Multitask Command

/multitask starts a bounded collaborative run from one prompt. It is for work that can be split into independent lanes, run in parallel, and synthesized into one answer.

Syntax

/multitask <request>
/multitask <lane-count> <request>

Examples:

/multitask fix the onboarding bugs and verify the flow
/multitask 6 audit the repo for performance issues and propose patches

The lane count defaults to 4 and is clamped to 2-8. If the command has no request after it, CoWork shows a validation error and does not create a task.

Runtime Behavior

When a desktop composer prompt starts with /multitask, CoWork:

• strips the /multitask prefix before saving the task prompt
• creates a normal task with collaborative multitask metadata
• creates an ephemeral agent team and AgentTeamRun
• plans lane-specific AgentTeamItem records
• dispatches each lane as a child task through the existing graph-backed team orchestrator
• keeps child tasks under the global task queue limit by setting bypassQueue=false
• synthesizes lane outputs through the existing collaborative synthesis phase

The task appears in the same task timeline and sidebar surfaces as other collaborative work, with a Multitask session label where available.

Jev team selection

If Jev Decision Support and automatic team selection are enabled, CoWork can use Jev to choose the best active roles and leader for the multitask run. Jev supplies only the structured team, lane, or strategy decision; the configured chat model still runs child tasks and synthesizes the result. With the JEV harness set to Active, Jev also selects bounded lane candidates and avoids the extra generative LLM lane-planning call.

Without Active adaptive strategy, Jev team selection is skipped when a run is already capped at one agent. With the JEV harness set to Active, an eligible foreground root task may still receive a bounded task-strategy decision while it is a single-agent candidate; Jev can keep it single agent or promote it when confidence, cost guards, and eligible roles permit. Explicit collaborative/multitask/verification work, child tasks, and background work keep their existing route.

If Jev is disabled, unavailable, times out, or returns an unusable decision, Observe mode continues with the existing chat-model selector and then deterministic keyword matching. Active mode skips the extra chat-model selector after a Jev failure and uses deterministic matching. The task timeline identifies which selection path was used.

Lane Planning

Lane assignment is best-effort and bounded:

1. Explicit bullet or numbered lists in the prompt become lane definitions first.
2. If no explicit lanes are present, Active JEV mode asks Jev to select from bounded lane focuses; otherwise CoWork asks the configured default LLM for JSON lane specs.
3. If the decision call fails or returns invalid output, CoWork falls back to deterministic lanes such as context/scope, implementation, risk review, and verification. Active mode does not spend a second generative LLM call after a Jev failure.

Each child prompt receives the root task context plus its lane title/details and is instructed to work only inside that lane, report files changed, and call out risks or blockers.

Worktree Safety

For coding requests, /multitask is safest with Git Worktree Isolation enabled. Existing worktree settings still apply to each child task, so parallel lanes can receive separate branches/worktrees when the workspace supports it.

If a prompt looks code-related and worktree isolation is unavailable, CoWork logs a non-blocking warning on the root task instead of failing the run.

Scope

V1 is command-first. It does not yet promote already queued tasks into one multitask run, and it is not a separate Agents Window. Use normal Agent Teams or Collaborative Mode when you want persistent team management rather than one prompt split into lanes.

Implementation Landmarks

Focused Checks

npx vitest \
  src/shared/__tests__/multitask-command.test.ts \
  src/electron/agents/__tests__/MultitaskLanePlanner.test.ts \
  src/electron/agents/__tests__/AgentTeamOrchestrator.test.ts

For broader validation after multitask changes:

npx vitest src/shared/__tests__ src/electron/agents/__tests__ src/renderer/components/__tests__
npm run type-check