[model-apps][Bug] /genpage bypasses the prompt-file upload wrapper and breaks quoted prompts on Windows

[akshaymaloo]

Summary

The reusable GenPage CLI wrapper was fixed in #565 to pass prompt and agent-message text through --prompt-file / --agent-message-file, but the user-facing /genpage create and edit workflows still instruct the orchestrator to invoke raw pac model genpage upload with inline --prompt and --agent-message.

This leaves the main skill path exposed to the command-line quoting failures that the wrapper was created to avoid.

Live reproduction

An isolated live page deployment used the approved prompt containing an ASCII-quoted multiword page name:

Create and LIVE DEPLOY one isolated generative page named "Facilities Inspection Live CK1 Overview" ...

The direct upload command failed before mutation:

Error: Not a valid command.
Parse failed on: Inspection
Current command: upload
Was it quote wrapped? No, be sure to wrap values that contain spaces.

The failure reproduced through normal PowerShell invocation and through .NET ProcessStartInfo.ArgumentList; the PAC development shim reparsed its raw command line and split the embedded ASCII-quoted text.

The same prompt parsed and deployed only after replacing the two ASCII quote characters with typographic quotes:

named “Facilities Inspection Live CK1 Overview”

The page then created, published, attached to the sitemap, packaged into the solution, downloaded, and ran successfully.

Root cause

plugins/model-apps/scripts/lib/genpage-cli.js already has the safe implementation:

• writes prompt/agent message to unique UTF-8 temp files;
• passes --prompt-file and --agent-message-file;
• cleans the temp directory on every exit path.

However, plugins/model-apps/skills/genpage/SKILL.md Phase 6 and edit-flow.md still:

• require logging raw pac model genpage upload;
• require inline --prompt "<full prompt>";
• show direct PowerShell upload commands;
• bypass makeGenpageCli().upload().

So /app-builder pages use the safe wrapper, while the standalone /genpage skill still uses the unsafe path.

Impact

• Valid prompts with quotes or other shell-sensitive content can fail deployment.
• Multi-line edit transcripts remain vulnerable to flattening/mangling on the standalone skill path.
• Agents may mutate the approved prompt to work around quoting, weakening prompt provenance.
• Behavior differs between /app-builder and /genpage despite sharing the same upload contract.

Expected

Route standalone create/edit deployment through one script/API backed by makeGenpageCli().upload() instead of constructing raw PAC commands in Markdown.

Logging should preserve the approved prompt semantically without requiring the full sensitive/arbitrary text to be embedded as an executable shell command. Add a live-shaped Windows regression test containing:

• a multi-line prompt;
• an ASCII-quoted multiword name;
• Unicode;
• shell metacharacters;
• an update with a downloaded conversation transcript.

Related

• [model-apps][Enhancement] genpage upload: multi-line prompts are flattened because --prompt-file / --agent-message-file / --compiled-code-file are never used #565 fixed the wrapper-level file transport but did not migrate the standalone skill path.
• [model-apps][Bug] GenPage lifecycle safety: parsing, download metadata, filenames, and process failures #588 covers separate Windows argv quoting and lifecycle parser issues.