📉 Skill: Token Squeezer (Context & Cost Optimizer)

1. When to Trigger (Routing)

• TRIGGER WHEN:
  • The agent is about to load large source files, multi-module directories, or verbose logs into context.
  • The user requests: "optimize context", "compress this file for LLM", "reduce token cost", or "make prompt fit context limit".
• DO NOT TRIGGER WHEN:
  • Exact literal string matching or character-for-character formatting must be preserved (e.g. formatting linters).

2. Fundamental Philosophy

"LLM context windows are expensive and suffer from 'lost in the middle' syndrome. Token Squeezer distills codebases to their semantic essence: structural signatures, type annotations, and core logic, eliminating 60%+ of token waste."

3. Operational Protocol

Phase 1: Compression Strategy Selection

Select one of two modes based on agent needs:

1. interface mode (Maximum Compression, ~70% reduction):
   • Preserves classes, function signatures, docstrings (first line only), and type hints.
   • Replaces function bodies with ... (ellipsis).
   • Ideal for agent architecture exploration and cross-file API discovery.
2. compact mode (Full Logic, ~40% reduction):
   • Preserves all logical execution.
   • Strips inline comments, trailing whitespace, empty lines, and normalizes imports.

Phase 2: Execution

Run the Squeezer script:

python skills/optimization/token-squeezer/scripts/squeezer.py --file <path_to_code.py> --mode interface

Phase 3: Telemetry & Injection

The script emits the compressed payload along with exact token reduction metrics:

• Original tokens vs Compressed tokens
• Percentage saved
• The compressed representation is then fed into the agent's prompt context.

4. Guardrails & Negative Constraints

• ❌ NEVER discard type annotations or return types during interface compression.
• ❌ NEVER mutate runtime logic when using compact mode.