OpenCode Harness v0.1.0

Initial clean-room release for model-agnostic coding-agent evaluation.

Highlights:

• Claude Code-class coding-agent runtime.
• DeepSeek/Qwen/OpenAI/Claude/local provider presets, including vLLM, SGLang, and Ollama-compatible endpoints.
• Native OpenAI-compatible tool calls, native Anthropic tool use, and JSON fallback tool protocol.
• Permission-gated file, shell, patch, repository map, context, todo, finish, and MCP tools.
• Stdio MCP runtime with resources, prompts, diagnostics, per-server approvals, and namespace collision handling.
• JSONL traces, provider transcripts, session state, replay, terminal timeline viewer, HTML trace viewer, and eval dashboard.
• Eval runner with JSON, Markdown, and HTML reports plus failure taxonomy.
• DeepSeek, Qwen, Claude, OpenAI, and Local Model Labs.
• GitHub Actions CI, package release workflow, model-eval workflow example, changelog, provider benchmark runbook, and release demo manifest.

Clean-room boundary:
This release does not contain, derive from, or depend on leaked or proprietary Claude Code source code.