feat(benchmarks): Support local Claude UI benchmark suites

.gitignore

@@ -118,3 +118,4 @@ DerivedData
 /repros
 /.xcodebuildmcp
 /out.nosync
+/benchmarks/claude-ui/local/

benchmarks/claude-ui/parse_claude_conversation.py

@@ -1,12 +1,12 @@
 #!/usr/bin/env python3
 """Parse a Claude Code NDJSON session log into per-turn files.
 
-Filters to: user prompts, assistant text replies, and XcodeBuildMCP tool
+Filters to: user prompts, assistant text replies, and configured tool
 calls/results. Strips screenshot image blobs to a short placeholder.
 
 Usage:
     parse_claude_conversation.py <session.jsonl> [output_dir] \\
-        [--tool-prefix=mcp__xcodebuildmcp]
+        [--tool-prefix=mcp__xcodebuildmcp] [--tool-name=Bash]
 """
 
 from __future__ import annotations
@@ -101,10 +101,14 @@ def extract_user_text(entry: dict) -> str:
     return "\n\n".join(parts)
 
 
-def parse(path: Path, out_dir: Path, tool_prefix: str) -> bool:
+def matches_tool_name(name: str, tool_prefixes: list[str], tool_names: set[str]) -> bool:
+    return name in tool_names or any(name.startswith(prefix) for prefix in tool_prefixes)
+
+
+def parse(path: Path, out_dir: Path, tool_prefixes: list[str], tool_names: set[str]) -> bool:
     out_dir.mkdir(parents=True, exist_ok=True)
 
-    # Track tool_use_ids that target our prefix so we keep matching results.
+    # Track tool_use_ids that target configured tools so we keep matching results.
     tracked_ids: set[str] = set()
     tool_name_by_id: dict[str, str] = {}
     counter = 0
@@ -185,7 +189,7 @@ def next_path(kind: str, label: str | None = None) -> Path:
                         )
                     elif btype == "tool_use":
                         name = block.get("name", "")
-                        if not name.startswith(tool_prefix):
+                        if not matches_tool_name(name, tool_prefixes, tool_names):
                             continue
                         tool_id = block.get("id", "")
                         tracked_ids.add(tool_id)
@@ -220,17 +224,25 @@ def main() -> int:
     )
     ap.add_argument(
         "--tool-prefix",
-        default="mcp__xcodebuildmcp",
+        action="append",
+        default=None,
         help="Only include tool calls whose name starts with this prefix",
     )
+    ap.add_argument(
+        "--tool-name",
+        action="append",
+        default=[],
+        help="Also include tool calls whose name exactly matches this value",
+    )
     args = ap.parse_args()
 
     if not args.jsonl.is_file():
         print(f"error: not a file: {args.jsonl}", file=sys.stderr)
         return 1
 
     out = args.output or args.jsonl.with_name(f"{args.jsonl.stem}_conversation")
-    return 0 if parse(args.jsonl, out, args.tool_prefix) else 1
+    tool_prefixes = args.tool_prefix or ["mcp__xcodebuildmcp"]
+    return 0 if parse(args.jsonl, out, tool_prefixes, set(args.tool_name)) else 1
 
 
 if __name__ == "__main__":

benchmarks/claude-ui/suites/contacts.yml

@@ -9,38 +9,38 @@ firstRunPromptDismissals:
     - Continue
     - Not Now
     - OK
-  timeoutSeconds: 8
+  timeoutSeconds: 30
 baseline:
-  totalToolCalls: 14
-  mcpToolCalls: 13
-  uiAutomationCalls: 11
-  wallClockSeconds: 97
+  totalToolCalls: 19
+  trackedToolCalls: 18
+  mcpToolCalls: 18
+  uiAutomationCalls: 16
+  wallClockSeconds: 102.94
   tools:
     session_show_defaults: 1
     launch_app_sim: 1
-    snapshot_ui: 1
+    snapshot_ui: 6
     tap: 5
     type_text: 5
-allowedVariance:
-  totalToolCalls: 3
-  mcpToolCalls: 3
-  uiAutomationCalls: 3
-  wallClockSeconds: 45
-  toolCalls: 2
-expectedToolSequence:
+baselineToolSequence:
   - session_show_defaults
   - launch_app_sim
   - snapshot_ui
   - tap
+  - snapshot_ui
   - tap
   - type_text
+  - snapshot_ui
   - type_text
   - type_text
   - tap
+  - snapshot_ui
   - type_text
   - tap
+  - snapshot_ui
   - type_text
   - tap
+  - snapshot_ui
 failurePatterns:
   - STALE_ELEMENT_REF
   - SNAPSHOT_MISSING

benchmarks/claude-ui/suites/reminders.yml

@@ -8,31 +8,29 @@ firstRunPromptDismissals:
   labels:
     - Continue
     - Not Now
-  timeoutSeconds: 12
+  timeoutSeconds: 30
 baseline:
-  totalToolCalls: 15
-  mcpToolCalls: 14
-  uiAutomationCalls: 12
-  wallClockSeconds: 85
+  totalToolCalls: 17
+  trackedToolCalls: 16
+  mcpToolCalls: 16
+  uiAutomationCalls: 14
+  wallClockSeconds: 92.79
   tools:
     session_show_defaults: 1
     launch_app_sim: 1
     snapshot_ui: 1
-    tap: 4
+    tap: 5
+    wait_for_ui: 1
     type_text: 4
     key_press: 2
     batch: 1
-allowedVariance:
-  totalToolCalls: 5
-  mcpToolCalls: 5
-  uiAutomationCalls: 5
-  wallClockSeconds: 60
-  toolCalls: 3
-expectedToolSequence:
+baselineToolSequence:
   - session_show_defaults
   - launch_app_sim
   - snapshot_ui
   - tap
+  - wait_for_ui
+  - tap
   - type_text
   - tap
   - tap

benchmarks/claude-ui/suites/weather.yml

@@ -6,25 +6,20 @@ sessionDefaults:
   scheme: Weather
   simulatorName: iPhone 17 Pro Max
 baseline:
-  totalToolCalls: 13
-  mcpToolCalls: 12
-  uiAutomationCalls: 10
-  wallClockSeconds: 98
+  totalToolCalls: 14
+  trackedToolCalls: 13
+  mcpToolCalls: 13
+  uiAutomationCalls: 11
+  wallClockSeconds: 100.03
   tools:
     session_show_defaults: 1
     build_run_sim: 1
-    snapshot_ui: 1
+    snapshot_ui: 2
     tap: 6
     batch: 1
     type_text: 1
     swipe: 1
-allowedVariance:
-  totalToolCalls: 2
-  mcpToolCalls: 2
-  uiAutomationCalls: 2
-  wallClockSeconds: 45
-  toolCalls: 2
-expectedToolSequence:
+baselineToolSequence:
   - session_show_defaults
   - build_run_sim
   - snapshot_ui
@@ -36,6 +31,7 @@ expectedToolSequence:
   - tap
   - tap
   - swipe
+  - snapshot_ui
   - tap
 failurePatterns:
   - STALE_ELEMENT_REF