Pi Team Orchestrator

A Pi extension for running multi-role team workflows from the normal CLI chat interface.

It supports:

• Team loading from shared-role manifests (teams/<team>/team.json) and legacy markdown role directories (teams/<team>/roles/*.md)
• Project context under ~/.pi/projects
• Session-backed orchestration state (team/project/phase/handoffs/blockers/decisions)
• Structured coordination tools (status, handoff, blocker, decision events)
• File-backed workflow docs under <project>/documents
• One active task packet per workflow lane under <project>/documents/working/tasks/
• Compact on-disk task registry under <project>/documents/working/task-registry.json
• Strict role gates so agents only advance when the brief, checkpoint packet, task packet, and ownership are in place
• Workflow modes (lean, delivery, recovery) to cap default read bundles
• Live workflow/status indicators

Default behavior (important)

This extension is chat-first and state-first:

• You communicate to the team through normal CLI chat prompts.
• Coordination is recorded as structured events (not chat-thread inboxes).
• pi_messenger is blocked by default.

You can override messenger policy per session:

/messenger-mode allowed

Return to default:

/messenger-mode blocked

Quick start

Start Pi with the extension from this repo root:

pi -e ./src/index.ts

Set context:

/team-list
/team-load web-app
/project-init my-project --bind-active-team
/project-brief Initial brief for this project.
/workflow-resume
/task-status

If project already exists:

/project-switch my-project
/project-bind-team web-app
/project-migrate
/workflow-resume

For an older project created before the file-backed workflow upgrade, run /project-migrate once after switching/binding. That seeds the new workflow packet and task registry from the old log where possible.

Copy/paste session script (new project smoke test)

Use this directly in a fresh Pi session (replace my-new-project and optionally team id):

/messenger-mode blocked
/team-list
/team-load web-app
/project-init my-new-project --bind-active-team
/team-status
/project-status

Scout this project and produce a concise brief: current state, top risks, and unknowns. Then hand off to planner.
Planner: create a milestone plan with acceptance criteria and explicit dependencies.
/session-signoff --role planner --status handoff --next architect
/workflow-status
/checkpoint-log

For game workflows, swap web-app with game-dev.

Checkpoint/handoff events are also written to a project-local file for cross-session recovery:

• <project>/.pi-orchestrator/checkpoints.md

Copy/paste resume script (existing project)

/messenger-mode blocked
/project-switch my-new-project
/team-load web-app
/project-bind-team web-app
/project-migrate
/workflow-status
/task-status

/workflow-resume is now the primary resume command. It points to the exact role, task packet, latest checkpoint, gate issues, and role-scoped read bundle. /workflow-status is the broader operator view.

How to communicate with the agents

After team/project context is set, type normal prompts in CLI.

Examples:

Scout this repo for auth and payment touchpoints, then hand off to planner with top risks.

Planner: break this into milestones with acceptance criteria and dependencies.

Architect + coder: propose a contract-first API change plan and implement the first safe slice.

Reviewer: evaluate current changes for release blockers and provide go/no-go recommendation.

The extension injects orchestration context at turn start so these prompts are interpreted as team workflow instructions.

Command reference

Primary daily commands

• /workflow-resume
• /project-brief <summary>
• /workflow-status
• /task-status
• /workflow-mode [lean|delivery|recovery]
• /workflow-next
• /session-signoff [--role ... --status ... --next ... --task ... --summary ...]

These are the commands you should need most of the time. The other workflow commands are mainly for inspection or manual recovery.

Team

• /team-list
• /team-load <team-id> (or /team-load for picker)
• /team-status (or /team-status <team-id>)
• /footer-mode <rich|compact>
• /team-board <on|off>
• /messenger-mode <blocked|allowed>

Project

• /project-list
• /project-init <project-id> [--bind-active-team]
• /project-switch <project-id> (or /project-switch for picker)
• /project-bind-team <team-id> (or /project-bind-team for picker)
• /project-brief <summary>
• /project-migrate
• /project-status (or /project-status <project-id>)

Workflow visibility

• /workflow-resume
• /workflow-status
• /workflow-reseed
• /workflow-mode [lean|delivery|recovery]
• /task-status
• /agent-status
• /handoff-log
• /blockers
• /decision-log
• /checkpoint-log
• /session-signoff [--role ... --status ... --next ... --task ... --summary ...]
• /workflow-next

Manual checkpoint signing

• /checkpoint-sign [--summary <text>] [--role <role-id>] [--status in_progress|handoff|done|blocked] [--task <task-id>] [--next <role-id>]

If --summary is omitted, the extension auto-generates a concise checkpoint summary.

End-of-session signoff (recommended)

• /session-signoff [--summary <text>] [--role <role-id>] [--status in_progress|handoff|done|blocked] [--task <task-id>] [--next <role-id>]

This signs a checkpoint (default status: handoff) and prints a compact resume card with the exact commands to continue in a new session.

Structured coordination primitives

The orchestrator prefers these tools internally:

• team_role_status
• team_handoff
• team_blocker
• team_decision_log
• team_checkpoint_sign

This keeps team coordination as a workflow/event log instead of conversational message history.

File-backed workflow model

Every bound project gets a small workflow filesystem:

<project>/
  documents/
    index.json
    canonical/
      project-brief.md
      decision-log.md
      workflow-status.md
    checkpoints/
      <timestamp>-<role>-<status>.md
    supporting/
      active-blockers.md
    working/
      tasks/
        <task-id>.md
      task-registry.json
    archive/
  .pi-orchestrator/
    checkpoints.md

The intended usage is:

• project-brief.md: concise canonical overview of the project
• workflow-status.md: previous/current/next workflow lane plus active task, gates, stale docs, and read bundle
• checkpoint packet: the latest resumable handoff packet for the active task
• task packet: the live working packet for the one active task lane
• task-registry.json: compact machine-readable task ownership and next-step state
• decision-log.md: concise approved decisions only
• checkpoints.md: compatibility audit log, not the main resume surface

Efficiency rules:

• Agents are prompted to read only the role-scoped relevant files by default.
• /workflow-resume and the active task packet are the default restart surface, not broad project history.
• Archived and superseded checkpoint packets are not part of the default read set.
• The task registry is compact on purpose; it should stay operational, not narrative.
• If a role is missing its required brief/checkpoint/task packet/ownership, gates stop it from advancing.
• Handoffs and done checkpoints require evidence references, not just narrative text.

Shared Roles And Custom Teams

Teams can now be composed from a shared role library.

Primary layout:

roles/<role-id>.md
teams/<team-id>/team.json

Example:

{
  "id": "my-team",
  "name": "My Team",
  "roles": ["scout", "planner", "architect", "coder", "reviewer"]
}

Each role id in team.json resolves to roles/<role-id>.md unless a custom source is set.

Team manifests can override selected role fields:

• name
• phase
• model
• thinking
• deliverables
• handoffTo

Backward compatibility remains:

• If teams/<team>/team.json does not exist, the loader still falls back to teams/<team>/roles/*.md.

Project-specific teams are supported, for example:

• teams/moonglow/team.json

See also:

• docs/06-quickstart-usage.md
• docs/07-file-backed-workflow.md

UI indicators

Footer now stays compact and project-focused.

A card-style Workflow Lane widget is rendered below the editor to show:

• previous agent
• current agent
• next agent
• status/task/note for each slot

Switch modes:

/footer-mode rich
/footer-mode compact

Toggle board widget:

/team-board on
/team-board off

Development

Run locally

pi -e ./src/index.ts

Smoke test

npm run smoke

Optional: use just command shortcuts

A justfile is included with Moonglow helpers:

just --list
just moonglow-new
just moonglow-resume
just moonglow-signoff-auto

Package + release

Verify package contents

npm run pack:check

Optional local package install test

pi install /home/openclaw/.pi/projects/pi-team-orchestrator
pi list

Release

Recommended GitHub release flow:

1. Bump package.json version and update CHANGELOG.md.
2. Push the commit.
3. Push a matching tag such as v0.2.6.
4. GitHub Actions creates the GitHub Release.
5. The npm publish workflow also runs from that tag push and publishes from GitHub.

Local dry run:

npm run release:dry

Local publish:

npm run release

Local publish with npm OTP:

NPM_OTP=123456 npm run release

Install from npm:

pi install npm:@gericomaverick/[email protected]

GitHub workflows:

• .github/workflows/github-release.yml: creates a GitHub Release on v* tag push
• .github/workflows/npm-publish.yml: publishes to npm on v* tag push, GitHub Release publish, or manual dispatch

npm package setting:

• If you want GitHub trusted publishing with an emergency token fallback, choose Require two-factor authentication or a granular access token with bypass 2fa enabled.
• If you want only trusted publishing and never want token-based fallback, choose Require two-factor authentication and disallow tokens.

This repo now supports both trusted publishing and NPM_TOKEN fallback in GitHub Actions, so the first option is the pragmatic choice.

Additional docs

• docs/06-quickstart-usage.md
• docs/02-cli-workflow-and-indicators.md
• docs/03-shared-memory-and-drift-prevention.md
• docs/04-data-model.md
• docs/05-implementation-plan.md
• docs/08-release-and-publishing.md