Choose the execution layer before acting. The same physical computer can expose different hosts, shells, permissions, projects, and browser capabilities.
Preserve These Boundaries
- Keep the Windows ChatGPT desktop app's Agent Environment set to Windows native. This owner's Windows Chrome-extension and Computer Use path is verified in that mode. Do not claim that it works with a WSL agent unless current official documentation establishes that behavior.
- Treat Integrated Terminal Shell as independent from Agent Environment. Changing one is not a fix for the other.
- Keep WSL2 available in parallel for Linux development and remote Codex projects. A Mac-to-WSL app SSH connection does not require changing the Windows desktop app's native agent environment.
- Distinguish ChatGPT Remote, raw SSH, and an app SSH project. They can use SSH under the hood or reach the same computer without being interchangeable.
- Apply the permissions, sandbox, tools, browser setup, and trust state of the selected host. Do not assume configuration or SSH trust copies between devices.
- Never inspect or publish credential-capable files, network coordinates, key material, fingerprints, account identifiers, or machine-specific absolute paths. Use configured SSH aliases and environment variables.
Route The Task
Read topology and routing whenever choosing a device, connection, shell, browser, or troubleshooting layer. It contains the owner-verified device inventory, interface table, preconditions, and diagnostic rules. It also contains the concrete Codex task-dispatch procedure for starting work on another connected desktop host without reusing an unrelated task.
Read synchronization and public-safety rules when changing repositories, skills, global guidance, refresh automation, or push policy. Remote Git repositories remain canonical; generated skill installs and rendered global instructions are not editable sources.
Global AGENTS.md may use lazy-loaded runtime-mode references in a sibling
agents-md-references/ directory. Keep only runtime detection and conditional links in the
main file; keep detailed mode-specific instructions in separate Markdown files
and load only the file for the identified product mode. The private context
repository owns personal runtime content, while SkillPort materializes managed
agents-md-references/*.md copies next to the generated AGENTS.md. Do not put private
runtime instructions in this public AgentDesk repository.
For detailed monitor, dock, KVM, or physical cabling work, use
dell-monitor-kvm-setup when installed. This skill retains only the digital
routing context needed to select a host and interface.
Before Acting
- Name the exact execution layer: desktop host, raw native-Windows SSH, raw WSL SSH, app SSH project, browser extension/Computer Use, SkillPort, or Git.
- Confirm its preconditions and gather evidence from that layer before changing settings.
- Make the narrowest layer-local correction. Do not change a system-wide shell or Agent Environment to repair a different layer.
- Verify observable behavior on the selected host and report what was actually checked versus what remains intended or unverified.