fix(windows): avoid unspawnable Deep Scan launchers

[kmbroai]

Summary

• Normalize CODEX_CLI_PATH at the nested Deep Scan MCP environment boundary.
• On Windows, preserve explicit direct executables while rejecting extensionless npm shims, .cmd shims that require a shell, and protected WindowsApps binaries.
• Fall back to the existing resolveCodexCommand() platform-binary resolver instead of adding another launcher resolver.

Root cause

Deep Scan workers inherit CODEX_CLI_PATH through the plugin MCP environment. Before this change, any non-empty value bypassed the canonical npm platform-binary resolver. On Windows that value can be the extensionless npm shim or a protected MSIX WindowsApps executable; CodexExec passes it directly to child_process.spawn without a shell, which can fail with spawn EPERM even when the installed @openai/codex platform package contains a valid executable.

This keeps case-insensitive Windows environment propagation intact, preserves explicit spawnable .exe / .com launchers, and uses the existing bundled platform executable when the inherited value cannot be launched.

Related: CLI-18410, codex#35872.

Validation

• bun test tests-ts/runtime.test.ts --test-name-pattern 'bundled Codex through|explicit Codex executable override|spawnable Windows' — 3 passed
• pnpm run types
• pnpm run build
• pnpm run format
• pnpm pack --pack-destination ../../dist
• pnpm run test:package — validated the installed package and a nested worker without global codex
• Native prepublish review gate: 3 independent passes plus verifier; 0 findings

The full runtime.test.ts module was also attempted, but this host reports / and /home as UID 65534, so 24 unrelated trusted-owner tests fail closed. The launcher-focused tests and package smoke pass without weakening those checks.

[ting-hong-shieh]

I validated exact head 8eb9493d and a conflict-resolved replay of its single commit onto current main at 216212b7.

The exact head passes the three focused launcher tests (11 assertions), generated-model check, TypeScript check, build, Prettier, and whitespace check. The replay passes the four focused launcher tests inherited from the PR and current main (15 assertions), generated-model check, TypeScript check, build, Prettier, and whitespace check.

The replay also exposes one relative-path regression that the existing tests do not cover. The inline note includes the reproduction and a tested fix direction. These were local unit and static checks on macOS; I did not run a native Windows process or use credentials or an external API.

[ting-hong-shieh]

Please preserve current main's absolute-path normalization when resolving this conflict. pluginExecutionEnvironment() now delegates to resolveCodexCommand(environment), which turns a relative CODEX_CLI_PATH into an absolute path before passing it to a nested worker. This helper instead returns the trimmed relative value unchanged on non-Windows platforms, and does the same for a relative .exe or .com path on Windows. The worker can then resolve that value from a different working directory and fail to launch Codex.

On the conflict-resolved replay at 216212b7, resolveNestedCodexPath({ CODEX_CLI_PATH: "./bin/codex" }, "linux") returned "./bin/codex" instead of resolve("./bin/codex"). Normalizing an accepted override with the platform-specific path resolver made this reproduction and all four focused launcher tests pass. Please retain that normalization and add relative-path coverage for accepted non-Windows and Windows executables.