pstack-pi

A pi port of pstack, Lauren Tan's Cursor plugin. Original content is MIT-licensed (see LICENSE); this port keeps its structure and wording and adapts Cursor-specific mechanics to pi.

Acknowledgment: all skills, workflows, and prose originate from Lauren Tan's pstack. This repository only ports them to pi.

Quick start

pi install npm:@0xrsydn/pstack-pi

What ships

• skills/ — all 45 original workflow and principle skills (poteto-mode, swarm, arena, interrogate, how, why, reflect, the principle-* set, etc.), rewritten where they referenced Cursor-specific mechanics.
• agents/ — subagent definitions discovered by the bundled extension: poteto-agent, comment-sicko, and general-purpose (replaces Cursor's built-in generalPurpose task agent).
• extensions/subagent/ — pi's example subagent tool, extended for pstack:
  • Per-task model override (with pstack's inherit-parent / auto aliases), so one parallel call can mix models.
  • Per-task readonly flag (restricts to read,grep,find,ls).
  • Discovers the bundled agents/ directory from inside the package. User (~/.pi/agent/agents/) and project (.pi/agents/) definitions win over bundled ones.
  • Single, parallel (tasks[]), and chain modes are unchanged otherwise.

Install

pi install npm:@0xrsydn/pstack-pi        # from npm (recommended, version-pinned)
pi install git:github.com/0xrsydn/pstack-pi   # or from git

After installing, confirm resources load with pi list.

Model configuration

Run the setup-pstack skill ("/skill:setup-pstack", or ask the agent to configure pstack models). It:

1. Detects your models with pi --list-models.
2. Asks which model each role uses (code delegates, judgment/prose, review panels, swarm workers).
3. Writes ~/.pi/agent/pstack-models.md.

The parent agent reads that file before spawning and passes each role's value as the model field of its subagent task entry. Delete a line to fall back to that skill's inline default; delete the file to fall back everywhere. There is no requirement to run setup — defaults described inline in the skills apply, but they reference placeholder model names from the upstream plugin until you write real ones.

Adaptations from the original

Cursor mechanic	pi equivalent
Task tool, subagent_type	subagent tool, agent field
generalPurpose built-in agent	bundled general-purpose agent
Task-level model slug	task-level model field (provider/model-id form)
run_in_background, cloud workers	parallel tasks[] calls run concurrently with isolated contexts; all local
environment: "cloud" / cloud VMs	removed; lanes run in worktrees on this machine
~/.cursor/rules/pstack-models.mdc	~/.pi/agent/pstack-models.md (plain markdown, read before spawns)
~/.cursor/skills/ paths	~/.pi/agent/skills/; project .cursor/skills/ → .pi/skills/
Session transcripts under ~/.cursor/projects/<slug>/agent-transcripts/	~/.pi/agent/sessions/<dashed-repo-path>/*.jsonl
Cursor's built-in create-skill / babysit skills	generic authoring guidance; pstack's own playbooks cover these requests
deslop, control-ui/control-cli (cursor-team-kit) references	degraded gracefully: steps drop the plugin dependency or drive the surface directly
Bugbot review-bot comments	generalized to any bot reviewer comments
Slash commands like /bro, /why	/skill:bro, /skill:why (enable skill commands in settings)

Not ported:

• Cursor Cloud Automations (automations/benny/) and the grokbot automation-webhook skill — no pi equivalent background-runner target.
• docs/guide/ — read it upstream; most prose still applies apart from setup details.
• Skills that rely on MCP servers (for example parts of why) expect you to have equivalent data sources reachable through other tools.

Requirements

• The bundled extension shells out to the pi binary on PATH (subagent tool).
• Panel/swarm fan-out works best with more than one provider configured.