Feature: Importance-weighted truncation #24

New issue

Open

opened 2025-12-11 00:06:47 +00:00 by blightbow · 0 comments

blightbow commented

2025-12-11 00:06:47 +00:00

Owner

Overview

Implement truncation that preserves narrative coherence by scoring importance rather than using simple FIFO.

Truncation Strategy

When perception stream exceeds token budget:

Never truncate: Most recent N events (configurable, default 10)
Summarize first: Events with narrative_text longer than threshold
Drop by importance: Remove lowest importance events first
Preserve causality: Keep action→result pairs together

Algorithm

def truncate_perception_stream(stream, max_tokens, model):
    """
    Truncate stream while preserving coherence.

    Priority order:
    1. Recent events (always keep)
    2. High importance events
    3. Self-action chains (keep together)
    4. Older/lower importance events (drop first)
    """

Importance Preservation Rules

Self-action + result pairs stay together
@mention events preserved longer than ambient
Causal chains (trigger → response) kept intact

Files to Modify

llm_interaction.py - Truncation logic in prompt assembly
api/serializers/workbench.py - Update for new model
Tests for truncation edge cases

Phase

Phase 4 of Milestone #4

## Overview Implement truncation that preserves narrative coherence by scoring importance rather than using simple FIFO. ## Truncation Strategy When perception stream exceeds token budget: 1. **Never truncate**: Most recent N events (configurable, default 10) 2. **Summarize first**: Events with `narrative_text` longer than threshold 3. **Drop by importance**: Remove lowest importance events first 4. **Preserve causality**: Keep action→result pairs together ## Algorithm ```python def truncate_perception_stream(stream, max_tokens, model): """ Truncate stream while preserving coherence. Priority order: 1. Recent events (always keep) 2. High importance events 3. Self-action chains (keep together) 4. Older/lower importance events (drop first) """ ``` ## Importance Preservation Rules - Self-action + result pairs stay together - @mention events preserved longer than ambient - Causal chains (trigger → response) kept intact ## Files to Modify - `llm_interaction.py` - Truncation logic in prompt assembly - `api/serializers/workbench.py` - Update for new model - Tests for truncation edge cases ## Phase Phase 4 of Milestone #4