Environment
Windows 11, Codex VS Code extension, Codex CLI 0.136.0-alpha.2
Detected shell (from environment_context): powershell
Webwright skill invoked via /webwright:run
Summary
The webwright skill instructs the agent to "drive the browser through one bash command at a time" and its reference (reference/playwright_patterns.md, the inline-Python example) demonstrates a bash heredoc (python - <<'PY' … PY). On Windows the shell is PowerShell, which cannot parse bash heredocs. Whether a run succeeds therefore depends entirely on the model being able to recognize this and deviate from the skill's examples.
Observed behavior (same machine, same skill, same shell)
gpt-5.5: recognizes it's on PowerShell, writes the script to a .py file and runs python file.py. Completes tasks in a few steps.
gpt-5-codex / gpt-5.3-codex: follow the skill's bash heredoc example literally, hit the PowerShell parse error, and retry the same broken syntax in a loop. A simple Wikipedia task took ~18 minutes; a moderately complex booking task looped for hours without converging.
Root cause
The skill's instructions and examples assume bash. Code-specialist models are tuned for tight, literal adherence to skill instructions, so they reproduce the bash heredoc verbatim instead of adapting to the reported powershell shell. There is no shell-awareness in the skill and no guardrail when the harness runs under a model that doesn't self-correct.
Environment
Windows 11, Codex VS Code extension, Codex CLI 0.136.0-alpha.2
Detected shell (from environment_context): powershell
Webwright skill invoked via /webwright:run
Summary
The webwright skill instructs the agent to "drive the browser through one bash command at a time" and its reference (reference/playwright_patterns.md, the inline-Python example) demonstrates a bash heredoc (python - <<'PY' … PY). On Windows the shell is PowerShell, which cannot parse bash heredocs. Whether a run succeeds therefore depends entirely on the model being able to recognize this and deviate from the skill's examples.
Observed behavior (same machine, same skill, same shell)
gpt-5.5: recognizes it's on PowerShell, writes the script to a .py file and runs python file.py. Completes tasks in a few steps.
gpt-5-codex / gpt-5.3-codex: follow the skill's bash heredoc example literally, hit the PowerShell parse error, and retry the same broken syntax in a loop. A simple Wikipedia task took ~18 minutes; a moderately complex booking task looped for hours without converging.
Root cause
The skill's instructions and examples assume bash. Code-specialist models are tuned for tight, literal adherence to skill instructions, so they reproduce the bash heredoc verbatim instead of adapting to the reported powershell shell. There is no shell-awareness in the skill and no guardrail when the harness runs under a model that doesn't self-correct.