dsh-pwsh-quoting-guard

The hardest part of LLM-driven PowerShell on Windows is rarely the script logic itself, but the nested quotes. The official pwsh tool packs the entire command into a single string passed via pwsh -Command, and Windows PowerShell 5.1 silently swallows embedded double quotes when forwarding arguments to native executables. The output is wrong, yet no error is raised and token metrics show nothing. dsh-pwsh-quoting-guard eliminates this class of silent failure at the structural level. It exposes two model-visible tools: pwsh_script delivers the script body as a single argv element and binds data parameters as a structured array to $DSH_ARGS, with all escaping handled by plugin code so the model never touches quotes; run_argv bypasses the shell entirely, passing the argument vector straight to the child process and supporting an explicit working directory.

The two tools coexist with the official pwsh tool rather than replacing it. Short one-liners, background tasks, and sandbox-permission escalation still go through the official channel. For quoted arguments, paths with spaces or non-ASCII characters, or multi-line scripts, reaching for these tools first yields correct results at the same token cost. The plugin has zero runtime dependencies and adds roughly 313 tokens of fixed overhead per request, paying for itself once it prevents even one or two quoted-argument failures per call.

Who is this for? If you run DSH sessions on Windows and the model frequently invokes PowerShell scripts or external programs, and you have already been burned by swallowed quotes, CLIXML noise, or retry storms, this plugin is built for exactly that pain point. Install it, restart the DSH web server, and the two tools appear in the tool registry with no extra configuration or dependencies required.