fix(agents): grade Delegate's Windows PowerShell tool as Bash

.claude/notes/agents.md

@@ -580,11 +580,16 @@ Four cases are worth knowing:
 - **Pi's search tool is `find`** (glob-by-pattern), not `glob`; there is no `glob` tool in
   its built-in set, so mapping `find` to the canonical `Glob` is what keeps
   `command_executed` and `commands_efficiency` comparable.
-- **Delegate's host maps most names itself, but not `LoadSkill`.** `delegate-stdio` reports
-  `ReadFile` as `Read` and `ExecuteSkillApi` as `Skill`, so `DelegateAgent` maps only
-  `LoadSkill {name}` to `Skill {skill}`; without it `skill_triggered` never sees a Delegate skill
-  load. The rename is keyed by the host's name: keyed by `Skill`, it would also turn
-  ExecuteSkillApi's `name`, an API call, into a skill engagement.
+- **Delegate's host maps most names itself, but not `LoadSkill` or its Windows shell.**
+  `delegate-stdio` reports `ReadFile` as `Read`, `ExecuteBashCommand` as `Bash` and
+  `ExecuteSkillApi` as `Skill`, so `DelegateAgent` maps only `LoadSkill {name}` to
+  `Skill {skill}` and `ExecutePowershellCommand` to `Bash`. Without the first,
+  `skill_triggered` never sees a Delegate skill load. Without the second, every Windows shell
+  call misses `tool_name: Bash` criteria and is matched as JSON, where an escaped quote
+  (`UiPath.Upgrade.exe\" analyze`) breaks the pattern and the call's plain-English `steps`
+  can trip an exclusion (coder_eval_uipath nightly 13817754 scored correct migrator and genome
+  runs 0.8–0.94 this way). The rename is keyed by the host's name: keyed by `Skill`, it would
+  also turn ExecuteSkillApi's `name`, an API call, into a skill engagement.
 
 Antigravity additionally strips the result payload out of a tool call's arguments: the
 harness folds result fields into the same `args` dict at DONE. Beyond a static key list,

src/coder_eval/agents/delegate_agent.py

@@ -150,11 +150,12 @@ def _is_config_init_error(message: str) -> bool:
 """``toolStatus`` values the SDK's ``tool_result`` carries for a tool that did not complete."""
 
 # The host reports most tools under their canonical (Claude) names already
-# (ReadFile -> Read, ...), but not LoadSkill, which loads a catalog skill by name.
+# (ReadFile -> Read, ExecuteBashCommand -> Bash, ...), but not LoadSkill, which loads a
+# catalog skill by name, nor ExecutePowershellCommand, the Windows shell tool.
 # Keyed by the host's name, not the canonical one: the host also reports
 # ExecuteSkillApi as `Skill`, and that call's `name` is an API call, not a skill load.
 # Rationale: .claude/notes/agents.md § Tool-name and argument normalization
-_TOOL_NAME_MAP: dict[str, str] = {"LoadSkill": "Skill"}
+_TOOL_NAME_MAP: dict[str, str] = {"LoadSkill": "Skill", "ExecutePowershellCommand": "Bash"}
 _DELEGATE_ARG_RENAME: dict[str, dict[str, str]] = {"LoadSkill": {"name": "skill"}}
 
 

tests/test_delegate_agent.py

@@ -413,7 +413,7 @@ async def test_happy_path_text_and_tool(self, patch_exec, tmp_path):
         events = [
             _ev(type="session_start", sessionId="sess-1"),
             _ev(type="thinking", content="let me think"),
-            _tool_call("tool-1", "ExecutePowershellCommand", command="ls"),
+            _tool_call("tool-1", "Bash", command="ls"),
             _tool_result("tool-1", "file.txt"),
             _ev(type="thinking", content=""),
             _ev(type="message", content="here is my answer", isStepStart=True),
@@ -428,7 +428,7 @@ async def test_happy_path_text_and_tool(self, patch_exec, tmp_path):
         assert not record.crashed
         assert len(record.commands) == 1
         command = record.commands[0]
-        assert command.tool_name == "ExecutePowershellCommand"
+        assert command.tool_name == "Bash"
         assert command.parameters == {"command": "ls"}
         assert command.result_status == "success"
         assert command.result_summary == "file.txt"
@@ -618,6 +618,19 @@ async def test_load_skill_is_recorded_as_the_canonical_skill_call(self, patch_ex
         assert (command.tool_name, command.parameters) == ("Skill", {"skill": "uipath-rpa", "plugin": ""})
         assert _engaged_skill_names(command) == {"uipath-rpa"}
 
+    async def test_powershell_command_is_recorded_as_the_canonical_bash_call(self, patch_exec, tmp_path):
+        """Shell criteria read `Bash` + `command`; on Windows the host sends `ExecutePowershellCommand`."""
+        command = '& "C:\\Tools\\UiPath.Upgrade.exe" analyze --output-format sarif'
+        events = [
+            _tool_call("p1", "ExecutePowershellCommand", command=command, steps=["Run analyze"]),
+            _tool_result("p1"),
+            _result(response="done"),
+        ]
+        agent, _ = await _started_agent(patch_exec, events, tmp_path)
+        record = await agent.communicate("hi")
+        [recorded] = record.commands
+        assert (recorded.tool_name, recorded.parameters) == ("Bash", {"command": command, "steps": ["Run analyze"]})
+
     async def test_skill_api_call_keeps_its_name_arg(self, patch_exec, tmp_path):
         """The host already reports ExecuteSkillApi as `Skill`; its `name` is not a skill load."""
         events = [