[EXAMPLE] [WIP] Add Terminus Pi TRL training

examples/end_to_end/tbench2_pi_trl/README.md

@@ -0,0 +1,9 @@
+# Terminus + TRL Async GRPO
+
+Start a Terminus server and a vLLM server with weight transfer enabled, then run:
+
+```bash
+TERMINUS_ENV_URL=http://localhost:8000 \
+TERMINUS_VLLM_SERVER_URL=http://localhost:8001 \
+uv run train_terminus_grpo.py
+```

examples/end_to_end/tbench2_pi_trl/agents/terminus/.pi/skills/terminus-terminal-task/SKILL.md

@@ -0,0 +1,26 @@
+---
+name: terminus-terminal-task
+description: Use inside a Terminus environment session when solving one sandboxed terminal task with the terminal tool.
+---
+
+# Terminus Terminal Task
+
+Use this skill only inside a Terminus task session.
+
+## Workflow
+
+1. Read the task.
+2. Use the `terminal` tool for each terminal action.
+3. Pass `command` to inspect and modify the sandbox.
+4. Check command output before choosing the next command.
+5. When the task is complete, pass `final_answer` exactly once.
+
+## Guardrails
+
+- Do not change hidden checks or task configuration.
+- Do not claim completion until the visible task requirements are satisfied.
+- Stay focused on the current task and terminal outputs.
+- Do not include both `command` and `final_answer` in the same tool call.
+- For simple file writes, prefer commands like `printf %s 'text' > path`.
+- If a command fails, inspect the error and continue with a smaller diagnostic
+  command.

examples/end_to_end/tbench2_pi_trl/agents/terminus/AGENTS.md

@@ -0,0 +1,15 @@
+# Terminus Task Instructions
+
+You are solving one task inside a Terminus terminal environment.
+
+Use the available `terminal` tool to inspect and modify the sandbox. Prefer
+short, direct shell commands. Read command output before deciding the next
+action. While working, call `terminal` with a `command`. When done, call
+`terminal` with a `final_answer`. Do not include both arguments in the same
+tool call.
+
+When the requested task is complete, submit exactly one final answer. The final
+answer should be concise and should not include implementation notes.
+
+Do not change hidden task checks or environment configuration. Stay focused on
+the current task and the available terminal tool.