Pipa

Pipa is an orchestration and execution-discipline package for the Pi coding agent.

It extends Pi with structured task execution, decentralized teammates, persistent project planning, knowledge management, permission guards, and context hygiene.

Installation

From npm

pi install npm:@aelinrezende/pipa

To pin a specific version:

pi install npm:@aelinrezende/[email protected]

From GitHub

pi install git:github.com/aelinrezende/pipa

Project installation

Use --local to register Pipa in the current project's .pi/settings.json:

pi install --local npm:@aelinrezende/pipa

What Pipa Adds

• Interactive onboarding that configures Pipa for any domain.
• Domain-specific plan, execute, and finish skills generated for the project.
• Decentralized, persistent teammates with direct collaboration.
• Structured tasks with ownership, dependencies, squads, and lifecycle enforcement.
• Session-level todo lists for small execution steps.
• A persistent project backlog for epics, stories, tasks, and bug fixes.
• A project knowledge base with Markdown documents and fuzzy search.
• Permission guards for terminal commands and direct file modifications.
• Automatic cleanup of noisy model context.
• English and Brazilian Portuguese interface messages.

Onboarding and Profiles

Pipa does not ship with a predefined professional focus. When .pi/PROFILE.md does not exist, it starts an interactive onboarding process to learn what kind of work the environment will support, how agents should communicate, which constraints they must respect, and how quality should be evaluated.

Pipa uses a single language-neutral onboarding prompt. It first tries to infer the conversation language from the user's initial message. If the message is too short, language-neutral, ambiguous, or intentionally multilingual, Pipa asks which language the user prefers before continuing. During discovery, it also confirms the working language for generated agents and artifacts; this may differ from both the first-message language and the configured UI locale.

The predefined answers are only discovery aids. Free-form answers and follow-up questions allow Pipa to configure any specialized or mixed workflow.

Example profiles

Onboarding creates the project's operating system:

• .pi/PROFILE.md: shared purpose, domain, language, vocabulary, global rules, constraints, and artifact conventions.
• .pi/skills/plan/SKILL.md: how work is understood, investigated, decomposed, risk-assessed, and approved.
• .pi/skills/execute/SKILL.md: how approved work is delegated, performed, validated, and recorded.
• .pi/skills/finish/SKILL.md: how results are reviewed, reported, preserved, and closed or handed off.
• .pi/teammates/*.md: initial specialist roles proposed from the profile and approved by the user.

flowchart LR
    Start[First message] --> Language{Language clear?}
    Language -->|Yes| Discovery[Discover domain, working language, constraints]
    Language -->|No| Ask[Ask preferred language]
    Ask --> Discovery
    Discovery --> Profile[Create PROFILE.md]
    Profile --> Skills[Generate plan, execute, finish]
    Skills --> Roles[Propose specialist roles]
    Roles --> Team[Create approved teammates]
    Team --> Restart[Restart Pi]
    Restart --> Ready[Domain-configured Pipa]

plan, execute, and finish are generated for the selected domain; they are not generic software-development instructions. The profile, skills, and teammate definitions are regular Markdown files and can evolve with the environment.

Teammates use locale-neutral, role-based identifiers by default. If human names are preferred, onboarding asks for a language, culture, or naming style instead of inferring one from the UI locale or from Pipa's origin.

An Opinionated Workflow

Most agent orchestrators use a central coordinator. Work flows through one planner or router, and every sub-agent is created for a single job and then discarded. That is simple, but it makes the coordinator a bottleneck, and every handoff loses context.

Pipa takes a different approach: teammates are decentralized, autonomous, and persistent.

Autonomy

A teammate does not wait for the main agent to approve each step. It configures its assigned task, works toward the goal it was given, and reports the result. The task guard keeps this autonomy accountable: teammates must hold an active task before they can use non-management tools.

Collaboration without a central router

Teammates can talk to each other directly.

• teammates → chat opens a synchronous conversation with another teammate by ID. It is the default for status checks, validation, and adjustments.
• teammates → send-inbox leaves an asynchronous message for a teammate who is busy and should not be interrupted.
• spawnableTeammates lets a teammate delegate to specific specialists, each with a stated reason. A reviewer can call a developer for a correction or a tester for more validation without going back to the main agent.

The main agent stays the supervisor. It can instantiate any teammate, but it is not on the critical path of every exchange.

Teammates that stay while they are useful

Teammates are not disposable workers. Once a teammate has context about the project, task history, and earlier decisions, its session stays available.

• The online action shows which teammates have an active session.
• A teammate is released only through dismiss, with an explicit reason, once it is no longer needed.
• Optional session compaction after task actions keeps long-lived teammates from filling their own context.

flowchart LR
    Main[Main agent / supervisor]
    Specialist[Domain specialist]
    Reviewer[Reviewer]
    Researcher[Researcher]

    Main -->|goal| Specialist
    Main -->|goal| Reviewer
    Specialist <-->|chat| Reviewer
    Reviewer -->|spawnableTeammates| Researcher
    Researcher <-->|inbox / chat| Specialist

Tools

Task Execution

A task can contain a short name, a detailed description, hard or soft dependencies, a parent task, and a squad.

Dependencies can be:

• hard: completely blocks execution until the dependency is completed.
• soft: records an informational dependency without blocking execution.

Teammates must configure an assigned task before calling non-management tools. The task, teammates, and todo tools remain available so an agent can organize its work and recover from a missing task. The main agent acts as the supervisor and is not subject to this guard.

Teammates

Teammates are specialized local agents defined through Markdown files. Pipa loads teammates bundled with the package and project teammates stored in .pi/teammates/.

Before delegating work, an agent can list the available teammates. A teammate is instantiated with a concrete goal containing the necessary context, constraints, and acceptance criteria.

Teammate definition

Create a Markdown file such as .pi/teammates/reviewer.md:

---
name: reviewer
description: Reviews deliverables, verifies requirements, and identifies risks.
thinking: medium
spawnableTeammates:
  specialist: Requests a domain-specific correction
  researcher: Requests additional evidence
---

You are a rigorous reviewer for the domain defined in `.pi/PROFILE.md`.

Review the assigned artifact against its requirements. Identify correctness
issues, missing validation, risks, and unnecessary complexity.

Return actionable findings ordered by severity.

Supported frontmatter fields:

The main agent can instantiate any teammate. Other teammates can instantiate only themselves or teammates listed in spawnableTeammates.

Model concurrency

Pipa can distribute teammates across candidate models to avoid sending every parallel agent through the same model queue. Model cycling can be triggered by error or concurrency events and is controlled through pipa.config.ts.

Todo Lists

The todo tool tracks small steps within the current execution flow. Items support the statuses pending, in_progress, and completed. Use todo for short checklists, task for delegated work, and backlog for persistent project planning.

Project Backlog

The backlog tool maintains persistent project work as Markdown documents with structured metadata.

• Item types: epic, story, task, bugfix.
• Priorities: low, medium, high, critical.
• Items can have a title, Markdown body, status, domains, tags, a parent item, and arbitrary JSON metadata.

The tool supports filtering, sorting, selection by title, code, or tags, metadata updates, body updates, and removal.

Project Documentation

The docs tool maintains a local Markdown knowledge base. Documents support titles, short summaries, tags, parent-child relationships, Markdown content, arbitrary metadata, fuzzy search, sorting, and filtering.

The publish action exports the documents to Markdown and generates a static site.

Permissions and Guardrails

Pipa adds guardrails around tool execution, but it is not a security sandbox.

• Terminal commands are validated against a command policy, and commands classified as dangerous are blocked.
• Teammates must configure an assigned task before using non-management tools.

These checks complement Pi's own security model. They do not replace operating system permissions, repository protections, isolated environments, or human review.

Context Hygiene

Long-running agent workflows can accumulate blocked calls and stale messages. When a tool call is blocked because no task is active, Pipa collapses the repeated attempts into a single notice so the model context stays clean.

Configuration

Create .pi/pipa.config.ts in the project. Every option is optional.

export default {
  // Interface and built-in Pipa message language: 'en' or 'pt-BR'.
  locale: 'en',

  // Persist teammate messages.
  persistMessages: false,

  teammate: {
    // How a teammate processes an incoming nudge.
    nudgeMode: 'steer',

    // Hide teammate tool calls from the terminal interface.
    hideTools: false,

    // Hide the model name used by teammates.
    hideModelName: false,

    // Models allowed to run concurrently.
    concurrentlyModels: [],

    cycling: {
      // Events that trigger model rotation: 'error' and/or 'concurrency'.
      events: [],

      // Candidate model IDs used during rotation.
      models: []
    },

    // Compact the teammate session after a task action.
    compactAfterTaskAction: false
  },

  tasks: {
    // Maximum number of tasks displayed simultaneously in the TUI.
    maxVisible: 5
  }
};

Project configuration is merged over the defaults above. Arrays replace the corresponding default array instead of being concatenated.

Localization

Pipa supports en (English) and pt-BR (Brazilian Portuguese) for its interface and built-in messages. Set locale in .pi/pipa.config.ts.

Onboarding is driven by one language-neutral prompt. The agent infers or asks for the conversation language, then confirms the working language used by the profile, generated skills, teammates, and artifacts. The working language is not limited to Pipa's supported interface locales.

Package Contents

The distributed Pi package includes the compiled Pipa extension, a language-neutral onboarding prompt, shared orchestration instructions, and templates for profiles, skills, teammates, tasks, and backlog items.

Project-specific skills and teammates are generated during onboarding and stored under the project's .pi/ directory.

Distribution

The aelinrezende/pipa repository contains generated distribution artifacts. The source implementation is maintained in pipa-source. Releases build the extension and publish the result to GitHub and npm. Do not edit the distribution repository directly.

License

Pipa is licensed under the MIT License.