samba-plugin

Skills

Claude Code skills for managing SambaNova models and running coding sub-agents.

Installation

Step 1: Add the marketplace

In Claude Code, run:

/plugin marketplace add <this repository link>

Step 2: Install the plugin

/plugin install samba-plugin

Prerequisites

• Python 3 with venv support
• A SAMBANOVA_API_KEY environment variable (required by /model-info, /code, and others.)
• Run /setup before using any other skill
• At least one of continue or opencode installed (CLI available — no model configuration required)

How it works

The plugin maintains its own local SQLite database (model_params.db) that stores all model configurations. This database is the sole source of truth — the plugin never reads your personal Continue or OpenCode config files. When you use /code, the plugin generates isolated, runtime-only configurations for the sub-agent, ensuring a clean separation from your existing IDE setups.

Skills Overview

Skill	Command	Description
setup	/setup	Initialize the virtual environment and install agent_shims
code	/code <tool> <model> <prompt> <cwd> [files...]	Delegate a coding task to a sub-agent
list-models	/list-models	List all models in the local parameters database
model-info	/model-info	Show all models available on the SambaNova platform
update-model	/update-model <name> <ctx> <max_tokens> [params_json]	Add or update a model in the database
reset-model-db	/reset-model-db	Clear all entries from the model database

Skill Details

code

Runs a coding tool (continue or opencode) as a sub-agent with a specified model and prompt. Useful for delegating tasks like code review, implementation, or ideation.

/code <tool> <model> <cwd> <prompt> [--max-tokens <n>] [--tool-arg <arg>...]

Arguments:

Argument	Required	Description
tool	Yes	continue or opencode (see tool selection guide below)
model	Yes	Bare model ID from the database (e.g. MiniMax-M2.7, not sambanova/MiniMax-M2.7)
cwd	Yes	Working directory for the tool (defaults to project root if unspecified)
prompt	Yes	The prompt to send, quoted as a single shell argument
--max-tokens	No	Override the model's max_completion_tokens for this run
--tool-arg	No	Extra args passed to the underlying tool (repeatable, use = syntax for flags: --tool-arg="--file")

Tool selection guide:

Scenario	Tool
Large code generation (1,000+ lines)	continue — iterates across turns, tested up to 3,700 lines
Code review of large files (1,500+ lines)	continue — reads in chunks; opencode truncates at ~1,500 lines
Qwen3-235B + C++	continue (cleaner) or opencode
Side-effect-free output needed	opencode — doesn't modify filesystem

See skills/code/prompting/ for detailed model and tool guides.

Common model aliases:

User says	Maps to
MiniMax, sambanova/MiniMax-M2.7	MiniMax-M2.7
gpt-oss, sambanova/gpt-oss-120b	gpt-oss-120b

list-models

Lists all models currently stored in the local parameters database along with their context length, max completion tokens, and sampling parameters.

/list-models

model-info

Queries the SambaNova API (https://api.sambanova.ai/v1/models) to display the full catalog of available models with their context length and max completion tokens. Requires SAMBANOVA_API_KEY to be set.

This shows what models can be used, as opposed to /list-models which shows what is stored locally.

/model-info

update-model

Inserts or updates a model entry in the local parameters database. Looks up model details from /model-info and sampling parameters from HuggingFace documentation before writing.

/update-model <name> <context_length> <max_completion_tokens> [sampling_parameters_json]

The sampling parameters argument is a JSON string (e.g. '{"temperature": 0.7, "topP": 0.9}').

reset-model-db

Deletes all entries from the model parameters database. Prompts for confirmation before proceeding.

/reset-model-db

Architecture

skills/
├── setup/                  # Environment initialization
├── code/                   # Sub-agent coding tool runner
├── list-models/            # Database model listing
├── model-info/             # SambaNova platform model catalog
├── update-model/           # Database model upsert
└── reset-model-db/         # Database reset