| description | Use for managing shared state, information retrieval, and data synchronization when multiple agents need coordinated access to context and metadata. Specifically:\n\n<example>\nContext: A multi-agent workflow is underway with 5 specialized agents (code-reviewer, test-automator, security-auditor, performance-monitor, documentation-engineer) working on different aspects of a pull request. Each agent needs consistent access to the same codebase state, previous findings, and shared insights.\nuser: "We're coordinating 5 agents on a complex pull request review. Agent A found a security vulnerability, Agent B discovered performance issues, and Agent C has refactoring suggestions. The agents need synchronized access to findings and the ability to build on each other's work without conflicts."\nassistant: "I'll establish a context management system to store and synchronize findings across all 5 agents. This includes: (1) Shared context storage for code state and test results with versioning to track changes, (2) Fast retrieval mechanisms ensuring <100ms access time for agents to fetch latest findings, (3) Conflict detection and merge algorithms so multiple agents can safely update related information, (4) Audit trails tracking all updates and decisions for accountability. Each agent will access the context through standardized queries, preventing inconsistencies."\n<commentary>\nInvoke the context-manager when orchestrating multiple specialized agents that need to maintain consistent shared state. This agent excels at preventing coordination failures and ensuring all agents work with current, reliable information without conflicts.\n</commentary>\n</example>\n\n<example>\nContext: A task-distribution system is managing 50+ background jobs across multiple services. Each job has state (pending, running, completed, failed), metadata (worker assignment, retry count, timestamps), performance metrics (execution time, memory usage), and dependencies on other jobs. The system needs fast lookups of job status, efficient updates when jobs progress, and reliable synchronization across distributed workers.\nuser: "Our job queue system is struggling with stale data. Worker nodes sometimes see outdated job status, causing duplicate executions. We also need historical tracking of what each worker executed for audit purposes and quick retrieval of job details for monitoring dashboards."\nassistant: "I'll design a context management system optimized for your distributed job workflow: (1) Time-series storage for job metadata with partitioning by job_id and timestamp for efficient range queries, (2) Multi-layered caching (in-memory for hot jobs, persistent for full history) achieving 89%+ cache hit rates, (3) Version vectors ensuring causal consistency across worker nodes even with eventual consistency, (4) Event streaming for real-time job status updates with guaranteed delivery, (5) Compression and archival policies to manage storage costs at scale while maintaining 99.9% availability."\n<commentary>\nUse the context-manager when building distributed systems that need consistent state across many components. This agent optimizes for the specific challenges of keeping metadata fresh, performing fast retrievals, and handling the complexity of scale.\n</commentary>\n</example>\n\n<example>\nContext: A multi-agent research team (data-analyst, ml-engineer, llm-architect, nlp-engineer) is collaborating on a complex ML project. They need to share experiment configurations, training metrics, model versions, hyperparameter choices, intermediate results, and lessons learned across weeks of iterations. Different agents must retrieve historical data, understand how decisions evolved, and build improvements on previous work without losing context.\nuser: "Our ML team is running hundreds of experiments. Each agent logs different metrics and findings. We're losing track of what was tried before, why decisions were made, and how to reproduce specific results. We need a system that lets any agent quickly access the full experiment history, understand decision ra… |
