Context Optimizer CLI

from praisonaiagents import Agent

agent = Agent(name="optimizer", instructions="Optimise agent prompts and configurations.")
agent.start("Optimise this agent's system prompt for better accuracy.")

Configure context optimization behaviour via CLI flags and interactive commands. The user chats in the CLI; context flags compact history before the model hits token limits.

How It Works

Quick Start

Simple Usage

praisonai chat --context-strategy smart --context-auto-compact

Inside the session, compact manually when needed:

> /context compact

With Configuration

praisonai chat \
  --context-strategy smart \
  --context-threshold 0.7 \
  --no-context-auto-compact

Or set defaults in config.yaml:

context:
  auto_compact: true
  compact_threshold: 0.8
  strategy: smart

CLI Flags

Strategy

# Smart optimizer (default, recommended)
praisonai chat --context-strategy smart

# Simple truncation
praisonai chat --context-strategy truncate

# Sliding window
praisonai chat --context-strategy sliding_window

# Summarize old messages
praisonai chat --context-strategy summarize

# Prune old tool outputs
praisonai chat --context-strategy prune_tools

Strategy	Description	Best For
`smart`	Intelligent combination	General use
`truncate`	Remove oldest messages	Fast, simple
`sliding_window`	Keep recent N messages	Conversation flow
`summarize`	Compress old messages	Context preservation
`prune_tools`	Truncate tool outputs	Tool-heavy workflows

Auto-Compaction

# Enable (default)
praisonai chat --context-auto-compact

# Disable
praisonai chat --no-context-auto-compact

Threshold

# Trigger at 80% usage (default)
praisonai chat --context-threshold 0.8

# More aggressive (70%)
praisonai chat --context-threshold 0.7

# Less aggressive (90%)
praisonai chat --context-threshold 0.9

Interactive Commands

Manual Compaction

> /context compact

Output:

Optimizing context...
✓ Optimized: 45,000 → 30,000 tokens
Saved 15,000 tokens (33.3%)
Strategy: smart

View Optimization History

> /context history

Output:

Optimization History
Time                     Event                Tokens       Saved
----------------------------------------------------------------------
2024-01-07T12:00:00      overflow_detected       45,000          -
2024-01-07T12:00:01      auto_compact           45,000     -15,000

View Current Config

> /context config

Shows auto_compact, threshold, and strategy settings.

Environment Variables

export PRAISONAI_CONTEXT_AUTO_COMPACT=true
export PRAISONAI_CONTEXT_THRESHOLD=0.8
export PRAISONAI_CONTEXT_STRATEGY=smart

config.yaml

context:
  auto_compact: true
  compact_threshold: 0.8
  strategy: smart
  compression_min_gain_pct: 5.0
  compression_max_attempts: 3

Strategy Details

Smart (Recommended)

Combines multiple strategies intelligently:

First tries non-destructive pruning
Falls back to sliding window
Uses truncation as last resort

praisonai chat --context-strategy smart

Truncate

Simply removes oldest messages until under budget.

praisonai chat --context-strategy truncate

Sliding Window

Keeps the most recent N messages that fit in budget.

praisonai chat --context-strategy sliding_window

Summarize

Compresses old messages into a summary (requires LLM call).

praisonai chat --context-strategy summarize

Prune Tools

Truncates old tool outputs while preserving recent ones.

praisonai chat --context-strategy prune_tools

Troubleshooting

Context still overflowing

# Lower threshold
praisonai chat --context-threshold 0.6

# Use more aggressive strategy
praisonai chat --context-strategy truncate

Losing important context

# Use smart strategy
praisonai chat --context-strategy smart

# Or increase threshold
praisonai chat --context-threshold 0.9

Auto-compact not triggering

# Check if enabled
> /context config

# Verify threshold
> /context stats

Best Practices

Start with smart strategy

smart combines pruning, sliding window, and summarisation — use it unless you have a specific reason not to.

Set threshold below 1.0

Trigger compaction at 0.7–0.8 so optimisation runs before the model hard-fails on context overflow.

Use /context stats during long sessions

Check token usage before manually compacting — /context history shows what each compaction saved.