diff --git a/environments/agent_loop.py b/environments/agent_loop.py
index 891ce42f448..880f2ca797b 100644
--- a/environments/agent_loop.py
+++ b/environments/agent_loop.py
@@ -225,20 +225,35 @@ class HermesAgentLoop:
                 chat_kwargs["extra_body"] = self.extra_body
 
             # Make the API call -- standard OpenAI spec
+            # Retry on timeout/connection errors (provider queuing, rate limits)
             api_start = _time.monotonic()
-            try:
-                response = await self.server.chat_completion(**chat_kwargs)
-            except Exception as e:
-                api_elapsed = _time.monotonic() - api_start
-                logger.error("API call failed on turn %d (%.1fs): %s", turn + 1, api_elapsed, e)
-                return AgentResult(
-                    messages=messages,
-                    managed_state=self._get_managed_state(),
-                    turns_used=turn + 1,
-                    finished_naturally=False,
-                    reasoning_per_turn=reasoning_per_turn,
-                    tool_errors=tool_errors,
-                )
+            response = None
+            max_retries = 3
+            for attempt in range(max_retries):
+                try:
+                    response = await self.server.chat_completion(**chat_kwargs)
+                    break
+                except Exception as e:
+                    api_elapsed = _time.monotonic() - api_start
+                    is_retryable = "timeout" in type(e).__name__.lower() or "connection" in type(e).__name__.lower()
+                    if is_retryable and attempt < max_retries - 1:
+                        wait = 2 ** attempt
+                        logger.warning(
+                            "[%s] API call timed out on turn %d attempt %d (%.1fs), retrying in %ds: %s",
+                            self.task_id[:8], turn + 1, attempt + 1, api_elapsed, wait, e,
+                        )
+                        await asyncio.sleep(wait)
+                        api_start = _time.monotonic()
+                        continue
+                    logger.error("API call failed on turn %d (%.1fs): %s", turn + 1, api_elapsed, e)
+                    return AgentResult(
+                        messages=messages,
+                        managed_state=self._get_managed_state(),
+                        turns_used=turn + 1,
+                        finished_naturally=False,
+                        reasoning_per_turn=reasoning_per_turn,
+                        tool_errors=tool_errors,
+                    )
 
             api_elapsed = _time.monotonic() - api_start
 
diff --git a/environments/benchmarks/terminalbench_2/default.yaml b/environments/benchmarks/terminalbench_2/default.yaml
index eb675b12e70..4a781cc3212 100644
--- a/environments/benchmarks/terminalbench_2/default.yaml
+++ b/environments/benchmarks/terminalbench_2/default.yaml
@@ -19,11 +19,11 @@ env:
   max_token_length: 32000
   agent_temperature: 0.8
   terminal_backend: "modal"
-  terminal_timeout: 300        # 5 min per command (builds, pip install)
-  tool_pool_size: 128          # thread pool for 89 parallel tasks
-  dataset_name: "NousResearch/terminal-bench-2"
+  terminal_timeout: 300 # 5 min per command (builds, pip install)
+  tool_pool_size: 128 # thread pool for 89 parallel tasks
+  dataset_name: "sidbin/terminal-bench-2-verified-flattened"
   test_timeout: 600
-  task_timeout: 1800           # 30 min wall-clock per task, auto-FAIL if exceeded
+  task_timeout: 1800 # 30 min wall-clock per task, auto-FAIL if exceeded
   tokenizer_name: "NousResearch/Hermes-3-Llama-3.1-8B"
   use_wandb: true
   wandb_name: "terminal-bench-2"
@@ -36,7 +36,8 @@ env:
 
 openai:
   base_url: "https://openrouter.ai/api/v1"
-  model_name: "anthropic/claude-opus-4.6"
+  model_name: "openai/gpt-oss-120b:nitro"
   server_type: "openai"
   health_check: false
+  timeout: 300 # 5 min per API call (default 1200s causes 20min stalls)
   # api_key loaded from OPENROUTER_API_KEY in .env
diff --git a/environments/benchmarks/terminalbench_2/run_eval.sh b/environments/benchmarks/terminalbench_2/run_eval.sh
index ffbe4848065..27eeef41ce7 100755
--- a/environments/benchmarks/terminalbench_2/run_eval.sh
+++ b/environments/benchmarks/terminalbench_2/run_eval.sh
@@ -32,8 +32,8 @@ export PYTHONUNBUFFERED=1
 # These go to the log file; tqdm + [START]/[PASS]/[FAIL] go to terminal
 export LOGLEVEL=INFO
 
-python terminalbench2_env.py evaluate \
-  --config default.yaml \
+uv run python environments/benchmarks/terminalbench_2/terminalbench2_env.py evaluate \
+  --config environments/benchmarks/terminalbench_2/default.yaml \
   "$@" \
   2>&1 | tee "$LOG_FILE"
 
diff --git a/environments/benchmarks/terminalbench_2/terminalbench2_env.py b/environments/benchmarks/terminalbench_2/terminalbench2_env.py
index c7eaff6c4c2..4036890aaa0 100644
--- a/environments/benchmarks/terminalbench_2/terminalbench2_env.py
+++ b/environments/benchmarks/terminalbench_2/terminalbench2_env.py
@@ -354,6 +354,16 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
         for i, task in enumerate(self.all_eval_items):
             self.category_index[task.get("category", "unknown")].append(i)
 
+        # Pre-compute which tasks need Modal's add_python (avoids re-decoding
+        # multi-MB environment_tar blobs during per-task rollouts).
+        self._needs_add_python: Dict[str, bool] = {
+            task["task_name"]: self._image_needs_add_python(task)
+            for task in self.all_eval_items
+        }
+        add_py_count = sum(self._needs_add_python.values())
+        if add_py_count:
+            print(f"  {add_py_count} tasks need add_python (non-python base image)")
+
         # Reward tracking for wandb logging
         self.eval_metrics: List[Tuple[str, float]] = []
 
@@ -414,6 +424,36 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
     # Docker image resolution
     # =========================================================================
 
+    @staticmethod
+    def _image_needs_add_python(item: Dict[str, Any]) -> bool:
+        """Check if the task's base image lacks `python` on PATH.
+
+        Parses the Dockerfile FROM line in environment_tar. Returns True
+        for non-python base images (ubuntu, debian, etc.) that need
+        Modal's add_python parameter.
+        """
+        environment_tar = item.get("environment_tar", "")
+        if not environment_tar:
+            return False
+        try:
+            raw = base64.b64decode(environment_tar)
+            buf = io.BytesIO(raw)
+            with tarfile.open(fileobj=buf, mode="r:gz") as tar:
+                for member in tar:
+                    if not member.isfile() or "Dockerfile" not in member.name:
+                        continue
+                    f = tar.extractfile(member)
+                    if not f:
+                        continue
+                    for line in f.read().decode("utf-8", errors="ignore").splitlines():
+                        stripped = line.strip()
+                        if stripped.upper().startswith("FROM "):
+                            base = stripped.split()[1].lower()
+                            return not base.startswith("python:")
+        except Exception:
+            pass
+        return False
+
     def _resolve_task_image(
         self, item: Dict[str, Any], task_name: str
     ) -> Tuple[str, Optional[Path]]:
@@ -505,11 +545,14 @@ class TerminalBench2EvalEnv(HermesAgentBaseEnv):
             # --- 2. Register per-task image override ---
             # Set both modal_image and docker_image so the task image is used
             # regardless of which backend is configured.
-            register_task_env_overrides(task_id, {
+            overrides = {
                 "modal_image": modal_image,
                 "docker_image": modal_image,
                 "cwd": "/app",
-            })
+            }
+            if self._needs_add_python.get(task_name, False):
+                overrides["add_python"] = "3.12"
+            register_task_env_overrides(task_id, overrides)
             logger.info(
                 "Task %s: registered image override for task_id %s",
                 task_name, task_id[:8],
diff --git a/tools/environments/modal.py b/tools/environments/modal.py
index 1cb8e47969e..89937afeb3b 100644
--- a/tools/environments/modal.py
+++ b/tools/environments/modal.py
@@ -153,6 +153,7 @@ class ModalEnvironment(BaseEnvironment):
         modal_sandbox_kwargs: Optional[Dict[str, Any]] = None,
         persistent_filesystem: bool = True,
         task_id: str = "default",
+        add_python: Optional[str] = None,
     ):
         super().__init__(cwd=cwd, timeout=timeout)
 
diff --git a/tools/file_tools.py b/tools/file_tools.py
index 186a9d052c6..2c017994486 100644
--- a/tools/file_tools.py
+++ b/tools/file_tools.py
@@ -214,6 +214,7 @@ def _get_file_ops(task_id: str = "default") -> ShellFileOperations:
                 image = ""
 
             cwd = overrides.get("cwd") or config["cwd"]
+            add_python = overrides.get("add_python")
             logger.info("Creating new %s environment for task %s...", env_type, task_id[:8])
 
             container_config = None
@@ -252,6 +253,7 @@ def _get_file_ops(task_id: str = "default") -> ShellFileOperations:
                 local_config=local_config,
                 task_id=task_id,
                 host_cwd=config.get("host_cwd"),
+                add_python=add_python,
             )
 
             with _env_lock:
diff --git a/tools/terminal_tool.py b/tools/terminal_tool.py
index 243127a2958..fb2f50071ab 100644
--- a/tools/terminal_tool.py
+++ b/tools/terminal_tool.py
@@ -458,6 +458,7 @@ def register_task_env_overrides(task_id: str, overrides: Dict[str, Any]):
         - modal_image: str -- Path to Dockerfile or Docker Hub image name
         - docker_image: str -- Docker image name
         - cwd: str -- Working directory inside the sandbox
+        - add_python: str -- Python version for Modal's add_python (for images without python on PATH)
 
     Args:
         task_id: The rollout's unique task identifier
@@ -584,7 +585,8 @@ def _create_environment(env_type: str, image: str, cwd: str, timeout: int,
                         ssh_config: dict = None, container_config: dict = None,
                         local_config: dict = None,
                         task_id: str = "default",
-                        host_cwd: str = None):
+                        host_cwd: str = None,
+                        add_python: str = None):
     """
     Create an execution environment for sandboxed command execution.
     
@@ -682,6 +684,7 @@ def _create_environment(env_type: str, image: str, cwd: str, timeout: int,
             image=image, cwd=cwd, timeout=timeout,
             modal_sandbox_kwargs=sandbox_kwargs,
             persistent_filesystem=persistent, task_id=task_id,
+            add_python=add_python,
         )
     
     elif env_type == "daytona":
@@ -1057,6 +1060,7 @@ def terminal_tool(
             image = ""
 
         cwd = overrides.get("cwd") or config["cwd"]
+        add_python = overrides.get("add_python")
         default_timeout = config["timeout"]
         effective_timeout = timeout or default_timeout
 
@@ -1133,6 +1137,7 @@ def terminal_tool(
                             local_config=local_config,
                             task_id=effective_task_id,
                             host_cwd=config.get("host_cwd"),
+                            add_python=add_python,
                         )
                     except ImportError as e:
                         return json.dumps({