fix: preserve exploration tool history and cap context budget

artemis/llm/reliability.py

@@ -81,6 +81,11 @@ _AUTH_MARKERS = ("unauthorized", "forbidden", "api key", "credential")
 # "deadline exceeded"/"deadline_exceeded" cover google-genai SDK timeouts that
 # surface without an HTTP status code attached.
 _TIMEOUT_MARKERS = ("timed out", "timeout", "deadline exceeded", "deadline_exceeded")
+_TOOL_HISTORY_ERRORS = (
+    "no tool call found for function call output",
+    "no tool output found for function call",
+    "tool_call_id not found",
+)
 
 
 def classify_failure(error: BaseException) -> Failure:
@@ -91,6 +96,12 @@ def classify_failure(error: BaseException) -> Failure:
     code = extract_status_code(error)
     message = str(error).lower()
 
+    # Compatible Responses endpoints can report input validation inside an
+    # HTTP 200 stream as APIError without a status code. Retrying the same
+    # broken call/output history cannot repair it, even on a fallback model.
+    if any(marker in message for marker in _TOOL_HISTORY_ERRORS):
+        return Failure(FailureCategory.BAD_REQUEST, False, False)
+
     if code == 429 or any(marker in message for marker in _RATE_LIMIT_MARKERS):
         return Failure(FailureCategory.RATE_LIMIT, True, True)
     if code in {500, 502, 503, 504} or any(marker in message for marker in _UNAVAILABLE_MARKERS):