Knuckles-Team Agent Spawner
Run one temporary isolated agent using deployment-neutral AgentConfig:
python scripts/spawn_agent.py --prompt "Inspect the configured tool catalog"
CHAT_MODELS, MCP_CONFIG, MCP_URL, the custom-skills root, TLS policy, and
runtime secret resolution come from AgentConfig/XDG configuration. Provider and
model identifiers may be selected explicitly when they already exist in the
deployment's model catalog:
python scripts/spawn_agent.py \
--provider openai \
--model-id configured-model \
--prompt "Perform the bounded delegated task"
The compatibility flags that accepted API keys, arbitrary dotenv/config paths, base URLs, personal agent names, or disabled TLS verification are retired. Never put credentials in arguments. The child uses strict tool guards and MCP isolation, emits no prompt/name/path/endpoint details, and relies on the shared transport profile. Use GraphOS delegation when policy, capability provenance, or trace-linked skill execution is required.