Removing old patches

environments/__init__.py

@@ -13,6 +13,7 @@ Core layers:
 Concrete environments:
     - terminal_test_env/: Simple file-creation tasks for testing the stack
     - hermes_swe_env/: SWE-bench style tasks with Modal sandboxes
+    - endless_terminals/: Terminal tasks from HuggingFace dataset with Apptainer containers
 
 Benchmarks (eval-only):
     - benchmarks/terminalbench_2/: Terminal-Bench 2.0 evaluation

environments/endless_terminals/__init__.py

@@ -0,0 +1,5 @@
+"""Endless Terminals Environment - Terminal task training from HuggingFace dataset."""
+
+from .endless_terminals_env import EndlessTerminalsEnv, EndlessTerminalsEnvConfig
+
+__all__ = ["EndlessTerminalsEnv", "EndlessTerminalsEnvConfig"]

environments/endless_terminals/tinker_qwen.yaml

@@ -0,0 +1,91 @@
+# Endless Terminals - Qwen3-4B-Instruct-2507
+# Single config for both trainer (launch_training.py) and env (endless_terminals_env.py serve)
+#
+# Usage:
+#   Terminal 1: run-api
+#   Terminal 2: cd tinker-atropos && python launch_training.py --config ../environments/endless_terminals/tinker_qwen.yaml
+#   Terminal 3: python environments/endless_terminals/endless_terminals_env.py serve --config environments/endless_terminals/tinker_qwen.yaml
+
+env:
+  # Toolsets
+  enabled_toolsets: ["terminal", "file"]
+
+  # Model / tokenizer
+  tokenizer_name: "Qwen/Qwen3-4B-Instruct-2507"
+
+  # Agent configuration
+  max_agent_turns: 16
+  max_token_length: 2048
+  agent_temperature: 0.6
+  extra_body:
+    chat_template_kwargs:
+      enable_thinking: false
+  tool_call_parser: "hermes"
+
+  # Terminal backend
+  terminal_backend: "docker"
+
+  # Dataset settings
+  use_dataset: true
+  dataset_name: "obiwan96/endless-terminals"
+  dataset_split: "train"
+  dataset_cache_dir: "~/.cache/huggingface/datasets"
+  tasks_base_dir: "/Users/samherring/Desktop/Projects/Hermes-Agent/endless-terminals"
+
+  # Test execution
+  test_timeout_s: 180
+  default_docker_image: "ubuntu:22.04"
+  max_concurrent_containers: 16
+
+  # Training configuration
+  group_size: 16
+  batch_size: 64          # 4 groups × 16 rollouts per step
+  total_steps: 500
+  steps_per_eval: 5
+  min_items_sent_before_logging: 1
+  ensure_scores_are_not_same: true
+  max_num_workers: 2048
+  worker_timeout: 3600
+  inference_weight: 1.0
+  eval_limit_ratio: 0.1
+  rollout_server_url: "http://localhost:8000"
+
+  # Evaluation configuration
+  num_eval_tasks: 20
+  eval_split_ratio: 0.1
+
+  # Logging
+  use_wandb: true
+  wandb_name: "endless-terminals-qwen3-4b"
+
+  # System prompt
+  system_prompt: >
+    You are a skilled Linux system administrator and programmer.
+    You have access to a terminal and file tools to complete system administration
+    and programming tasks. Use the tools effectively to solve the given task,
+    and verify your solution works correctly before finishing.
+    Keep each command short and focused — break complex tasks into multiple steps
+    rather than writing long one-liners.
+
+tinker:
+  lora_rank: 32
+  learning_rate: 0.0000005
+  max_token_trainer_length: 32768
+  checkpoint_dir: "./temp/"
+  save_checkpoint_interval: 50
+  wandb_project: "endless-terminals"
+  wandb_group: null
+  wandb_run_name: "qwen3-4b"
+  tool_call_parser: "hermes"
+
+openai:
+  - model_name: "Qwen/Qwen3-4B-Instruct-2507"
+    base_url: "http://localhost:8001/v1"
+    api_key: "x"
+    weight: 1.0
+    num_requests_for_eval: 64
+    timeout: 600
+    server_type: "sglang"
+
+slurm: false
+testing: false

environments/hermes_base_env.py

@@ -298,7 +298,6 @@ class HermesAgentBaseEnv(BaseEnv):
             return False
 
         server = self.server.servers[0]
-        # If the server is an OpenAI server (not VLLM/SGLang), use direct mode
         from atroposlib.envs.server_handling.openai_server import OpenAIServer
         return not isinstance(server, OpenAIServer)
 

environments/tool_call_parsers/hermes_parser.py

@@ -48,7 +48,13 @@ class HermesToolCallParser(ToolCallParser):
                 if not raw_json.strip():
                     continue
 
+                try:
                     tc_data = json.loads(raw_json)
+                except json.JSONDecodeError:
+                    # Fix invalid backslash escapes from shell commands in JSON strings
+                    # e.g. \s \w \d \n (unescaped) → \\s \\w \\d \\n
+                    fixed = re.sub(r'\\([^"\\/bfnrtu0-9\n])', r'\\\\\1', raw_json)
+                    tc_data = json.loads(fixed)
                 tool_calls.append(
                     ChatCompletionMessageToolCall(
                         id=f"call_{uuid.uuid4().hex[:8]}",