NanmiCoder/dsh-agent-teams

English · 简体中文

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, 14 coordination tools, persistent state, an automatic shared-task scheduler, and a live Web UI—without requiring a separate workflow engine.

v0.1.23 fixes team identity, member model routing, quality and repair gates, and cold fork recovery, and adds owner-configured member tool policies. This plugin release uses npm latest; the recommended host remains Harness 0.2.0-rc.2 (its next channel). See the exact support matrix.

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.

Recommended pair: DeepSeek Harness 0.2.0-rc.2 + AgentTeams 0.1.23. The host remains an RC.

In the app sidebar, open Plugins → Add plugin and enter the npm package spec:

@nanmicoder/dsh-agent-teams@0.1.23

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 in 0.2.0-rc.2. Check the embedded core against the compatibility list. Harness 0.2.0-rc.2 is now the exact published target; its RC status is unchanged. The earlier source-only 0.2.0 prediction is no longer a support declaration. See the release migration evidence.

npm install --global @deepseek-ai/dsh@0.2.0-rc.2
dsh --version

Skip this if you already run this version. Alpha is opt-in: select an exact Alpha version from the table and lock all host dependencies as described in the maintenance guide.

Install or upgrade from npm (pinned to this latest release):

dsh plugin --profile web add --save-exact @nanmicoder/dsh-agent-teams@0.1.23

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.

For Git installation, use a Git package spec pinned to a commit containing the verified server, client, and type build outputs in lib/, rather than an npm webpage URL. 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.

For older 0.1.0-* / 0.1.1-* or unlisted hosts, keep a working pair and follow the older-version and diagnostic guide.

See the full compatibility matrix, source installation and Alpha testing guide, and verification coverage 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.

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. Older hosts keep the activity panel.

1. 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_status when current state needs checking. When no team exists, the goal becomes a staged plan for review.
2. The captain adds role-specific members backed by continuable sub-agents.
3. The goal becomes tasks with owners and explicit dependencies.
4. The shared scheduler uses real running / idle / ready state 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.
5. 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.
6. 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 reasoning effort, while a member on a requested alternative route snapshots the target model's default effort; later continuations restore the resolved snapshot. 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 the original 14 native team tools 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.

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.

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]

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.

• deny hides the named global tools from that member's schemas and rejects their invocation.
• allow keeps 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_status shows the selected policy id per member; the complete resolved filter is kept in team.json for 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_policy on agent_teams_edit_plan clears a staged member's policy.

This is model tool visibility and invocation policy for global tools, not an OS sandbox or an authorization boundary.

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, while a changed provider or model automatically uses the target model's default. 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.

• 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/usage.md for the full tool reference, state model, Web UI behavior, configuration, and known limits.

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 NanmiCoder/dsh-agent-teams --skill dsh-plugin-development

pnpm install
pnpm build
pnpm verify

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/reasoning choices, 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.

MIT