Feature Request

Allow sub-agent spawns to specify a tool scope to reduce system prompt size for small/local models.

Problem

Sub-agents receive the full tool definition set (~20+ tools) in their system prompt, consuming thousands of tokens. For local models with limited context (8-16K) and output caps (4K), this leaves insufficient room for reasoning and response generation.

A Qwen3 30B sub-agent tasked with "edit one line in a file" gets definitions for browser, cron, message, nodes, canvas, tts, etc. — none of which it needs. The model then spends its limited output budget on thinking tokens before it can even make a tool call.

Proposed Solution

Add a tools or toolScope parameter to sub-agent spawning:

{
  "task": "Edit the DND section in AGENTS.md",
  "model": "ollama/qwen3:30b-a3b-16k",
  "toolScope": "files"
}

Possible scopes:

• files — read, write, edit only (~minimal prompt overhead)
• shell — read, write, edit, exec, process
• web — web_search, web_fetch, browser
• full — everything (current default)
• none — no tools, text-in text-out only
• Or explicit list: ["read", "write", "edit"]

Impact

• Makes local models (8-16K context) viable for tool-using sub-agent tasks
• Reduces token waste on irrelevant tool definitions
• Faster prefill = faster time-to-first-token
• Could also benefit cloud model sub-agents (cheaper, faster)

Current Workaround

Use cloud models (Mistral, Sonnet) for tool-using sub-agents, reserve local models for text-only tasks.