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

[Mihaiii]

Summary

• DelegateAgent now records Delegate's Windows shell tool, ExecutePowershellCommand, as the canonical Bash tool. It sits in _TOOL_NAME_MAP next to the existing LoadSkill → Skill mapping.
• With this change, command_executed treats a Windows Delegate shell call like any other shell call. tool_name: Bash filters match it, and patterns run against the raw command string instead of the JSON-dumped parameters.

Why

delegate-stdio renames ExecuteBashCommand to Bash itself, but it leaves ExecutePowershellCommand as is. On the Windows slice of the coder_eval_uipath nightly (build 13817754), the agent ran the right commands and still lost points:

Task	Score	Check that failed	Why
skill-genome-execute-component	0.938	ran uip rpa build	The check filters on tool_name=Bash, so it can never match ExecutePowershellCommand.
skill-migrator-migrate-classic-windows	0.867	ran analyze --output-format sarif / upgrade --output-path	In the JSON the command reads UiPath.Upgrade.exe\" analyze. The escaped quote breaks the ["']?\s+analyze pattern.
same		"no analyze/upgrade without SARIF output"	The exclusion matched the call's plain-English steps text ("Run UiPath.Upgrade.exe analyze with SARIF output…"), not the command.

Those were the only checks that failed on either task, so with this change both pass.

Why here, not in delegate-stdio

Adding ExecutePowershellCommand: 'Bash' to TOOL_NAME_ALIAS in delegate-stdio would also work. But the fix would then only reach runs that pick up a new host release. This side covers every pinned delegate-stdio version, and follows the pattern already used for LoadSkill.

Tests

• New: test_powershell_command_is_recorded_as_the_canonical_bash_call. It failed before the change ('ExecutePowershellCommand' != 'Bash').
• test_happy_path_text_and_tool used ExecutePowershellCommand as a generic tool name and asserted it came through unchanged. It now uses Bash, so it keeps testing the happy path rather than the name mapping.
• tests/test_delegate_agent.py, tests/test_command_executed.py, tests/test_custom_lint.py, ruff format/check and pyright (delegate_agent.py) pass locally on Windows. The process-group teardown test is skipped as POSIX-only, and CE033 needs PYTHONUTF8=1 on a cp1252 console.

🤖 Generated with Claude Code

[Mihaiii]

Superseded: the alias now lives where the rest of Delegate's tool-name aliases do, in delegate-stdio's TOOL_NAME_ALIAS (UiPath/Autopilot@87fbbfc1e8 on fix/sdk-interop-log-path). It reaches the nightly with the next delegate-stdio release, since DELEGATE_STDIO_VERSION defaults to latest.