start-work

Start Work - Atlas Orchestrated Execution

Execute a plan using agent teams. Follow these EXACT steps:

Step 1: Locate and read the plan

If "$ARGUMENTS" is provided, read .sisyphus/plans/$ARGUMENTS.md. If not provided, list files in .sisyphus/plans/ and ask the user which plan to execute. If no plans exist, suggest running /oh-my-team:plan first and stop.

You MUST read the entire plan file content now. Store it — you'll embed it in the Atlas briefing.

Step 2: Extract plan name and tasks

The plan name is the filename without .md extension (e.g., auth-api from auth-api.md).

Parse all TODO items from the plan. Each - [ ] line is a task.

Step 3: Create the team

TeamCreate(team_name="exec-{plan-name}", description="Executing plan: {plan-name}")

Step 4: Create tasks from the plan

For each TODO item in the plan:

TaskCreate(subject="[TODO title from plan]", description="[TODO details from plan]")

Set up dependencies between tasks using TaskUpdate(taskId="...", addBlockedBy=["..."]) where the plan indicates ordering.

Step 5: Spawn Atlas as the conductor

CRITICAL: Embed the FULL plan content in Atlas's prompt. Atlas cannot read files outside its briefing.

Agent(
  prompt="You are Atlas, conducting the execution of a work plan.

  ## THE PLAN (embedded — this is your source of truth)

  [PASTE THE ENTIRE PLAN FILE CONTENT HERE]

  ## YOUR PROTOCOL

  1. Read the plan above carefully. Identify all tasks and their dependencies.
  2. Group independent tasks into parallel waves.
  3. For each task, spawn the appropriate teammate:
     - Implementation/coding → subagent_type='hephaestus'
     - Research/investigation → subagent_type='explorer'
     - Architecture decisions → subagent_type='oracle'
  4. Brief each teammate with the 6-section format:
     TASK, EXPECTED OUTCOME, MUST DO, MUST NOT DO, CONTEXT
  5. After each teammate completes: READ the changed files yourself. Verify the work.
  6. Only mark a task complete after verification passes.
  7. If verification fails: send feedback to the teammate (max 3 retries).
  8. After all tasks: run final verification (build/test commands if available).
  9. Report completion to the team lead.

  ## RULES
  - NEVER write code yourself. Delegate everything.
  - NEVER skip verification. Read every changed file.
  - NEVER proceed if a critical task failed.",
  subagent_type="atlas",
  team_name="exec-{plan-name}",
  name="atlas"
)

Step 6: Monitor

Atlas handles the orchestration from here. Monitor progress via the task list (Ctrl+T). Atlas will spawn workers, assign tasks, and verify results.

When Atlas reports completion, optionally suggest: "Run /oh-my-team:review-work for a comprehensive 5-agent review."