Multi-Agent

One system, three repos: an 8-phase AI development pipeline, the MCP server that drives real devices for it, and the stack skill plugins it dispatches to.

Drives a Jira issue or GitHub URL to a merged PR in one command. It ships as three repos because they release on different clocks: the pipeline changes weekly, the MCP server only when a device capability lands, and each stack plugin only when its own skills move. Installing the pipeline is enough - it pulls the other two.

Private RepositoryThis is a private repository. You need collaborator access to view the source code on GitHub.
8
Phases
53
Commands
210
Skills
84
MCP tools
6
Plugins
8
Sub-agents
Stable
Status

Tech Stack

Node.jsClaude CodeCopilot CLICodex CLIMCP ProtocolJSON Schemas

Features

8-Phase Pipeline

Issue → analysis → plan (approval gate) → TDD dev → parallel review + triage → test → commit/PR → report. Every retry loop is bounded (hard-stop at 3).

CLI-Aware Parallel Review + Fable Triage

2 models on Claude Code, 3 on Copilot CLI; Fable triage filters false positives. Findings cite stable rule IDs, so a review says SEC-01 instead of a preference. Blocking findings can be confirmed by an auto-written repro test.

Deterministic Safety Gates

Build, lint, test and secret gates run before AI review, and a 'passed' claim needs a log behind it. The applicable rule IDs are written to disk first, so "was this applied completely" has a fixed denominator.

Two Analysis Standards

The analysis command asks which standard the document follows: a development handoff, or a requirements document with an IG → UC → FG spine and three traceability matrices. Both read the same evidence, so they cannot drift apart.

Device Control Over MCP

The toolkit server gives the pipeline eyes and hands on a running app: tap, swipe and type on iOS Simulator and Android Emulator, read the accessibility tree, diff screenshots, pull crash logs, audit an archive before submission.

Stack Skills as Versioned Plugins

Engineering knowledge ships as marketplace plugins rather than inside the pipeline, enabled per repo. Their knowledge layer is generated from one authoring source and version-bumped only when a skill set changes.

Multi-Repo Orchestration

Per-repo worktrees, platform identity routing, and file-based state so tasks resume across sessions. A task worktree removes itself once its PR is open, keeping the branch without checking it out, so your own working copy is never disturbed.

Cost, Telemetry and Memory

Per-phase token ledger, a cost breakdown naming the top driver, budget ceilings, and a model fallback contract. A learnings ledger keeps rejected review preferences from being re-flagged.

Available Commands

Claude Code (colon-form)Copilot CLI (dash-form)Description
Development
/multi-agent "task"multi-agent "task"

Full 8-phase pipeline with worktree + interactive confirmations.

/multi-agent:autopilot "task"multi-agent-autopilot "task"

Same pipeline, zero confirmations - fire and forget.

/multi-agent:jiramulti-agent-jira

Browse your open Jira issues in a terminal picker and launch.

/multi-agent:issuemulti-agent-issue

Browse unassigned GitHub issues, auto-assign, and launch.

Analysis
/multi-agent:analysis "feature"multi-agent-analysis "feature"

A document, not code: Init → Analysis → Synthesis → Review → Publish → Report.

/multi-agent:review-analysis [path | page | KEY]multi-agent-review-analysis [path | page | KEY]

Reviews the document instead of the diff: deterministic gates first, then a parallel model review. Findings cite the rule they break.

Shared
/multi-agent:feedback "message"multi-agent-feedback "message"

Tell the maintainer something broke. Only the text you type is sent - no logs, no paths - and the payload is shown before it goes.

/multi-agent:reviewmulti-agent-review

Standalone parallel review on a branch diff or PR, with triage.

/multi-agent:complaint-analysis [--file path]multi-agent-complaint-analysis [--file path]

Customer-complaint triage: log evidence per transaction id + read-only repo correlation, then a client/bff root cause with a fix plan and dev prompt, or a routing recommendation to the service owner.

/multi-agent:resume [#id]multi-agent-resume [#id]

Resume a paused/failed task from its last completed phase (file-based state).

/multi-agent:kill [#id]multi-agent-kill [#id]

Delete a task's worktree + branch; logs are preserved.

/multi-agent:prune-logsmulti-agent-prune-logs

Delete per-task logs (audit + metrics kept). Dry-run first.

/multi-agent:garbage-collectmulti-agent-garbage-collect

Sweep leftover /tmp scratch + worktree residue. Dry-run first.

/multi-agent:purgemulti-agent-purge

Full reset - every worktree, branch, state file and log. Double confirmation.

/multi-agent:setupmulti-agent-setup

First-run wizard - tokens, Git identity, per-project routing.

/multi-agent:helpmulti-agent-help

Bilingual usage guide - 47 live commands, the Full/Short depth question, preferences.