From lt-dev
Provides auto-detection heuristics, coordination patterns, and worktree isolation guidance for parallel Claude Code operations. Covers Agent Teams (independent sessions with messaging) and parallel subagent spawning (Agent tool with isolation worktree). Activates when user mentions "agent team", "parallel review", "parallel agents", "team debug", "parallel worktrees", "batch rebase", "parallel backend frontend", "implement in parallel", or when commands evaluate team suitability via CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS. Also activates when spawning multiple file-modifying subagents concurrently. NOT for single sequential subagent invocations.
How this skill is triggered — by the user, by Claude, or both
Slash command
/lt-dev:coordinating-agent-teamsThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Claude Code Agent Teams (`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1`) coordinate multiple independent Claude Code sessions with inter-agent messaging and a shared task list. Unlike subagents (Agent tool), teammates communicate directly and challenge each other.
Claude Code Agent Teams (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1) coordinate multiple independent Claude Code sessions with inter-agent messaging and a shared task list. Unlike subagents (Agent tool), teammates communicate directly and challenge each other.
Every team-capable command follows this decision tree:
1. Is CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 set?
No → Single Agent Mode (existing behavior)
2. Did user pass --no-team?
Yes → Single Agent Mode (forced)
3. Did user pass --team?
Yes → Team Mode (forced)
4. Does complexity heuristic match?
Yes → Team Mode (auto-detected)
No → Single Agent Mode (overhead not justified)
| Command | Team Trigger |
|---|---|
/review | >100 changed lines AND >3 files, OR changes in both projects/api/ and projects/app/ |
/create-story (TDD) | Fullstack monorepo detected AND story involves backend + frontend |
/rebase-mrs | >2 branches selected |
/debug | Always team (the workflow requires it) |
| Task Type | Team Advantage | Token Overhead |
|---|---|---|
| Multi-dimension review | Independent analysis prevents anchoring bias | ~3x |
| Fullstack test writing | Parallel backend + frontend, contract sharing | ~2x |
| Adversarial debugging | Competing hypotheses with falsification | ~3-5x |
| Batch rebase | True parallelism via worktrees | ~1.5x per branch |
Each pattern is described in detail in patterns.md. Summary:
Agent Teams cost approximately 3-5x a single agent run. This is justified when:
Not justified when:
When spawning multiple file-modifying subagents concurrently via the Agent tool (not Agent Teams), use isolation: "worktree" to prevent file conflicts:
Agent tool call:
subagent_type: "lt-dev:backend-dev"
isolation: "worktree" ← each gets its own working copy
prompt: "Implement feature X in projects/api/..."
Agent tool call:
subagent_type: "lt-dev:frontend-dev"
isolation: "worktree"
prompt: "Implement feature X in projects/app/..."
isolation: "worktree"| Scenario | Isolation needed? |
|---|---|
| Multiple file-modifying agents in parallel | Yes |
| Single agent (sequential) | No |
| Read-only agents (reviewers) in parallel | No |
| Agent modifying lockfiles/dependencies | No — needs in-place access |
| Supports worktree | No worktree (in-place only) |
|---|---|
backend-dev, frontend-dev, devops, branch-rebaser | fullstack-updater, nest-server-updater, npm-package-maintainer, all reviewers |
See worktree-guide.md for setup, cleanup, naming conventions, performance settings (worktree.sparsePaths, worktree.symlinkDirectories), dependency isolation, and known limitations.
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 environment variable| Element | Relationship |
|---|---|
/lt-dev:debug | Always uses team (Adversarial Convergence pattern) |
/lt-dev:review | Auto-detects team for large/fullstack changes |
/lt-dev:create-story | Auto-detects team for fullstack TDD |
/lt-dev:git:rebase-mrs | Auto-detects team for batch operations |
Note: /lt-dev:debug REQUIRES Agent Teams (no single-agent fallback). All other commands auto-detect based on complexity heuristics and fall back to single-agent mode gracefully.
npx claudepluginhub lennetech/claude-code --plugin lt-devOrchestrates teams of Claude Code sessions for parallel work on research, features, debugging, and cross-layer tasks using shared tasks, inter-agent messaging, and team lead coordination.
Coordinates multiple Claude Code instances as agent teams with shared tasks, inter-agent messaging, and management for parallel research, feature development, debugging, and cross-layer coordination.
Coordinates multiple Claude Code sessions as a team with shared task lists, mailbox messaging, and file-lock claiming. Includes patterns for team sizing (3-5 recommended), task decomposition strategies, and dependency handling.