SDK should support generic local inference for subagent/tool-call tasks (BYOM limited to Anthropic only)

[ian-morgan99]

Problem

The Copilot SDK does not expose a configuration path to route subagent and tool-call tasks through generic local inference endpoints (e.g., LM Studio, Ollama). While BYOM providers are supported, they are limited to Anthropic providers — there is no mechanism for arbitrary local model routing.

What BYOM Does Today

• BYOM supports Anthropic provider configuration
• Subagents inherit the session model by design
• The multiplier guard prevents subagent model escalation (see #3565)

What BYOM Does Not Do

• Register generic local inference servers as providers
• Route subagent tasks through LM Studio or other local endpoints
• Support non-Anthropic external model providers

Impact

• Privacy: workspace context leaks to cloud when local routing unavailable
• Security: sensitive code transmitted to third-party inference when no local path exists
• Cost: untracked cloud API usage from subagent tasks that could run locally
• Trust: silent cloud routing violates local-only expectations

Requested Behavior

1. SDK configuration for generic local model routing in subagent/tool-call paths (not just Anthropic BYOM)
2. Per-request cost/privacy warnings when cloud fallback triggers
3. Tiered privacy levels with explicit opt-in for cloud routing
4. Default to local-first routing where available

Related

• Task tool silently downgrades subagent model to session model via multiplier guard (frontmatter and explicit model override both ignored) copilot-cli#3565 — multiplier guard silently downgrades subagent model
• Feature request: local web interface (like opencode web) copilot-cli#3301 (local model feature request)
• Proposal: External model provider API surface for Copilot CLI (flags/env/config + BYOK) copilot-cli#2710 (external model proposal)
• Does Copilot SDK support both subscription and local models together? #807 (local model support question)

[ian-morgan99]

Seems this functionality is present, but requires local models to be registered in a particular way. Will close issue now whilst exploring if we have not registered correctly.