Thanks for tracking this down — the bug in #2930 is real and the behavior here is correct: Codex can't load symlinked SKILL.md files, so writing them as real files is the right outcome.

My concern is with how it's special-cased rather than the fix itself. The change hardcodes agent_name == "codex" in CommandRegistrar._write_registered_output (agents.py) and selected_ai != "codex" in _register_extension_skills (extensions.py). Both are generic, agent-agnostic code paths, and we have an integration architecture specifically so per-agent knowledge lives in one declarative place. AGENT_CONFIGS is already derived from INTEGRATION_REGISTRY via _build_agent_configs(), and CodexIntegration is a clean self-contained class — "Codex can't use dev symlinks" really belongs there, not in shared plumbing. (The existing agent_name == "copilot" carve-out is the same debt; I'd rather not grow it.)

Could we route this through integration metadata instead? Roughly:

1. Add dev_no_symlink = True to CodexIntegration and propagate it into the config dict in _build_agent_configs() the same way invoke_separator already is.
2. In _write_registered_output, swap the string check for a metadata lookup:

   if not link_outputs or agent_config.get("dev_no_symlink"):

3. In _register_extension_skills, replace selected_ai != "codex" with link_outputs and not agent_config.get("dev_no_symlink").

This follows the existing inject_name convention (opt-in flag, absent means the established default of symlinking), keeps the constraint owned by the Codex integration as the single source of truth, and means any future agent with the same limitation just sets the flag — no edits to shared code. The test you added is great and would carry over almost unchanged.