DSH | dsh-agent-team-gui | Reusable planner, implementer and reviewer teams

[lzh19162600626-design]

Unofficial community plugin / 非官方社区插件

Give DeepSeek Harness a team you can reuse: a planner, an implementation engineer, and a reviewer, with a model and tool policy for each member.

dsh-agent-team-gui adds saved multi-model teams to the DSH Web profile. Select one beside the ordinary composer, then inspect the plan, member outputs, review/repair stages, and provider-reported Tokens in the Run Center.

80-second screenshot-based UI guide with Chinese synthetic narration. It uses example data from product captures, not a live-model task recording or a performance benchmark.

Try v1.0.1

Use the compiled release package to avoid the plugin's Git prepare build and its allowBuilds step:

dsh plugin --profile web add -w https://github.com/toolclub/dsh-agent-team-gui/releases/download/v1.0.1/dsh-agent-team-gui-1.0.1.tgz
dsh --profile web

Restart an existing DSH Web process after installing or updating. You need a working Web profile, pnpm, a configured model route, and Node.js >=22.19.0 <23 or >=24.0.0. Declared DSH range: >=0.1.0-rc.5 <0.2.0; the v1.0.1 release was verified on DSH 0.1.1-rc.2.

If a marketplace selected Git and installation failed, use this release path directly. An affected user confirmed it works in #27.

A first team you can import

1. Download the three-member recipe.
2. In Settings → Teams → Recipes & data, preview it, map its placeholder routes to your models, and import a copy. All three members can use the same working model initially.
3. Follow the English first-run guide or 中文上手教程. The exercise creates a dependency-free to-do list in an empty scratch project and asks for verifiable files and node --test results.

The exercise is ready to reproduce on your own model configuration; no measured task completion, speedup, or cost saving is claimed.

What can be reused

• Saved members with individual models, roles, fallback routes, and tool policies.
• Teams across conversations/projects, with explicit Team/Solo/Inherited modes.
• Credential-free JSON recipes, version snapshots, and preview-first imports.
• Run records with bounded planning, cancellation, linked retries, and optional review/repair.

Provider usage can be partial or unavailable. The team Token budget is a soft scheduling limit, and recipes use local JSON files. See the README and v1.0.1 release notes for details.

欢迎正在使用 DSH 的朋友试用

这次补齐了预编译安装入口、中文首次任务教程，以及带示例数据标记的界面导览。可以先导入规划／实现／审核三人小队，映射自己的模型，在空临时项目里跑一个待办清单任务，再查看文件、测试与运行记录。

希望收到具体反馈：是否安装成功、做了什么任务、在哪一步卡住、是否会再次使用这支小队。欢迎在配方与试用反馈帖分享。有帮助的话，欢迎给 GitHub 仓库点一个 Star。

[zoahdev]

Persistent multi-model squads is a strong direction — this is the kind of "workflow product" the ecosystem will need. Added to the ecosystem map (https://github.com/zoahdev/dsh-ecosystem).

Three concrete suggestions:

1. You already have a release — that puts you ahead of most community plugins. Add a CI badge and an install smoke (pack → fresh profile → dsh web boot) to make it verifiable; dsh-plugin-doctor preflight automates the checks.
2. Marketplace: dsh-marketplace (https://github.com/ydhrdh/dsh-marketplace) — one plugin.json PR; your release tarball makes the install spec easy to verify.
3. Cross-link with dsh-agent-teams (NanmiCoder) — you're solving adjacent problems; a "compare" note in both READMEs helps users choose.

Happy to run a preflight review on your repo and report findings.

[lzh19162600626-design]

thks,bro

[lzh19162600626-design]

Update: v0.4.0 is now released.

The plugin now ships guaranteed team orchestration for ordinary conversation sends, persistent project defaults, per-agent primary/fallback model and tool policies, fixed or leader-planned execution, durable run history, live progress/cancellation, version restore, diagnostics, and official per-member token-usage monitoring.

Validation completed: TypeScript checks, 36 tests, host/client builds, package inspection, and a live Web run with a three-member team. The README now includes fresh screenshots and the pinned install command.

I also opened awesome-dsh-plugin #1264 to remove the stale pre-rename duplicate and update the ecosystem description.

[lzh19162600626-design]

Update: v0.5.0 is now released and installed successfully from the public tag.

This release turns the original squad prototype into a bounded workflow product:

• the conversation's Main Agent dynamically assigns focused work to saved members and produces a validated DAG;
• Team / Solo / Inherited project default / one-message modes survive refreshes and Host restarts;
• foreground/background runs, cancellation, immutable retry, crash reconciliation, optional review/repair, versions, backups, and portable recipes;
• parent-scoped recursive-agent protection, including renamed subagent/workflow tools and an at-most-once latest-message claim;
• DSH's official provider Token projection across planner, members, review, and repair, with full/partial/unavailable coverage;
• redesigned Composer, Settings, Run Center, and Insights UI with EN/ZH live locale, keyboard/focus, light/dark, Axe, and 390 px layout coverage.

I followed the earlier review suggestions here: CI badge, immutable Actions, pack → fresh Web profile → restart smoke, public full-commit Git install, exact-tarball browser smoke, and dsh-plugin-doctor are all green. The release includes six reproducible screenshots, bilingual docs, an example recipe, governance/security files, and an attached compiled tarball plus SHA-256 checksum.

Validation: Node 22.19 + Node 24, 117 Host tests, 62 rendered Client tests, DSH 0.1.0-rc.6, and an upgrade of a real profile retaining 313 historical runs without re-saving or automatic retention loss.

Install:

dsh plugin --profile web add -w github:toolclub/dsh-agent-team-gui#v0.5.0
dsh --profile web

Feedback on real team recipes, planner behavior, and provider-specific Token coverage would be especially useful.

[lzh19162600626-design]

New onboarding material is now available: an 80-second Chinese UI guide, English / 中文 first-run guides, and the existing three-member recipe.

The introduction and v1.0.1 Release now recommend the compiled tarball, avoiding the plugin's Git build authorization step. The guide explains importing the recipe, mapping your models, enabling the team, and checking a small development task's files and tests.

The video is a screenshot-based walkthrough with clearly labelled example data, not a live-model recording or benchmark. 欢迎正在使用 DSH 的朋友试用；可以在反馈帖说说是否装上、首个任务是否跑通，以及卡住的步骤。有帮助的话，欢迎给项目点一个 Star。