Skip to content

Latest commit

 

History

History
54 lines (43 loc) · 3.36 KB

pstack.md

File metadata and controls

54 lines (43 loc) · 3.36 KB

pstack

pstack is Lauren Tan's set of rigorous agent workflows, such as architect, arena, how, why, and poteto-mode. T3 Code ships it built in and runs it through its orchestration tools across supported providers. Its multi-model steps can mix providers. For example, one $pstack:architect can run an Opus runner and a GPT runner side by side, whichever provider your thread uses.

Turn it off

pstack is on by default. To turn it off, open Settings → Labs → Zack and turn off pstack. The setting applies to one environment. Turning it off removes the pstack: skills from the composer and keeps your model choices.

Choose models once

Run $pstack:setup-pstack in any thread. It reads the providers and models this environment offers, asks for a reasoning budget, shows every pstack role with its models, and saves your choices for every provider's threads on this machine. Run it again to change them, or after you enable or add a provider. Until you run it, pstack matches its defaults to models advertised by this environment's available providers, including models offered through another harness. Models without a reasoning setting keep their own defaults.

List several models on a role to give pstack a pool that picks one per spawn. Arena runners, architect runners, and interrogate reviewers are panels that run every model listed.

Run a workflow

Type $ or / in the composer and pick a pstack: skill, for example $pstack:architect design the cache layer. Start rigorous work with $pstack:poteto-mode; it picks the other skills as it goes. Helper agents appear under the thread as tasks you can open, follow, or stop. Use $pstack:automate-maintainer or $pstack:automate-team to capture a maintainer’s or team’s engineering conventions. Mentioning pstack without a skill, such as “use pstack principles”, points the agent at T3 Code’s copy so it can read the skills, but runs no workflow.

When a model becomes unavailable, pools use their remaining choices; if none remain, they use your thread’s model. Panels keep every seat, using your thread’s model for unavailable entries. The reply explains each unavailable choice with Fallback: …; run $pstack:setup-pstack to change it. If your thread's provider cannot start a child either, pstack reports the unavailable seat and checks the current capabilities before choosing another target.

Good to know

  • pstack runs on the machine the thread runs on. A linked machine needs pstack on and its own $pstack:setup-pstack, since its providers differ.
  • OpenCode sessions using an external Server URL do not expose T3's orchestration tools. Leave that setting blank to let T3 manage the server before running a pstack workflow.
  • Long-running workflows use T3 Code's scheduled tasks and pull request watching. Remove schedules you no longer need from the scheduled tasks list.
  • Steps that need a Cursor-only feature, such as Bugbot setup or a Cursor cloud machine, are skipped with a one-line note in the reply.
  • If you also installed pstack as a Codex or Claude Code plugin, T3 Code turns that plugin off in the sessions it starts while pstack is on, so agents see one copy. Your own Codex and Claude Code sessions outside T3 Code keep it. A Cursor install is only replaced in the composer. Only the T3 Code skills can mix providers.