Integrate with baserun (#274)

agbenchmark/start_benchmark.py

@@ -214,7 +214,8 @@ def start(
 
     # when used as a library, the pytest directory to execute is in the CURRENT_DIRECTORY
     pytest_args.append(str(CURRENT_DIRECTORY))
-
+    if os.environ.get("BASERUN_API_KEY"):
+        pytest_args.extend(["--baserun"])
     return sys.exit(pytest.main(pytest_args))
 
 

agent/mini-agi

@@ -1 +1 @@
-Subproject commit 2fc70aa0032eec986dfb1020854a1b3b8aaf6780
+Subproject commit 667d834cfe7e250aa654e7b0c663650451278728

poetry.lock

@@ -258,6 +258,20 @@ files = [
     {file = "backcall-0.2.0.tar.gz", hash = "sha256:5cbdbf27be5e7cfadb448baf0aa95508f91f2bbc6c6437cd9cd06e2a4c215e1e"},
 ]
 
+[[package]]
+name = "baserun"
+version = "0.3"
+description = "Tools for testing, debugging, and evaluating LLM features."
+optional = false
+python-versions = ">=3.7.1"
+files = [
+    {file = "baserun-0.3-py3-none-any.whl", hash = "sha256:b4570d3d3ab64b2fc346d1bd41aad145404e3d521065778dfd942e84056a302a"},
+    {file = "baserun-0.3.tar.gz", hash = "sha256:e7b8a63d1039228973c4a93c4a3b3df31f4b979a5d7476df12c7dd55c9bed41f"},
+]
+
+[package.dependencies]
+requests = ">=2.31.0"
+
 [[package]]
 name = "black"
 version = "22.3.0"
@@ -2716,4 +2730,4 @@ multidict = ">=4.0"
 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "855fe892a4401b03bbda926df20c140ea5143574ff5d04b6168472a2d162b9dd"
+content-hash = "d3f4df7a26eacfe0b726142521f20f3c98ee721b4710502e1ed7bffda3c9dff0"

pyproject.toml

@@ -27,6 +27,7 @@ colorama = "^0.4.6"
 pyvis = "^0.3.2"
 selenium = "^4.11.2"
 agent-protocol = "^0.2.3"
+baserun = "^0.3"
 
 [tool.poetry.group.dev.dependencies]
 flake8 = "^3.9.2"