Token Economy

Canopy includes several built-in optimizations to reduce token usage without sacrificing quality.

System prompts

The three system prompts (default.txt, anthropic.txt, gemini.txt) were rewritten to be:

• 66% shorter — Same instructions, fewer tokens
• Zero dead references — No outdated OpenCode features mentioned
• Canopy-specific — Model routing instructions, skill references

Tool descriptions

Tool descriptions are sent to the model every turn. Canopy shortened the most verbose ones:

Tool	Before	After	Savings
task.txt	2,305 bytes	489	-79%
todowrite.txt	2,012 bytes	366	-82%
edit.txt	1,369 bytes	347	-75%

Total: ~30% reduction in tool description tokens per turn.

Context compaction

The compaction system aggressively manages context window usage:

• Tool output truncation: 2,000 → 1,500 chars (-25%)
• Protected context: 40,000 → 30,000 tokens
• Protected tools: read, grep, glob outputs preserved during pruning
• Recent context: 8,000 → 4,000 tokens kept verbatim

Configuration

{
  "compaction": {
    "auto": true,
    "tail_turns": 3
  },
  "tool_output": {
    "max_lines": 200,
    "max_bytes": 6000
  }
}