context-management

Context management is a specialized skill designed to optimize the performance and continuity of long-running agentic coding sessions. By providing tools to monitor token consumption in real-time, it enables agents to proactively manage the limited context window of LLMs. This prevents sudden information loss or system-triggered compaction that often disrupts complex software engineering workflows. The skill centers around three primary mechanisms: status monitoring, intelligent snapshotting, and multi-level rehydration.

• Monitors token usage percentages and provides actionable alerts (ok, consider, recommended, urgent) to preemptively avoid context truncation.

• Facilitates the creation of named context snapshots, which package conversation metadata, key decisions, and current state into persistent files located in ~/.amplihack/.claude/runtime/context-snapshots/.

• Supports granular rehydration through three distinct levels: Essential (requirements and current state), Standard (key decisions and open items), and Comprehensive (full metadata, tool history, and detailed logic).

• Integrates seamlessly with agent workflows by allowing users to restore state precisely when needed, such as after automated compaction or when shifting focus between complex tasks.

• Maintains project consistency by preserving high-level requirements and architectural decisions even when the active conversation window is forced to shrink.

• Use this tool when approaching token limits to secure progress before critical data is pruned by the underlying model.

• Ideal for long-duration coding tasks, team handoffs, or when switching between multiple implementation branches where state preservation is critical.

• The rehydration process is highly efficient, allowing users to move from minimal context to comprehensive documentation on demand, ensuring that the model always has exactly the right amount of information to perform its next task without wasting token space.

• Compatible with the broader amplihack ecosystem, complementing existing /transcripts and PreCompact hooks to ensure that no vital engineering context is ever permanently lost.