fix: correct is_action_valid logic, scroll_direction, stale refs, and DRY violation

Review fixes for the GPU training automation branch:

- Fix is_action_valid: was inverted (DONE()→invalid, garbage→valid), now uses
  regex match on original action string
- Fix scroll_direction: SCROLL parsing now populates BenchmarkAction.scroll_direction
- Fix stale repo URLs: mll-lab-nu/VAGEN → RAGEN-AI/VAGEN across vendored files and docs
- Fix stale branch ref: setup_gpu_training.sh referenced merged spike branch, now uses main
- Fix stale repo URL: langfengQ/verl-agent → RAGEN-AI/VAGEN in setup script
- Add --recurse-submodules to git clone (verl is a VAGEN submodule)
- Remove dead params from register_waa_env() (waa_server, task_id, max_steps)
- Deduplicate training command: vm_cli.py now delegates to launch_training()
- Update test count in docs: 21 → 40+
- Add 3 new tests for is_action_valid behavior
- Add scroll_direction assertion to existing scroll test

All 43 tests pass.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: abrichr

docs/verl_agent_decision.md

@@ -138,7 +138,7 @@ loop (~546 lines in `openadapt_ml/training/grpo/trainer.py`).
 
 ### D. VAGEN / VAGEN-Lite
 
-**Repository**: [mll-lab-nu/VAGEN](https://github.com/mll-lab-nu/VAGEN)
+**Repository**: [RAGEN-AI/VAGEN](https://github.com/RAGEN-AI/VAGEN)
 
 - Built on verl's `agent_loop` abstraction (same ecosystem as verl-agent)
 - **Bi-Level GAE** for turn-aware credit assignment
@@ -299,7 +299,7 @@ scaling to multi-VM env pools.
    Works, well-tested (56 unit tests + 5 E2E tests). Episode-level rewards only.
 
 2. **Spike complete**: `WAADesktopEnv` adapter in openadapt-evals (PR #84, merged).
-   21 tests passing. Implements GymImageEnv protocol.
+   40 tests passing. Implements GymImageEnv protocol.
 
 3. **Next**: Test end-to-end with verl-agent on a GPU machine. If successful,
    the standalone trainer becomes a reference implementation / fallback, and
@@ -360,8 +360,8 @@ To validate the verl-agent integration provides real value over standalone:
 
 ## References
 
-- [verl-agent](https://github.com/langfengQ/verl-agent) — GiGPO paper implementation
-- [VAGEN](https://github.com/mll-lab-nu/VAGEN) — Multi-turn VLM agent training
+- [VAGEN](https://github.com/RAGEN-AI/VAGEN) — Multi-turn VLM agent training (GiGPO)
+- [verl-agent](https://github.com/langfengQ/verl-agent) — GiGPO paper's original codebase
 - [verl](https://github.com/verl-project/verl) — Volcano Engine RL for LLMs
 - [GiGPO paper](https://arxiv.org/html/2505.10978) — Group-in-Group Policy Optimization
 - [VAGEN paper](https://arxiv.org/abs/2510.16907) — World Model Reasoning for VLM Agents

openadapt_evals/adapters/_vendored/__init__.py

@@ -1,6 +1,6 @@
 """Vendored pure-abstract base classes from VAGEN.
 
-These are copied from https://github.com/mll-lab-nu/VAGEN so that
+These are copied from https://github.com/RAGEN-AI/VAGEN so that
 openadapt-evals can implement the GymImageEnv protocol without
 requiring the full VAGEN package (and its heavy transitive
 dependencies) to be installed.

openadapt_evals/benchmarks/vm_cli.py

@@ -7919,58 +7919,33 @@ def cmd_gpu_train(args):
                 print(f"ERROR: Setup failed")
                 return 1
 
-    # Launch training via the E2E script (handles data prep, env registration, config)
-    # Import here to avoid circular deps
+    # Delegate to the E2E script's launch_training() which handles:
+    # - env registry registration
+    # - training data preparation
+    # - training config generation
+    # - training launch
+    # This avoids duplicating the Hydra overrides in two places.
     import sys
     sys.path.insert(0, str(Path(__file__).parent.parent.parent / "scripts"))
-    from train_verl_e2e import prepare_training_data, register_waa_env
-
-    print("Registering WAADesktopEnv in VAGEN env registry...")
-    register_waa_env(ip, args.waa_server, args.task_id, username=username)
-
-    print("Preparing training data...")
-    prepare_training_data(ip, group_size=8, username=username)
-
-    # Hydra overrides for verl training loop + VAGEN env config.
-    # WAADesktopEnv is registered in VAGEN's env_registry.yaml as 'WAADesktop'.
-    # The env connects to WAA server via HTTP (GymImageEnv protocol).
-    train_cmd = (
-        f"cd ~/verl-agent && "
-        f"conda run -n verl-agent python3 -m verl.trainer.main_ppo "
-        f"algorithm.adv_estimator={args.algorithm} "
-        f"algorithm.gamma={'0.95' if args.algorithm == 'gigpo' else '1.0'} "
-        f"actor_rollout_ref.model.path={args.model} "
-        f"actor_rollout_ref.rollout.name=vllm "
-        f"actor_rollout_ref.rollout.tensor_model_parallel_size={args.n_gpus} "
-        f"actor_rollout_ref.rollout.gpu_memory_utilization=0.6 "
-        f"actor_rollout_ref.rollout.enable_chunked_prefill=False "
-        f"actor_rollout_ref.actor.ppo_mini_batch_size=64 "
-        f"actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=8 "
-        f"data.train_files=$HOME/data/verl-agent/visual/train.parquet "
-        f"data.val_files=$HOME/data/verl-agent/visual/test.parquet "
-        f"data.train_batch_size=8 "
-        f"data.val_batch_size=128 "
-        f"data.max_prompt_length=2048 "
-        f"data.max_response_length=512 "
-        f"data.return_raw_chat=True "
-        f"data.filter_overlong_prompts=True "
-        f"trainer.n_gpus_per_node={args.n_gpus} "
-        f"trainer.nnodes=1 "
-        f"trainer.total_epochs={args.epochs} "
-        f"trainer.test_freq=5 "
-        f"trainer.experiment_name={args.algorithm}_waa_desktop "
-        f"trainer.logger=['console','wandb'] "
-        f"trainer.project_name=openadapt-waa-rl"
-    )
+    from train_verl_e2e import launch_training
 
     print(f"Launching {args.algorithm} training on {args.n_gpus} GPU(s)...")
     print(f"Model: {args.model}")
     print(f"WAA server: {args.waa_server}")
     print(f"Task: {args.task_id}")
 
     try:
-        result = ssh_run(ip, train_cmd, username=username, stream=True)
-        return result.returncode
+        exit_code = launch_training(
+            ip=ip,
+            waa_server=args.waa_server,
+            task_id=args.task_id,
+            algorithm=args.algorithm,
+            model=args.model,
+            n_gpus=args.n_gpus,
+            epochs=args.epochs,
+            username=username,
+        )
+        return exit_code
     finally:
         if args.cleanup and not args.gpu_ip:
             print(f"Deallocating GPU VM '{gpu_vm_name}'...")

scripts/setup_gpu_training.sh

@@ -23,9 +23,11 @@ set -euo pipefail
 CONDA_ENV="verl-agent"
 VLLM_VERSION="0.11.0"
 FLASH_ATTN_VERSION="2.7.4.post1"
-VERL_AGENT_REPO="https://github.com/langfengQ/verl-agent.git"
+# VAGEN includes verl as a submodule + the GymImageEnv protocol, env registry,
+# and GymAgentLoop that we integrate with.
+VERL_AGENT_REPO="https://github.com/RAGEN-AI/VAGEN.git"
 OPENADAPT_EVALS_REPO="https://github.com/OpenAdaptAI/openadapt-evals.git"
-OPENADAPT_EVALS_BRANCH="spike/verl-agent-integration"
+OPENADAPT_EVALS_BRANCH="main"
 
 log() { echo "=== [$(date '+%H:%M:%S')] $*"; }
 
@@ -69,13 +71,13 @@ pip3 install "flash-attn==$FLASH_ATTN_VERSION" --no-build-isolation --no-cache-d
 
 # --- Clone and install verl-agent ---
 if [ ! -d "$HOME/verl-agent" ]; then
-    log "Cloning verl-agent..."
-    git clone "$VERL_AGENT_REPO" "$HOME/verl-agent"
+    log "Cloning VAGEN..."
+    git clone --recurse-submodules "$VERL_AGENT_REPO" "$HOME/verl-agent"
 else
-    log "verl-agent already cloned, pulling latest..."
-    cd "$HOME/verl-agent" && git pull
+    log "VAGEN already cloned, pulling latest..."
+    cd "$HOME/verl-agent" && git pull && git submodule update --init --recursive
 fi
-log "Installing verl-agent..."
+log "Installing VAGEN..."
 cd "$HOME/verl-agent"
 pip install -e .
 
@@ -103,10 +105,10 @@ python -c "from openadapt_evals.adapters.verl_env import WAADesktopEnv; print('W
 
 log "Setup complete! Activate with: conda activate $CONDA_ENV"
 log ""
-log "To start training:"
-log "  cd ~/verl-agent"
-log "  python3 -m verl.trainer.main_ppo \\"
-log "    algorithm.adv_estimator=gigpo \\"
-log "    env.env_name=openadapt_evals.adapters.verl_env.WAADesktopEnv \\"
-log "    trainer.n_gpus_per_node=$GPU_COUNT \\"
-log "    ..."
+log "To start training, use the orchestration script:"
+log "  python scripts/train_verl_e2e.py --gpu-ip \$(hostname -I | awk '{print \$1}') --task-id <WAA_UUID>"
+log ""
+log "Or via oa-vm CLI:"
+log "  oa-vm gpu-train --gpu-ip \$(hostname -I | awk '{print \$1}') --task-id <WAA_UUID>"
+log ""
+log "GPU count: $GPU_COUNT"

scripts/train_verl_e2e.py

@@ -166,7 +166,7 @@ def prepare_training_data(ip: str, group_size: int = 8, username: str = "ubuntu"
         raise RuntimeError("Data preparation failed")
 
 
-def register_waa_env(ip: str, waa_server: str, task_id: str, max_steps: int = 15, username: str = "ubuntu"):
+def register_waa_env(ip: str, username: str = "ubuntu"):
     """Register WAADesktopEnv in VAGEN's env registry on the GPU VM.
 
     VAGEN uses a YAML registry (vagen/configs/env_registry.yaml) to dispatch
@@ -354,7 +354,7 @@ def launch_training(
     4. Launch training via VAGEN's entry point
     """
     # Step 1: Register WAADesktopEnv in VAGEN's env registry
-    register_waa_env(ip, waa_server, task_id, max_steps=max_turns, username=username)
+    register_waa_env(ip, username=username)
 
     # Step 2: Prepare parquet data files (required by VAGEN's AgenticDataset)
     prepare_training_data(ip, group_size=group_size, username=username)

tests/test_verl_env.py

@@ -247,6 +247,27 @@ def test_protocol_has_required_methods(self):
             assert hasattr(env, method)
             assert callable(getattr(env, method))
 
+    def test_is_action_valid_on_good_action(self):
+        """Valid actions (parseable) should have is_action_valid=True."""
+        env = _make_mock_env()
+        asyncio.run(env.reset(seed=42))
+        _, _, _, info = asyncio.run(env.step("CLICK(x=0.5, y=0.5)"))
+        assert info["is_action_valid"] is True
+
+    def test_is_action_valid_on_done(self):
+        """Explicit DONE() should have is_action_valid=True."""
+        env = _make_mock_env()
+        asyncio.run(env.reset(seed=42))
+        _, _, _, info = asyncio.run(env.step("DONE()"))
+        assert info["is_action_valid"] is True
+
+    def test_is_action_valid_on_garbage(self):
+        """Unparseable actions should have is_action_valid=False."""
+        env = _make_mock_env()
+        asyncio.run(env.reset(seed=42))
+        _, _, _, info = asyncio.run(env.step("random garbage text"))
+        assert info["is_action_valid"] is False
+
     def test_obs_contains_image_placeholder(self):
         """Test that observations with screenshots include <image> placeholder."""
         env = _make_mock_env()
@@ -281,28 +302,6 @@ def test_health_check_busy_mid_episode(self):
         result = asyncio.run(env.health_check())
         assert result["status"] == "busy"
 
-    # --- is_action_valid tests ---
-
-    def test_is_action_valid_for_parsed_action(self):
-        """Actions that parse successfully should be marked valid."""
-        env = _make_mock_env()
-        asyncio.run(env.reset(seed=42))
-        _, _, _, info = asyncio.run(env.step("CLICK(x=0.5, y=0.5)"))
-        assert info["is_action_valid"] is True
-
-    def test_is_action_valid_for_done(self):
-        """Explicit DONE() should be marked valid."""
-        env = _make_mock_env()
-        asyncio.run(env.reset(seed=42))
-        _, _, _, info = asyncio.run(env.step("DONE()"))
-        assert info["is_action_valid"] is True
-
-    def test_is_action_invalid_for_garbage(self):
-        """Unparseable input should be marked invalid."""
-        env = _make_mock_env()
-        asyncio.run(env.reset(seed=42))
-        _, _, _, info = asyncio.run(env.step("random garbage"))
-        assert info["is_action_valid"] is False
 
 
 # --- VAGEN registration helpers tests ---