3.3 KiB
agent-harness
Standalone grok-class coding agent on QVAC. Sample → tools → compact → repeat. No BridgeSwarm, no Chrome native messaging, no extension.
Inference is @qvac/sdk (Node worker). Only Hugging Face GGUFs that already exist as QVAC constants load. Arbitrary HF repos will not.
Install
Node.js 20 or newer. @qvac/sdk itself wants 22.17+ for inference.
cd ~/dev/agent-harness
npm install
Weights download on first load into the QVAC cache (see qvac.config.json / cacheDirectory).
CLI
node bin/cli.js --list-models
node bin/cli.js --cwd ~/src/myapp "list the workspace and summarize"
node bin/cli.js --model qwen3.5-4b --yes
--yes skips write/shell prompts. Default model is qwen3.5-4b. On a 16 GB card, gemma4-4b and qwen3-8b also fit.
Sessions and permission rules live in ~/.agent-harness (or $AGENT_HARNESS_HOME).
Embed
const Agent = require('~/dev/agent-harness') // or require('/root/dev/agent-harness')
await Agent.engine.load({ model: 'qwen3.5-4b', tools: true, device: 'auto' })
const session = await Agent.create({
cwd: process.cwd(),
model: 'qwen3.5-4b',
permissionMode: 'ask', // ask | allowlist | always-approve
tools: [
{
name: 'get_time',
description: 'Current unix time',
parameters: { type: 'object', properties: {} },
execute: () => Date.now(),
},
],
})
session.on('agent_message_chunk', (ev) => process.stdout.write(ev.text || ''))
session.on('permission', (ev) => session.permit(ev.jobId, ev.toolCallId, 'allow'))
await session.prompt('List the workspace and summarize')
await session.dispose()
await Agent.engine.close()
hostWorkspace: false drops host FS/shell/memory so you can register your own read_file / run_terminal_cmd (same pattern Pip uses against a container).
Tools
Host builtins (cwd-jailed): read_file, write_file, search_replace, grep, list_dir, run_terminal_cmd, todo_write, web_search, web_fetch (opt-in), memory_*, plan mode, ask_user_question, update_goal, subagents (task), MCP HTTP (search_tool / use_tool).
Custom tools with execute run in-process. Without execute, the loop emits tool_request and waits for session to call the loop resolver (embedder-owned handlers).
Catalog
| id | tools | vision | ~weights |
|---|---|---|---|
| qwen3.5-0.8b | yes | yes | 0.7 GB |
| qwen3.5-2b | yes | yes | 1.6 GB |
| qwen3.5-4b | yes | yes | 2.8 GB |
| qwen3.5-9b | yes | yes | 5.5 GB |
| gemma4-2b | yes | yes | 3.5 GB |
| gemma4-4b | yes | yes | 5 GB |
| qwen3-8b | yes | no | 5 GB |
| qwen3vl-2b | yes | yes | 1.5 GB |
| qwen3.6-27b / gemma4-31b / gpt-oss-20b / … | yes | varies | larger card |
Not added: Qwen3-14B, Gemma 3 12B, Llama 3.1 8B, and similar chats with no QVAC constants.
Layout
index.js Agent.create / load / catalog
bin/cli.js REPL and one-shot prompt
lib/qvac.js @qvac/sdk load + completion
lib/catalog.js id → QVAC constant
agent/loop.js turn loop
agent/tools.js sandbox execute
agent/*.js compact, plan, goal, MCP, permissions, …
This is a port of the BridgeSwarm grok-class loop onto Node + @qvac/sdk. It is not the Pip panel client and does not speak native messaging.
Test
npm test
Unit tests cover catalog, compaction, tools policy, and path jail. They do not download a GGUF.