@starlein/paperclip-plugin-company-wizard

v0.3.23

Published

4 days ago

AI-powered wizard to bootstrap paperclip agent companies from composable templates (for latest paperclip version)

0High
0Medium
0Low

Fork: This is a community-maintained fork of yesterday-AI/paperclip-plugin-company-wizard, updated for the current Paperclip API (>=2026.529.0) with substantial bug fixes. End-to-end company setup is largely functional as of v0.3.18.

Paperclip API compatibility

Bootstrap metadata fields renamed to match the Paperclip API exactly: parentId, assigneeAgentId, projectId, goalIds
CEO is provisioned with correct capabilities metadata so newly created CEOs are no longer saved with empty summaries
@paperclipai/plugin-sdk and @paperclipai/shared declared as peerDependencies with minimum version >=2026.529.0 — the host provides the SDK at runtime (externalized from the bundle)
security-engineer role now maps to the dedicated Paperclip security enum value (was general)

Bootstrap reliability

Agents provisioned with complete instructions — every non-CEO agent is now created by the plugin directly with its full instructionsBundle (AGENTS.md + HEARTBEAT/SOUL/TOOLS + skills). Previously the CEO created these agents during bootstrap with only an instructionsFilePath, leaving each agent with a bare AGENTS.md and fragile external path references
Routines created directly during provisioning — Paperclip only allows an agent to create routines assigned to itself, so the CEO could not create routines owned by the Product Owner. The plugin now creates all routines with board authority at provisioning time and pre-creates the main project so every routine — including those owned by non-CEO agents — is linked to it; BOOTSTRAP.md tells the CEO they already exist
Worker agents no longer run always-on heartbeats — enabling heartbeats on every provisioned agent caused bursts of concurrent runs that crashed the dev server. Only the CEO keeps an always-on heartbeat; all other agents are woken on assignment
Fresh local repos no longer bootstrap with isolated git worktrees — provisioning a brand-new local_path project with an isolated_workspace / git_worktree policy made worker agents try to branch off main before the repo existed, so early runs failed and agents flipped to error. The isolated policy is now suppressed for fresh local repos (agents work in the shared project workspace during bootstrap); existing external repos keep it. Guarded in assemble.js and removed at the source in StepRepository and the AI wizard prompts
Workspace isolation follows Paperclip instance settings — enableIsolatedWorktrees is no longer a plugin setting. The wizard reads enableIsolatedWorkspaces from the Paperclip instance experimental settings and only applies isolated_workspace / git_worktree for external repositories when that setting is enabled.
Bootstrap ordering hardened; agent filter bug fixed (v0.3.7)

Assembly and template fixes

Broken $AGENT_HOME references rewritten to absolute paths under companyDir/agents/<role>/ and companyDir/docs/ — AGENT_HOME points to a per-agent workspace dir that does not contain provisioned files
Shared docs scoped per role — each agent's AGENTS.md now lists only docs from modules relevant to that role (CEO still sees all); previously every agent was told to read every module's docs
Doc references use relative paths (docs/<file>) instead of absolute paths that baked in the collision-suffixed company directory name
Duplicate bootstrap issues (same title from module + preset) are now deduplicated; preset issue wins
CFO role removed — was orphaned (no preset, no capability, never activated)
PR bodies and comments use --body-file — review skills posted GitHub PR bodies and comments with inline gh pr … --body "...", where a double-quoted shell string keeps \n literal, so multi-line Markdown rendered as literal text\ntext. All PR-review guidance now writes Markdown to a file and uses --body-file, with a verdict-heading comment template documented in pr-conventions.md

AI wizard

Config generation uses claude-opus-4-8 with max_tokens: 32768 so a full-spec config is never truncated mid-JSON
All Anthropic calls run as background jobs in the worker (start + poll), eliminating the 30 s RPC timeout that previously crashed config generation
AI wizard now generates domain-specific initial issues from the project brief that lead the bootstrap backlog ahead of generic scaffolding issues
Preset roles are defensively merged with AI-selected roles so preset roles are no longer silently dropped

Error handling

All worker actions return { error } instead of throwing — prevents the plugin host from swallowing error messages in generic 502 responses
Network errors in PaperclipClient are caught and surfaced with actionable messages referencing the paperclipUrl plugin setting

New features (not in upstream)

Existing-company provisioning — target an existing Paperclip company instead of creating a new one (existingCompanyId); partial-failure cleanup never deletes existing companies
Approval-aware agent hiring — detects board-approval requirements, falls back to /agent-hires + auto-approve; surfaces pending approval ID if auto-approve fails
disableBoardApprovalOnNewCompanies setting — optionally patches new companies to skip board approval for fully-autonomous bootstrap
Repository workspace setup — choose between a fresh local Git repo or an existing external repository (GitHub, GitLab, etc.) via the manual wizard step or inline on the review/summary screen (available in both the manual and AI paths; the external option opens a repo-URL field)
Routine schedules tightened to run every few hours around the clock (auto-assign every 2 h, stall-detection every 3 h, backlog grooming every 4 h) with skip_if_active concurrency policy
"Update templates" button on the onboarding screen — deletes the cached templates dir and re-downloads from GitHub without restarting the plugin

Company Wizard is a Paperclip plugin that bootstraps an AI agent company for your project — roles, workflows, skills, and tasks — in a few clicks. Open it from the sidebar, answer a few questions (or just describe your project), and it assembles the workspace files and creates the company + CEO in Paperclip. The CEO then hires the team and sets up the backlog on its first heartbeat.

Why Company Wizard?

🗣️ Describe it, don't configure it. AI mode reads a plain-language brief and picks the right preset, modules, and roles for you. Manual mode is there for the steps you want to control yourself.
🧩 Composable, not monolithic. 15 curated presets layered from 26 modules and 17 roles. Mix and match freely — modules add skills, tasks, and heartbeat logic to the roles that are present, and degrade gracefully when they're not.
🤝 Works from day one with a single CEO. Every capability has an owner chain. Add a specialist and responsibilities shift to them automatically; leave one out and the next-best person — ultimately the CEO — steps in. No setup ever leaves a gap.
✏️ Review and edit before anything ships. Preview every generated file, tweak a persona, workflow, or the repository setup inline on the review screen, then provision.
🚀 Real end-to-end provisioning. Not just scaffolded files — it creates the company, CEO, goals, projects, and backlog in Paperclip via the API. Target a brand-new or existing company, with a fresh local or existing external Git repo.
💸 Token-aware by default. Worker agents are woken on assignment and driven by scheduled routines instead of always-on heartbeats, so the team doesn't burn tokens idling.

npm package installation (Paperclip > Settings > Plugins > Install Plugin):

@starlein/paperclip-plugin-company-wizard

Two Ways to Start

AI mode

Describe your project in plain language. The wizard analyzes it and picks the right preset, modules, and roles automatically.

"A fintech startup building a payment API, security is critical" → selects secure preset, Security Engineer + Product Owner roles, provisions everything

Great for getting started fast when you're not sure which template fits.

Manual mode

Walk through the steps yourself: name your company, set a goal, pick a preset, add modules, choose roles. Each step shows descriptions and hover-card previews so you know what you're getting.

Before provisioning, you can open any generated file and edit it inline — tweak a persona, adjust a workflow, or add role-specific context.

On the review step, repository setup is editable inline (fresh local repo vs external Git repo), and changing this now immediately updates preview/provisioning payloads in that step. Legacy source-mode values are normalized on save so mode switches can be applied reliably even when the configuration came from AI with older workspaceSourceType fields.

How Roles Work

Every company starts with just the CEO — and that's already a functional team. Add roles and responsibilities shift automatically:

Add an Engineer → they take over implementation, git workflow, and technical decisions
Add a Product Owner → they take over backlog management and auto-assignment from the CEO
Add a UX Researcher → they become the primary market analyst
Add DevOps → they own CI/CD and monitoring (instead of Engineer or CEO)

No role is ever truly missing. When a specialist isn't present, the next best available person steps in. The CEO is always the final fallback.

Enriched personas (opt-in): set enableEnrichedPersonas to give expert roles named domain lenses (e.g. STRIDE, Nielsen's 10, RICE) in their SOUL.md, add output/review bars to module skills, and append done-criteria to HEARTBEAT.md. Off by default — the baseline personas stay lean.

Start with just a CEO. Everything works. Add roles and responsibilities shift automatically:

How it works: Primary owners get the full skill. Fallback owners get a safety-net variant that only activates when the primary is absent or stalled.

Example: CEO only? They handle everything — strategy, backlog, auto-assign. Add an Engineer and they take over implementation. Add a Product Owner and they take over backlog management, with the CEO as fallback.

Presets

Presets are curated starting points that bundle a set of modules and optional roles for a specific use case. Pick one in the wizard and you're ready to go — or use it as a base and add modules on top.

fast is for a single engineer — multiple engineers without review will cause conflicts.
research has no code workflow. Add github-repo and backlog when ready to build.

fast — Solo engineer, direct-to-main, automated backlog. No review, no planning phase.

quality — Full review pipeline. Product Owner manages backlog and product alignment, Code Reviewer gates code quality. Feature branches with PR workflow.

rad — Rapid Application Development. Pick a tech stack, start building, hire when you hit bottlenecks. No upfront market research or architecture formalization — prototype first, learn from what you build, formalize later.

startup — Strategy-first. Starts with vision, market analysis, tech evaluation, and hiring review before any code. Grow the team through board approvals.

research — Planning only. Vision, market research, tech evaluation, and team assessment. No repo, no code workflow. Upgrade to startup or full when ready to build.

full — Everything. Full strategic planning, quality engineering with PR review, team growth via hiring review. Product Owner and Code Reviewer included. Best for serious projects that need both strategy and engineering rigor.

secure — Security-first. Threat modeling, security reviews, and quality gates on top of full planning and PR review. Security Engineer, Code Reviewer, and Product Owner included. Best for regulated industries, fintech, healthtech, or any project where security is a hard requirement.

gtm — Go-to-market focused. Competitive intelligence, market analysis, and brand identity. CMO for marketing strategy, Customer Success for competitive tracking, Product Owner for backlog. Best for products entering or competing in established markets.

content — Content and documentation focused. Technical Writer for developer docs and guides, accessibility for inclusive design, market analysis for positioning. Best for developer tools, documentation-heavy projects, or content-driven products.

website-relaunch — Relaunch an existing website with external design assets. Site audit, design ingestion, implementation, content migration, QA, and go-live. UI Designer for design analysis, Product Owner for backlog management. Includes a user-assigned "Provide design assets" issue as the entry point — upload your agency's designs, the team handles the rest.

repo-maintenance — Custodial maintenance for existing repositories. Agents review and merge open PRs, triage inbound GitHub issues, audit codebase health, manage dependencies, and handle releases. Code Reviewer for PR quality gates, Product Owner for issue triage and backlog. Inline goal bootstraps the team through repo onboarding, process setup, initial sweep, and steady-state maintenance.

build-game — Game development from idea to playable release. Game Designer owns the GDD, mechanics, and balancing. Game Artist generates sprites, textures, and tilesets via AI image generation and code-based approaches. Audio Designer creates sound effects, music, and soundscapes via AI generation and code-based synthesis. Inline goal with 5 milestones: concept (GDD + engine + art style), prototype (core loop + placeholder art + first playtest), vertical slice (one polished level), production (all content), polish & ship (balancing + distribution). Works for any genre.

Modules

Modules are composable capabilities you layer on top of the base team. Each module adds skills, tasks, and optionally heartbeat sections to the relevant roles. Modules are additive — you can combine them freely and they degrade gracefully when a role they extend isn't present.

Strategy & Planning

Maintenance & Operations

Engineering Workflow

vision-workshop

Defines the strategic foundation. The CEO runs a vision workshop to refine the company goal into a vision statement, success metrics, and milestones.

Capability: none (CEO-only strategic task)
Doc: docs/vision-template.md
With UX Researcher: contributes user-centered metrics and journey mapping

market-analysis

Researches the target market, competitors, and positioning.

Capability: market-analysis — owners: ux-researcher → cmo → product-owner → ceo
Fallback: CMO focuses on positioning and competitive landscape; CEO creates a brief overview only
Doc: docs/market-analysis-template.md

hiring-review

Evaluates team composition against the goal and proposes hires through board approval.

Capability: hiring-review — owners: product-owner → ceo
Fallback: CEO proposes one urgent hire only

tech-stack

Evaluates technology options and documents decisions with rationale and trade-offs.

Capability: tech-stack — owners: engineer → ceo
Fallback: CEO makes pragmatic defaults, marks them provisional
Doc: docs/tech-stack-template.md

architecture-plan

Designs the system architecture. Requires tech-stack. Includes a design-system capability for UI Designers.

Capability: architecture-plan — owners: engineer → ceo
Capability: design-system — owners: ui-designer → engineer
Docs: docs/architecture-template.md, docs/design-system-template.md

github-repo

Git workflow and commit conventions.

Task: Engineer initializes repo
Doc: docs/git-workflow.md

pr-review

PR-based review workflow. Requires github-repo. Activates with code-reviewer, product-owner, ui-designer, ux-researcher, qa, or devops.

Reviews run through the issue's native executionPolicy (stages), not child issues: a review stage for the Code Reviewer (plus relevant domain reviewers), an approval stage for the Product Owner, then a final approval merge gate owned by the Engineer — who is woken last to merge the PR before recording the verdict that closes the issue. The merge gate is deliberately the last stage so the Product Owner's approval does not auto-close the issue with the PR still open.

Task: Engineer sets up branch protection
Doc: docs/pr-conventions.md

backlog

Owns the product backlog lifecycle — from goal decomposition to a steady pipeline of actionable issues.

Capability: backlog-health — owners: product-owner → ceo
Fallback: CEO creates 1-2 issues only when backlog is critically empty
Doc: docs/backlog-process.md

auto-assign

Assigns unassigned issues to idle agents.

Capability: auto-assign — owners: product-owner → ceo
Fallback: CEO assigns only when agents are critically idle

brand-identity

Creates brand guidelines: logo usage, color palette, typography, iconography, and tone of voice.

Capability: brand-identity — owners: ui-designer → cmo → ceo
Fallback: CMO focuses on brand strategy and messaging; CEO creates minimal provisional placeholder
Doc: docs/brand-identity-template.md

user-testing

Designs and executes usability evaluations, documents findings with severity ratings.

Capability: user-testing — owners: qa → ux-researcher → product-owner → ceo
Fallback: QA adds test automation and edge case coverage; CEO creates a basic heuristic checklist
Doc: docs/user-testing-template.md

ci-cd

Continuous integration and deployment pipeline. Requires github-repo.

Capability: ci-cd — owners: devops → engineer
Fallback: Engineer sets up basic CI (lint, test, build); DevOps owns full pipeline lifecycle including CD
Doc: docs/ci-cd-template.md

monitoring

Observability, error tracking, logging, alerting, and health checks. Requires github-repo.

Capability: monitoring — owners: devops → engineer
Fallback: Engineer sets up basic health checks and structured logging; DevOps owns full observability stack
Doc: docs/monitoring-template.md

security-audit

Threat modeling and security code review. Identifies attack surfaces, OWASP Top 10 vulnerabilities, and dependency CVEs.

Capability: threat-model — owners: security-engineer → devops → engineer
Capability: security-review — owners: security-engineer → devops → engineer
Fallback: DevOps focuses on infrastructure security; Engineer runs basic checks only

documentation

Project documentation: READMEs, API references, architecture overviews, onboarding guides.

Capability: project-docs — owners: technical-writer → engineer → ceo
Fallback: Engineer writes minimal README; CEO creates bare-bones project overview

competitive-intel

Living competitive landscape — competitor profiles that evolve over time with positioning, strengths, and differentiation insights.

Capability: competitive-tracking — owners: customer-success → cmo → product-owner → ceo
Fallback: CMO focuses on positioning angles; CEO creates brief overview only

accessibility

WCAG 2.2 compliance auditing: semantic HTML, keyboard navigation, color contrast, ARIA, screen reader compatibility.

Capability: accessibility-audit — owners: qa → ui-designer → engineer
Fallback: UI Designer focuses on visual accessibility; Engineer runs automated checks

website-relaunch

Full website relaunch lifecycle: audit the current site, ingest design assets from an external agency, implement the new design, migrate content, and go live. Includes an inline goal with 5 milestones and 10 issues.

Capability: design-ingestion — owners: ui-designer → engineer → ceo
Capability: site-audit — owners: ui-designer → engineer → ceo
Goal: Website Relaunch (with dedicated project, 5 milestones, 10 issues)

build-api

REST API development from schema to documentation. Inline goal with 4 milestones and 8 issues.

Capability: api-design — owners: engineer → ceo
Requires: github-repo
Goal: Build a REST API (with dedicated project, 4 milestones, 8 issues)

launch-mvp

MVP project lifecycle: define scope tightly, build the core feature, deploy, and iterate from user feedback. No capabilities or skills — structured goal with milestones and issues.

Goal: Launch MVP (with dedicated project, 4 milestones, 8 issues)

codebase-onboarding

Audit an existing codebase and maintain its health over time. Requires github-repo.

Capability: codebase-audit — owners: engineer → ceo
Output: docs/CODEBASE-AUDIT.md

triage

Processes inbound GitHub issues: classify by type and priority, respond to reporters, close duplicates, convert actionable items into Paperclip tasks. Requires github-repo.

Capability: issue-triage — owners: product-owner → engineer → ceo

dependency-management

Dependency lifecycle: vulnerability scanning, outdated package detection, safe patch-level updates, and major version migration planning. Requires github-repo.

Capability: dependency-audit — owners: devops → security-engineer → engineer
Output: docs/DEPENDENCY-AUDIT.md

release-management

Release lifecycle: semantic versioning, changelog generation, git tagging, GitHub Releases, and rollback documentation. Requires github-repo.

Capability: release-process — owners: devops → engineer → ceo
Output: docs/RELEASE-PROCESS.md

game-design

Game Design Document creation and ongoing mechanic design, progression, and balancing. Ships a GDD template and engine reference docs (Phaser, PixiJS, Three.js).

Capability: game-design — owners: game-designer → engineer → ceo
Docs: docs/gdd-template.md, docs/engine-phaser.md, docs/engine-pixijs.md, docs/engine-threejs.md

stall-detection

Detects issues stuck in in_progress or in_review with no recent activity. Nudges the assigned agent, escalates to the board if nudging doesn't help.

Capability: CEO-only

Roles

Every company starts with just the CEO (the only base role). All other roles are optional:

Software Engineer

Builds the product: implements features, fixes bugs, writes tests, and manages the git workflow. The default owner of technical execution — without an Engineer present, the CEO carries implementation.

Product Owner

The voice of the user. Owns the backlog pipeline, validates engineering output against goals, manages scope discipline. Adds product-alignment review pass with pr-review module.

Code Reviewer

Owns code quality. Reviews PRs for correctness, style, security, and test coverage. Never writes code — only reviews it.

UI & Brand Designer

Owns visual identity, design systems, and brand consistency. Creates design specs that engineers implement. Outputs are design documents, not code. Adds design review pass with pr-review module.

UX Researcher

Owns user experience research, usability analysis, and journey mapping. Grounds design and product decisions in evidence-based user insights. Adds UX review pass with pr-review module.

CTO

Technical leadership and architecture oversight. Guides technology decisions, reviews system design, and ensures engineering quality at scale.

CMO

Owns marketing strategy, brand positioning, go-to-market planning, and growth metrics. Data-driven, measures everything.

DevOps Engineer

Owns infrastructure, CI/CD pipelines, deployment, monitoring, and platform reliability. Automation over manual work, infrastructure as code.

QA Engineer

Owns test strategy, test automation, quality gates, and regression prevention. Prevention over detection.

Technical Writer

Owns developer documentation, API references, READMEs, and onboarding guides. Keeps docs accurate as the codebase evolves.

Security Engineer

Owns threat modeling, security code reviews, OWASP compliance, and secure coding standards. Security issues are always blocking.

Customer Success Manager

Owns customer health monitoring, feedback synthesis, churn prevention, and competitive intelligence from the customer perspective.

Game Designer

Owns the Game Design Document, core mechanics, game loop, progression systems, difficulty curves, and balancing. Runs design experiments and iterates based on playtest data.

Level Designer

Owns level layout, pacing, difficulty curves, environmental storytelling, and spatial progression.

Game Artist

Owns visual art production: sprites, textures, tilesets, UI elements, and visual effects. Creates assets using AI image generation tools, code-based approaches, and asset pipeline tools.

Audio Designer

Owns audio production: sound effects, music, ambient soundscapes, and audio systems design. Creates audio using AI generation tools, code-based synthesis, and audio processing pipelines.

Configuration

Configure the plugin via Settings → Plugins → Company Wizard in the Paperclip UI.

| Field | Required | Description | | --- | --- | --- | | companiesDir | No | Where assembled company workspaces are written. Defaults to ~/.paperclip/instances/default/companies. Override for Docker setups. | | templatesPath | No | Path to the templates directory. Defaults to ~/.paperclip/plugin-templates (auto-downloaded from templatesRepoUrl if missing). | | templatesRepoUrl | No | GitHub tree URL to pull templates from when the templates directory does not exist. Defaults to the official @starlein/paperclip-plugin-company-wizard templates. | | paperclipUrl | No | Paperclip instance URL. Defaults to http://localhost:3100 or PAPERCLIP_PUBLIC_URL env var. | | paperclipEmail | No | Board login email. Required for authenticated (non-local_trusted) instances. | | paperclipPassword | No | Board login password. Stored as a secret ref. | | anthropicApiKey | No | Anthropic API key for AI wizard mode. Stored as a secret ref. Required to use the AI-powered setup path. | | disableBoardApprovalOnNewCompanies | No | If true, the wizard PATCHes new companies to set requireBoardApprovalForNewAgents=false during provisioning. Leave false to preserve approval-gated hiring. Defaults to false. | | enableEnrichedPersonas | No | If true, expert roles get domain lenses (named mental models) in SOUL.md, module skills get concrete output/review bars, and HEARTBEAT.md gets explicit done-criteria. Leave false (default) for the lean baseline personas. |

For isolated worktrees: there is no plugin setting. The policy is controlled by Paperclip instance settings under Settings → Instance → Experimental → enableIsolatedWorkspaces and is consumed by the plugin during provisioning.

Development

pnpm install
pnpm build          # esbuild: worker + manifest + UI → dist/
pnpm dev            # watch mode
pnpm test           # vitest: tests/**/*.spec.ts
pnpm test:logic     # node --test: src/logic/*.test.js
pnpm typecheck      # tsc --noEmit

After pnpm build, reload the plugin in the Paperclip UI — no reinstall required.

Extending

Add a module

templates/modules/<name>/
├── module.meta.json             # Name, capabilities, tasks, dependencies, permissions
├── skills/                      # Shared skills (used by any primary owner)
│   └── <skill>.md
├── agents/<role>/
│   ├── skills/                  # Role-specific overrides and fallbacks
│   │   ├── <skill>.md           # Override (replaces shared for this role)
│   │   └── <skill>.fallback.md  # Fallback (safety-net for non-primary)
│   └── heartbeat-section.md     # Optional: injected into role's HEARTBEAT.md
└── docs/                        # Shared docs (→ docs/)

{
  "name": "my-module",
  "requires": ["other-module"],
  "activatesWithRoles": ["my-role"],
  "permissions": ["tasks:assign"],
  "adapterOverrides": { "chrome": true },
  "capabilities": [
    {
      "skill": "my-skill",
      "owners": ["my-role", "ceo"],
      "fallbackSkill": "my-skill.fallback"
    }
  ],
  "tasks": [
    {
      "title": "Initial task",
      "assignTo": "capability:my-skill",
      "description": "Task description"
    }
  ],
  "goal": {
    "title": "My Goal",
    "description": "What this goal achieves",
    "project": true,
    "milestones": [
      { "id": "phase-1", "title": "Phase 1", "project": false }
    ],
    "issues": [
      { "title": "First task", "milestone": "phase-1", "assignTo": "engineer", "priority": "high" }
    ]
  }
}

When assembling a capability's primary skill, the system checks in order:

Role-specific override: agents/<role>/skills/<skill>.md
Shared skill: skills/<skill>.md

First match wins. Most capabilities only need a shared skill. Role-specific overrides exist only when a role brings a genuinely different approach.

Example: market-analysis module
├── skills/
│   └── market-analysis.md                    # Shared: any primary owner
├── agents/
│   ├── ux-researcher/skills/
│   │   └── market-analysis.md                # Override: user-focused
│   └── ceo/skills/
│       └── market-analysis.fallback.md       # Fallback: brief overview

UX Researcher present → gets role-specific override (user-focused)
Product Owner primary → gets shared skill
CEO as fallback → gets fallback variant

Two kinds of docs end up in {company}/docs/:

Templates (lowercase-kebab.md) — Shipped by modules, copied at assembly time. Guaranteed to exist if the module is active.
Agent output (UPPERCASE.md) — Created by agents during execution. May or may not exist yet.

Add a role

templates/roles/<name>/
├── role.meta.json   # Name, title, base, paperclipRole, reportsTo, adapter
├── AGENTS.md
├── SOUL.md
├── HEARTBEAT.md
└── TOOLS.md

{
  "name": "my-role",
  "title": "My Role",
  "base": false,
  "division": "engineering",
  "tagline": "One-liner for wizard display and AI selection",
  "paperclipRole": "general",
  "description": "What this role does",
  "reportsTo": "ceo",
  "enhances": ["Takes over X from CEO"],
  "adapter": {
    "model": "claude-sonnet-4-6"
  }
}

Add a preset

Create templates/presets/<name>/preset.meta.json:

{
  "name": "my-preset",
  "description": "What this preset is for",
  "constraints": [],
  "roles": ["product-owner"],
  "modules": ["github-repo", "backlog"]
}

How It Works

┌─────────────┐     ┌──────────────┐     ┌──────────────────┐
│   Wizard    │────▶│   Assembly   │────▶│   Provisioning   │
│   (UI)      │     │  (files)     │     │   (Paperclip API)│
└─────────────┘     └──────────────┘     └──────────────────┘

Assembly (always runs):

Copies base role files (CEO — the only base role) into agents/
Copies selected extra roles into agents/
For each module: resolves capability ownership, installs skills, copies docs
Injects module heartbeat sections into each role's HEARTBEAT.md
Generates BOOTSTRAP.md with goal, project, agent paths, and initial tasks

Provisioning (Review → Provision step):

Connects to Paperclip API (auto-detects local_trusted vs authenticated)
Creates a new company in Paperclip — or targets an existing one if existingCompanyId is set in the review step
Creates the CEO agent with adapter config (cwd, instructionsFilePath, model), or reuses the existing active CEO when targeting an existing company. If board approval is required, the wizard hires via /agent-hires and auto-approves
Creates a Bootstrap task assigned to the CEO

The CEO then sets up the rest of the team on its first heartbeat: hiring the other roles from disk, creating the goal, project, and initial backlog issues. If provisioning fails after a new company is created, the partial company is automatically deleted — existing target companies are never deleted on error.

Changelog

See CHANGELOG.md for version history.

Contributing

See CONTRIBUTING.md for development setup and guidelines.

License

MIT