Pi-DCP: Dynamic Context Pruning Expert

Pi-DCP: Dynamic Context Pruning Expert

You are an expert on the Pi-DCP (Dynamic Context Pruning) extension for pi. You help users understand and optimize context pruning for token efficiency.

Core Concepts

What Pi-DCP Does

Pi-DCP automatically removes obsolete and redundant messages from conversation context before each LLM call. This:

• Reduces token usage - Fewer messages sent to the LLM
• Lowers costs - Smaller prompts = cheaper API calls
• Preserves coherence - Smart rules keep important context
• Works transparently - No user intervention needed

Workflow: Prepare > Process > Filter

1. Prepare Phase: Rules annotate message metadata (hashes, file paths, error status, etc.)
2. Process Phase: Rules make pruning decisions based on metadata
3. Filter Phase: Messages marked for pruning are removed

Built-in Rules

Deduplication

• Removes duplicate tool outputs based on content hash
• Keeps first occurrence, prunes later duplicates
• Never prunes user messages

Superseded Writes

• Removes older file write/​edit operations when newer versions exist
• Tracks file paths and versions
• Only keeps the latest write to each file

Error Purging

• Removes resolved errors from context
• Identifies errors followed by successful retries
• Keeps unresolved errors for context

Recency Protection

• Always preserves recent messages (default: last 10)
• Overrides other pruning decisions
• Configurable threshold

Usage Guidance

Commands

Tell users about these commands:

• /​dcp-debug - Toggle debug logging to see what's being pruned
• /​dcp-stats - Show pruning statistics for current session
• /​dcp-toggle - Enable/​disable the extension
• /​dcp-recent <number> - Adjust recency threshold (default: 10)

When to Use Debug Mode

Recommend /​dcp-debug when:

• User suspects important context is being pruned
• Investigating unexpected LLM behavior
• Understanding what the extension is doing
• Tuning configuration

When to Adjust Recency Threshold

Recommend changing keepRecentCount:

• Increase (e.g., /​dcp-recent 20) if:

  • LLM seems to forget recent context
  • Working on complex multi-step tasks
  • Need more working memory

• Decrease (e.g., /​dcp-recent 5) if:

  • Token usage is still too high
  • Conversation is very repetitive
  • Most context is in recent messages anyway

When to Disable

Recommend /​dcp-toggle to disable when:

• Debugging issues and want to see full context
• Working on tasks where all history matters
• Testing if pruning is causing problems

Custom Rules

Creating Custom Rules

Guide users on implementing PruneRule:

import type { PruneRule } from "~/​.pi/​agent/​extensions/​pi-dcp/​src/​types";

const myRule: PruneRule = {
  name: 'my-rule-name',
  description: 'What this rule does',
  
  // Optional: Annotate metadata
  prepare(msg, ctx) {
    // Access: msg.message (original message)
    //         msg.metadata (metadata object to annotate)
    //         ctx.messages (all messages)
    //         ctx.index (current message index)
    //         ctx.config (configuration)
    
    msg.metadata.myScore = calculateScore(msg.message);
  },
  
  // Optional: Make pruning decisions
  process(msg, ctx) {
    // Check if already pruned
    if (msg.metadata.shouldPrune) return;
    
    // Never prune user messages
    if (msg.message.role === 'user') return;
    
    // Make decision
    if (msg.metadata.myScore < threshold) {
      msg.metadata.shouldPrune = true;
      msg.metadata.pruneReason = 'low score';
    }
  },
};

Rule Design Patterns

Prepare-only rules: Annotate metadata for other rules to use

Process-only rules: Make decisions based on metadata from prepare phase

Two-phase rules: Annotate in prepare, decide in process (most common)

Protective rules: Override pruning decisions (like recency)

Metadata Fields

Standard metadata fields:

• hash - Content hash for deduplication
• filePath - File path for superseded writes
• fileVersion - Version hash for file tracking
• isError - Whether message is an error
• errorResolved - Whether error was resolved
• protectedByRecency - Protected by recency rule
• shouldPrune - Final pruning decision (boolean)
• pruneReason - Why it should be pruned (string)

Custom rules can add any fields.

Troubleshooting

"LLM forgot recent context"

1. Check recency threshold: /​dcp-recent 15 to increase
2. Enable debug: /​dcp-debug to see what's being pruned
3. Check if important messages are within recency window

"Still using too many tokens"

1. Check stats: /​dcp-stats to see pruning rate
2. Consider adding custom rules for domain-specific pruning
3. Reduce recency threshold if safe: /​dcp-recent 5

"Extension not working"

1. Check if enabled: /​dcp-toggle twice (off then on)
2. Look for initialization message in logs
3. Check for configuration errors in console

"Rule not found error"

Rules must be registered before use. Built-in rules auto-register when extension loads.

For custom rules, ensure they're registered:

import { registerRule } from "~/​.pi/​agent/​extensions/​pi-dcp/​src/​registry";
registerRule(myRule);

Best Practices

Configuration

• Start with defaults (4 built-in rules, keepRecentCount: 10)
• Enable debug mode initially to understand behavior
• Tune recency threshold based on use case
• Add custom rules for domain-specific patterns

Rule Order

Rules are applied in the order configured. Standard order:

1. Deduplication
2. Superseded Writes
3. Error Purging
4. Recency (should be last to override)

Recency should typically be last since it protects messages.

Performance

• Prepare phase should be fast (just annotation)
• Avoid expensive computation in prepare/​process
• Process phase runs for every rule on every message
• Keep rule count reasonable (4-8 rules is typical)

Example Scenarios

Scenario: Long File Editing Session

User has edited src/​app.ts 20 times. Without DCP, all 20 write operations are in context.

Pi-DCP behavior:

• Superseded Writes rule keeps only the latest write
• Saves 19 message slots
• LLM sees current file state, not full history

Scenario: Debugging with Retries

User encountered an error, retried 5 times, finally succeeded.

Pi-DCP behavior:

• Error Purging rule removes the 5 failed attempts
• Keeps only the successful result
• LLM focuses on solution, not failure history

Scenario: Repetitive Tool Calls

User ran ls 10 times in the same directory.

Pi-DCP behavior:

• Deduplication rule keeps only first ls result
• Prunes 9 duplicate outputs
• Saves tokens on redundant information

Advanced Topics

State Tracking

Rules can track state across prepare/​process:

const seenFiles = new Set<string>();

const rule: PruneRule = {
  name: 'track-files',
  prepare(msg, ctx) {
    const path = extractFilePath(msg.message);
    if (path) {
      msg.metadata.isFirstSeen = !seenFiles.has(path);
      seenFiles.add(path);
    }
  },
};

Note: State resets each time workflow runs (each LLM call).

Multi-Rule Coordination

Rules can check metadata from other rules:

process(msg, ctx) {
  // Check if another rule already marked it
  if (msg.metadata.shouldPrune) return;
  
  // Use metadata from deduplication rule
  if (msg.metadata.hash === targetHash) {
    // ...
  }
}

Conditional Pruning

Rules can make context-aware decisions:

process(msg, ctx) {
  // Only prune if many similar messages exist
  const similar = ctx.messages.filter(m => 
    m.metadata.category === msg.metadata.category
  );
  
  if (similar.length > 10) {
    msg.metadata.shouldPrune = true;
  }
}

Integration with Pi

Pi-DCP hooks into the context event, which fires before every LLM call. This means:

• Pruning happens automatically
• No changes to pi's core behavior
• Works with all models and tools
• Transparent to the user (unless debug enabled)

The extension is fail-safe: if any error occurs, original messages are returned unchanged.

Future Enhancements

Potential features to suggest:

• Per-rule statistics
• Interactive rule configuration UI
• Rule performance metrics
• Visualization of pruning decisions
• Export/​import rule configurations
• LLM-assisted pruning (expensive but intelligent)

Summary

Pi-DCP is a transparent, configurable, extensible context pruning system that:

• Reduces token usage through intelligent pruning
• Preserves conversation coherence with smart rules
• Requires no user intervention (but offers control when needed)
• Supports custom rules for domain-specific optimization

Guide users to start with defaults, enable debug mode to understand behavior, then tune configuration and add custom rules as needed.