FEATURE: SDK should support generic local inference endpoints as model providers

[ian-morgan99]

Problem Statement

The Copilot SDK does not expose a configuration path for generic local inference endpoints (non-Anthropic BYOM). This blocks extensions and integrations from routing agent tasks through local models.

Current State

• BYOM support exists but is limited to Anthropic-specific configurations
• No mechanism to register arbitrary OpenAI-compatible endpoints as model providers
• Subagent dispatch has no local routing path

Impact on Extensions

Extensions like VSCode-LMStudio-Bridge cannot expose local models to the Copilot ecosystem because:

1. No SDK API for registering custom model providers
2. No configuration schema for generic inference endpoints
3. Subagent tasks bypass any local routing entirely

How Other Agents Handle This

Opencode (anomalyco/opencode)

• ProviderConfig abstraction with explicit provider registration
• Supports llama.cpp, Ollama, and generic OpenAI-compatible endpoints
• Provider configuration flows through all dispatch paths including subagents

Codex (openai/codex)

• ModelProvider abstraction with routing layer (models_endpoint.rs)
• Clean separation between tool execution and model inference
• Supports multiple providers including local endpoints

What We Need in the SDK

1. Model provider registration API: Allow extensions to register custom model providers
2. Generic endpoint support: Support arbitrary OpenAI-compatible inference endpoints
3. Provider priority/fallback: Support tiered routing (local first, cloud fallback)
4. Subagent model inheritance: Ensure subagent tasks respect the session model

Security Implications

• Data leakage: Workspace context transmitted to cloud without user consent
• Cost opacity: Untracked cloud API usage from subagent tasks
• Trust erosion: Silent cloud routing violates local-only expectations
• Compliance risk: Sensitive code may be processed by third-party inference services

Related Issues

• 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
• ian-morgan99/VSCode-LMStudio-Bridge#337 — bridge-level gap analysis
• ian-morgan99/VSCode-LMStudio-Bridge#336 — zero BYOM integration in bridge

Acceptance Criteria

1. SDK exposes API for registering custom model providers
2. Generic OpenAI-compatible endpoints can be configured as providers
3. Provider priority/fallback configuration is supported
4. Subagent tasks respect session model when it is a local endpoint
5. Documentation updated with provider registration and routing behavior

[stephentoub]

No mechanism to register arbitrary OpenAI-compatible endpoints as model providers

This is not the case. Both OpenAI chat completion and responses endpoints are supported, in addition to Anthropic messages endpoints.

[stephentoub]

Details available at https://github.com/github/copilot-sdk/blob/main/docs/auth/byok.md