Codex capability boundary

This is source guidance, not an installed connector or a runtime proof. Resolve tools against the current host and read their actual schemas before using them. Preserve the named product and requested outcome. If a required operation is unavailable, report its exact name and leave that step incomplete. Do not invent tool names, connect accounts, install helpers, spend money, or broaden access from this document alone. Current operator authorization and native permissions govern every action.

Use native Codex delegation and the operator model policy. Missing named reviewers produce an incomplete review, never a clean verdict. Cursor plugins and built-in commands are not installed by this mirror. Read the adapted requirements reference before dependent work. Poteto owns its inline principle routing; no additional principle-pack-router pass is needed inside that workflow. System and developer instructions retain their priority. Existing operator authorization persists across turns unless withdrawn; do not ask again solely because a turn changed.

No sticky mode is claimed. Use native completion notifications for running workers. A recurring continuation requires the operator request and a supported native scheduling tool; do not emulate Cursor /loop.

The Algorithm

Five steps, in order, for any change bigger than an edit you can see at a glance. The order is the principle. Each step is cheap only after the one before it, and the common failure is doing a late step on something an early step would have removed. Adapted from Elon Musk's five-step engineering process.

1. Make the requirements less dumb. Assume the requirement is wrong and make it less wrong. It comes from a person, not a department. "The ticket", "the linter", "the reviewer", or "the old code did it" is a source, not a reason; find the person and their reasoning. If you do not agree with the reasoning, do not accept the requirement, however smart the person who gave it, the human included. No is a valid outcome.
2. Delete the part or process step. Try hardest to delete the part, flag, layer, retry, or step entirely before touching step 3. Optimizing something that should not exist is the biggest mistake smart people make. If you never add anything back, you did not delete enough. Leaf: pstack cursor-principle-subtract-before-you-add.
3. Optimize what survived. Third step, not first. Have one part do many things instead of a parallel copy per caller. Look hardest at boundaries between owners and layers, where a wrapper wraps a wrapper. Leaves: pstack cursor-principle-laziness-protocol, cursor-principle-minimize-reader-load.
4. Accelerate. Be the part: walk one change through edit, build, check, and review yourself and note where you wait, crawl, or bounce. Tighten that. Leaf: pstack cursor-principle-sequence-verifiable-units.
5. Automate last, or you automate something that should not exist. Volume alone does not justify it; precision and reviewability do. pstack cursor-principle-build-the-lever still applies. It runs on what survived.

Run it again. This is not a one-time gate. Rerun it on the result before you ship; the second pass deletes more.