Agent: Detect RESOURCE_EXHAUSTED when parsing LLM response and dont

src/agent/loop.py

@@ -6,16 +6,17 @@
 import logging
 
 from ..config.settings import Settings
+from ..llm.base import ResourceExhaustedError
 from ..store.factory import create_checkpointer
 from .graph import build_graph
 
 logger = logging.getLogger(__name__)
 
 # Per-1M-token pricing: (input, output)
 _PRICING: dict[str, tuple[float, float]] = {
-    "anthropic": (3.0, 15.0),   # Claude Sonnet
-    "gemini": (1.25, 10.0),     # Gemini 2.5 Pro
-    "codex": (2.50, 10.0),      # Codex
+    "anthropic": (3.0, 15.0),  # Claude Sonnet
+    "gemini": (1.25, 10.0),  # Gemini 2.5 Pro
+    "codex": (2.50, 10.0),  # Codex
 }
 
 
@@ -87,7 +88,11 @@ async def run(self) -> str:
                     "repo_url": self.repo_config.url,
                 }
 
-            final_state = await graph.ainvoke(initial_state, config=config)
+            try:
+                final_state = await graph.ainvoke(initial_state, config=config)
+            except ResourceExhaustedError as e:
+                logger.error("LLM resource exhausted. Aborting run: %s", e)
+                return "Agent stopped due to LLM resource exhaustion."
 
             token_usage = final_state.get("token_usage", {})
             input_tokens = token_usage.get("input_tokens", 0)

src/llm/base.py

@@ -9,6 +9,10 @@
 from langchain_core.output_parsers import BaseOutputParser
 
 
+class ResourceExhaustedError(Exception):
+    """Raised when the LLM response indicates a resource exhaustion error."""
+
+
 class RobustJsonOutputParser(BaseOutputParser[dict[str, Any]]):
     """Extract JSON from LLM responses, handling markdown fences and surrounding text.
 
@@ -33,6 +37,9 @@ def parse(self, text: str | list) -> dict[str, Any]:
                     parts.append(item.get("text", ""))
             text = "".join(parts)
 
+        if "RESOURCE_EXHAUSTED" in text:
+            raise ResourceExhaustedError("LLM response indicates resource exhaustion")
+
         text = text.strip()
 
         # Try direct parse first (fast path for well-behaved responses)
@@ -51,7 +58,7 @@ def parse(self, text: str | list) -> dict[str, Any]:
                     continue
 
         # Brute-force: find every '{' and try json.loads from it, longest match first.
-        for m in re.finditer(r"\{", text):
+        for m in re.finditer(r"{", text):
             start = m.start()
             end = text.rfind("}", start)
             while end > start:

tests/test_executor.py

@@ -2,7 +2,10 @@
 
 from pathlib import Path
 
+import pytest
+
 from src.agent.executor import list_files, read_file, run_command
+from src.llm.base import ResourceExhaustedError, RobustJsonOutputParser
 
 
 def test_run_command_success(tmp_path: Path):
@@ -40,3 +43,11 @@ def test_read_file(tmp_path: Path):
 def test_read_file_missing(tmp_path: Path):
     result = read_file(tmp_path / "nope.txt")
     assert "Error" in result
+
+
+def test_robust_json_parser_resource_exhausted():
+    """Verify parser raises ResourceExhaustedError on matching error text."""
+    parser = RobustJsonOutputParser()
+    error_text = "Something went wrong. RESOURCE_EXHAUSTED. Please try again."
+    with pytest.raises(ResourceExhaustedError):
+        parser.parse(error_text)