fix: revert opencode_env config regression and move RLM logic out of cli_agent_env

[mikasenghaas]

Summary

• Reverts opencode_env.py config builder to use shell variable expansion (${OPENAI_MODEL%%/*} etc.) instead of hardcoded "intercepted/model" to include the correct model name in the native oc system prompt — regression from opencode_rlm_env #1023
• Moves is_sub_llm_call-aware first-turn prompt check from CliAgentEnv.add_model_response into OpenCodeRLMEnv, restoring the simple len(trajectory) == 0 check in the base class
• OpenCodeRLMEnv.add_model_response uses super(CliAgentEnv, self) to skip the base class check and call MultiTurnEnv directly

🤖 Generated with Claude Code

---

Note

Medium Risk
Changes how OpenCode model/provider config is generated and how first-turn prompts are recorded when sub-LLM calls occur, which can affect model selection and trajectory/prompt correctness during rollouts.

Overview
Fixes an OpenCode config regression by switching build_opencode_config back to shell-expanded OPENAI_MODEL parsing so the generated config uses the real provider/model ($OPENAI_MODEL, ${OPENAI_MODEL%%/*}, ${OPENAI_MODEL##*/}) instead of a hardcoded intercepted/model.

Moves the sub-LLM aware “first main turn” prompt update logic out of CliAgentEnv.add_model_response into OpenCodeRLMEnv.add_model_response, ensuring state["prompt"] is set based on the first non-sub call while keeping the base CLI env’s first-turn check simple.

Updates tests to validate the rendered config after shell expansion (provider key, model name, and baseURL) using an explicit OPENAI_MODEL=myprovider/mymodel.

^{Written by Cursor Bugbot for commit b38775c. This will update automatically on new commits. Configure here.}