other-sub-agents

A Claude Code skill that delegates development sub-tasks to local and cloud LLMs.

Claude acts as the orchestrator -- it decides what to delegate, picks the right model, crafts the prompt, and integrates the results. Tasks like generating boilerplate, writing tests, reviewing code, and drafting docs can be handled by cheaper or local models, saving API costs and leveraging your hardware.

Providers

Provider	Privacy	Cost	Context	Offline
Ollama (local)	Full -- data stays on machine	Free (your GPU)	2K-32K typical	Yes
Gemini (cloud)	Data sent to Google	Free tier available	Up to 1M tokens	No

Supported Tasks

• Code generation -- functions, classes, modules from specs
• Debugging -- find and fix bugs
• Code review -- style, performance, correctness
• Test writing -- unit tests, edge cases
• Documentation -- docstrings, READMEs, API docs
• Architecture -- design discussions, planning
• Refactoring -- restructure and clean up code

Installation

As a Claude Code Skill

Copy or symlink this directory into your Claude Code skills folder:

# Clone the repo
git clone https://github.com/bh679/other-sub-agents.git

# Symlink into Claude Code skills
ln -s "$(pwd)/other-sub-agents" ~/.claude/skills/other-sub-agents

Prerequisites

Ollama (for local models):

# Install Ollama (macOS)
brew install ollama

# Start the server
ollama serve

# Pull a code model
ollama pull qwen2.5-coder:32b

Gemini (for cloud models):

# Get an API key from https://aistudio.google.com/apikey
export GEMINI_API_KEY="your-key-here"

Usage

Discover Available Models

python3 scripts/discover_models.py                    # All providers
python3 scripts/discover_models.py --provider ollama   # Local only
python3 scripts/discover_models.py --provider gemini   # Cloud only
python3 scripts/discover_models.py --task code --json  # Best model for code tasks

Run a Sub-Agent Task

Local (Ollama):

python3 scripts/agent_runner.py \
  --model "qwen2.5-coder:32b" \
  --task code \
  --prompt "Write a Python function that validates email addresses."

Cloud (Gemini):

python3 scripts/agent_runner.py \
  --provider gemini \
  --model "gemini-2.5-flash" \
  --task code \
  --prompt "Write a Python function that validates email addresses."

With file context:

python3 scripts/agent_runner.py \
  --provider gemini \
  --model "gemini-2.5-pro" \
  --task review \
  --prompt "Review this code for bugs and performance issues:" \
  --file "/path/to/code.py"

Key Flags

Flag	Description
--provider ollama|gemini	Which provider (default: ollama)
--model	Model name (e.g. qwen2.5-coder:32b, gemini-2.5-flash)
--task	Task type: code, review, debug, test, docs, architecture, refactor, general
--file	Include a file as context
--temperature	Override default temperature
--stats	Print generation stats to stderr
--timeout	Timeout in seconds (default: 120)
--max-tokens	Max output tokens (Gemini only)
--no-stream	Wait for full response instead of streaming

Routing Logic

The skill uses an Ollama-first strategy:

1. Use Ollama when: code is private/sensitive, batch processing needed, suitable local model exists
2. Escalate to Gemini when: large context needed (>32K tokens), no suitable local model, no local GPU
3. Fall back to Claude when: neither provider is available

See references/model-profiles.md for detailed routing decisions.

File Structure

other-sub-agents/
  SKILL.md                       # Orchestration instructions for Claude Code
  scripts/
    agent_runner.py              # Core execution engine (both providers)
    discover_models.py           # Unified model discovery
  references/
    model-profiles.md            # Model capabilities and routing matrix

License

MIT