dsh-agent-teams
IdentifiedAgentTeams for DeepSeek Harness — an independent design branch of NanmiCoder/dsh-agent-teams; do not install both, since both register the same `agent-teams` row. Multi-agent team collaboration with a
English · 简体中文
One prompt. A working team.
dsh-agent-teams turns the current DeepSeek Harness session into a captain that can assemble durable sub-agents, split a goal into dependency-aware tasks, and coordinate work through direct messages.
Ask in natural language. The plugin provides the team protocol, 15 coordination tools, persistent state, an automatic shared-task scheduler, and a live Web UI—without requiring a separate workflow engine.
About this fork
This repository is an independent design of that idea, maintained by ChenXu233 and started from NanmiCoder/dsh-agent-teams. The upstream author still publishes @nanmicoder/dsh-agent-teams, and credit for the originals stays with them — see Acknowledgements.
The fork is built out of the DSH features that already exist instead of patching around them:
- Everything is a plugin. Member-page controls are additive slot contributions (
conversation.input.right); no host service is patched, and the single-ownerconversation.input.modelseat is left to the host. - One compatibility floor.
compatibility.jsondeclares onlyminimumHost; every DSH peer stays*, CI exercises the floor, and a newer Harness release installs without a plugin release. - One verification entry.
pnpm verifyruns 24 named phases in order and stops at the first failure. - One home per fact. Guides live under
docs/in three reader-oriented folders; a fact kept in two places is a bug, not redundancy. - Teams end, archives can be deleted. An ended team keeps its full history — and can be removed from the team view once you no longer need it.
Releases
v0.2.2 is the stable release on npm latest. An ended team can now be deleted from the team view: a confirm dialog states that the removal is permanent, and the host deletes the archived directory after refusing an unsafe id, a missing archive, or an id a live team still owns. The same release fixes where that control appeared — the workspace tab, not only the floating panel — and keeps the archive list loaded across reloads. The install floor is unchanged (minimumHost: 0.2.1-rc.2), so any newer Harness release still installs without a plugin release. The previous stable was v0.2.1. See the floor declaration and verification scope.
Why AgentTeams?
| Capability | What it changes |
|---|---|
| Captain-led delegation | The current session creates the team, assigns roles, and consolidates the final result. |
| Durable members | Members are continuable DSH sub-agents that can be woken for focused follow-up turns. |
| Dependency-aware tasks | Tasks move through explicit states and cannot be claimed before their dependencies finish. |
| Automatic reuse and safe takeover | Idle members claim the next ready task; reassignment revokes stale attempts before new work starts, and cold recovery retries stranded open attempts. |
| Direct messaging | Members send durable mailbox messages directly to teammates or the captain—no relay required. |
| Live activity panel | The Web UI combines segmented progress, a collapsible roster, and an interactive task DAG; running tasks show the member's model, and completed archives retain their full member and task history. |
| Plan before execution | Normal /agent-teams runs stage an unspawned roster and DAG first. The Web panel uses the host model catalog for member routes. Returning to chat stops the planning turn, asks what should change, and revises the same draft; discarding archives the draft, aborts the turn, and explicitly prevents automatic recreation. Only Approve & Run enables scheduling; each member starts with its first ready task. |
| Quality gates | Opt-in quality tasks support requirements → implementation → verification → review → integration contracts, automatic repair/re-review, and explicit resume. Scope control is a completion-time audit, not host write interception. See docs/3-reference/quality-gates.md. |
The conversation card and activity panel use Harness's official locale service. They follow live language changes between English and Simplified Chinese—including status labels, dynamic summaries, controls, archive markers, and accessibility text—without a page reload or a separate plugin setting.
Install and choose versions
Supported pair: DeepSeek Harness >=0.2.1-rc.2 + AgentTeams v0.2.2. The plugin declares one floor and nothing else.
| Use case | DeepSeek Harness | AgentTeams plugin |
|---|---|---|
| The floor (oldest supported host) | 0.2.1-rc.2 | v0.2.2 |
| Any later RC, prerelease or stable release | >=0.2.1-rc.2 | v0.2.2 |
There is no enumerated tested list to keep in sync. engines.dsh carries the floor and the host reads it before peers, so newer hosts install directly; CI exercises the floor, and everything above it is assumed compatible. A host below the floor is rejected by the installer.
Desktop app
In the app sidebar, open Plugins → Add plugin. You can install directly from GitHub by pasting this repository URL into the plugin input, without a version, tag, or commit suffix:
@chenxu233/dsh-agent-teams
The package name resolves npm latest. To pin a plugin version, append @0.2.2, or the tag from the table above. Install it, then choose Enable now. If the host asks for a restart, restart the desktop app. Use the package spec above for the published build; an npm or GitHub webpage URL is not the same install source. For Git installation, use a commit containing the verified build outputs described below.
The desktop app supplies its own Harness core and package manager. Upgrading the global CLI does not upgrade that core, and an ordinary standalone CLI cannot manage its desktop profile. The CLI carrier installed by the desktop app can manage that profile from 0.2.1-rc.2 on. Compare the embedded core with the floor 0.2.1-rc.2: anything equal or newer installs, and no per-version exemption exists to maintain. See the compatibility floor.
CLI / Web: 1. Install DeepSeek Harness
npm install --global @deepseek-ai/dsh
dsh --version
The command selects the host's npm latest; after upgrading, dsh --version only has to be at or above the floor 0.2.1-rc.2. Alpha is opt-in: select an exact Alpha version from the table and lock all host dependencies as described in the maintenance guide.
CLI / Web: 2. Install the AgentTeams plugin
Install or upgrade from npm using the package name (resolves npm latest; the preview channel next carries the newest prerelease, which can be older than latest once a stable release of the same line exists):
dsh plugin --profile web add --save-exact @chenxu233/dsh-agent-teams
For a CLI-managed installation, replace web with your active CLI profile. Stop and restart that profile's Harness process, then refresh the browser. Installing the plugin does not upgrade Harness; the host and plugin have independent latest channels.
Alternatively, install or update from the latest GitHub default branch without a tag or commit suffix:
dsh plugin --profile web add github:NanmiCoder/dsh-agent-teams
To update, run the command again with the new version, then restart the profile; a bare package name can retain an already installed exact version. For 24 hours after a release, a bare package name or @latest installs the previous version. If a registry mirror has not synced the release, append --registry=https://registry.npmjs.org to the npm command. GitHub default-branch code may be newer than npm latest. The default branch includes the verified server, client, and type build outputs in lib/. The repository distributes these outputs so installation does not need a prepare hook or build-script approval. Source maintainers must run pnpm build and update the committed outputs whenever the source changes. Existing npm releases remain immutable.
For a manually cloned checkout, run pnpm install --frozen-lockfile and pnpm build after pulling, then restart the corresponding Harness process. Updating Git alone does not update local build output. See the maintenance guide.
A host below the floor cannot install this plugin. Keep the exact plugin version that worked on that host, or raise both ends together; the rule is in the compatibility floor.
See the compatibility floor, source installation and Alpha testing guide, and verification scope and platform limits.
Then ask for a team directly:
Use AgentTeams to review the commits after v0.5.3 from performance, security, and product perspectives. Return one consolidated report.
View teams in the workspace
Click View team in the current chat header or the team card below a reply to open the native Team collaboration tab. The entry belongs to its chat, rather than the general workspace start page.
Team and task content starts immediately: wide panes use columns and narrow panes stack them. Select a task to locate its owner; member icons open their conversations. Closing the tab does not stop the team, and its chat can reopen it. Completed teams retain their archives; an archived team can be deleted from the panel, which removes its state directory and cannot be undone. Older hosts keep the activity panel.
How it works
- For a request to use AgentTeams, the captain follows the core protocol already in its system instructions. It continues an existing team and uses
agent_teams_statuswhen current state needs checking. When no team exists, the goal becomes a staged plan for review. - The captain adds role-specific members backed by continuable sub-agents.
- The goal becomes tasks with owners and explicit dependencies.
- The shared scheduler uses real
running / idle / readystate to atomically claim one ready task per idle member and wake it. An interrupted resident attempt stays parked and can resume through a direct message without losing its capability; after a cold process restart, the scheduler retries stranded open work with a fresh attempt. - Members update with the current
attempt_id; reassignment or captain takeover revokes the old attempt and waits for the old worker to quiesce before a new attempt starts. - The captain presents the combined result, then archives the complete team record.
Team state is stored under <workspace>/.agent-teams/; the Web panel reads that disk truth and combines it with live sub-agent activity.
Member creation is zero-interaction by default: a member on the captain's current LLM route snapshots that provider, model, and the captain's chosen reasoning effort, while a member on any other route keeps following that model's own default effort; later continuations restore the pinned selection. Only an explicit heterogeneous-team request (for example, “backend on provider A/model X, frontend on provider B/model Y”) supplies a member-specific provider + model; there is no per-member model or reasoning prompt.
Captain sessions keep the concise core protocol and all 15 native team tools (the original 14 plus agent_teams_run_command) from their first request. All business tools are directly available; no loading tool or extra activation call is needed. Configured profiles retain their bounded directory in the fixed system prompt. Creating, approving, continuing or ending a team does not rewrite the system prompt or tool schemas. Core rules remain available after history compaction or discarded code-mode tool results. Members receive four team tools, fixed member instructions, and their ordinary coding/research tools. Web approval wakes the captain with a control message; later member reports wake it again. See the fixed protocol and benchmark contract.
Slash command
No “use AgentTeams” phrasing required. The plugin registers the
closed-namespace /agent-teams host command, so the Web GUI slash menu shows
an agent-teams placeholder with an input hint: pick it (or type the
command), describe the goal, and press Enter.
/agent-teams research the pricing pages of three competitors
The command pipeline claims the line, then preserves that exact input as an
ordinary user follow-up so it remains visible in the main chat. The gesture
boundary adds the deterministic activation directive at pre-step, so the
first model request follows the staged planning protocol without a mandatory helper call. The invocation is also durably
logged (command/run / command/done).
Surfaces without command adjudication (for example the headless CLI) get the
same deterministic activation through a gesture boundary: any genuine user
message starting with /agent-teams activates the protocol for the rest of
the text. Mid-sentence mentions stay ordinary prose.
Historical panels require saved team state or archives. Sessions from early versions that deleted teams without retaining archives do not yet support reconstructing the full panel from logs.
Configuration
Defaults work without extra setup. A trusted profile can override member behavior:
- id: agent-teams
config:
stateDir: .agent-teams
memberProvider: spawn
memberModel: deepseek-v4
memberMaxDepth: 0
maxMembers: 8
toolPolicies:
researcher:
allow: [read, glob, grep]
implementer:
deny: [web_search]
# Optional: per-member context budget (auto-compacts once an idle member
# reaches its threshold)
memberContext:
maxTokens: 150000 # with maxRatio: either one, or both (smaller wins)
# maxRatio: 0.15 # share of the member model's contextWindow; without a
# known window this field has no effect
cooldownMs: 60000
minTurnGap: 1
onLimit: compact # compact | notify | off
memberMaxDepth defaults to 0: team members cannot create nested subagents. Set 1 to explicitly permit one descendant level; the limit also covers runtime/code-tool calls. Default members report through team messages only, avoiding duplicate native parent reports. Idle roster members make no model requests. Task assignments start distinct turns; coordination joins the nearest model step. Acceptance and consumption are tracked separately. Removal/archive drains the selected branch and its pending input before reporting success.
toolPolicies is owner configuration: reusable per-member tool policies keyed by id. A member selects one by id — tool_policy on agent_teams_add_member, a member row of an inline plan, or a profile member row. The Captain model can select a declared policy but can never submit raw allow/deny lists or invent one at runtime.
denyhides the named global tools from that member's schemas and rejects their invocation.allowkeeps only the named global tools. The four coordination tools (agent_teams_claim_task,agent_teams_update_task,agent_teams_send_message,agent_teams_status) are always added to an allow list.- The captain-only tool list is always denied and cannot be reopened by a policy; a policy may not deny a required coordination tool.
- Unknown ids, empty policies, and empty lists fail at mount or before a member is persisted.
agent_teams_statusshows the selected policy id per member; the complete resolved filter is kept inteam.jsonfor audit, and the child's durable descriptor reapplies it after a restart.- Teams and members created without a policy keep their previous tool set and record shape. An empty
tool_policyonagent_teams_edit_planclears a staged member's policy.
This is model tool visibility and invocation policy for global tools, not an OS sandbox or an authorization boundary.
memberCommandDeny is the deny list for the captain-only agent_teams_run_command tool (default ["agent-teams"], this plugin's own activation surface). Write command names without the leading slash: entries are normalized (surrounding whitespace, any number of leading slashes and case are ignored), so compact, /Compact and /COMPACT are the same switch; every entry also blocks its whole <entry>-* family (review denies /review-tools). An explicit [] denies nothing — the built-in agent-teams protection included — because the array is a literal replacement rather than an addition. An entry with no usable command name fails the plugin load with a readable first sentence instead of silently denying nothing. The activity panel's command bar is the human path and deliberately does not consult this list (the operator is the authorizer).
memberContext gives each member a plugin-level context budget: when a member is idle and its measured context usage reaches the threshold, the plugin compacts it once and records the event. maxTokens (absolute) and maxRatio (a share of the member model's contextWindow; without a known window the ratio is ignored rather than guessed) — give at least one, and the smaller budget wins when both apply. cooldownMs (default 60000) and minTurnGap (default 1) debounce repeats; onLimit (default compact) also accepts notify (logs only, never compacts) or off. With the whole memberContext block absent the feature is fully off (no watchdog, no new fields) and behavior matches previous versions; to restore the old behavior either remove the block or keep a threshold and set onLimit: off (the block still needs at least one maxTokens/maxRatio, otherwise mounting fails). A member-level context_max_tokens (on a profile member row, an inline plan member row, agent_teams_add_member, or agent_teams_edit_plan update_member; pass null to clear; the pre-run plan review panel's member card also edits it directly — the input prefills the member's declared value, and saving it blank/cleared sends null to fall back to the plugin default) joins the same resolution — the smallest of the member value, the plugin maxTokens, and the window-derived ratio wins, so it only makes that member compact earlier and never loosens the plugin bound. A busy member or one with queued waking input is skipped and re-evaluated at the next idle boundary. agent_teams_status members can carry context { tokens, limit, ratio, compactions, measured_at } plus context_live, and the activity panel shows each member's usage/limit and compaction count. See docs/3-reference/usage.md for details.
memberProvider is the sub-agent runtime backend (spawn / fork), not an LLM provider. Cross-LLM-provider routing uses the optional provider + model fields of agent_teams_add_member; memberModel is only a model default for all members. A member on the captain's current provider/model inherits the captain's reasoning effort when the captain actually chose one; a changed provider or model — or a captain running on its model's own default — leaves the effort unpinned, so the member keeps following the target model's default instead of freezing today's value into team.json. To request a particular effort, pass the optional reasoning_effort field — one of the target model's supported effort ids, or "default" to force the model's own default.
slashCommand: false disables the deterministic /agent-teams activation surfaces (slash command and gesture boundary), leaving the natural-language trigger as the only entry point.
Boundaries
- One captain leads one active team at a time.
- Idle members with no open task are automatically reused for ready work. An idle member that still owns an open attempt is parked until messaged or explicitly reassigned; messages that cannot be delivered live remain durable and are retried at a later status boundary.
- State is file-backed and serialized within one DSH process; concurrent processes editing the same team are not coordinated.
- The activity panel reports persisted state as-is. Models may occasionally finish work without performing the expected task-state update.
See docs/3-reference/usage.md for the full tool reference, state model, Web UI behavior, configuration, and known limits.
Plugin development Skill
Community upgrade, audit, benchmark, testing and release skills are vendored with a pinned source revision. See skills/README.md for local rules and CONTRIBUTING.md for the contribution workflow.
The repository also ships the open Agent Skills package dsh-plugin-development:
npx skills add ChenXu233/dsh-agent-teams --skill dsh-plugin-development
Documentation
docs/README.md is the index: tutorials, how-tos, and references, with each fact kept in one place.
| Guide | Covers |
|---|---|
| Usage | Architecture, UI behavior, tools, configuration, limits, and validation |
| Verification | Offline, composition, real e2e, and GUI verification |
| Maintain and release | Report → implementation → acceptance → publish |
| Compatibility | The single install floor, and why there is no host list |
| Tool surface | Tool inventory, session identity, approval notification |
| Plugin development | Human-readable guide built from this plugin |
| README writing | Repository documentation conventions |
Development
pnpm install
pnpm build
pnpm verify
Named multi-role profiles
Configure one or more complete team profiles in cordis.patch.yml. A profile always supplies the roster (independent provider/model/role/reasoning effort). Set taskPlanning: captain when the Captain should derive the DAG from the user's goal; omit it or set taskPlanning: seed to keep a fixed template workflow:
profiles:
demo-delivery:
description: Ship a small feature
protocol: Discuss requirements, review, test, then prepare release; do not deploy automatically.
members:
- name: analyst
model: gpt-5.6-sol
role: Analyze requirements
- name: implementer
model: gpt-5.6-terra
role: Implement the approved solution
tasks:
- id: requirements
subject: Requirements discussion
assignee: analyst
- id: implementation
subject: Implement solution
assignee: implementer
dependencies: [requirements]
Use an explicit profile flag: /agent-teams --profile demo-delivery implement the feature. The first ordinary token is never treated as an implicit profile. Normal command runs call agent_teams_create({ profile, approval: "required" }): the roster and seed/Captain-designed DAG remain staged, no child session is created, and no task is claimed. Edit the plan in the activity panel using the host model catalog, return to chat so the Captain asks what to revise and then atomically updates the same draft, discard it, or click Approve & Run. Return/discard actions cancel any planning turn still running; discard also parks model-facing context that forbids silently creating a replacement team. Approval resolves the final provider/model and any pinned reasoning choice, commits the roster, and creates each member session only when its first task is ready. A running team is stopped from its own panel header through a confirmation dialog rather than from the composer. Direct tool clients may pass approval: "automatic" for the legacy immediate path. Failed review/test tasks do not unlock downstream work; automatic repair/review tasks do not depend on the failed review.
Acknowledgements
This project is a fork of NanmiCoder/dsh-agent-teams by 程序员阿江 (Relakkes), published on npm as @nanmicoder/dsh-agent-teams. All credit for the original design and implementation belongs to the upstream author and contributors. This fork is maintained independently by ChenXu233 and released under the same MIT license; the upstream copyright notice is preserved in the LICENSE file.
License
Comments
Loading…
Similar plugins
by limuyang2
Multi-agent team collaboration for DeepSeek Harness, with independent models, skills, MCP tools, contexts, and a shared workspace.
★ 53
↓ 590/wk
MIT
TypeScript
Aug 26, 2026
dsh plugin --profile web add @limuyang2/dsh-agent-teamby toolclub
Persistent multi-model workflow teams for DeepSeek Harness — dynamic lead planning, bounded DAGs, per-agent model/tools, Run Center and Token insights.
★ 288
MIT
TypeScript
Nov 1, 2026
dsh plugin --profile web add dsh-agent-team-guiby Contexera
dsh-agent-team gives DeepSeek Harness agents that don't reset: durable Members with their own memory, notes, and skills across sessions, rollovers, and restarts. You set the direction; agents coordina
★ 106
↓ 577/wk
MIT
TypeScript
Nov 18, 2026
dsh plugin --profile web add @contexera/dsh-agent-teamby Fish-under-sea
DSH 插件(上游 NanmiCoder/dsh-agent-teams 的补充版):AgentTeams 多智能体团队协作 —— 自然语言组队、任务依赖 DAG、信箱通信、右侧栏树状监测;随包自带 17 厂商 × 10 岗位头像与 17 个厂商徽标(286 个文件),支持自定义美术目录与中文岗位名。
★ 4
↓ 193/wk
MIT
JavaScript
Nov 8, 2026
dsh plugin --profile web add dsh-agent-teams-fishby huxint
Agent teams for DeepSeek Harness: named long-lived teammates over ctx.subagents, a shared task list, a member-to-member mailbox, virtual workspaces, and a live team room in the conversation view
★ 14
↓ 155/wk
MIT
TypeScript
Sep 28, 2026
dsh plugin --profile web add dsh-teamby wowyuarm
Give DeepSeek Harness a persistent agent team for long-running collaboration
★ 69
↓ 64/wk
MIT
TypeScript
Oct 6, 2026
dsh plugin --profile web add @wowyuarm/dsh-agent-team