prime-pstack

pstack skills ported to prime-agent. Rigorous agent workflows: fewer, higher-quality lines, verifiable work, and subagents you can parallelize with confidence.

This is a port, not the original: Cursor-specific mechanics (Task subagents, ~/.cursor/rules, AskQuestion, Cursor model slugs) are translated to prime-agent equivalents (rlm children, agent_message, references/models.md, session logs). The methodology text is preserved verbatim wherever it did not reference Cursor.

Install

Via npm (recommended)

prime-agent package install npm:prime-pstack                 # all your sessions
prime-agent package install --local npm:prime-pstack         # one project only

prime-agent installs the package, registers it under packages in ~/.prime/agent/settings.json (or .prime/settings.json with --local), and loads its 45 skills. Update with prime-agent package update npm:prime-pstack, remove with prime-agent package remove npm:prime-pstack.

New skills load on session start; in an interactive session use /reload.

Note: a plain npm install prime-pstack into a project's node_modules does not register the package with prime-agent — the package install command is what adds the skills/ directory to your settings.

From the GitHub repo

git clone https://github.com/eryzerz/prime-pstack ~/.prime/agent/skills/prime-pstack

The loader scans nested skill directories, so cloning the repo into your skills dir works as-is. Per-project: clone or symlink skills/ into .prime/agent/skills/ at your project root.

Start here

• /skill:poteto-mode — the hub skill; routes a task to one of 23 playbooks, wraps code in the principle skills. This is the one to use daily.
• /skill:setup-pstack — configure per-role models by editing each skill's references/models.md (default: children inherit the parent model). You can also edit those files by hand — one role: <model selector> line per role, resolved via await rlm.find_models(...).

What was ported

All 45 skills (markdown + references + playbooks) plus the poteto-mode helper scripts, with Cursor mechanics translated:

| pstack (Cursor) | prime-agent | |---|---| | Task calls, subagent_type, run_in_background | await rlm(prompt, name=...) children, spawned in parallel | | model: <cursor slug> defaults | references/models.md per skill; omit model to inherit parent | | findings in Task response body | await agent_message.send(..., receiver_role='parent') | | ~/.cursor/rules/pstack-models.mdc | references/models.md in each skill that reads model config | | agent-transcripts/ | ~/.prime/agent/sessions/*.jsonl (child rollouts in session_dir) | | MCP discovery via Cursor env | MCP servers configured via prime CLI; children keep MCP access | | Cursor built-in create-skill | prime-agent skill-creator | | cursor-team-kit (deslop, control-cli/ui) | not ported — hooks removed |

Not ported (deliberately)

• automations/benny (cron issue triage → use prime-agent heartbeats)
• agents/* (Cursor agent definitions) — except comment-sicko, folded into no-comments/references/comment-sicko.md
• docs/guide/ images

The poteto-mode helper scripts are ported (scripts/): watch-pr (GitHub PR watcher, needs bun), orch (orchestrate store, needs bun), check-plan.mjs (plan validator, needs node), worktree-audit.sh (pure bash). All are runtime-agnostic aside from their runtimes; worktree-audit.sh and check-plan.mjs were translated for prime-agent paths.

Contributing

PORTING.md in this repo is the translation table used for the port. When you fix or extend a skill, keep the conventions there so the port stays coherent.

License

MIT. Port of cursor/plugins pstack by Lauren Tan, which remains MIT-licensed; see LICENSE.