| name | context-engineering |
| description | Dynamic context injection, mode switching (dev/review/research), selective loading, and strategic compaction for token optimization. |
| allowed-tools | Read, Write, Edit, Bash, Grep, Glob |
Context Engineering
Overview
Context management methodology adapted from the Everything Claude Code project. Optimizes context window usage through dynamic injection, mode switching, selective loading, and strategic compaction.
Context Modes
Dev Mode
- Load: architecture docs, active code files, test files, recent changes
- Skip: historical discussions, completed milestones, research notes
- Priority: implementation speed
Review Mode
- Load: code diff, coding standards, security rules, test coverage
- Skip: architecture docs, planning notes, research
- Priority: thoroughness and accuracy
Research Mode
- Load: requirements, existing patterns, external research, alternatives
- Skip: implementation details, test files, CI configs
- Priority: breadth of information
Dynamic Injection
- Detect project context automatically (language, framework, tools)
- Load relevant skills based on detected context
- Inject domain-specific patterns and conventions
- Adjust tool allowlists per context mode
Selective Loading
- Load only files relevant to the current task
- Use glob patterns to scope file reading
- Prioritize recently modified files
- Skip binary files and generated code
Strategic Compaction
- Monitor context token usage
- Suggest compression for resolved/completed items
- Archive to memory files (activeContext, patterns, progress)
- Pre-compaction state preservation
- Automated compaction triggers at token thresholds
Cross-Platform Detection
- Package manager: npm (package-lock.json), pnpm (pnpm-lock.yaml), yarn (yarn.lock), bun (bun.lockb)
- Language: TypeScript (tsconfig.json), Go (go.mod), Python (pyproject.toml), Java (pom.xml)
- Test runner: vitest, jest, pytest, go test
- CI/CD: GitHub Actions, Dockerfile, docker-compose
When to Use
- Session initialization (detect context)
- Before each phase (inject relevant context)
- Token budget warnings (strategic compaction)
- Mode transitions (dev to review to research)
Agents Used
- Used by all agents indirectly through context detection
context-engineering agent for explicit compaction analysis
