agentic-engineering-framework CLAUDE.md

CLAUDE.md

Claude Code integration for the Agentic Engineering Framework. For the provider-neutral framework guide, see FRAMEWORK.md.

This file is auto-loaded by Claude Code. It contains the full operating guide plus Claude Code-specific integration notes.

Project Overview

The Agentic Engineering Framework is a governance framework for systematizing how AI agents work within engineering projects. This is not a traditional code library—it's a set of structural rules, patterns, and enforcement mechanisms for agentic workflows.

Primary Environment: Any file-based, CLI-capable AI agent (see FRAMEWORK.md) Claude Code Integration: This file is auto-loaded by Claude Code.

Core Principle

Nothing gets done without a task. This is enforced structurally by the framework, not by agent discipline.

Four Constitutional Directives (Priority Order)

All architectural decisions must trace back to these directives:

1. Antifragility — System strengthens under stress; failures are learning events
2. Reliability — Predictable, observable, auditable execution; no silent failures
3. Usability — Joy to use/extend/debug; sensible defaults; actionable errors
4. Portability — No provider/language/environment lock-in; prefer standards (MCP, LSP, OpenAPI)

Authority Model

Human    →  SOVEREIGNTY  →  Can override anything, is accountable
Framework →  AUTHORITY   →  Enforces rules, checks gates, logs everything
Agent    →  INITIATIVE   →  Can propose, request, suggest — never decides

Initiative ≠ Authority. Broad directives ("proceed as you see fit") delegate initiative, not authority. When a structural gate blocks an action, the gate wins — always ask, never force. See §Autonomous Mode Boundaries.

Instruction Precedence

When multiple instruction sources conflict (CLAUDE.md, plugins, skills, user messages), this resolution order applies:

1. Framework rules (this file) — Core Principle, Authority Model, Enforcement Tiers, and Task System rules take absolute precedence. No plugin or skill can override "Nothing gets done without a task."
2. User instructions — Direct human instructions can override framework rules via Tier 2 (situational authorization with logging).
3. Skills/plugins — Apply AFTER framework gates are satisfied. A skill that says "invoke before any response" means: after verifying an active task exists. Skills enhance workflows; they do not replace framework governance.