reload

The reload skill is a specialized utility designed for Claude Code power users who manage long-running, resource-intensive development sessions. As Claude Code accumulates metadata, large tool outputs, thinking blocks, and duplicate file reads, context bloat can lead to degraded performance, high API token consumption, and eventual context overflow. The reload command streamlines the recovery process by executing a comprehensive session treatment—which prunes unnecessary data while preserving vital agent history—and subsequently spawning a clean, auto-resume watcher in a new terminal window to maintain workflow continuity.

This skill is ideal for engineers who maintain complex agent teams or long-duration projects where session state needs to be reset periodically without losing the underlying progress or behavioral digest. By automating the transition from a 'bloated' session to a 'reloaded' session, it protects your workspace from context loss and ensures consistent interaction quality. Users simply invoke the command, confirm the token-saving metrics provided by the diagnostic check, and allow the daemon to manage the handoff.

• Performs a holistic session diagnosis using cozempic current --diagnose before any modifications to identify bloat points.

• Applies customizable prescription tiers (gentle, standard, or aggressive) to balance pruning intensity with context retention.

• Automatically spawns an auto-resume watcher that detects the exit signal and restores your session environment seamlessly.

• Implements session compaction without manual intervention, removing redundant items like progress ticks, stale file reads, and oversized thinking blocks.

• Guards against session corruption and orphaned tool results while keeping your behavioral digest and team state checkpoints intact.

• Always run the command without --execute flags first; the reload action handles the treatment internally, so avoid redundant executions of treat --execute to prevent double-pruning.

• Use the argument hint (gentle, standard, or aggressive) to tailor the pruning level based on current token usage and model threshold limits.

• The workflow expects an /exit command post-treatment, which triggers the watcher to launch the new, optimized terminal instance.

• This tool is designed for users of the Claude Code CLI who rely on continuous agent interaction and need to manage the 200K or 1M context windows effectively.