tests(claude_code): bugbot — raise MIN_STREAM_EVENTS for tool-use streaming on remaining 4 providers

Per Bugbot, MIN_STREAM_EVENTS=4 is below the buffered-proxy baseline for
tool-use streaming: even a buffered wire produces ~5 stream-json records
(system + assistant w/ tool_use + user tool_result + assistant text +
result), so the < 4 check would silently pass for a regression that
collapses the upstream stream. Bring azure / bedrock_invoke /
bedrock_converse / vertex_ai in line with test_anthropic.py (MIN=8,
already raised in a253bd940b) so every provider catches the regression
this row is designed to detect.

Co-authored-by: Mateo Wang <mateo-berri@users.noreply.github.com>

tests/claude_code/tool_use_streaming/test_azure.py

@@ -50,7 +50,18 @@ TOOL_USE_ARGS = [
     "dontAsk",
 ]
 
-MIN_STREAM_EVENTS = 4
+# Floor on the number of stream-json records we expect to see for a
+# tool-use turn. A buffered (non-streamed) wire for this multi-turn
+# flow collapses to roughly: one `system` init + one `assistant` with
+# the `tool_use` block + a `user` tool_result + one `assistant` final
+# text + one `result`, i.e. ~5 records (the CLI executes the tool
+# locally and sends the result back, producing a second model turn
+# even on a fully buffered proxy). Real fine-grained streaming
+# produces many more (incremental input_json_delta events,
+# intermediate assistant deltas, etc., typically 15+). We pick a
+# floor comfortably above the buffered case so the assertion catches
+# the regression without being flaky on short responses.
+MIN_STREAM_EVENTS = 8
 
 
 def _has_tool_use_event(events: Sequence[Mapping[str, Any]]) -> bool:

tests/claude_code/tool_use_streaming/test_bedrock_converse.py

@@ -56,7 +56,18 @@ TOOL_USE_ARGS = [
     "dontAsk",
 ]
 
-MIN_STREAM_EVENTS = 4
+# Floor on the number of stream-json records we expect to see for a
+# tool-use turn. A buffered (non-streamed) wire for this multi-turn
+# flow collapses to roughly: one `system` init + one `assistant` with
+# the `tool_use` block + a `user` tool_result + one `assistant` final
+# text + one `result`, i.e. ~5 records (the CLI executes the tool
+# locally and sends the result back, producing a second model turn
+# even on a fully buffered proxy). Real fine-grained streaming
+# produces many more (incremental input_json_delta events,
+# intermediate assistant deltas, etc., typically 15+). We pick a
+# floor comfortably above the buffered case so the assertion catches
+# the regression without being flaky on short responses.
+MIN_STREAM_EVENTS = 8
 
 
 def _has_tool_use_event(events: Sequence[Mapping[str, Any]]) -> bool:

tests/claude_code/tool_use_streaming/test_bedrock_invoke.py

@@ -54,7 +54,18 @@ TOOL_USE_ARGS = [
     "dontAsk",
 ]
 
-MIN_STREAM_EVENTS = 4
+# Floor on the number of stream-json records we expect to see for a
+# tool-use turn. A buffered (non-streamed) wire for this multi-turn
+# flow collapses to roughly: one `system` init + one `assistant` with
+# the `tool_use` block + a `user` tool_result + one `assistant` final
+# text + one `result`, i.e. ~5 records (the CLI executes the tool
+# locally and sends the result back, producing a second model turn
+# even on a fully buffered proxy). Real fine-grained streaming
+# produces many more (incremental input_json_delta events,
+# intermediate assistant deltas, etc., typically 15+). We pick a
+# floor comfortably above the buffered case so the assertion catches
+# the regression without being flaky on short responses.
+MIN_STREAM_EVENTS = 8
 
 
 def _has_tool_use_event(events: Sequence[Mapping[str, Any]]) -> bool:

tests/claude_code/tool_use_streaming/test_vertex_ai.py

@@ -53,7 +53,18 @@ TOOL_USE_ARGS = [
     "dontAsk",
 ]
 
-MIN_STREAM_EVENTS = 4
+# Floor on the number of stream-json records we expect to see for a
+# tool-use turn. A buffered (non-streamed) wire for this multi-turn
+# flow collapses to roughly: one `system` init + one `assistant` with
+# the `tool_use` block + a `user` tool_result + one `assistant` final
+# text + one `result`, i.e. ~5 records (the CLI executes the tool
+# locally and sends the result back, producing a second model turn
+# even on a fully buffered proxy). Real fine-grained streaming
+# produces many more (incremental input_json_delta events,
+# intermediate assistant deltas, etc., typically 15+). We pick a
+# floor comfortably above the buffered case so the assertion catches
+# the regression without being flaky on short responses.
+MIN_STREAM_EVENTS = 8
 
 
 def _has_tool_use_event(events: Sequence[Mapping[str, Any]]) -> bool: