context-engine

Context Engine

An agentic orchestration framework for Claude Code. Structured workflows with explicit phase hand-offs, Agent Teams for parallel execution, 23 progressive-disclosure skills, persistent project knowledge, enforced safety hooks, and auto-learning from every session.

Context is the bottleneck, not intelligence. Skills load domain expertise on-demand.

v0.1.7 | MIT License

Quick Start

Option A: Project-Level Install (recommended for teams)

./install.sh /path/to/your/project
cd /path/to/your/project
claude
/init                   # auto-detect project, populate .context/
/research [feature]     # begin first feature

Option B: Plugin Install (recommended for personal use)

# Install plugin
claude plugin add --path ./dist/context-engine-plugin

# In any project:
/context-engine:init

See docs/PLUGIN.md for the full plugin distribution guide.

How It Works

/init ──> /adapt ──> /research ──> /planner ──> /implement ──> /validate ──> commit/PR
(once)   (optional)  researcher    planner      Agent Team     Agent Team
                     subagent      subagent     or subagent    or subagent
                        │             │              │               │
                     NOTES.md      PRP.md      Code + Tests     Learnings
                                                    │            captured
                                [/clear + /resume between phases as needed]

Each command ends with the exact next command to run. Hand-offs are explicit. Commands choose Agent Teams or subagents automatically based on task complexity.

Project Structure

commands/              21 user-facing commands (slash commands)
agents/                6 agent role definitions
skills/                23 progressive-disclosure skills (loaded on-demand)
hooks/                 11 enforcement scripts + hooks.json
.claude/
  agent-memory/        Per-role persistent memory
  instructions/        Shared framework instructions (5 files)
.context/
  architecture/        OVERVIEW.md, TECH_STACK.md, DIRECTORY_MAP.md
  features/            FEATURES.md + per-feature NOTES.md and PRP.md
  patterns/            CODE_PATTERNS.md, ANTI_PATTERNS.md
  decisions/           Architecture Decision Records (ADR-NNN.md)
  errors/              INDEX.md + detail/ for complex errors
  knowledge/           LEARNINGS.md + libraries/ + stack/ + dependencies/
  checkpoints/         MANIFEST.md + CP-NNN snapshots
  metrics/             HEALTH.md (velocity, errors, knowledge, agents, context)
  templates/           PRP and NOTES templates
docs/                  PLUGIN.md, CHEATSHEET.md, WALKTHROUGH.md
dist/                  Built plugin output

Commands

Core Workflow

Command	Phase	Purpose
/init	Setup	Bootstrap .context/, detect tech stack
/adapt	Setup	Context-aware research for new projects
/research [topic]	1	Explore codebase, produce NOTES.md
/planner [notes]	2	Create PRP with testing strategy
/implement [PRP]	3	Execute PRP steps (Agent Team or subagent)
/validate [PRP]	4	Review, simplify, capture learnings, ship prompt

Utility Commands

Command	Purpose
/plan-quick [task]	Quick plan + implement for small tasks
/debug [error]	Diagnose and fix bugs
/refactor [goal]	Restructure code with safety checks
/security-review	Standalone security review
/simplify	Dead code, duplication, over-abstraction pass
/status	Project briefing (add onboard for extended version)
/resume	Reload state after /clear or new session
/cancel	Abandon a feature in progress

Knowledge & Metrics

Command	Purpose
/learn [insight]	Route knowledge to deep reference layer
/knowledge	Browse, search, promote knowledge base
/checkpoint [action]	Create, list, rollback hybrid checkpoints
/health [action]	Metrics dashboard with deep-dive modes
/update-arch	Refresh architecture docs
/create-skill	Create a new skill
/update-skill	Update an existing skill

Orchestration

Two modes. Commands pick automatically based on complexity.

Agent Teams

Used by /implement, /validate, /debug, /refactor when work has 3+ parallel tracks. Teammates work independently on owned files, communicate directly, and coordinate via shared task list with dependency ordering.

Subagents

Used by /research, /planner, and as fallback for small tasks. Single specialist works in isolated context, returns a summary.

Agent Roles