synth-ai 0.2.16__py3-none-any.whl → 0.2.17__py3-none-any.whl

examples/qwen_vl/configs/eval_gpt4o_mini_vision.toml

@@ -1,43 +1,26 @@
 # Evaluation config for gpt-4o-mini with vision
-# Stronger teacher than gpt-5-nano, use for high-quality distillation
+# Higher-quality teacher for Crafter SFT distillation
 
 [eval]
-model = "gpt-4o-mini-2024-07-18"
-provider = "openai"  # Use OpenAI API
-
-# Task app endpoint
+app_id = "grpo-crafter-task-app"
 task_app_url = "https://synth-laboratories--grpo-crafter-task-app.modal.run"
-
-# Vision settings (auto-detected from "gpt-4o" in model name)
-use_vision = true
-image_only_mode = false  # Include both text + images
-
-# Rollout settings
-num_episodes = 100
-max_steps_per_episode = 50
-seeds = "200-299"  # Different seeds for comparison
-
-# Sampling parameters
-temperature = 0.6  # Lower temperature for more consistent behavior
+model = "gpt-4o-mini-2024-07-18"
+seeds = "200-299"
+max_turns = 50
+concurrency = 5
+env_name = "crafter"
+policy_name = "crafter-react"
+trace_format = "structured"
+return_trace = true
+
+[eval.env_config]
+env_params = {max_steps_per_episode = 50}
+
+[eval.policy_config]
+provider = "openai"
+model = "gpt-4o-mini-2024-07-18"
+temperature = 0.6
 max_tokens = 512
-
-# Trace collection
-collect_traces = true
-trace_db = "traces/gpt4o_mini_vision/rollouts.db"
-
-# Tools
+use_vision = true
+image_only_mode = false
 use_tools = true
-
-# Parallel rollouts
-parallel_episodes = 5
-
-[task]
-name = "crafter"
-environment = "crafter-classic"
-
-# Task-specific settings
-[task.config]
-seed_start = 200
-max_episode_length = 256
-render_size = [64, 64]  # 64x64 PNG images
-

examples/qwen_vl/configs/eval_gpt5nano_vision.toml

@@ -1,45 +1,26 @@
-# Evaluation config for gpt-4o-mini with vision
-# Collects vision traces for SFT training
-# Note: gpt-5-nano doesn't support tool calling yet, use gpt-4o-mini instead
+# Evaluation config for gpt-4o-mini (vision)
+# Collects traces for SFT training; legacy gpt-5-nano naming kept for convenience
 
 [eval]
-model = "gpt-4o-mini-2024-07-18"  # Changed from gpt-5-nano (no tool support)
-provider = "openai"  # Use OpenAI API
-
-# Task app endpoint (local or hosted)
-# task_app_url = "http://localhost:8000"  # Local
-task_app_url = "https://synth-laboratories--grpo-crafter-task-app.modal.run"  # Hosted
-
-# Vision settings (auto-detected from "gpt-5" in model name)
-use_vision = true
-image_only_mode = false  # Include both text + images
-
-# Rollout settings
-num_episodes = 100
-max_steps_per_episode = 50
-seeds = "0-99"  # Seeds 0 through 99
-
-# Sampling parameters
+app_id = "grpo-crafter-task-app"
+task_app_url = "https://synth-laboratories--grpo-crafter-task-app.modal.run"
+model = "gpt-4o-mini-2024-07-18"
+seeds = "0-99"
+max_turns = 50
+concurrency = 5
+env_name = "crafter"
+policy_name = "crafter-react"
+trace_format = "structured"
+return_trace = true
+
+[eval.env_config]
+env_params = {max_steps_per_episode = 50}
+
+[eval.policy_config]
+provider = "openai"
+model = "gpt-4o-mini-2024-07-18"
 temperature = 0.7
 max_tokens = 512
-
-# Trace collection
-collect_traces = true
-trace_db = "traces/gpt5nano_vision/rollouts.db"
-
-# Tools
+use_vision = true
+image_only_mode = false
 use_tools = true
-
-# Parallel rollouts (speeds up collection)
-parallel_episodes = 5  # Run 5 episodes in parallel
-
-[task]
-name = "crafter"
-environment = "crafter-classic"
-
-# Task-specific settings
-[task.config]
-seed_start = 0
-max_episode_length = 256
-render_size = [64, 64]  # 64x64 PNG images
-

examples/qwen_vl/configs/eval_qwen3vl_vision.toml

@@ -0,0 +1,26 @@
+# Evaluation config for Qwen3-VL vision rollouts
+# Collects traces for SFT training via synth-ai hosted inference
+
+[eval]
+app_id = "grpo-crafter-task-app"
+task_app_url = "https://synth-laboratories--grpo-crafter-task-app.modal.run"
+model = "Qwen/Qwen3-VL-8B-Instruct"
+seeds = "100-199"
+max_turns = 50
+concurrency = 5
+env_name = "crafter"
+policy_name = "crafter-react"
+trace_format = "structured"
+return_trace = true
+
+[eval.env_config]
+env_params = {max_steps_per_episode = 50}
+
+[eval.policy_config]
+provider = "synth"
+model = "Qwen/Qwen3-VL-8B-Instruct"
+temperature = 0.7
+max_tokens = 512
+use_vision = true
+image_only_mode = false
+use_tools = true

examples/qwen_vl/configs/{filter_qwen2vl_sft.toml → filter_qwen3vl_sft.toml}

@@ -1,9 +1,9 @@
-# Filter Qwen2-VL vision traces for SFT training
-# Same settings as gpt5nano filter but for Qwen2-VL traces
+# Filter Qwen3-VL vision traces for SFT training
+# Mirrors the GPT-4o mini filter configuration for vision data
 
 [filter]
-input_db = "traces/qwen2vl_vision/rollouts.db"
-output_dir = "traces/qwen2vl_vision/sft"
+input_db = "traces/qwen3vl_vision/rollouts.db"
+output_dir = "traces/qwen3vl_vision/sft"
 
 # Quality filters
 min_steps_per_episode = 5
@@ -47,4 +47,3 @@ val_file = "val.jsonl"
 save_stats = true
 stats_file = "filter_stats.json"
 save_filtered_episode_ids = true
-

examples/qwen_vl/configs/filter_vision_sft.toml

@@ -2,8 +2,8 @@
 # Applies quality filters and exports to SFT JSONL format
 
 [filter]
-input_db = "traces/gpt4o_vision_test/rollouts.db"
-output_dir = "traces/gpt4o_vision_test/sft"
+input_db = "traces/gpt4omini_vision/rollouts.db"
+output_dir = "traces/gpt4omini_vision/sft"
 
 # Quality filters
 min_steps_per_episode = 5        # Remove very short episodes
@@ -50,4 +50,3 @@ val_file = "val.jsonl"
 save_stats = true
 stats_file = "filter_stats.json"
 save_filtered_episode_ids = true
-

examples/qwen_vl/crafter_qwen_vl_agent.py

@@ -2,7 +2,7 @@
 """
 Crafter agent using Qwen-VL models via synth-ai's hosted inference.
 
-This demonstrates vision-language models (Qwen2-VL, Qwen3-VL) playing Crafter
+This demonstrates vision-language models (Qwen3-VL family) playing Crafter
 with image observations. The CrafterPolicy automatically detects vision capability
 from the model name and includes base64-encoded PNG frames in the prompt.
 
@@ -12,7 +12,7 @@ Requirements:
 
 Usage:
   uv run python examples/qwen_vl/crafter_qwen_vl_agent.py \
-      --model Qwen/Qwen2-VL-7B-Instruct --seeds 10 --steps 20
+      --model Qwen/Qwen3-VL-8B-Instruct --seeds 10 --steps 20
 """
 
 from __future__ import annotations
@@ -142,7 +142,7 @@ async def _run_episode(
     env = CrafterClassicEnvironment(task_instance)
     wrapper = CrafterEnvironmentWrapper(env, seed=seed)
     
-    # Policy will auto-detect vision from model name (qwen-vl, qwen2-vl, qwen3-vl)
+    # Policy will auto-detect vision from model name (qwen-vl and qwen3-vl tokens)
     policy = CrafterPolicy(inference_url="synth://inference", model=model)
     await policy.initialize({
         "use_tools": True,
@@ -235,8 +235,8 @@ async def main() -> None:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument(
         "--model",
-        default="Qwen/Qwen2-VL-7B-Instruct",
-        help="Qwen-VL model name (e.g., Qwen/Qwen2-VL-7B-Instruct, Qwen/Qwen3-VL-8B)",
+        default="Qwen/Qwen3-VL-8B-Instruct",
+        help="Qwen-VL model name (e.g., Qwen/Qwen3-VL-2B-Instruct, Qwen/Qwen3-VL-8B-Instruct)",
     )
     parser.add_argument("--seeds", type=int, default=10, help="Number of random seeds to evaluate")
     parser.add_argument("--steps", type=int, default=20, help="Max steps per seed")

examples/qwen_vl/run_vision_comparison.sh

@@ -37,13 +37,13 @@ uv run python examples/qwen_vl/crafter_gpt5nano_agent.py \
 
 echo ""
 echo "======================================"
-echo "2. Running Qwen2-VL-7B (synth-ai)"
+echo "2. Running Qwen3-VL-8B (synth-ai)"
 echo "======================================"
 uv run python examples/qwen_vl/crafter_qwen_vl_agent.py \
-    --model Qwen/Qwen2-VL-7B-Instruct \
+    --model Qwen/Qwen3-VL-8B-Instruct \
     --seeds $SEEDS \
     --steps $STEPS \
-    --output-dir "$OUTPUT_DIR/qwen2vl"
+    --output-dir "$OUTPUT_DIR/qwen3vl"
 
 echo ""
 echo "======================================"
@@ -53,10 +53,9 @@ echo ""
 echo "gpt-5-nano (OpenAI):"
 cat "$OUTPUT_DIR/gpt5nano/gpt5nano_summary.json" | python -m json.tool
 echo ""
-echo "Qwen2-VL-7B (synth-ai):"
-cat "$OUTPUT_DIR/qwen2vl/qwen_vl_summary.json" | python -m json.tool
+echo "Qwen3-VL-8B (synth-ai):"
+cat "$OUTPUT_DIR/qwen3vl/qwen_vl_summary.json" | python -m json.tool
 echo ""
 echo "Frames saved in:"
 echo "  - $OUTPUT_DIR/gpt5nano/gpt5nano_frames/"
-echo "  - $OUTPUT_DIR/qwen2vl/qwen_vl_frames/"
-
+echo "  - $OUTPUT_DIR/qwen3vl/qwen_vl_frames/"

examples/rl/README.md

@@ -5,8 +5,8 @@ This example trains a reinforcement learning policy on single-step math problems
 ## Quick Commands
 
 ```bash
-# Serve locally with tracing
-uvx synth-ai serve math-single-step --port 8101 --env-file examples/rl/.env --trace traces/math
+# Serve locally with tracing (uvicorn runtime)
+uvx synth-ai deploy --runtime uvicorn math-single-step --port 8101 --env-file examples/rl/.env --trace traces/math
 
 # Modal deployment
 uvx synth-ai deploy --name synth-math-single-step --env-file examples/rl/.env
@@ -45,10 +45,10 @@ The task app is defined in `synth_ai/task/apps/math_single_step.py` and register
   - `-0.5` if the tool call omits an answer or uses the wrong tool
   - `-1.0` when no tool call is provided
 
-Serve locally with tracing to capture trajectories:
+Run locally (uvicorn runtime) with tracing to capture trajectories:
 
 ```bash
-uvx synth-ai serve math-single-step \
+uvx synth-ai deploy --runtime uvicorn math-single-step \
   --port 8101 \
   --env-file examples/rl/.env \
   --trace traces/math \
@@ -162,7 +162,7 @@ For broader background on Synth task apps, CLI commands, and tracing, see the ne
 
 
 uv run python examples/rl/run_eval.py --toml examples/rl/configs/eval_base_qwen.toml
-uvx synth-ai serve math-single-step \
+uvx synth-ai deploy --runtime uvicorn math-single-step \
     --port 8101 \
     --env-file examples/rl/.env \
     --trace traces/math \

examples/rl/configs/rl_from_base_qwen.toml

@@ -1,10 +1,15 @@
-type = "rl"
+[algorithm]
+type = "online"
+method = "policy_gradient"
+variety = "gspo"
 
 [services]
 task_url = "https://your-math-task.modal.run"
 
 [model]
 base = "Qwen/Qwen3-4B"
+trainer_mode = "full"
+label = "math-single-step-qwen3-4b"
 
 [policy]
 model = "Qwen/Qwen3-4B"
@@ -20,6 +25,8 @@ evaluation_split = "validation"
 evaluation_episodes = 256
 
 [training]
+num_epochs = 1
+iterations_per_epoch = 20
 max_turns = 1
 ops = ["agent", "env"]
 batch_size = 128
@@ -33,5 +40,23 @@ learning_rate = 5e-6
 gpu_type = "A10G"
 gpu_count = 4
 
+[topology]
+type = "single_node_split"
+gpus_for_vllm = 2
+gpus_for_training = 2
+gpus_for_ref = 0
+tensor_parallel = 1
+
+[rollout]
+env_name = "math"
+policy_name = "math-single-step"
+max_turns = 1
+episodes_per_batch = 256
+
+[evaluation]
+instances = 256
+every_n_iters = 10
+seeds = [0, 1, 2, 3, 4]
+
 [tags]
 experiment = "math_single_step"

examples/rl/configs/rl_from_base_qwen17.toml

@@ -1,5 +1,3 @@
-type = "rl"
-
 [algorithm]
 type = "online"
 method = "policy_gradient"
@@ -10,6 +8,8 @@ task_url = "http://localhost:8101"
 
 [model]
 base = "Qwen/Qwen3-1.7B"
+trainer_mode = "full"
+label = "math-single-step-qwen3-1.7b"
 
 [policy]
 model = "Qwen/Qwen3-1.7B"
@@ -25,6 +25,8 @@ evaluation_split = "validation"
 evaluation_episodes = 50
 
 [training]
+num_epochs = 1
+iterations_per_epoch = 20
 max_turns = 1
 ops = ["agent", "env"]
 batch_size = 2
@@ -63,6 +65,7 @@ health_max_wait_s = 180
 health_interval_ms = 300
 
 [rollout]
+env_name = "math"
 policy_name = "math-single-step"
 max_turns = 1
 episodes_per_batch = 32  # group_size * batch_size

examples/rl/task_app/README.md

@@ -3,7 +3,7 @@
 This directory hosts the legacy entrypoint for the math single-step task app. Prefer starting the app via:
 
 ```bash
-uvx synth-ai serve math-single-step --env-file examples/rl/.env --port 8101
+uvx synth-ai deploy --runtime uvicorn math-single-step --env-file examples/rl/.env --port 8101
 ```
 
 If you need to run it directly (e.g., for Modal `modal deploy` compatibility), use:
@@ -19,4 +19,3 @@ Environment variables:
 - `MATH_DATASET_DEFAULT_SPLIT`, `MATH_DATASET_VALIDATION_SPLIT`, `MATH_DATASET_TEST_SPLIT`
 
 The task app enforces a single `math_submit` tool call per episode, enabling RL to reward correct final answers and penalise missing or malformed submissions.
-

examples/rl/task_app/math_single_step.py

@@ -800,7 +800,7 @@ def build_dataset() -> tuple[TaskDatasetRegistry, MathDataset]:
 def _base_task_info() -> TaskInfo:
     return TaskInfo(
         task={"id": "math_single_step", "name": "Math Single Step", "version": "1.0.0"},
-        environments=["math"],
+        environment="math",
         action_space={
             "type": "tool_call",
             "tools": [
@@ -891,7 +891,7 @@ def provide_task_instances(dataset: MathDataset, seeds: Sequence[int]) -> Iterab
         sample = dataset.sample(split=DEFAULT_SPLIT, index=seed)
         yield TaskInfo(
             task=info.task,
-            environments=info.environments,
+            environment=info.environment,
             action_space=info.action_space,
             observation={**info.observation, "sample_index": sample["index"]},
             dataset={

examples/run_crafter_demo.sh

@@ -4,7 +4,7 @@
 # This script demonstrates a reactive agent in the Crafter environment
 
 echo "🚀 Starting Crafter agent demo with Gemini 1.5 Flash..."
-echo "Make sure the synth-ai service is running: uvx synth-ai serve"
+echo "Make sure the synth-ai service is running: uvx synth-ai deploy --runtime uvicorn"
 echo ""
 
-uv run python -m synth_ai.environments.examples.crafter_classic.agent_demos.test_crafter_react_agent --model gemini-1.5-flash
+uv run python -m synth_ai.environments.examples.crafter_classic.agent_demos.test_crafter_react_agent --model gemini-1.5-flash

examples/sft/README.md

@@ -25,7 +25,7 @@ You can generate traces with the Crafter task app and then export them to SFT JS
 
 ```bash
 # Serve the task app locally with tracing enabled (example)
-uvx synth-ai serve grpo-crafter \
+uvx synth-ai deploy --runtime uvicorn grpo-crafter \
   --trace traces/v3 \
   --trace-db traces/v3/task_app_traces_<timestamp>.db \
   --port 8001

examples/sft/configs/crafter_fft_qwen0p6b.toml

@@ -1,4 +1,7 @@
-type = "sft"
+[algorithm]
+type = "offline"
+method = "sft"
+variety = "fft"
 
 [job]
 model = "Qwen/Qwen3-0.6B"

examples/sft/configs/crafter_lora_qwen0p6b.toml

@@ -1,4 +1,7 @@
-type = "sft"
+[algorithm]
+type = "offline"
+method = "sft"
+variety = "qlora"
 
 [job]
 model = "Qwen/Qwen3-0.6B"

examples/swe/task_app/README.md

@@ -28,13 +28,13 @@ endpoints.
 ## Using the task app
 
 ```
-uvx synth-ai serve swe-mini --port 8020
+uvx synth-ai deploy --runtime uvicorn swe-mini --port 8020
 ```
 
 ### Recommended: non-interactive serve + .env
 
 ```bash
-uvx synth-ai serve swe-mini \
+uvx synth-ai deploy --runtime uvicorn swe-mini \
   --port 8020 \
   --env-file .env \
   --trace traces/v3 \
@@ -60,6 +60,36 @@ Execution is handled by mini-swe's environment classes. Configure execution via
 `SWE_MINI_ENVIRONMENT_CLASS` (`local`, `docker`, `singularity`, …) and pass
 additional keyword arguments with `SWE_MINI_ENVIRONMENT_KWARGS` (JSON).
 
+### Morph Cloud backend
+
+The task app now ships with a Morph-powered environment class so you can run
+mini-SWE rollouts in managed sandboxes. When `MORPH_API_KEY` is present the app
+defaults to this backend automatically unless you override
+`SWE_MINI_ENVIRONMENT_CLASS`.
+
+1. Install the optional dependencies: `pip install "synth-ai[swe]"`.
+2. Export your API key: `export MORPH_API_KEY=...`.
+3. Point the task app at Morph by setting:
+
+   ```bash
+   export SWE_MINI_ENVIRONMENT_CLASS=morph
+   export SWE_MINI_ENVIRONMENT_KWARGS='{
+     "snapshot_id": "snap_your_pre_baked_swebench_image",
+     "cwd": "/workspace/swebench",
+     "env": {"PIP_PROGRESS_BAR": "off"},
+     "metadata": {"project": "synth-ai", "task": "swe-mini"}
+   }'
+   ```
+
+   If you do not have a pre-built snapshot, provide `"image_id"` (defaults to
+   `morphvm-minimal`) along with resource hints (`"vcpus"`, `"memory_mb"`,
+   `"disk_mb"`). You can also set `SWE_MINI_MORPH_SNAPSHOT_ID` globally.
+
+During cleanup the backend deletes the remote workspace and stops the Morph
+instance automatically. All shell commands (including submissions) now execute
+inside the Morph sandbox, enabling RL workflows that require persistent remote
+compute.
+
 ### Tracing & SFT
 
 Tracing works the same as Crafter; pass `--trace` / `--trace-db` to the CLI or

examples/swe/task_app/grpo_swe_mini.py

@@ -404,6 +404,10 @@ def _ensure_env_has_task(
         if not instance_id:
             raise ValueError("mini-swe rollout request requires env.config.instance_id")
         config["task"] = dataset.get(instance_id)
+    env_cfg = dict(config.get("environment") or {})
+    if "environment_class" not in env_cfg and os.getenv("MORPH_API_KEY"):
+        env_cfg["environment_class"] = "morph"
+    config["environment"] = env_cfg
     return env_spec.model_copy(update={"config": config})
 
 

examples/swe/task_app/hosted/envs/crafter/react_agent.py

@@ -46,7 +46,7 @@ class CrafterReActAgent:
             "- Always return a single tool call: interact_many({actions: [...]})\n"
             "- Use 2–5 actions per call; prefer long movement sequences to explore.\n"
             "- Mix in 'do' only when it makes sense (tree, stone, animal, enemy nearby).\n"
-            "- Do not spam the same exact sequence twice in a row—explore in varied directions.\n\n"
+            "\n"
             "Available actions: noop, move_up, move_down, move_left, move_right, do (interact), sleep, "
             "place_stone, place_table, place_furnace, place_plant, make_wood_pickaxe, make_stone_pickaxe, "
             "make_iron_pickaxe, make_wood_sword, make_stone_sword, make_iron_sword\n"

examples/swe/task_app/hosted/envs/mini_swe/environment.py

@@ -18,6 +18,7 @@ from typing import Any
 from minisweagent.environments import get_environment
 from synth_ai.environments.environment.tools import EnvToolCall
 
+from examples.swe.task_app.morph_backend import MorphSandboxBackend
 from .shared import summarise_history
 from .tools import TOOLS_SCHEMA
 
@@ -25,8 +26,9 @@ logger = logging.getLogger(__name__)
 
 
 def _environment_type_from_config(config: dict[str, Any]) -> str:
+    default = "morph" if os.getenv("MORPH_API_KEY") else "local"
     value = (config or {}).get("environment_class") or os.getenv(
-        "SWE_MINI_ENVIRONMENT_CLASS", "local"
+        "SWE_MINI_ENVIRONMENT_CLASS", default
     )
     return str(value).strip() or "local"
 
@@ -91,6 +93,7 @@ class MiniSweEnvironmentWrapper:
         self._local_workspace_dir: Path | None = None
         self._remote_workspace: str | None = None
         self._cleanup_workspace = False
+        self._using_morph_backend = False
 
         if self.environment_type == "local":
             workspace = self._prepare_local_workspace(kwargs)
@@ -117,11 +120,11 @@ class MiniSweEnvironmentWrapper:
             timeout = self.env_config.get("timeout")
             if timeout and "timeout" not in kwargs:
                 kwargs["timeout"] = int(timeout)
-            if self.repo_url and "image" not in kwargs:
+            if self.environment_type in {"docker", "bubblewrap"} and self.repo_url and "image" not in kwargs:
                 image = self.metadata.get("image_name") or os.getenv("SWE_MINI_DOCKER_IMAGE")
                 if image:
                     kwargs["image"] = image
-            if self.environment_type in {"docker", "bubblewrap"}:
+            if self.environment_type in {"docker", "bubblewrap", "morph"}:
                 remote_env = dict(kwargs.get("env") or {})
                 remote_env.setdefault("GIT_TERMINAL_PROMPT", "0")
                 kwargs["env"] = remote_env
@@ -131,13 +134,34 @@ class MiniSweEnvironmentWrapper:
             self.environment_type,
             kwargs,
         )
-        self.env = get_environment(
-            {
-                "environment_class": self.environment_type,
-                **kwargs,
-            },
-            default_type="local",
-        )
+        if self.environment_type == "morph":
+            morph_kwargs = dict(kwargs)
+            image_value = morph_kwargs.pop("image", None)
+            if image_value and "image_id" not in morph_kwargs:
+                morph_kwargs["image_id"] = image_value
+            timeout_value = morph_kwargs.pop("timeout", None)
+            if timeout_value is not None and "startup_timeout" not in morph_kwargs:
+                try:
+                    morph_kwargs["startup_timeout"] = int(timeout_value)
+                except Exception:
+                    logger.warning("Invalid timeout value for morph backend: %r", timeout_value)
+            metadata_override = morph_kwargs.pop("metadata", {}) or {}
+            metadata_payload = {
+                "app": "swe-mini",
+                "instance_id": self.instance_id,
+            }
+            metadata_payload.update({str(k): str(v) for k, v in dict(metadata_override).items()})
+            morph_kwargs["metadata"] = metadata_payload
+            self.env = MorphSandboxBackend(**morph_kwargs)
+            self._using_morph_backend = True
+        else:
+            self.env = get_environment(
+                {
+                    "environment_class": self.environment_type,
+                    **kwargs,
+                },
+                default_type="local",
+            )
 
         if self.environment_type != "local":
             self._bootstrap_remote_workspace()
@@ -181,6 +205,9 @@ class MiniSweEnvironmentWrapper:
             with contextlib.suppress(Exception):
                 self.env.execute(f"rm -rf {shlex.quote(self._remote_workspace)}")
         self._remote_workspace = None
+        if self._using_morph_backend and hasattr(self.env, "close"):
+            with contextlib.suppress(Exception):
+                self.env.close()
 
     def _resolve_repo_url(self, metadata: dict[str, Any]) -> str | None:
         candidates = [

examples/swe/task_app/hosted/inference/openai_client.py

@@ -156,13 +156,13 @@ class OpenAIClient:
             keys_preview = sorted(processed_request.keys())
             logger.info(f"Request keys: {keys_preview}")
 
-        # Final hard-guard for OpenAI: ensure unsupported field is not present
+        # Final hard-guard for OpenAI/Groq: ensure unsupported field is not present
         try:
-            if "openai" in url.lower() and "stop_after_tool_calls" in processed_request:
+            low_url = url.lower()
+            if ("openai" in low_url or "groq.com" in low_url or "/proxy/groq" in low_url) and "stop_after_tool_calls" in processed_request:
                 processed_request.pop("stop_after_tool_calls", None)
-                logger.info("Removed stop_after_tool_calls for OpenAI request")
+                logger.info("Removed stop_after_tool_calls for Groq/OpenAI request")
             # Groq-specific requirement: when using JSON mode, one of the messages must contain the word 'json'
-            low_url = url.lower()
             if ("groq.com" in low_url or "/openai" in low_url) and isinstance(
                 processed_request, dict
             ):