Grok's 'Potato Sister' Open-Sources pstack: Agent Harness for 1150 PRs in 4 Days

Background: The Grok Bot Engineer's Open-Source Agent Harness

Grok Bot engineering lead Lauren Tan (known as poteto or "Potato Sister") recently open-sourced pstack, an agent harness designed for Cursor and Claude Code. The release follows a high-profile 4-day sprint where she merged roughly 1,150 pull requests to ship a playable survivor game, using pstack as the underlying infrastructure.

What Is pstack?

pstack

(hosted at cursor/plugins/tree/main/pstack) is a battle-tested framework that prioritizes depth before breadth: "if you want to go fast, go deep first." It includes:

58 Skills (34 functional + 24 principle-based)

23 Playbooks covering research, bug fixes, new features, performance analysis, metric optimization, and PR merging

2 Sub-agents : poteto-agent and comment-sicko A harness/pstack/ adapter layer for zero-config use in both Cursor and Claude Code

The core philosophy: embed specifications directly into file suffixes (e.g., .hot.ts), enforce hard lint rules, model state machines with TLA+ to catch concurrency edge cases, and keep audio/video fully code-driven so agents can verify themselves in a closed loop.

Core Workflow: Single Entry Point /poteto-mode

Instead of writing step-by-step prompt instructions ("read file A, call B, run C"), users invoke a single entry point:

/poteto-mode 导出接口重试时会写入重复行，先复现再修，拿出前后行对比证据

Given a goal, acceptance criteria, and required evidence, /poteto-mode automatically:

Restates the problem and minimal change scope in plain language for immediate correction.

Selects one of the 23 playbooks, copies it verbatim into a todo list, and marks skipped steps with skip: reason.

Can be locked as a persistent Custom Mode in Cursor via Option+Enter (macOS) or Alt+Enter (Windows).

Four High-Leverage Tool Groups

1. Understand Before You Touch: /how + /why + /teach + /recall

/how

: Read-only decomposition of module call chains and core types. /why: Parallel search of git history, issues, team chats, and Sentry errors to explain original design intent. /teach: Step-by-step mechanism explanation, including self-proof of why alternatives were rejected. /recall: Retrieves context from prior conversations.

2. No First-Draft Designs: /architect + /arena + /interrogate

/architect

+ /arena: Define caller-side contracts first; for tough decisions, spin up multiple models in independent worktrees to produce competing designs, then have a judge model pick the best base and graft highlights. /interrogate: Adversarial review — different models act as critics, each required to attach a real failing command and output; pure speculation is discarded.

3. Overnight Runs with Audit Trails: autonomous-run + /show-me-your-work

Paired with /loop for overnight iteration: each round makes the smallest possible change, validates with actual run results, commits on progress, discards on no gain, and logs every decision to a local decisions.tsv.

Morning review via /show-me-your-work dispatches a heterogeneous model to replay the night's trajectory and flag high-risk decisions; /ketchup or /ketchup-video can generate a quick slideshow or subtitled video summary.

4. One-Liner Corrections & TLA+ Safety Net

24 engineering principles carry fixed English phrase names for instant correction:

Stacking compatibility layers → use subtract before you add Relying only on compilation → apply prove it works For concurrency retries or state-machine corner cases, pstack uses mise to provision Java and TLC, then /create-formal-verification-skill and /statechart to exhaustively enumerate interleavings and visualize violating paths.

Scaling to Hundreds of PRs/Day: Heterogeneous Roles + Narrow PR Stacks + Comment Sicko

The system solves "how to run multiple agents in parallel without them colliding" through three hard mechanisms:

Heterogeneous model assignment — Pain point: single model is expensive, slow, or has overlapping blind spots. pstack 's hard rule: /setup-pstack defaults assign Grok to write features/bugs/run swarm; high-stakes architecture and cross-review go to Opus 5.5 (or inherit-parent to follow the primary model).

Ultra-narrow PR stacks + independent judges — Pain point: concurrent agents modifying the same repo likely collide. pstack 's hard rule: opening-a-pr enforces "5 narrow PRs beat 1 fat PR"; shipping forbids the writing agent from self-review — every PR gets a fresh agent that actually runs verification, merged bottom-up.

Comment Sicko — Pain point: writing agents love to defend their own verbose comments. pstack 's hard rule: /no-comments assigns a read-only reviewer that first checks /how and /why for external constraints; only comments that can be encoded as types or lint rules are removed.

Core Philosophy: Encode Lessons in Structure

The system's most valuable sentence lives in Encode Lessons in Structure:

What can be fixed structurally must only be fixed structurally — the written instruction is merely a symptom.

Typical agent-taming loops add warnings to CLAUDE.md after each mistake, bloating the file to hundreds of lines that still get ignored. /correct documentation states bluntly: "Nothing fails when an agent skips them."

Tan organizes a four-layer defense, investing effort top-down:

Code structure & architecture — single source of truth, encapsulation; wrong patterns become inexpressible. Highest priority, zero overhead.

Types & static checks — .hot.ts bans assignment, size budgets, errors include fix hints. Hard interception, self-healing on misreads.

Tests & state modeling — assert external behavior; TLA+ exhausts 17 concurrency rules to catch races.

Docs & human review — CLAUDE.md only for subjective judgment; skipping produces no error. Lowest priority.

Only the top two layers can hard-block; lint catches violations instantly, while agents skipping docs face zero consequences.

/correct : Five-Step Closed Loop

The /correct skill automates the loop:

Mine history for recurring error patterns.

Classify and deduplicate.

Propose structural fixes (architecture, types, tests).

Verify the check would have caught the original failure.

Commit the fix as a lint rule or type constraint.

Two disciplines govern it:

Same error twice before a rule — one-off flukes don't add complexity.

New check must fail on real historical error first — a check that never turned red proves nothing; formal verification artifacts must also error on a deliberately broken variant before acceptance.

5-Minute Onboarding

1. Install & Run /setup-pstack

In Cursor: /add-plugin pstack In Claude Code (official adapter at harness/pstack/, commands prefixed with /pstack:, model rules written to ~/.claude/rules/pstack-models.md):

/plugin marketplace add cursor/plugins
/plugin install pstack@cursor-plugins
/pstack:setup-pstack

The wizard offers four budget tiers ( unlimited, large, medium, small); single-model accounts set the role to inherit-parent to reuse the current session model.

When prompted to generate /create-verification-skill, accept — it auto-discovers project start commands, test accounts, and entry points, writing a .cursor/skills/verify-<app>/ skill into the repo. The official guide ranks this as the #1 value driver: agents that can drive the software and record video proof truly free human hands.

2. Run /correct First to Codify Frequent Pitfalls

Before large tasks, run on existing codebases: /pstack:correct It scans commits, reverts, and review comments, converting twice-hit traps into architectural boundaries or static checks. You can also adopt Tan's .hot.ts suffix convention.

3. Dislike the Default Style? /automate-me Clones Your Workflow

Run /pstack:automate-me to concurrently scan recent workspace conversations, extract communication, verification, and PR-splitting habits that appear at least twice, strip AI-flavored phrasing, and open a PR with a personalized <your-name>-mode entry skill.