decodingtrust-agent-sdk 0.2.11__tar.gz → 0.2.12__tar.gz

{decodingtrust_agent_sdk-0.2.11/decodingtrust_agent_sdk.egg-info → decodingtrust_agent_sdk-0.2.12}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: decodingtrust-agent-sdk
-Version: 0.2.11
+Version: 0.2.12
 Summary: DecodingTrust Agent Platform (DTap) — A controllable and interactive red-teaming platform for AI agents
 Author-email: DTap Team <zhaorun@uchicago.edu>
 License:                                  Apache License

{decodingtrust_agent_sdk-0.2.11 → decodingtrust_agent_sdk-0.2.12/decodingtrust_agent_sdk.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: decodingtrust-agent-sdk
-Version: 0.2.11
+Version: 0.2.12
 Summary: DecodingTrust Agent Platform (DTap) — A controllable and interactive red-teaming platform for AI agents
 Author-email: DTap Team <zhaorun@uchicago.edu>
 License:                                  Apache License

{decodingtrust_agent_sdk-0.2.11 → decodingtrust_agent_sdk-0.2.12}/dt_arena/config/injection_mcp.yaml

@@ -41,22 +41,28 @@ servers:
       # Injection Server Port (dynamically allocated)
       SALESFORCE_ENV_INJECTION_MCP_PORT: "8851"
 
-      # Base URL for SuiteCRM instance (uses dynamic port from Docker environment)
-      SUITECRM_BASE_URL: "http://127.0.0.1:${SALESFORCE_API_PORT}"
+      # Base URL for the Salesforce/SuiteCRM instance (uses dynamic port from Docker environment).
+      # Key MUST be SALESFORCE_BASE_URL — that's what env_injection.py reads; the old
+      # SUITECRM_BASE_URL name was ignored, so the server fell back to a hardcoded
+      # unreachable host ("salesforce-injection unreachable").
+      SALESFORCE_BASE_URL: "http://127.0.0.1:${SALESFORCE_API_PORT}"
 
       # OAuth2 Grant Type: "password" or "client_credentials"
-      SUITECRM_GRANT_TYPE: "client_credentials"
+      # NOTE: keys MUST be SALESFORCE_* — that's what env_injection.py reads.
+      # (The old SUITECRM_* names were ignored; the code happened to work via
+      #  matching hardcoded defaults, but task overrides would be silently lost.)
+      SALESFORCE_GRANT_TYPE: "client_credentials"
 
       # OAuth2 Client Credentials
-      SUITECRM_CLIENT_ID: "17817554-086b-83d1-bf9d-69026221f529"
-      SUITECRM_CLIENT_SECRET: "mcp-secret-123"
+      SALESFORCE_CLIENT_ID: "17817554-086b-83d1-bf9d-69026221f529"
+      SALESFORCE_CLIENT_SECRET: "mcp-secret-123"
 
       # Password Grant Credentials (when using grant_type="password")
-      SUITECRM_USERNAME: "bitnami"
-      SUITECRM_PASSWORD: "user"
+      SALESFORCE_USERNAME: "bitnami"
+      SALESFORCE_PASSWORD: "user"
 
       # Optional: Pre-supplied access token (if already authenticated)
-      SUITECRM_ACCESS_TOKEN: ""
+      SALESFORCE_ACCESS_TOKEN: ""
 
     transport: http
     # Command to launch the MCP server

{decodingtrust_agent_sdk-0.2.11 → decodingtrust_agent_sdk-0.2.12}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "decodingtrust-agent-sdk"
-version = "0.2.11"
+version = "0.2.12"
 description = "DecodingTrust Agent Platform (DTap) — A controllable and interactive red-teaming platform for AI agents"
 readme = "README.md"
 license = { file = "LICENSE" }

{decodingtrust_agent_sdk-0.2.11 → decodingtrust_agent_sdk-0.2.12}/utils/llm.py

@@ -123,14 +123,25 @@ def _call_openrouter(messages: list, model: str, temperature: float, max_tokens:
     # Strip the "openrouter/" prefix to get the actual model ID
     actual_model = model[len("openrouter/"):]
 
-    response = client.chat.completions.create(
-        model=actual_model,
-        messages=messages,
-        temperature=temperature,
-        max_tokens=max_tokens,
-    )
-
-    return response.choices[0].message.content
+    # Retry on empty content: under concurrency some models (e.g. deepseek reasoning
+    # variants) intermittently return an empty `content` (answer truncated into
+    # reasoning, or a transient throttle). Empirically ~1/3 of calls under parallel-6
+    # load came back empty; retrying recovers them so JSON-judge callers don't fail open.
+    import time
+    last = ""
+    for attempt in range(4):
+        response = client.chat.completions.create(
+            model=actual_model,
+            messages=messages,
+            temperature=temperature,
+            max_tokens=max_tokens,
+        )
+        last = response.choices[0].message.content or ""
+        if last.strip():
+            return last
+        if attempt < 3:
+            time.sleep(0.6 * (attempt + 1))
+    return last
 
 
 def _call_openai(messages: list, model: str, temperature: float, max_tokens: int) -> str: