Skip to content
/

Cursor Rules Pack v2

cursor-rules-pack-v2 · cursor/plugins · ★ 853 · last commit 2026-05-26

Official Cursor plugin marketplace providing CI workflows, engineering principles, parallel cloud-agent orchestration, transcript-based memory, and iterative completion loops.

Best whenThe planner-worker-verifier separation is load-bearing: planners that code drift; a script with a JSON state file keeps the tree converging.
Skip ifPlanners editing files instead of publishing tasks, Agents sharing state between siblings
vs seeds
Primitive shape 71 total
Skills 64 Subagents 4 Hooks 3
00

Summary

Cursor Rules Pack v2 (cursor/plugins) — Summary

cursor/plugins is Cursor's official plugin marketplace repository, shipping 12 plugins authored by the Cursor team and one community plugin (pstack by poteto). Each plugin is a standalone directory containing skills (SKILL.md files with YAML frontmatter), agents (subagent definition files), rules (.mdc files), hooks (hooks.json), and scripts, all bundled under a .cursor-plugin/plugin.json manifest. Plugins install with /add-plugin <name> from within Cursor. The 12 plugins span: cursor-team-kit (18 CI/dev workflow skills), pstack (32 engineering principle skills + 14 playbooks), orchestrate (parallel cloud-agent spawning via Cursor SDK), continual-learning (transcript-based AGENTS.md memory), ralph-loop (self-referential iterative loops), agent-compatibility (repo audit), cli-for-agent, pr-review-canvas, docs-canvas, cursor-sdk, create-plugin, and teaching. This is the most architecturally mature framework in this batch: it has a first-party SDK for spawning cloud agents (orchestrate), a memory system backed by transcript mining (continual-learning), a Stop hook for automated memory updates, and a defined plugin schema. It differs from superpowers (single plugin, skills-only behavioral framework) by shipping a multi-plugin marketplace with heterogeneous component types including cloud-agent orchestration — the closest analog in the seed corpus is a hybrid of superpowers (skills architecture) and claude-flow (parallel agent spawning), but implemented natively inside Cursor's own runtime.

01

Overview

Cursor Rules Pack v2 (cursor/plugins) — Overview

Origin

Officially created and maintained by Cursor (the company, plugins@cursor.com). One plugin (pstack) is a community contribution by poteto (Lauren Tan, React Core Team, formerly at Meta/Netflix). The repository was created in January 2026 and is actively maintained (pushed May 26, 2026). 853 stars, 90 forks.

Philosophy

From the orchestrate README:

"Workers don't talk to each other; they talk up through structured handoffs. The spawn, wait, and handoff loop lives in scripts/cli.ts. The planner writes plan.json, the script executes it, and the planner reads handoffs to decide what's next. Long-running agent loops drift; a script with a JSON state file keeps its footing."

From the pstack README:

"there's a growing sense that ai writes too much slop code. i agree. i don't want to ship like a team of twenty slop artists. throughput without quality is not a goal i aspire to. if you want to go fast, go deep first."

"pstack gives you fearless parallelism. when you can go deep on one agent and trust it to write good, verifiable code, you can truly parallelize with confidence."

From the ralph-loop README:

"Two hooks drive the loop. An afterAgentResponse hook watches each response for a <promise> tag matching the completion phrase. A stop hook fires when Cursor finishes a turn. If the promise hasn't been detected and the iteration limit hasn't been reached, the stop hook sends the original prompt back as a followup_message, starting the next iteration."

Key Design Opinions

  1. Plugin modularity — each plugin is independently installable; users compose their own tool set
  2. Cloud-agent parallelismorchestrate is the canonical pattern for parallelizing work across Cursor cloud agents using the Cursor SDK
  3. Quality over throughput — pstack's philosophy explicitly rejects "slop code" generated for velocity
  4. Memory from evidencecontinual-learning mines transcript history to update AGENTS.md rather than asking the user to curate it
  5. Self-improving loopsralph-loop implements the Ralph Wiggum technique for iterative convergence on a goal

Marketplace Structure

{
  "name": "cursor-plugins",
  "metadata": {
    "description": "Official Cursor plugin marketplace: developer tools, framework rules, MCP integrations, and agent skills"
  },
  "plugins": [12 entries]
}

The root .cursor-plugin/marketplace.json lists all plugins; each plugin has its own .cursor-plugin/plugin.json manifest with skills, agents, rules, hooks pointers.

02

Architecture

Cursor Rules Pack v2 (cursor/plugins) — Architecture

Distribution Type

claude-plugin (Cursor plugin format) — installed via Cursor's /add-plugin <name> command.

Install

# Within Cursor chat:
/add-plugin cursor-team-kit
/add-plugin pstack
/add-plugin orchestrate
/add-plugin continual-learning
/add-plugin ralph-loop
# etc.

Plugins are fetched from the marketplace and installed into the Cursor project or global config.

Repository Structure

cursor/plugins/
├── .cursor-plugin/
│   └── marketplace.json          # Marketplace manifest listing all plugins
├── .github/
│   └── workflows/
├── schemas/                       # Plugin JSON schemas
├── scripts/                       # Repo maintenance scripts
│
├── cursor-team-kit/               # Plugin 1 — Cursor internal workflows
│   ├── .cursor-plugin/plugin.json
│   ├── skills/                    # 18 skills (SKILL.md each)
│   ├── agents/                    # 2 subagents
│   ├── rules/                     # 2 .mdc rules
│   └── README.md
│
├── pstack/                        # Plugin 2 — Engineering principles
│   ├── .cursor-plugin/plugin.json
│   ├── skills/                    # 32 skills (poteto-mode + 14 principles + utilities)
│   └── agents/                    # subagents
│
├── orchestrate/                   # Plugin 3 — Parallel cloud agent spawning
│   ├── .cursor-plugin/plugin.json
│   ├── skills/
│   │   └── orchestrate/SKILL.md
│   └── scripts/
│       └── cli.ts                 # Bun script for Cursor SDK orchestration
│
├── continual-learning/            # Plugin 4 — Transcript-based memory
│   ├── .cursor-plugin/plugin.json
│   ├── skills/
│   │   └── continual-learning/SKILL.md
│   ├── agents/
│   │   └── agents-memory-updater.md
│   └── hooks/
│       ├── hooks.json             # Stop hook
│       └── continual-learning-stop.ts
│
├── ralph-loop/                    # Plugin 5 — Iterative completion loop
│   ├── .cursor-plugin/plugin.json
│   ├── skills/                    # 3 skills (ralph-loop, cancel-ralph, ralph-loop-help)
│   └── hooks/                     # afterAgentResponse + stop hooks
│
├── agent-compatibility/           # Plugin 6 — Repo audit
├── cli-for-agent/                 # Plugin 7 — CLI design patterns
├── create-plugin/                 # Plugin 8 — Plugin scaffolding
├── cursor-sdk/                    # Plugin 9 — Cursor SDK patterns
├── docs-canvas/                   # Plugin 10 — Documentation canvas
├── pr-review-canvas/              # Plugin 11 — PR review canvas
└── teaching/                      # Plugin 12 — Skill mapping + practice plans

Required Runtime

  • Cursor (latest version with plugin support)
  • Bun (for orchestrate scripts)
  • CURSOR_API_KEY (for orchestrate)
  • SLACK_BOT_TOKEN (optional, for orchestrate Slack integration)

Target AI Tools

  • Cursor exclusively (/add-plugin is Cursor-native)

Plugin Manifest Format

{
  "name": "cursor-team-kit",
  "displayName": "Cursor Team Kit",
  "version": "1.1.0",
  "description": "...",
  "author": {"name": "Cursor", "email": "plugins@cursor.com"},
  "skills": "./skills/",
  "agents": "./agents/",
  "rules": "./rules/"
}
03

Components

Cursor Rules Pack v2 (cursor/plugins) — Components

Plugin Inventory (12 plugins)

cursor-team-kit (18 skills, 2 agents, 2 rules)

Skill Purpose
loop-on-ci Watch CI runs and iterate on failures until checks pass
review-and-ship Run structured review, commit, open PR
pr-review-canvas Generate interactive HTML PR walkthrough
verify-this Prove/disprove claims with baseline/treatment artifacts
control-cli Build/adapt harness to drive interactive CLIs or TUIs
control-ui Build/adapt browser/CDP harness for web or Electron UIs
make-pr-easy-to-review Clean noisy PR history, improve descriptions
run-smoke-tests Run Playwright smoke tests and triage failures
fix-ci Find failing CI jobs, inspect logs, apply fixes
new-branch-and-pr Create fresh branch, complete work, open PR
get-pr-comments Fetch and summarize review comments
check-compiler-errors Run compile/type-check commands
what-did-i-get-done Summarize authored commits over time period
weekly-review Generate weekly recap with highlights
fix-merge-conflicts Resolve merge conflicts, validate build/tests
deslop Remove AI-generated slop, clean code style
workflow-from-chats Extract working preferences from chats into skills/rules
thermo-nuclear-code-quality-review Unusually strict maintainability review

Agents: ci-watcher, thermo-nuclear-code-quality-review
Rules: typescript-exhaustive-switch.mdc, no-inline-imports.mdc

pstack (32 skills)

Primary skills:

Skill Purpose
poteto-mode Main entry point: matches task to 14 playbooks, routes to sub-skills
how Walkthrough of how a subsystem works
why Why something was built; multi-MCP parallel evidence gathering
architect Settle types/module shape before implementing (parallel design exploration)
arena N parallel attempts at same thing; extract best parts
interrogate 4-model adversarial review before shipping
automate-me Mine transcripts, draft custom -mode skill
reflect After long task: capture recipe as skill edit
tdd Bug fix with failing test first
typescript-best-practices TypeScript type-system discipline
figure-it-out Design rigorous playbook when no preset fits
show-me-your-work Log decisions to reviewable TSV
unslop Remove AI tells from writing

Principle skills (19): principle-laziness-protocol, principle-foundational-thinking, principle-redesign-from-first-principles, principle-subtract-before-you-add, principle-minimize-reader-load, principle-outcome-oriented-execution, principle-experience-first, principle-exhaust-the-design-space, principle-build-the-lever, principle-boundary-discipline, principle-type-system-discipline, principle-make-operations-idempotent, principle-migrate-callers-then-delete-legacy-apis, principle-separate-before-serializing-shared-state, principle-prove-it-works, principle-fix-root-causes, principle-guard-the-context-window, principle-never-block-on-the-human, principle-encode-lessons-in-structure

orchestrate (1 skill, scripts)

Component Purpose
orchestrate/SKILL.md Plan + spawn parallel Cursor cloud agent workers/verifiers/subplanners
scripts/cli.ts Bun CLI for kickoff, run, spawn, respawn, kill, tail, comment, andon subcommands

continual-learning (1 skill, 1 agent, 1 hook)

Component Purpose
continual-learning/SKILL.md Orchestrates memory update: delegates to agents-memory-updater
agents/agents-memory-updater.md Mines transcripts, updates AGENTS.md, maintains incremental index
hooks/hooks.json (stop event) Triggers continual-learning-stop.ts on Cursor stop

ralph-loop (3 skills, hooks)

Component Purpose
ralph-loop Start self-referential loop with prompt + options
cancel-ralph Remove state file, stop loop
ralph-loop-help Explain technique and usage

Other Plugins (6)

Plugin Skills Purpose
agent-compatibility 1 CLI-backed repo compatibility scans
cli-for-agent 1 CLI design patterns for agents
create-plugin 2 Scaffold and validate new plugins
cursor-sdk 1 Cursor SDK patterns (auth, streaming, MCP)
docs-canvas 1 Render docs as navigable Canvas
pr-review-canvas 1 Render PR diffs as Canvas
teaching 2 Skill mapping + practice plans

Total Primitive Count

  • Plugins: 12
  • Skills (total across all plugins): ~64
  • Agents: ~4
  • Rules (.mdc): 2
  • Hooks: 1 (continual-learning stop hook) + ralph-loop hooks
  • Scripts: 1 (orchestrate/scripts/cli.ts — Bun/TypeScript)
05

Prompts

Cursor Rules Pack v2 (cursor/plugins) — Prompts

Excerpt 1: orchestrate/SKILL.md — Role Definition and Core Principles

---
name: orchestrate
description: Use only when the user explicitly types `/orchestrate <goal>` to decompose a large task, spawn a tree of parallel cloud-agent workers/subplanners/verifiers via the Cursor SDK, and collect structured handoffs; do not invoke autonomously.
disable-model-invocation: true
---

# Orchestrate

## Core principles

These rules make the tree self-converging without global coordination.

1. **Planners own scopes and publish tasks. They do no coding.** Writing `plan.json`, reading handoffs, and deciding what's next are planner work. Editing files, running `git merge`, and fixing conflicts inline are not. If a planner feels the urge to code, it publishes a task for a worker instead.
2. **Planners don't know who picks up their tasks.** The script routes each task to a cloud agent. The planner's mental model stays at the task level.
3. **Workers are isolated.** One task, one clone of the repo, no channel to any other agent. One handoff when done.
4. **Subplanners are recursive planners.** A planner publishes a "subplan this slice" task; the subplanner fully owns that slice and hands back an aggregated handoff.
5. **Continuous motion via handoffs.** A planner that thought it was done can receive a late handoff and replan. No "finished" state until the planner decides to stop publishing.
6. **Propagation, not synchronization.** No cross-talk between siblings. No shared state between levels. Each level sees only its children's handoffs.

Prompting technique: disable-model-invocation: true in YAML frontmatter prevents autonomous activation — this skill only fires on explicit /orchestrate. The principles are numbered and bold-named. The role-separation principle ("planners don't code, workers don't plan") is enforced via instruction rather than technical constraint.


Excerpt 2: continual-learning/agents/agents-memory-updater.md

---
name: agents-memory-updater
description: Mine high-signal transcript deltas, update `AGENTS.md`, and keep the incremental transcript index in sync.
model: inherit
---

# AGENTS.md memory updater

## Workflow

1. Read existing `AGENTS.md` first. If it does not exist, create it with only:
   - `## Learned User Preferences`
   - `## Learned Workspace Facts`
2. Load the incremental index if present.
3. Inspect only transcript files under `~/.cursor/projects/<workspace-slug>/agent-transcripts/` that are new or have newer mtimes than the index.
4. Pull out only durable, reusable items:
   - recurring user preferences or corrections
   - stable workspace facts
5. Update `AGENTS.md` carefully:
   - update matching bullets in place
   - add only net-new bullets
   - deduplicate semantically similar bullets
   - keep each learned section to at most 12 bullets

Prompting technique: Strict minimalism guardrails ("keep each learned section to at most 12 bullets") prevent runaway growth. The incremental index (mtime-based) ensures idempotent runs — the agent only processes what's new. The model: inherit frontmatter defers model selection to the parent context.


Excerpt 3: pstack/skills/poteto-mode/SKILL.md — Subagent Dispatch Pattern

## Subagents

**Use `subagent_type: "poteto-agent"` for any subagent you spawn directly inside a playbook step** (code-writing delegates, ad-hoc helpers). `/poteto-mode` and `subagent_type: "poteto-agent"` route through the same wrapper. Routed workflow skills (`how`, `why`, `interrogate`, `reflect`) configure their own `subagent_type` for diverse-model review and exploration; respect what the skill prescribes rather than overriding it to `poteto-agent`.

**Defaults for every `Task` call.** `run_in_background: true`, agent mode (readonly strips MCP), file pointers not inlined context, and explicit model (`composer-2.5-fast` for code, `claude-opus-4-7-thinking-xhigh` for prose and judgment).

Prompting technique: Multi-model routing embedded in the subagent spawn spec — composer-2.5-fast for code tasks, claude-opus-4-7-thinking-xhigh for prose/judgment. This is one of the few frameworks in the corpus with explicit per-task model assignment. The run_in_background: true default enables true parallel execution.


Excerpt 4: ralph-loop — Completion Promise Pattern

Build a REST API for todos.

When complete:
- All CRUD endpoints working
- Input validation in place
- Tests passing (coverage > 80%)
- Output: <promise>COMPLETE</promise>

Prompting technique: XML-tagged <promise> marker in the prompt body that hooks monitor for. The hook system converts this into a programmatic completion signal. The pattern separates verifiable completion criteria from prose description — the hook watches for the tag, not the quality of the code.

09

Uniqueness

Cursor Rules Pack v2 (cursor/plugins) — Uniqueness

Differs From Seeds

This is the most architecturally complete framework in this batch and rivals the top-tier seeds. The closest seed analogs are:

  • superpowers (skills-only behavioral framework): cursor/plugins shares the SKILL.md format but goes far beyond — it has hooks, agents, multi-plugin composition, a CLI-backed orchestration layer, and a Stop-hook memory system.
  • claude-flow (MCP-anchored, parallel agent spawning, SQLite memory): orchestrate achieves similar parallel fan-out but through Cursor's cloud agent SDK rather than an MCP server. Where claude-flow has 305 MCP tools, orchestrate has a purpose-built CLI and a hierarchical planner/worker/verifier model with explicit handoff contracts.

The continual-learning plugin is architecturally novel vs. all seeds: it mines actual agent transcript files to extract durable preferences, uses an mtime-based incremental index to avoid reprocessing, and triggers automatically via a Stop hook. ccmemory (seed) uses Neo4j + vector storage for memory; continual-learning uses file-based AGENTS.md with 12-bullet compaction limits.

The pstack plugin's multi-model dispatch (composer-2.5-fast for code, claude-opus-4-7-thinking-xhigh for prose) is one of the most explicit model-role mappings in the entire corpus. Only taskmaster-ai approaches this, and its model routing is via config file, not embedded in skill frontmatter.

The ralph-loop <promise> tag pattern — where a prompt's output contains an XML tag that a hook monitors programmatically — is unique in the corpus. It converts natural-language completion criteria into a machine-detectable signal.

Positioning

  • This is Cursor's own official distribution of agent capabilities — the canonical example of what the Cursor plugin format can do
  • The multi-plugin marketplace architecture (each plugin independently installable, composable) is unique among all analyzed frameworks
  • pstack is the only community-contributed plugin; its principles-as-skills architecture (19 named engineering principles each as a leaf skill) is a novel granularity

Observable Failure Modes

  1. Cursor-only — entirely non-portable; no value for Roo Code, Claude Code, or other tools
  2. orchestrate requires CURSOR_API_KEY — no fallback for non-API-key users
  3. Bun dependency for orchestrate — not universally available
  4. continual-learning transcript path is cursor-version-sensitive~/.cursor/projects/<workspace-slug>/agent-transcripts/ may change between Cursor versions
  5. pstack model names may become stalecomposer-2.5-fast, claude-opus-4-7-thinking-xhigh hardcoded in SKILL.md; will break when models are retired

Cross-References

  • ralph-loop credits Geoffrey Huntley's "Ralph Wiggum technique" and mikeyobrien/ralph-orchestrator
  • orchestrate references cursor-sdk plugin as required reading
  • pstack's deslop skill relies on cursor-team-kit's deslop skill across plugin boundaries
04

Workflow

Cursor Rules Pack v2 (cursor/plugins) — Workflow

Plugin Installation Workflow

Step Action Artifact
1 /add-plugin cursor-team-kit in Cursor Plugin installed to project
2 Plugin skills appear in skill picker Available skills
3 User invokes skill: /review-and-ship Skill executes

pstack Workflow (Primary Entry Point)

  1. User types /poteto-mode (or describes task)
  2. Skill opens a todolist; first item: read Principles section
  3. Matches task to one of 14 playbooks (investigation, bug-fix, perf, feature, refactoring, etc.)
  4. Copies playbook steps verbatim into execution plan
  5. Routes to leaf skills as steps fire (how, architect, interrogate, deslop, etc.)
  6. Before commit: runs /deslop from cursor-team-kit
  7. After PR opened: uses Cursor's built-in babysit skill

orchestrate Workflow (Multi-Agent)

Step Actor Output
1 Dispatcher (local IDE) runs /orchestrate <goal> Kicks off cloud root planner
2 Root planner Writes plan.json, publishes tasks
3 cli.ts kickoff Spawns cloud agents via Cursor SDK
4 Worker agents Execute tasks, write handoff files
5 Planner Reads handoffs, replans, publishes more tasks
6 Verifier agents Verify acceptance criteria
7 Planner Decides when done, opens optional PR

Node types: Planner, Subplanner, Worker, Verifier, Git
CLI subcommands: kickoff, run, spawn, respawn, kill, tail, comment, andon

continual-learning Workflow

Step Actor Output
1 User session ends (Cursor Stop event) Hook fires
2 continual-learning-stop.ts Invokes agents-memory-updater subagent
3 agents-memory-updater Reads ~/.cursor/projects/<workspace>/agent-transcripts/
4 Extracts high-signal items Durable preferences + workspace facts
5 Updates AGENTS.md 2 sections: Learned User Preferences + Learned Workspace Facts
6 Updates incremental index .cursor/hooks/state/continual-learning-index.json

ralph-loop Workflow

  1. User provides prompt + --completion-promise <text> + --max-iterations N
  2. Skill starts Cursor with the prompt
  3. After each response: afterAgentResponse hook checks for <promise> tag
  4. stop hook fires: if promise not seen + iterations < max, sends same prompt as followup_message
  5. Loop continues until promise detected or max iterations reached

Approval Gates Per Workflow

Workflow Gate
pstack User confirmation for irreversible writes (force-push, deploy, data delete)
orchestrate Dispatcher kickoff (manual); Andon (🚨 reaction on Slack) pauses run
ralph-loop Max iterations limit (user-specified)
continual-learning None (fully automated on Stop)
06

Memory Context

Cursor Rules Pack v2 (cursor/plugins) — Memory & Context

Primary Memory System: continual-learning

The continual-learning plugin provides automatic memory management:

Input: Cursor agent transcript files at ~/.cursor/projects/<workspace-slug>/agent-transcripts/
Processing: Incremental index tracks which files have been processed (mtime-based)
Output: AGENTS.md updated with 2 sections:

  • ## Learned User Preferences (recurring corrections, style preferences)
  • ## Learned Workspace Facts (stable workspace-specific facts)

Limits: Each section capped at 12 bullets to prevent context bloat.
Index: .cursor/hooks/state/continual-learning-index.json tracks processed transcripts.

Memory Activation

The continual-learning Stop hook fires automatically when a Cursor session ends:

{
  "hooks": {
    "stop": [
      {
        "command": "bun run ${CURSOR_PLUGIN_ROOT}/hooks/continual-learning-stop.ts"
      }
    ]
  }
}

orchestrate Memory

The orchestrate skill uses file-based state:

  • plan.json — the current task decomposition tree
  • handoffs/ — worker/verifier result files
  • .cursor/hooks/state/ — hook state files

Git serves as the audit log (all workers commit to branches).

pstack Memory

show-me-your-work skill logs decisions to a TSV file (commitable artifact). The automate-me skill mines transcripts to generate a custom -mode skill. The reflect skill captures task recipes as skill file edits.

Cross-Session Handoff

continual-learning provides automated cross-session memory via AGENTS.md updates on every Stop event.

orchestrate provides cross-session continuity via plan.json + handoffs/ directory — the script can resume a tree from disk.

Memory Type Summary

Plugin Memory Type Storage Persistence
continual-learning File-based (AGENTS.md + index.json) ~/.cursor/projects/ Global
orchestrate File-based (plan.json + handoffs/) Project Project
pstack (show-me-your-work) File-based (TSV) Project Project

Compaction

continual-learning enforces 12-bullet limits per section — this IS a compaction mechanism. Old low-signal bullets are replaced or deduplicated when new higher-signal items arrive.

07

Orchestration

Cursor Rules Pack v2 (cursor/plugins) — Orchestration

Multi-Agent Pattern

The orchestrate plugin implements hierarchical parallel-fan-out orchestration:

Dispatcher (local IDE) → Root Planner (cloud) → Workers + Subplanners + Verifiers (cloud)

Node types:

Node Runs loop? Scope Output
Planner yes Entire user goal User-facing message + optional PR
Subplanner yes One slice of parent scope Handoff to parent
Worker no One concrete task Handoff to spawning planner
Verifier no Acceptance criteria for one target Verdict handoff
Git n/a Shared medium Branches + handoffs/

Pattern: hierarchical (planner tree) + parallel-fan-out (workers run concurrently)

Multi-Model Routing (pstack)

pstack's poteto-mode skill specifies per-task models:

  • Code-writing delegates: composer-2.5-fast
  • Prose and judgment tasks: claude-opus-4-7-thinking-xhigh
  • how, why, interrogate skills: configure their own subagent_type for diverse-model review

This is one of the only frameworks in the corpus with explicit per-task model assignment.

Isolation Mechanism

orchestrate workers each get "one clone of the repo" — process-level isolation via Cursor cloud agent spawning. Each worker operates independently with no shared state to sibling agents.

ralph-loop Execution Mode

continuous-ralph — the stop hook resubmits the same prompt as a followup_message until the <promise> tag is detected or max iterations reached. This implements a stateful completion loop within a single agent session.

Execution Modes

Plugin Mode
orchestrate parallel-fan-out (cloud agents)
ralph-loop continuous-ralph
continual-learning event-driven (Stop hook)
cursor-team-kit one-shot per skill invocation
pstack interactive-loop (poteto-mode orchestrates sub-skills)

Consensus Mechanism

None formal. orchestrate uses "handoffs" (structured text files in handoffs/) as the coordination medium. No raft/quorum protocol.

Prompt Chaining

Strong prompt chaining in pstack: poteto-mode reads task → matches playbook → dispatches leaf skills whose outputs feed back into the main thread via subagent handoffs.

Strong in orchestrate: planner reads worker handoffs → generates new planner prompts → spawns next wave of workers.

Subagent Spawn Mechanism

orchestrate: Cursor SDK (@cursor/sdk) + CURSOR_API_KEY
pstack: Cursor's built-in Task tool with subagent_type: "poteto-agent"
continual-learning: Cursor's built-in subagent mechanism for agents-memory-updater

08

Ui Cli Surface

Cursor Rules Pack v2 (cursor/plugins) — UI / CLI Surface

Dedicated CLI Binary

The orchestrate plugin ships a Bun/TypeScript CLI at orchestrate/scripts/cli.ts:

bun skills/orchestrate/scripts/cli.ts kickoff "<goal>" \
  [--repo <url>] [--ref main] [--model claude-opus-4-7] \
  [--slack-channel <id>] [--dispatcher-name "<name>"]

Subcommands: kickoff, run, spawn, respawn, kill, tail, comment, andon

This CLI wraps the Cursor SDK and manages the orchestration tree state. It is a "thin wrapper" in the sense that it delegates execution to Cursor cloud agents, but it is not a wrapper around the cursor CLI binary itself.

Local Web Dashboard

None. The pr-review-canvas and docs-canvas plugins render HTML artifacts (Cursor Canvas) but these are inline in Cursor's chat, not standalone web dashboards.

IDE Integration

Cursor (exclusive):

  • Skills appear in Cursor's skill picker
  • Agents appear in Cursor's agent mode
  • Rules (.mdc) activate on file open
  • Hooks (hooks.json) fire on lifecycle events
  • Plugins install via /add-plugin <name>

Slack Integration (orchestrate, optional)

When SLACK_BOT_TOKEN is set, orchestrate:

  • Posts kickoff thread to Slack channel
  • Updates task status as agents complete
  • Reads 🚨 Andon reaction to pause/halt runs
  • Uploads handoff artifacts as file attachments

Observability

Plugin Observability
orchestrate plan.json + handoffs/*.json files; optional Slack thread; git branch history
continual-learning AGENTS.md updates; .cursor/hooks/state/continual-learning-index.json
pstack (show-me-your-work) Decision log as TSV file
cursor-team-kit (what-did-i-get-done) Commit summary
ralph-loop State file tracking iteration count + promise detection

Plugin Installation UI

Plugins install via /add-plugin <name> in Cursor chat. The marketplace manifest at .cursor-plugin/marketplace.json defines available plugins. No web store UI exists in this repository — the marketplace is consumed by Cursor directly.

Related frameworks

same archetype · same primary tool · same memory type

Claude-Flow / Ruflo ★ 55k

Eliminates single-agent context limits and sequential bottlenecks by orchestrating fault-tolerant swarms of specialized AI agents…

Hermes Agent (NousResearch) ★ 168k

Self-improving personal AI agent with closed learning loop, 7 terminal backends, and messaging gateway — not tied to any AI…

OpenCode ★ 165k

Terminal-first AI coding agent with multi-model routing, native desktop app, and a typed .opencode/ configuration system for…

OpenHands ★ 75k

Open-source AI software development platform (open-source Devin alternative) with Docker sandbox isolation, 77.6% SWE-bench…

DeerFlow ★ 70k

Long-horizon superagent that researches, codes, and creates by orchestrating parallel sub-agents with isolated contexts in Docker…

oh-my-openagent (omo) ★ 60k

Multi-provider AI agent orchestration for OpenCode: escape vendor lock-in by routing Sisyphus (Claude/Kimi/GLM) and Hephaestus…