context-management_skill

Overview

This skill helps manage the context window, survive automatic compaction, and persist actionable state across turns and sessions. It provides disciplined task-based state management, delegation patterns, and pre-compaction checklists so agents can resume work without losing progress or decisions.

How this skill works

The skill treats Tasks as the persistent memory layer: create, update, list, and get tasks to record progress, decisions, blockers, and agent handoffs that survive compaction. It recommends delegating heavy exploration or file-reading to subagents so the main context stays small, and offers patterns for saving/restoring cross-session state via an episodic memory tool when available.

When to use it

• Planning long-running or multi-step tasks that need resumability
• Coordinating multiple subagents or parallel workstreams
• Approaching context limits (slower responses, repetition, degraded reasoning)
• When you must preserve decisions, intermediate state, or handoffs through compaction
• Orchestrating workflows that will be interrupted or handed off

Best practices

• Use Tasks for any work over 2–3 steps and call TaskUpdate before significant actions
• Keep exactly one task marked in_progress at a time and mark tasks completed immediately
• Embed key decisions and outcomes in task descriptions so they survive compaction
• Delegate exploration and heavy file reading to subagents; record their IDs in task metadata
• Run the pre-compaction checklist when you see degraded responses or repetition
• Create blocker tasks and link them with blockedBy rather than marking progress incorrectly

Example use cases

• Implementing a token refresh flow: record current progress, decisions (library/algorithms), remaining work, and file pointers in the in_progress task
• Coordinating a cross-team review: create reviewer tasks with agentId metadata and link dependencies via blockedBy
• Running parallel audits: dispatch background agents for security and performance reviews and record outputs as task updates
• Recovering after compaction: use TaskList/TaskGet to find the in_progress task and continue from recorded details
• Saving multi-day work: persist full task state to episodic memory before pausing and restore on resume