Playwright MCP Output Compression (Layer 4)

Purpose

Reduces Playwright MCP tool output tokens by 94-96% using intelligent Haiku summarization while preserving ref= values for interactive flow continuity.

Architecture

MCP tool response (37K+ chars)
  ↓ PostToolUse hook
  ↓ playwright-compress.sh
  ↓ claude -p --model haiku (summarize)
  ↓ updatedMCPToolOutput (1.4K-1.9K chars)

Token Defense Stack Position

Layer	Component	Mechanism	Scope
1	cc-token-saver	Time-based budget alerts	Session
2	R013 Ecomode	Context-aware output compression	Agent
3	MAX_MCP_OUTPUT_TOKENS	Hard truncation (lossy)	Setting
4	playwright-compress	Intelligent summarization (lossless ref=)	Hook

Behavior

• Trigger: PostToolUse on mcp__playwright__.* tools
• Skip condition: Output < 3000 characters (not worth compressing)
• ref= preservation: All ref= attribute values are extracted and preserved in the summary
• Fallback: If Haiku summarization fails, original output is returned unchanged
• Auth: Uses Claude subscription auth (claude -p), no API key needed

Integration

Rule	Interaction
R001	No external data transmission — uses local claude -p
R013	Complements Ecomode (Layer 2) with MCP-specific compression
R021	Advisory PostToolUse hook — never blocks

Hook Configuration

Configured in .claude/hooks/hooks.json PostToolUse section:

{
  "matcher": "mcp_tool_name matches \"mcp__playwright__.*\" || mcp_tool_name matches \"mcp__claude-in-chrome__.*\"",
  "hooks": [{
    "type": "command",
    "command": "bash .claude/hooks/scripts/playwright-compress.sh"
  }],
  "description": "Layer 4: Compress Playwright/Chrome MCP output via Haiku summarization"
}

Source

Adapted from treesoop/claude-native-plugin playwright-optimizer (MIT).