DSH Auto Team

English | 中文

@oh-my-dsh/auto-team is a host-only DeepSeek Harness plugin (Cordis id auto-team). It gives long tasks a scored start and a team-shaped run, and it keeps the numeric score out of the model's context.

The plugin adds the first-round decision, the goal bridge, and the Lead and worker policy gates. The teammate, task-board, and worktree tools come from the host profile. It targets DSH 0.2.1-rc.2.

How a task runs

The Lead calls auto_team_score once in the first round. The call takes five factors from 1 to 5, covering scope, parallelism, ambiguity, risk, and expected_steps, plus the task goal in its ordinary meaning.

The plugin turns the factors into a weighted score from 0 to 1. The model gets only team or solo; the score and the effective threshold stay host-side. The default threshold is 0.55. If a user-created goal already exists at the first score, the threshold is multiplied by 0.6.

The first round must be scored. If the turn tries to close before the tool ran, an agent/turn-stopping listener steers the Lead back with a reminder.

The score is frozen for the task. It is written to the session log as a plugin:auto-team/score record and folded into the host-only autoTeam projection, so resuming a session keeps the decision and does not score again.

On a team decision the plugin:

• reuses the current goal, or starts one when none exists;
• ensures a user-facing reporting teammate named reporter, which is configurable; the Lead relays its updates to the user;
• tells the Lead to keep the main Git repository and the team, and not to write code;
• denies the Lead's write and edit calls by default;
• requires each execution teammate to complete create_worktree before its own write or edit is allowed, and checks that an edit path stays inside the registered worktree when the host returns one. The reporter is exempt.

Multi-round execution belongs to the goal runtime. The plugin starts or reuses the goal and does not run its own continuation loop.

On a solo decision the task runs as a normal single-agent task.

Requirements

The profile must compose these host services:

• agents, tools, systemPrompt, sessionProjections
• agentTeams from @deepseek-ai/dsh-experimental-agent-team
• goals from @deepseek-ai/dsh-goal

The plugin loads without them but logs a warning and installs nothing. When a service arrives later, an internal/service listener retries the install.

Install

Build and pack the package:

pnpm install
pnpm run lint
pnpm test
pnpm run build
pnpm pack

Install from npm once the package is published:

dsh plugin --profile <profile> add @oh-my-dsh/auto-team

Or install a locally packed tarball:

dsh plugin --profile <profile> add ./oh-my-dsh-auto-team-0.1.0.tgz

The package name is @oh-my-dsh/auto-team, published under the @oh-my-dsh scope.

Configuration

Every field is optional. Schemastery checks the values at load; invalid values fail activation.

Field	Default	Meaning
message	auto-team loaded	Activation log line.
threshold	0.55	Base score needed for Team mode.
manualGoalFactor	0.6	Multiplier applied to the threshold when a user-created goal exists.
weights	scope 0.3, parallelism 0.25, ambiguity 0.15, risk 0.15, expectedSteps 0.15	Relative weight of each difficulty factor.
reportingAgentName	reporter	Name of the reporting teammate.
reportingProvider	spawn	Subagent provider used for the reporter.
goalMaxRounds	256	Round cap used when the plugin starts a goal.
enforceLeadNoCode	true	Install the Lead write/edit gate.
leadDeniedTools	['write', 'edit']	Tool names denied to a Team-mode Lead.

Tool

The package adds one model-facing tool.

auto_team_score takes:

Parameter	Type	Meaning
scope	integer 1 to 5	Expected change scope.
parallelism	integer 1 to 5	How much independent work can run at once.
ambiguity	integer 1 to 5	How much requirement or design uncertainty remains.
risk	integer 1 to 5	How costly a wrong change would be.
expected_steps	integer 1 to 5	Expected task length.
objective	string	The task goal in its ordinary meaning. Used only when the plugin starts a goal.

It returns { decision, goal, reporter? }. The numeric score and threshold are not in the result. The host's Agent Teams tools handle spawning, messaging, and the shared task board. The host goal lifecycle handles completion and continuation.

Guards and limits

• The Lead gate is a tool-name policy, not a sandbox. pwsh, run_code, and direct filesystem effects can bypass it. The policy tells the Lead not to use them as a workaround.
• The worker gate checks that a create_worktree call succeeded and, when the host returns the worktree path, that each write or edit uses an absolute file_path inside it. If the host result has no path, the gate falls back to checking only that the call succeeded.
• The reporter is spawned with the configured provider. If the spawn fails, the score stays frozen and a later call retries the reporter.
• A real DSH profile composition test has not been run in this repository. tests/runtime.test.ts mounts a real Cordis context and a real detached Session with fake host services, and covers the install path, the scoring tool, the stop guard, the Lead gate, and the worker gate.

Development

pnpm install
pnpm run lint
pnpm test
pnpm run build
pnpm pack --dry-run --json

The tests cover the scoring policy, configuration, goal and reporter orchestration, the durable score projection, the runtime install path, and the Loader export shape.

Read the design note for the behavior and the flow chart, and the plugin contracts for the package rules.

Status

The core flow works and is covered by the repository tests. The release checks pass. Install the package in a profile that composes the required services, then run a long task to verify the behavior in your environment.