spawn-team

Spawn-Team Skill

Launches an isolated LLM instance with coordinated review from a secondary LLM. This enables automated code review workflows where one LLM implements changes and another reviews them.

Usage

When you need delegated task execution with automated code review:

1. Describe the task in the prompt
2. The primary LLM (default: claude-code) implements the changes
3. The reviewer LLM (default: gemini-cli) evaluates the changes
4. Based on coordination mode, either:
   • Sequential: Primary applies fixes once, then PR is created
   • Ping-pong: Primary and reviewer iterate until approval

Invocation

/spawn-team "implement the new authentication flow"

With coordination mode:

/spawn-team --sequential "straightforward feature"
/spawn-team --ping-pong "complex refactoring"

With custom LLMs:

/spawn-team --reviewer claude-code "use Claude as reviewer"
/spawn-team --primary gemini-cli --reviewer claude-code "swap roles"

Coordination Modes

Sequential (Default)

Primary LLM completes the task, reviewer provides feedback, primary applies fixes once. Best for straightforward tasks.

Primary → Reviewer → Primary (fixes) → PR

Ping-Pong

Primary and reviewer alternate until the reviewer approves or max iterations reached. Best for complex tasks requiring iteration.

Primary → Reviewer → Primary → Reviewer → ... → PR

Control iterations:

/spawn-team --ping-pong --max-iterations 5 "complex task"

Supported LLMs

Identifier	Description
claude-code	Claude Code CLI (default primary)
gemini-cli	Gemini CLI (default reviewer)

Configuration

Settings in .infinite-probability/improbability-drive.toml:

[spawn-team]
coordination = "sequential"  # or "ping-pong"
max_iterations = 3
reviewer_llm = "gemini-cli"

[spawn]
default_llm = "claude-code"

Review Process

The reviewer LLM receives:

1. The original task prompt
2. A git diff of the changes made
3. Instructions to respond with structured JSON

Review response format:

{
  "verdict": "approved",
  "suggestions": []
}

Or with suggestions:

{
  "verdict": "needs_changes",
  "suggestions": [
    {
      "file": "src/auth.rs",
      "line": 42,
      "issue": "Missing error handling",
      "suggestion": "Add Result return type"
    }
  ]
}

Fix Process

When the reviewer returns needs_changes, the primary LLM receives:

1. The original task prompt
2. The list of issues to fix with file paths and suggestions
3. Instructions to address each issue

Output

Returns a SpawnTeamResult containing:

• success: Whether the team operation succeeded
• iterations: Number of review iterations performed
• final_verdict: Approved, NeedsChanges, or Failed
• reviews: All review results with suggestions
• summary: Human-readable summary

When to Use Spawn-Team

Use spawn-team when:

• Task requires validation beyond basic completion
• You want automated code review before PR creation
• Task is complex enough to benefit from a second perspective
• You want to catch issues before human review

Use basic spawn when:

• Task is simple and well-defined
• Speed is more important than review
• You'll do manual code review anyway

Safety

Same isolation guarantees as /spawn:

• Sandbox runs in isolated git worktree
• No access to $HOME or config files
• --dangerously-skip-permissions never allowed
• Secrets injected as env vars, never logged
• Both primary and reviewer run in sandboxes

Example Workflow

User: /spawn-team --ping-pong "add rate limiting to the API"

Spawn-team:
1. Primary (claude-code) implements rate limiting
2. Reviewer (gemini-cli) finds edge case: "No handling for burst traffic"
3. Primary addresses feedback, adds burst handling
4. Reviewer approves changes
5. PR created with 2 review iterations documented

Result:
- Status: Success
- Iterations: 2
- Final verdict: Approved
- PR: https://github.com/repo/pull/123