@adamjen/pi-compact-fast

/compact-fast command for Pi — compact your session using a fast local model instead of your main conversation model.

What it does

Pi normally uses your current conversation model for compaction (summarizing the session). This extension adds /compact-fast which:

1. Looks up qwen-35b-moe from your configured models (models.json)
2. Uses that model directly via a separate API call to generate the summary
3. Returns the structured summary to Pi's built-in compaction system

This saves tokens on expensive models and speeds up compaction when using a smaller/faster local model.

Install

pi install npm:@adamjen/pi-compact-fast@latest

Or try without installing:

pi -e npm:@adamjen/pi-compact-fast

Usage

/compact-fast

That's it — Pi will compact the session using qwen-35b-moe instead of your current model.

Configure a different model

Edit extensions/index.ts and change line 24:

const COMPACT_MODEL_ID = "your-model-id";

The model must be defined in your ~/.pi/agent/models.json or provider config.

How it works

1. Intercepts session_before_compact event
2. Makes a direct API call to the target model via completeSimple() from @mariozechner/pi-ai
3. Returns { compaction: { summary, firstKeptEntryId, tokensBefore, details } } which Pi uses for compaction

Prompt structure (mirrors pi's native compaction)

• System prompt tells model to ONLY summarize, not continue conversation
• Conversation comes BEFORE instructions — the model follows the last pattern
• Previous summary in <previous-summary> tags between conversation and instructions (for incremental updates)
• Split turns generate two summaries merged with separator (mirrors pi's native behavior)
• File operations tracked in both summary text AND details object

Changelog

2.0.0 — Major rewrite

• ✅ Guard against empty content — aborts cleanly instead of writing garbage "empty conversation" summary
• ✅ File operations tracking — <read-files> and <modified-files> tags appended to summary, plus details object for pi's internal tracking
• ✅ Split turn handling — parallel dual summaries with proper merge separator when a single turn exceeds keepRecentTokens
• ✅ Correct token budgets — dynamic from settings.reserveTokens * 0.8 / * 0.5 instead of hardcoded 8192
• ✅ Added TURN_PREFIX_SUMMARIZATION_PROMPT for split-turn prefix summaries (different format: "Original Request", "Early Progress")
• ✅ Switched from complete() to completeSimple() — matches pi's internal API usage
• ✅ Added stopReason === "error" check — handles API failures gracefully instead of producing empty strings
• ✅ Removed debug logging (console.error)
• ⚠️ Default model changed from qwen3.6-35b to qwen-35b-moe (configure your own in models.json)

1.x — Initial release

• Basic compaction using a fast local model via direct API call

License

MIT