An unofficial Codex adaptation of PStack, the engineering skills and playbooks created by Lauren Tan (Poteto). This repository keeps PStack's workflows and principles while mapping Cursor-specific integration to Codex.
PStack's central idea is simple: make fewer changes, verify them against real behavior, and use parallel agents only when their work can be reviewed. $poteto-mode chooses a playbook for the task and calls the other skills as needed.
This is a community adaptation, not an official PStack release from Poteto or Cursor.
Add this repository as a Codex plugin marketplace, then install the plugin:
codex plugin marketplace add aldiz1711/pstack-codex
codex plugin add pstack-codex@pstack-codexStart a new Codex task so it loads the installed skills. Review the bundled hook if Codex asks, then trust it if you want PStack to load your model configuration at task and subagent startup. The hook reads ~/.codex/pstack-models.md when that file exists; PStack also has working defaults without it.
Run $setup-pstack to choose a reasoning budget and models for each role. Its default budget is unlimited (max). Then use $poteto-mode on a real task:
$poteto-mode reproduce this bug, fix its root cause, and verify the result in the running app.
See the setup guide for model configuration and pull request review setup. The full guide covers the workflows from investigation through verification and shipping.
- Routes work through playbooks.
$poteto-modecovers investigation, bug fixes, features, refactoring, performance, PR babysitting, shipping, and longer autonomous runs. It stays active in the task until you opt out. - Uses focused skills and principles.
$how,$why,$architect,$arena,$swarm,$interrogate, and the verification skills can also be called directly. The guide andpoteto-modeskill lead to the full instructions. - Makes review and proof explicit. Agents inspect diffs, check claims against the code, and exercise the real app where the playbook requires it.
$interrogateruns PStack's local review panel. - Lets you configure model roles.
$setup-pstackwrites~/.codex/pstack-models.md. The current defaults use GPT-6 Luna for scoped code work, GPT-6 Sol for selected review work, and GPT-6 Astra for the hardest judgment and design tasks.
For example:
$how does cancellation flow through this service?
$arena compare several designs for this API boundary.
$interrogate review the current diff.
PStack's playbooks, skill instructions, and engineering principles remain the source of the workflow. The integration points use Codex mechanisms:
| PStack need | Codex mechanism |
|---|---|
| Model choices loaded into agent sessions | $setup-pstack writes a role map; the bundled Codex hook loads it at task and subagent startup. |
| Durable objectives and recurring audits | Codex /goal and scheduled follow-ups. |
| Parallel workers and independent reviewers | Codex subagents and isolated worktrees; read-only roles use the bundled pstack-readonly agent. |
| Local code review | $interrogate keeps PStack's review workflow; Codex /review is also available. |
| Automated GitHub PR review | Enable Codex Cloud Code Review separately for the repository. PStack's babysit playbook then triages its review threads. |
Cursor supports a wider mix of model providers. To keep Arena candidates at comparable capability in Codex, the default Arena runs four independent GPT-6 Astra candidates at max reasoning and uses a separate, blinded judge. $setup-pstack can change this panel. Independent runs of one model do not provide the same provider diversity as the original setup.
Original PStack was created by Lauren Tan and is included under its MIT license. The bundled $control-cli, $control-ui, and $deslop skills come from Cursor Team Kit under its MIT license. The Codex adaptation is covered by this repository's MIT license.