Read AGENTS.md, CLAUDE.md, and the canonical
~/.agentskills/rules/agent-provider-policy.md before substantial work. Classify every work
unit before spending premium-model capacity: when OpenCode can complete it effectively, delegate
through LiteLLM to configured local Ollama or rotating DeepSeek routes and assess safe swarm
fan-out with ~/.agentskills/bin/opencode-swarm.mjs. Codex retains independent research,
architecture challenge, sensitive/shared-state judgment, integration review, and verification.
Do not conflate internal agent routing with product inference. The aid product runtime uses
the ProjectSites/no-LiteLLM route specified in AGENTS.md.