Merge pull request #119 from mpenn/mpenn_fix-llm-timeout

Fix LLM timeout handling and resolve symlinked skills in Docker builds

Author: hlinsen

cuopt-agent/cuopt_agent/configs/config-deepagent-eval.yml

@@ -38,6 +38,7 @@ llms:
     api_key: ${NVIDIA_API_KEY}
     base_url: https://integrate.api.nvidia.com/v1/
     max_tokens: 16384    
+    do_auto_retry: false
 
 functions:
   healthcheck:
@@ -57,17 +58,17 @@ workflow:
     - "AGENTS.md"
   workspace_dirs:
     - cuopt_agent/data/max_supply_what_ifs/input
-  max_retries: 2
+  max_retries: 16
   retry_backoff_factor: 2.0
   retry_initial_delay: 1.0
-  retry_max_delay: 60.0
+  retry_max_delay: 120.0
   system_prompt: |
-    You are a supply-chain optimization coordinator powered by NVIDIA cuOpt.
-    For optimization problems, delegate to the cuopt_sub_agent using the task() tool.
-    For general questions, respond directly.
+    You are a supply-chain optimization assistant powered by NVIDIA cuOpt.
+    Solve the following optimization problem using Python and the cuOpt solver.
+    Only write files in your working directory. Do not ask follow-up questions.
+    Reference your skills to understand latest API usage before getting started.
+    Answer simple questions directly.
     Report the optimal objective value as a single float in \boxed{...} on the last line.
-  subagents:
-    - cuopt_sub_agent
 
 eval:
   general:

cuopt-agent/cuopt_agent/configs/config-deepagent.yml

@@ -45,6 +45,7 @@ llms:
     api_key: ${NVIDIA_API_KEY}
     base_url: https://integrate.api.nvidia.com/v1/
     max_tokens: 16384
+    do_auto_retry: false
 
 functions:
   healthcheck:
@@ -63,17 +64,13 @@ workflow:
     - /skills/cuopt
   memory:
     - /AGENTS.md
-  max_retries: 2
+  max_retries: 16
   retry_backoff_factor: 2.0
   retry_initial_delay: 1.0
-  retry_max_delay: 60.0
+  retry_max_delay: 120.0 
   system_prompt: |
-    You are a supply-chain optimization coordinator powered by NVIDIA cuOpt.
-    For optimization problems, delegate to the cuopt_sub_agent using the task() tool.
-    For general questions, respond directly.
-  # system_prompt: |
-  #   You are a supply-chain optimization assistant powered by NVIDIA cuOpt.
-  #   Solve the following optimization problem using Python and the cuOpt solver.
-  #   Only write files in your working directory. Do not ask follow-up questions.
-  #   Reference your skills to understand latest API usage before getting started.
-  #   Answer simple questions directly.
+    You are a supply-chain optimization assistant powered by NVIDIA cuOpt.
+    Solve the following optimization problem using Python and the cuOpt solver.
+    Only write files in your working directory. Do not ask follow-up questions.
+    Reference your skills to understand latest API usage before getting started.
+    Answer simple questions directly.

cuopt-agent/cuopt_agent/docker/Dockerfile

@@ -31,7 +31,11 @@ ENV PATH="/app/cuopt_agent/.venv/bin:$PATH"
 COPY cuopt_agent/configs /app/cuopt_agent/configs
 COPY cuopt_agent/data    /app/cuopt_agent/data
 COPY skills              /app/skills
+<<<<<<< HEAD
+COPY external/cuopt/skills /app/skills/cuopt
+=======
 COPY external/cuopt/skills /app/external/cuopt/skills
+>>>>>>> upstream/main
 COPY AGENTS.md           /app/AGENTS.md
 
 # CWD must be cuopt_agent/ so config relative paths resolve correctly:

cuopt-agent/cuopt_agent/docker/Dockerfile.dev

@@ -21,6 +21,7 @@ WORKDIR /app
 # Copy package and repo assets (build context: repo root)
 COPY cuopt_agent /app/cuopt_agent
 COPY skills /app/skills
+COPY external/cuopt/skills /app/skills/cuopt
 COPY AGENTS.md /app/AGENTS.md
 
 # Install manually inside container (e.g. uv pip install --extra-index-url https://pypi.nvidia.com -e /app/cuopt_agent)

cuopt-agent/cuopt_agent/src/nat_cuopt_agent/function/deepagent_fn.py

@@ -117,7 +117,7 @@ class DeepAgentConfig(FunctionBaseConfig, name="deepagent_fn"):
         description="Initial delay in seconds before first retry.",
     )
     retry_max_delay: float = Field(
-        default=60.0,
+        default=120.0,
         description="Maximum delay cap in seconds between retries.",
     )
     strip_reasoning_pattern: str = Field(
@@ -142,6 +142,7 @@ async def deep_agent(config: DeepAgentConfig, builder: Builder):
         SANDBOX_AGENTS_MD,
         SANDBOX_SKILLS_DIR,
         FixToolNamesMiddleware,
+        ToolRetryMiddleware,
         kill_orphaned_children,
         populate_sandbox,
         resolve_skills_dirs,
@@ -224,6 +225,7 @@ async def _inner(chat_request_or_message: ChatRequestOrMessage) -> ChatResponse
             # Create a middleware chain for the agent to improve reliability and performance
             middleware = [
                 FixToolNamesMiddleware(),
+                ToolRetryMiddleware(),
                 ModelRetryMiddleware(
                     max_retries=config.max_retries,
                     backoff_factor=config.retry_backoff_factor,

cuopt-agent/cuopt_agent/src/nat_cuopt_agent/function/utils.py

@@ -13,6 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import asyncio
 import logging
 import re
 import shutil
@@ -69,6 +70,47 @@ async def awrap_model_call(self, request, handler):
         return self._patch(await handler(request))
 
 
+class ToolRetryMiddleware(AgentMiddleware):
+    """Retries failed tool calls with exponential backoff.
+
+    Provides uniform retry coverage for all tools. Some tools (e.g., Tavily)
+    have their own internal retry; this middleware wraps the outer call so
+    tools without retry (knowledge layer, paper search) are also covered.
+    """
+
+    def __init__(
+        self,
+        max_retries: int = 16,
+        backoff_factor: float = 2.0,
+        initial_delay: float = 1.0,
+    ):
+        self.max_retries = max_retries
+        self.backoff_factor = backoff_factor
+        self.initial_delay = initial_delay
+
+    async def awrap_tool_call(self, request, handler):
+        """Retry tool calls on failure with exponential backoff."""
+        delay = self.initial_delay
+        last_exception = None
+        for attempt in range(self.max_retries + 1):
+            try:
+                return await handler(request)
+            except Exception as e:
+                last_exception = e
+                if attempt < self.max_retries:
+                    tool_name = request.tool_call.get("name", "?") if hasattr(request, "tool_call") else "?"
+                    logger.warning(
+                        "Tool %s failed (attempt %d/%d): %s",
+                        tool_name,
+                        attempt + 1,
+                        self.max_retries + 1,
+                        e,
+                    )
+                    await asyncio.sleep(delay)
+                    delay *= self.backoff_factor
+        raise last_exception
+
+
 def strip_pattern(text: str, pattern: re.Pattern[str] | None) -> str:
     """Remove all regex matches from *text*.