Execute Stage

Owns Phase 0 (resolve plan + decompose + approve), Phase 3 (create team + init workspace), and Phase 4 (per-task adversarial review pipeline + error recovery). Auto-chains to the audit skill when work completes.

This skill is the entry point for the agent-team workflow. It accepts either a plan file path (from superpowers:writing-plans) or a free-form task description; if no plan is found, it invokes superpowers:writing-plans to produce one before continuing.

Overview

The execute stage takes an approved plan and brings it to life. It creates the team infrastructure, spawns teammates with precise role definitions and file ownership, then coordinates their parallel work through structured communication. When teammates hit errors, it applies bounded auto-recovery before escalating.

This stage takes a plan file (or the user's task description) and brings it to life. Phase 0 resolves the plan, decomposes into parallel streams, and gates on user approval. Phase 3 creates the workspace and team. Phase 4 runs the per-task adversarial review pipeline (Executor → Reviewer → Challenger) until all tasks are completed or abandoned.

Preconditions

1. Agent Teams feature flag is enabled — verify TeamCreate tool is available. If not, tell the user: "Agent Teams requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in your settings.json env or shell environment. Please enable it and restart."
2. Optional: A plan file at the path passed in $ARGUMENTS, or in docs/plans/, or superpowers:writing-plans is installed (Phase 0.1 will auto-invoke it).
3. Optional: An existing .agent-team/{team-name}/ workspace if resuming. Phase 3 Step 1a detects resumable workspaces via the detect-resume.sh hook and offers resumption.

If preconditions are not met (Agent Teams flag missing), exit with the message above and stop.