Implement OpenCode inference client

[Coldaine]

Origin: docs/todo.md entry dated 2026-04-01
Plan: docs/plans/plan-inference-engine.md (OpenCode client), docs/architecture.md §How the Shadow Model Works

Description

Create shadow-agent/src/inference/opencode-client.ts by porting sidecar's opencode-client.js. Responsibilities: start the OpenCode server (or attach to an existing one), create a session, send prompts, poll completions, and surface provider-neutral responses. The OpenCode SDK is the primary inference harness because it gives free provider abstraction (Claude, GPT-4, Gemini, OpenRouter) and inherits credentials from the user's existing OpenCode config.

Acceptance criteria

• Wraps @opencode-ai/sdk behind the shared InferenceClient seam at src/inference/inference-client.ts
• Start-or-attach logic avoids duplicate servers when one is already running
• Session creation, prompt send, completion polling all implemented
• Surfaces structured errors (auth, transport, model) for the orchestrator
• Honors model selection from config; defaults to a Claude model when unspecified
• Fake-client contract tests pass without reaching a real OpenCode server

[kilo-code-bot]

This issue appears to be a duplicate of #10.

Implement inference auth loader (#10)

Similarity score: 91%

This comment was generated by Kilo Auto-Triage.

[Coldaine]

Triage closeout: this finish-line merge campaign completed the currently shippable PR set on main. Remaining scope from this issue is either superseded by merged changes or deferred to a future planning cycle.