Download the PHP package rhazen35/agent-orchestration without Composer

What is this?

This package provides a complete Factory Pipeline workflow for AI agents. It includes skills, steering rules, templates, and governance guardrails — all as plain markdown files that any agent can read, regardless of language or tooling.

No runtime code. No dependencies. Works with PHP, Go, Rust, Python, Node, or anything else.

The Factory Pipeline

Every phase is manually triggered. Nothing runs automatically. You stay in control, and your token budget stays intact.

🧠 Phase 1 — Design

Trigger: "I want to build X" or "Start a new spec for Y"

The spec-architect skill orchestrates sub-skills in sequence:

Output: signed-off plan/{namespace}/plan.md Gate: you must approve the plan before anything enters the inbox.

📥 Phase 2 — Queue

Trigger: "queue tasks" or "load inbox"

Tasks are loaded into inbox/{namespace}/tasks.md in batches of max 3 (the Hydration Limit). The next batch only loads after a slot opens. A flawed plan fails fast and cheap, not after 20 API calls.

Gate: confirm the inbox before production starts.

⚙️ Phase 3 — Production

Trigger: "process next task"

The task-processor runs with JIT context — it only reads the task, the file it's editing, and steering/principles.md. Nothing else. Small context window, fast and cheap calls.

Output: task moved to outbox/{namespace}/pending-review.md

🛡️ Phase 3.5 — Sentinel Pre-Check

Trigger: runs as part of "review", before the quality reviewer

The sentinel catches cheap failures before they reach the expensive inspector:

• Syntax errors or lint failures
• Ghost comments (TODO, FIXME, unfinished notes)
• Changelog format errors
• Scope creep (files touched outside the task's scope)

Fail: task returns to inbox/ with [SENTINEL_FAIL]. Quality reviewer is never called.

🔍 Phase 4 — Quality Control

Trigger: "review"

The quality-reviewer performs deep reasoning against acceptance-criteria.md and steering/principles.md. Also runs integrity-check for ghost features and architectural drift.

If a task is rejected 4 times, the pipeline locks. Human intervention required.

📦 Phase 5 — Archiving

Trigger: "archive"

The changelog-architect skill:

1. Appends an entry to CHANGELOG.md under ## [Unreleased]
2. Appends the task + review notes to archive/{namespace}/decision-log.md
3. Cleans up outbox/{namespace}/completed.md
4. Invokes monitor to update dashboard.md

Quick Reference

Installation

This package contains no runtime code — only markdown files consumed by your AI agent. It works with any language or stack.

PHP — Composer

Copy the directories into your project root:

Node — npm

Go, Rust, Python, or any other language — Git Submodule

Point your agent tooling at .agent/steering/, .agent/skills/, etc.

Update to the latest version:

Clone a project that uses this as a submodule:

Quick Start (no package manager)

Namespacing

Every specification gets its own namespace folder — a short kebab-case name applied consistently across all directories.

Governance

steering/governance.md defines hard limits no agent can override:

Dashboard

dashboard.md is the control tower. Check it before starting any work.

Skills Reference

Steering Files

Always loaded into every agent context:

MCP Integration

This package is designed to work with Model Context Protocol (MCP) servers out of the box. MCP tools extend the agent's capabilities beyond local file editing — enabling live database inspection, web search, API fetching, and more.

How it works

The agent does not hardcode any MCP servers. Instead, it performs a handshake at session start:

1. Agent calls list_tools to discover what the host environment provides.
2. Agent matches the returned tool schemas against steering/mcp-policy.md.
3. Authorized tools are used purposefully — local files always take precedence.

Capability Requirements

The factory expects the host to provide tools for these capabilities:

If a capability is missing, the pipeline continues — it never blocks on absent tools.

Tool Selection Priority

When multiple tools fulfill the same capability, the agent picks in this order:

1. Local filesystem tools (lowest cost)
2. Specialized local servers (e.g., local Postgres, SQLite)
3. Remote API servers (e.g., GitHub, Brave Search, Slack)

The Local-First Rule

Before calling any MCP tool, the agent must check:

1. specification/{namespace}/ — does the spec already answer this?
2. archive/mcp-data/{namespace}/ — has this tool been called and cached?
3. cache/discovery-report.md — is this covered by the last discovery run?

Only if all three are empty or stale does the agent proceed with a tool call. Tool outputs are persisted to archive/mcp-data/ so the worker never re-calls a tool for the same data.

MCP in the Pipeline

Relevant Files

Customizing for Your Project

Every skill, steering file, and template can be overridden at the project level without touching the package files. The agent checks .agent/ first before falling back to the package defaults.

To extend rather than replace, start your override file with:

Two files should never be overridden: steering/workflow.md (the pipeline contract) and steering/overrides.md (the override system itself).

See steering/overrides.md for the full resolution order, a complete list of overridable files, and guidance on what to override with caution.

License

MIT — see LICENSE.