Split engine provider selection from harness execution and route Claude/Codex via reflected CAPI endpoints

[Copilot]

This change separates inference-provider concerns from harness-runner concerns in the agentic engine flow, enabling Claude Code and Codex to run against GitHub Copilot CAPI-backed endpoints discovered at runtime. It also introduces explicit provider override semantics so workflows can select provider behavior without coupling to a specific harness implementation.

• Provider override semantics (engine.provider)

  • Added engine.provider parsing as a first-class provider override for non-inline engines, while preserving existing engine.model-provider compatibility.
  • Kept inline engine.runtime + engine.provider object behavior intact (id/auth/request path).

• Engine abstraction split (provider vs harness)

  • Introduced explicit inference-provider and harness-runner interfaces (with compatibility aliases) to decouple model-provider resolution from CLI harness execution mechanics.
  • Clarified intent in engine abstractions and call sites where harness capability is optional.

• Known provider normalization

  • Consolidated provider alias normalization around known providers (github, anthropic, openai) and mapped GitHub aliases consistently (including github_models).
  • Ensures provider selection resolves predictably across parser/runtime paths.

• Claude/Codex runtime endpoint routing from /reflect

  • Extended shared reflect utilities to resolve configured provider endpoints from AWF /reflect payloads.
  • Updated Claude harness to derive provider base URL from reflected endpoint data at runtime.
  • Updated Codex harness to derive and apply reflected endpoint routing by rewriting openai-proxy base_url and exporting corresponding env overrides.

• Targeted test coverage updates

  • Added/updated parser and harness tests for:
    • engine.provider override behavior
    • provider-endpoint resolution from reflect data
    • Codex provider config rewrite/env wiring behavior

engine:
  id: claude
  provider: github   # runtime provider override (CAPI-backed endpoint selected from /reflect)
  model: claude-sonnet-4.6

[pelikhan]

@copilot add smoke-github-Claude that uses github to run Claude code

[Copilot]

@copilot add smoke-github-Claude that uses github to run Claude code

Implemented in 57546cf: added smoke-github-claude and wired centralized slash-command routing so it runs Claude with GitHub provider (model-provider: github).