • @trySynara Congrats ! Any plan soon regarding Hermes and Deepseek Harness integration ? 🙏x · ▲0 · 0 comments
Last active Nov 4, 2026
2 authors2 posts · 0 reactions0 comments
104
HARNESS
2hackernews1

Support WebMCP Tool Discovery and Reuse

Enable DSH agents and browser plugins to discover and invoke WebMCP tools, then build, verify, version, manage, and reuse them across websites. This includes the core WebMCP support requested for using page-provided capabilities while reducing repeated interactions and token costs.

Explore challenge ↗

The problem

Enable DSH agents and browser plugins to discover and invoke WebMCP tools, then build, verify, version, manage, and reuse them across websites. This includes the core WebMCP support requested for using page-provided capabilities while reducing repeated interactions and token costs.

Possibly addressed by:Web Bridge Mcp · Browser Mcp · Hermes Web Mcp

Start with these discussions

  • DeepSeek Harness + WebMCP + Builder = Efficient Computer Usegithub · ▲3 · 4 comments
  • Support webMCPgithub · ▲3 · 3 comments
105
HARNESS
3hackernews1

Configure Independent Presets for Subagents

Let each subagent use a dedicated provider, model, and preset instead of being forced to inherit the parent configuration. Preserve parent-model inheritance when no subagent-specific configuration is set.

Explore challenge ↗

The problem

Let each subagent use a dedicated provider, model, and preset instead of being forced to inherit the parent configuration. Preserve parent-model inheritance when no subagent-specific configuration is set.

Possibly addressed by:Configurable Subagents · Subagent Model Picker · Submodel Change

Start with these discussions

  • 每次切换主模型的时候，子agent模型会跟随主模型一起变更（同主模型），怎么才能固定子agent模型？github · ▲2 · 3 comments
  • 进一步扩大子代理自定义程度（如原会话使用极简模式而子代理使用PTC模式）github · ▲1 · 0 comments
106
HARNESS
2

Expose context limits and guide recovery

Display effective context and session limits, including output reserves, and warn before compaction or exhaustion. Clearly distinguish input-context exhaustion, output truncation, and automatic compaction, while offering recovery guidance such as /compact, /clear, a new session, or one-click compaction.

Explore challenge ↗

The problem

Display effective context and session limits, including output reserves, and warn before compaction or exhaustion. Clearly distinguish input-context exhaustion, output truncation, and automatic compaction, while offering recovery guidance such as /compact, /clear, a new session, or one-click compaction.

Possibly addressed by:Idle Compactor · Context Budget · Auto Compact

Start with these discussions

  • 上下文占满窗口时没有任何明确的用户提醒：不知道「输入上下文已满」，也没有 /compact、/clear 的引导github · ▲7 · 6 comments
  • Feedback: the line is not where the user thinks — one threshold that never fires, one that fires six times in under six minutesgithub · ▲1 · 0 comments
Last active Sep 24, 2026
107
HARNESS
3

Show live agent work and collapse it after completion

Keep reasoning, tool calls, and related activity visible during execution, then collapse or minimize them after completion while retaining access to detailed execution logs and keeping the final answer prominent.

Explore challenge ↗

The problem

Keep reasoning, tool calls, and related activity visible during execution, then collapse or minimize them after completion while retaining access to detailed execution logs and keeping the final answer prominent.

Possibly addressed by:Toolfold · Conversation Process Collapse · Trace Collapse

Start with these discussions

  • 0.1.7alpha版本建议github · ▲3 · 3 comments
  • Feature request: fold pre-result thinking/tool rows into a "work" group in the chat viewgithub · ▲1 · 3 comments
  • 优化建议：对话应聚焦文字消息，工具等调用不应占用高频阅读空间github · ▲1 · 0 comments
Last active Sep 22, 2026
108
HARNESS
2hackernews1

Notify on agent state changes

Add opt-in notifications for agent waits, approvals, escalations, errors, and task or round completion through configurable channels, including system, browser, and audio alerts. Expose these state transitions as public events for plugins and presets, with controls for hidden versus active tabs and optional notification actions.

Explore challenge ↗

The problem

Add opt-in notifications for agent waits, approvals, escalations, errors, and task or round completion through configurable channels, including system, browser, and audio alerts. Expose these state transitions as public events for plugins and presets, with controls for hidden versus active tabs and optional notification actions.

Possibly addressed by:Chime Alerts · Agent Notify · Notify

Start with these discussions

  • Feature request — automatic notifications when an agent pauses for user input or completes a long-running flowgithub · ▲2 · 6 comments
  • 浏览器声音+notification 提示完成一轮
109
HARNESS
2

Isolate host core packages and warn on plugin dependency conflicts

Prevent plugins from installing or overwriting host core packages, and detect dependency conflicts during installation with a clear warning or rejection. This would prevent silent profile corruption and delayed startup failures caused by plugins pinning incompatible core versions.

Explore challenge ↗

The problem

Prevent plugins from installing or overwriting host core packages, and detect dependency conflicts during installation with a clear warning or rejection. This would prevent silent profile corruption and delayed startup failures caused by plugins pinning incompatible core versions.

Possibly addressed by:Plugin Security Review · Plugin Guard · Plugin Allcrash

Start with these discussions

  • Third-party plugin pins an old @deepseek-ai/dsh-tools as a runtime dependency: install silently overwrites host core packages with no error at the time, crash on next startupgithub · ▲1 · 1 comments
  • 第三方插件把 @deepseek-ai/dsh-tools 钉旧版本声明为 dependencies 时，宿主核心包被静默覆盖（无任何报错）github · ▲1 · 0 comments
110
HARNESS
32

Add web-based workspace file browsing

Allow users to browse project and workspace directories, view source files and changes, and access web-viewable results directly in the web UI, including remote or headless deployments.

Explore challenge ↗

The problem

Allow users to browse project and workspace directories, view source files and changes, and access web-viewable results directly in the web UI, including remote or headless deployments.

Possibly addressed by:Fileview · Web Preview Panel · File Workbench

Start with these discussions

  • 可以提供类似opencode的web页面中源代码文件查看功能github · ▲1 · 4 comments
  • DSH的webUI看不到本地文件内容真的挺难受的github · ▲1 · 1 comments
111
HARNESS
3

Stack sidebar footer actions vertically

Make multiple sidebar footer actions or launchers occupy full-width rows stacked vertically instead of being compressed into one flex row. Apply this through the generic slot layout in web and desktop clients, including appropriate collapsed or rail-mode behavior.

Explore challenge ↗

The problem

Make multiple sidebar footer actions or launchers occupy full-width rows stacked vertically instead of being compressed into one flex row. Apply this through the generic slot layout in web and desktop clients, including appropriate collapsed or rail-mode behavior.

Possibly addressed by:Sidebar Footer Stack · Sidebar Stack · Footer Order

Start with these discussions

  • 【反馈】侧边栏底部插件入口横排拥挤（sidebar.footer.action），建议支持纵向堆叠github · ▲1 · 1 comments
  • [Bug][2.0.15-next] sidebar.footer.action is a row flex with a display:contents anchor — multiple launchers crush each other (Desktop-only workaround exists)github · ▲1 · 0 comments
  • Sidebar: stack multiple footer actions vertically (wide plugin footers get squeezed)
112
HARNESS
2

Sort the sub-agent dropdown by recent activity

Sort sub-agents in the dropdown by time in descending order so currently running or recently created agents remain visible without scrolling through hundreds of entries.

Explore challenge ↗

The problem

Sort sub-agents in the dropdown by time in descending order so currently running or recently created agents remain visible without scrolling through hundreds of entries.

Possibly addressed by:Subagent Mgm

Start with these discussions

  • 建议子智能体列表改为按照时间从后往前排github · ▲1 · 0 comments
  • sub agents 下拉列表UXgithub · ▲1 · 0 comments
Last active Oct 15, 2026
2 authors2 posts · 2 reactions0 comments
113
HARNESS
3

Configure each model's API protocol

Allow each model in a provider or catalog route to declare its wire protocol and use that declaration for discovery and request routing across Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses. Support adding protocols through the UI and provide a clear diagnostic when one cannot be inferred.

Explore challenge ↗

The problem

Allow each model in a provider or catalog route to declare its wire protocol and use that declaration for discovery and request routing across Anthropic Messages, OpenAI Chat Completions, and OpenAI Responses. Support adding protocols through the UI and provide a clear diagnostic when one cannot be inferred.

Possibly addressed by:Live Model Catalog · Provider Hub · Model Catalog Refresh

Start with these discussions

  • [Bug/UX] Cannot add a new model to a multi-protocol catalog route (opencode) from the GUI — "needs an api" with no way to set itgithub · ▲5 · 2 comments
  • llm-pi-ai: a mixed-protocol catalog route cannot serve a model the installed catalog has not shipped
114
PLUGIN
1hackernews12

Expand the context window and token capacity

Increase DSH’s effective context and token capacity so it can handle larger projects, very long documents, and codebases without truncation or segmentation, including support for a possible 2M-token trial or API option.

Explore challenge ↗

The problem

Increase DSH’s effective context and token capacity so it can handle larger projects, very long documents, and codebases without truncation or segmentation, including support for a possible 2M-token trial or API option.

Possibly addressed by:Billion Context · Context Budget · Context Enhancement

Start with these discussions

  • [Feature Request] 建议扩容至 2M 上下文，并申请今天开放内测github · ▲1 · 1 comments
115
HARNESS
1hackernews41

Enable vision-based visual validation

Accept images from native inputs or captured emulator, browser, and application environments, then use vision-capable models for iterative UI and rendering verification and correction.

Explore challenge ↗

The problem

Accept images from native inputs or captured emulator, browser, and application environments, then use vision-capable models for iterative UI and rendering verification and correction.

Possibly addressed by:Vision Toolkit · Llm Vision · Computer Use Vision

Start with these discussions

  • 建議：DeepSeek Harness 原生圖像輸入 + 眼驗證加權（支援自動長時間運行）github · ▲1 · 1 comments
116
HARNESS
2

Add blocking subagent wait and join

Provide a blocking wait or join primitive for parallel and continuable subagents, with optional timeouts, so parent agents can await results without premature responses or repeated polling.

Explore challenge ↗

The problem

Provide a blocking wait or join primitive for parallel and continuable subagents, with optional timeouts, so parent agents can await results without premature responses or repeated polling.

Possibly addressed by:Wait Subagent · Wait Guard · Background Agents

Start with these discussions

  • 【建议】后台/continuable 子代理提供 wait 语义（对齐 job_output wait）github · ▲1 · 6 comments
  • Agent Teams 下用 `subagent` 并行委派缺少 join 原语：主模型在子代理返回前就下结论，之后再补充和更正github · ▲1 · 0 comments
Last active Oct 2, 2026
2 authors2 posts · 2 reactions6 comments
117
HARNESS
2

Expose Structured MCP Results

Make MCP structuredContent visible to the model by serializing it into model-readable content, including when the regular content array is empty. Preserve authoritative fields such as element tokens and frames instead of returning a no-content placeholder.

Explore challenge ↗

The problem

Make MCP structuredContent visible to the model by serializing it into model-readable content, including when the regular content array is empty. Preserve authoritative fields such as element tokens and frames instead of returning a no-content placeholder.

Possibly addressed by:Mcp Lens · Mcp Manager · Mcp Apps Host

Start with these discussions

  • [Bug] MCP tool results drop structuredContent — the official computer-use plugin can't reach its own element_tokengithub · ▲3 · 2 comments
  • MCP tool results delivered only via `structuredContent` are invisible to the model ("no model-visible content" false timeout)github · ▲1 · 1 comments
Last active Sep 29, 2026
118
HARNESS
1hackernews2

Preserve context for interactive subagents

Allow asynchronously spawned or forked subagents to inherit the parent’s cached conversation context, including the current request and relevant tool state, while building their own context and safely handling inherited unfinished executions. The main agent should be able to monitor and interact with them dynamically.

Explore challenge ↗

The problem

Allow asynchronously spawned or forked subagents to inherit the parent’s cached conversation context, including the current request and relevant tool state, while building their own context and safely handling inherited unfinished executions. The main agent should be able to monitor and interact with them dynamically.

Possibly addressed by:Subagent Model · Fork Relink · Agent Chat

Start with these discussions

  • Proposal: Preserve current-turn context when forking subagentsgithub · ▲2 · 1 comments
  • Comment on: Pi 1.0hackernews · ▲0 · 0 comments
119
HARNESS
2

Support running the referenced tool on DeepSeek Harness

The author wants the referenced tool or agent to run on DeepSeek Harness, noting that it reportedly performs better there than in Claude Code.

Explore challenge ↗

The problem

The author wants the referenced tool or agent to run on DeepSeek Harness, noting that it reportedly performs better there than in Claude Code.

Possibly addressed by:Deepseek Harness Plugin · Harness Ops · Deepseek Kernel Harness

Start with these discussions

  • @PawelHuryn I agree it should run on the DeepSeek harness — per the official report, it performs far better there than Cx · ▲2 · 0 comments
  • @PawelHuryn I agree it should run on the DeepSeek harness — per the official report, it performs far better there than Cx · ▲1 · 0 comments
Last active Oct 17, 2026
1 authors2 posts · 3 reactions0 comments
120
HARNESS
1hackernews2

Continue Unfinished Tasks

Detect when a multi-step task or announced action remains unexecuted and keep the turn open to drive the remaining steps automatically. This prevents premature completion without requiring another manual turn.

Explore challenge ↗

The problem

Detect when a multi-step task or announced action remains unexecuted and keep the turn open to drive the remaining steps automatically. This prevents premature completion without requiring another manual turn.

Possibly addressed by:Loop Continue · Wait Guard · Auto Stop

Start with these discussions

  • [Bug] 模型以旁白收尾（宣布动作、无 tool-call）时 turn 被正确关闭，宣布的动作消失 —— 机制已存在，缺的是判据github · ▲2 · 1 comments
  • Comment on: GPT-6 Sol and Lunahackernews · ▲0 · 0 comments
  • Comment on: GPT-5.6 Luna vs. GPT-6 Astra: Is a $1.20 Model Good Enough for Code Review?hackernews · ▲0 · 0 comments
← PreviousPage 6 of 15Next →
hackernews · ▲0 · 0 comments
Last active Oct 2, 2026
3 authors3 posts · 2 reactions3 comments
Comment on: Amazon blocks Meta’s new Muse AI agent from shopping on amazon.com
hackernews · ▲0 · 0 comments
Last active Sep 21, 2026
3 authors3 posts · 6 reactions7 comments
Feature request: 子代理模型可配置（agent-child-model 设置分节）
github · ▲1 · 0 comments
Last active Sep 23, 2026
4 authors4 posts · 4 reactions3 comments
2 authors2 posts · 8 reactions6 comments
3 authors3 posts · 5 reactions6 comments
github · ▲2 · 2 comments
  • Comment on: Show HN: Share your AI Setup, Learn from othershackernews · ▲0 · 0 comments
    • Last active Sep 23, 2026
    3 authors3 posts · 4 reactions8 comments
    Last active Oct 19, 2026
    1 authors2 posts · 2 reactions1 comments
    网页端不能直接浏览项目文件。
    github · ▲1 · 1 comments
    Last active Sep 21, 2026
    4 authors5 posts · 3 reactions6 comments
    github · ▲1 · 0 comments
    Last active Oct 1, 2026
    3 authors3 posts · 3 reactions1 comments
    github · ▲1 · 3 comments
  • Custom provider defaults to Chat Completions only, missing Anthropic / Response compatibility — some models undiscoverablegithub · ▲1 · 2 comments
    • Last active Sep 18, 2026
    3 authors3 posts · 7 reactions7 comments
  • Used DeepSeek Harness, Open Weight Glimmer, and Qwen to replicate agent flows. Also built it to connect to my phone viax · ▲0 · 0 comments
  • @MLStreetTalk Totally share this frustration. There are so many exciting projects enabled by these new models for me, bux · ▲0 · 0 comments
    • Last active Oct 2, 2026
    4 authors4 posts · 1 reactions1 comments
    @ivanfioravanti After playing with deepseek harness, I could feel the intelligence of deepseek coming through... not lik
    x · ▲1 · 0 comments
  • Comment on: DeepSeek-v4-flash-vision-exphackernews · ▲0 · 0 comments
    • Last active Sep 22, 2026
    6 authors6 posts · 2 reactions1 comments
    2 authors
    2 posts · 4 reactions
    3 comments
  • Comment on: GPT 6.1 Sol: Near-Astra intelligence for a fifth of the pricehackernews · ▲0 · 0 comments
    • Last active Oct 2, 2026
    3 authors3 posts · 2 reactions1 comments
    Last active Oct 1, 2026
    3 authors3 posts · 2 reactions1 comments