@adamjen/pi-compact-fast

This Skill addresses the issues of high token consumption and slow speed when Pi users rely on expensive primary models for session compression. In practical development, developers frequently need to compress long conversations to free up context windows, but the cost of using primary models (e.g., GPT-4) for compression is high, and the speed is limited by model inference efficiency.

Usage is simple: just enter the /compact-fast command in Pi. It will automatically locate the configured qwen-35b-moe model (or another custom model), generate a session summary via an independent API call, and return the structured summary to Pi's built-in compression system. You can also switch the model used by modifying the COMPACT_MODEL_ID in extensions/index.ts.

It is ideal for Pi users who frequently manage long conversations, especially developers using high-cost models (e.g., GPT-4, Claude 3 Opus). Agent developers and backend engineers who accumulate substantial context during coding and need regular compression to maintain session continuity will see significant token savings and efficiency gains.

It is recommended to use this tool consistently after configuring a fast model (like qwen-35b-moe) in models.json. Note that it depends on the Pi platform and the @mariozechner/pi-ai library, and the compression model must be available in the local or remote configuration. If the fast model fails to respond, the command will exit gracefully without affecting the current session.