@casualjim/pi-pstack

pstack skills, Poteto Mode, and bundled agents for pi and omp. Delegates through pi-herdr-agents instead of its own subagent tool.

Packages

Package details

extensionskill

Install @casualjim/pi-pstack from npm and Pi will load the resources declared by the package manifest.

$ pi install npm:@casualjim/pi-pstack
Package
@casualjim/pi-pstack
Version
0.1.31
Published
Nov 20, 2026
Downloads
3,616/mo · 933/wk
Author
casualjim
License
MIT
Types
extension, skill
Size
3.1 MB
Dependencies
0 dependencies · 3 peers
Pi manifest JSON
{
  "skills": [
    "./skills"
  ],
  "extensions": [
    "./extensions/pstack"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

pi-pstack

A pi-mimir fork of pstack, for pi and omp. It keeps every upstream skill except the Cursor-platform-only make-bot-ui (57; setup-pstack is adapted to pi-herdr-agents routing instead of Cursor's model rule), the sticky Poteto Mode, and the bundled poteto-agent and comment-sicko agents — but it delegates through pi-herdr-agents instead of shipping its own subagent tool.

Install

pi install npm:@casualjim/pi-pstack      # pi
omp plugin install @casualjim/pi-pstack  # omp

Local development from this checkout uses pi install ./packages/pi-pstack or omp plugin link ./packages/pi-pstack.

Why this fork exists

Upstream pi-pstack registers a tool named subagent. pi-herdr-agents registers a tool with the same name. Loading both made one shadow the other. This fork drops the subagent tool registration entirely so pi-herdr-agents' subagent is the only one, and pstack's bundled agents become inputs to it.

How delegation works here

pi-herdr-agents resolves roles from bundled, role-pack, global, and project definitions. This extension publishes its bundled agents/ directory as a role pack, registering it on the pi-herdr-subagents:roles:discover:v1 event when the extension loads. The host reads and validates the definitions in place, so nothing is copied into the host agent directory's agents/.

After install, check the roles with subagents_list and delegate the pi-herdr-agents way:

subagent({ name: "Fix retry regression", agent: "poteto-agent", task: "investigate and fix the retry regression, then verify it" })

poteto-agent self-instructs to read the bundled poteto-mode skill in full before working, so the behavior upstream got via prompt injection is preserved without this fork injecting anything.

What's included

  • 57 skills under skills/, matching the upstream inventory minus Cursor-only make-bot-ui (setup-pstack adapted to subagents_write_task_models; prose adapted where the host differs). New in this sync: one-file why, how, reflect, and architect per upstream; worker.md kickoff rules for spawned agents; Orchestrate removed (figure-it-out covers program-scale work); statechart gained SANY error reporting with line/column and parent-border arrow drawing.
  • poteto-agent and comment-sicko agent definitions under agents/, published to pi-herdr-agents as a role pack.
  • Commands: /skill:poteto-mode (sticky Poteto Mode for the session).
  • Tools: none. Task tracking, session recall, and subagent model routing use the host environment (set_tasks family, recall, subagents_write_task_models).

What's removed vs upstream

  • The subagent tool and its child-process runner. Use pi-herdr-agents' subagent instead.
  • The old /setup-pstack, pstack_config, pstack_todo, pstack_sessions, and the pstack/models.json role map in the host agent directory. /skill:setup-pstack now configures pi-herdr-agents' models.tasks/models.agents (subagents_write_task_models) and offers the verification-skill generators. Workflow skills consult that config for the per-call model they pass to subagent. Panel roles take model lists; set those via models.agents per agent or task:<category> routing.

The statechart suite's differential tier (real TLC) needs the TLA+ tools. mise run test fetches tla2tools.jar into .tools/ and exports TLA2TOOLS for you; without it, those tests self-skip. CI fetches the same jar.

Safety

The extension requests confirmation for recognizable shell commands that push, alter pull requests, merge, deploy, mutate infrastructure, or recursively delete files. In non-interactive mode it blocks these commands. This is a guardrail, not a complete shell-security sandbox.

For agent questions, pstack prefers the active ask tool on omp and ask_user before ask_user_question on Pi. Shell approvals use the omp ask dialog when the tool is active and the UI supports it. Otherwise they use ctx.ui.confirm. Pi approval hooks cannot invoke the questionnaire tools and keep native confirmation. Cancelling or timing out does not authorize the command.

To run without approval prompts, set confirmExternalActions to false in the pstack config file: $PI_CODING_AGENT_DIR/pstack/config.json when that env var is set, otherwise ~/.omp/agent/pstack/config.json under omp or ~/.pi/agent/pstack/config.json under pi (same resolution as pi-headroom's settings). Restart the host after a config change. A missing, malformed, or non-boolean config keeps the guard enabled. With the guard disabled, agents are told to run recognized external commands without asking and still pause for irreversible writes (force-pushes to shared branches, deployments, data deletion, customer messages).

License and provenance

Derived from Cursor's pstack, licensed under MIT. See LICENSE.