Team Composition

This skill determines the optimal number of teammates and the correct agent type for a swarm run. It is invoked during the planning phase, before team creation.

Analysis Process

1. Read

   tasks.md

   from the spec directory.
2. Parse the File Manifest at the bottom of tasks.md (or the

   Files:

   section of each task).
3. Compute parallel batches by analyzing file conflicts:
   • Two tasks conflict if they share any file (CREATE or MODIFY).
   • Tasks with dependencies cannot be in the same batch as their prerequisites.
   • Group non-conflicting, dependency-free tasks into batches (same algorithm as the coordinator's Runtime Parallelism Computation).
4. Find the largest batch. That count is the maximum number of useful teammates — spawning more than that means idle agents burning tokens.
5. Apply the hard cap of 5 teammates. Beyond 5, coordination overhead and token costs grow faster than throughput. Diminishing returns hit hard.
6. Recommended teammate count:

   min(largest_batch_size, 4)

   . Using 4 instead of 5 gives a buffer for the coordinator to stay responsive.

Examples

Agent Type Selection

When

--agent-type

is explicitly provided

Use that type for all teammates. No analysis needed.

When

--agent-type

is

auto

or not provided

Analyze the project to determine the dominant language/framework:

1. Check for build system / config files (most reliable signal):

   • go.mod

     --> Go project -->

     golang-pro

   • package.json

     with TypeScript dependencies or

     tsconfig.json

     -->

     typescript-pro

   • package.json

     without TypeScript -->

     typescript-pro

     (JS projects benefit from TS tooling awareness)

   • Cargo.toml

     --> Rust project -->

     rust-pro

   • pyproject.toml

     ,

     setup.py

     ,

     setup.cfg

     ,

     requirements.txt

     -->

     python-pro

   • mix.exs

     --> Elixir project -->

     elixir-expert

2. If no config files found, check file extensions in the project root and

   src/

   directory:
   • Majority

     .go

     files -->

     golang-pro

   • Majority

     .ts

     /

     .tsx

     files -->

     typescript-pro

   • Majority

     .js

     /

     .jsx

     files -->

     typescript-pro

   • Majority

     .py

     files -->

     python-pro

   • Majority

     .rs

     files -->

     rust-pro

   • Majority

     .ex

     /

     .exs

     files -->

     elixir-expert

   • Majority

     .sql

     files -->

     sql-pro

3. If tasks span multiple languages (e.g., Go backend + TypeScript frontend), use

   general-purpose

   . Specialized agents struggle when they hit code outside their domain.

4. If unclear or mixed, default to

   general-purpose

   . It is the safest fallback.

Agent Type Availability Note

The language-specific types (

golang-pro

typescript-pro

python-pro

rust-pro

elixir-expert

sql-pro

Fallback chain (always apply):

1. Try detected language-specific agent (e.g.,

   golang-pro

   ,

   typescript-pro

   )
2. If unavailable → fall back to

   swarm-executor

   (included with this plugin)
3. If unavailable → fall back to

   general-purpose

   (always available)

Agent Type Reference

golang-pro

typescript-pro

python-pro

rust-pro

elixir-expert

sql-pro

swarm-executor

general-purpose

Cost Awareness

Exact costs depend on model, pricing, prompt length, and task complexity. No dollar estimates are provided — they would be speculative.

What affects cost:

• Number of teammates (each is an independent LLM session)
• Task complexity (more turns = more tokens)
• Verification cycles (failed tasks trigger retries)

Cost-saving tips:

• Sequential mode for < 4 tasks
• Fewer teammates with more tasks > many idle teammates

• --max-iterations

  to cap runaway sessions
• Review plan before execution (don't use

  --yolo

  unless you mean it)

Output Format

When reporting team composition to the user or the coordinator, always use this format:

Team: <N> x <agent-type> in worktrees
Computed batches: <B> (largest batch: <L> tasks)

Example Output

Team: 3 x golang-pro in worktrees
Computed batches: 4 (largest batch: 3 tasks)

If the user has not confirmed yet, present this and wait for approval before proceeding to team creation.