[PF-24] Sandbox PATH resolution and Windows EINVAL fix for Claude/Codex

[chucoding]

Purpose

Make both Claude and Codex CLIs run reliably in Cursor agent sandbox (PowerShell on Windows, macOS) by (1) resolving executables from common npm global paths when PATH is restricted, and (2) avoiding EINVAL on Windows by running .cmd via PowerShell instead of spawning them directly.

Description

Related Notion: Sandbox Path

Summary of changes:

Area	File(s)	Change
cli-js shell	packages/cli-js/src/utils/shell.ts	Added getNpmGlobalExeCandidates(cmd) and resolveCommandPathWithNpmFallback(cmd) for PATH + npm global fallback (Windows: APPDATA/LOCALAPPDATA npm; Unix: HOME-based paths).
cli-js Claude	packages/cli-js/src/providers/claude.ts	resolveClaudeExe() now uses the shared fallback. completeOneTurn and runClaudeStreaming on Windows use a temp file + PowerShell script (Get-Content ... | & 'exe') so .cmd is never spawned directly (avoids EINVAL / CVE-2024-27980).
cli-js Codex	packages/cli-js/src/providers/codex.ts	Added resolveCodexExe(). checkCodex / listModelsCodex use fallback. runCodexExec and runCodexExecStreaming resolve exe first; on Windows the script uses & 'escapedExe' exec -, on non-Windows spawn(exe, ["exec", fullPrompt], ...).
cli-py shell	packages/cli-py/planforge/utils/shell.py	Added get_npm_global_exe_candidates(cmd) and resolve_command_path_with_npm_fallback(cmd) (same path logic as cli-js).
cli-py Claude	packages/cli-py/planforge/providers/claude.py	All subprocess calls use resolved exe. On Windows (os.name == "nt"), complete_one_turn and _run_claude_streaming use temp file + PowerShell. run_implement delegates to _run_claude_streaming.
cli-py Codex	packages/cli-py/planforge/providers/codex.py	Added _resolve_codex_exe(). Windows script invokes codex by absolute path (& 'escaped_exe' exec -); non-Windows uses [exe, "exec", full_prompt].

• Root cause: In sandbox, PATH often omits npm global bin → ENOENT for claude. On Windows, spawning claude.cmd (or codex.cmd) directly from Node can yield EINVAL; running them via PowerShell avoids that.
• Design: Shared “PATH then npm global candidates” resolution in shell utils; both providers use it. On Windows, Claude and Codex are invoked only via powershell -Command "..." with temp-file or pipe, never by direct spawn of the .cmd.

How to test

1. Build: From repo root, run pnpm run build in packages/cli-js — should succeed.
2. Doctor (Claude on Windows): Run planforge doctor ai, choose claude. TC1–TC3 should pass without spawnSync ... EINVAL.
3. Plan/Implement in sandbox: In Cursor agent (or a restricted PATH), run planforge plan "simple goal" and planforge implement "simple task" with planner=claude and implementer=codex — both should run without ENOENT/EINVAL.
4. Python (optional): From packages/cli-py, run the same flows if you use the Python CLI; behavior should match.

Review Requirement

• Windows path escaping: In PowerShell scripts, exe and temp paths use single-quote escaping (' → '') for safety; please confirm there are no edge cases (e.g. paths with ').
• Provider symmetry: Claude and Codex now share the same resolution and Windows-invocation pattern; worth a quick cross-check between claude.ts/codex.ts and claude.py/codex.py.

Additional Info

• Related Notion: Sandbox Path (design and implementation notes).
• Aligns with .cursor/rules/provider-symmetry.mdc, node-python-symmetry.mdc, and cross-platform.mdc.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.