Summary

Revises opencode_agent runner backed by the OpenCode CLI and integrates it with the AssetOpsBench scenario suite.

This PR supports two execution modes:

• Default MCP-only mode: OpenCode can use AssetOpsBench MCP tools, while local files, bash, edits, and web access are disabled.
• Optional CLI workspace mode: benchmark runs can create a dedicated per-run workspace and allow file inspection and bash execution so the agent can run local Python/code-based analysis.

Changes

• Added OpenCodeAgentRunner and opencode-agent CLI entry point.
• Added OpenCode config generation for AssetOpsBench MCP servers.
• Added support for OpenCode model IDs including direct providers and router-backed models.
• Added scenario-suite integration via --agent_name opencode_agent.
• Added optional workspace flags:
  • --opencode-workspace-root
  • --opencode-allow-files
  • --opencode-allow-bash
  • --opencode-allow-edit
• Kept web access disabled by default.
• Denied OpenCode external_directory access by default.
• Added tests for runner config, permissions, workspace handling, and scenario-suite command construction.
• Added docs for OpenCode agent usage and CLI workspace mode.

Safety Notes

By default, the agent is MCP-only and cannot inspect local files, run bash, edit files, or access the web.

CLI capabilities are opt-in. When file or bash access is enabled, each scenario run receives a dedicated workspace directory. This gives CLI agents a place to write scripts, temporary files, and intermediate artifacts.

--opencode-allow-bash is not a hard OS-level sandbox. For strict filesystem isolation, the benchmark should be run inside Docker or another sandboxed environment.

Testing

.venv/bin/pytest src/agent/opencode_agent/tests/test_runner.py
.venv/bin/pytest src/benchmark/tests/test_scenario_suite_runner.py