Token Optimizer

Solving Context Bloat in Multi-turn Conversations

In complex AI agent workflows, multi-turn conversations often lead to rapid context bloat. Verbose historical messages, redundant user inputs, and unconstrained model outputs not only increase token costs but also dilute the LLM's attention, reducing the success rate of core tasks.

Core Capabilities and Execution Mechanisms

The skill employs a structured compression and filtering mechanism to manage conversation context, featuring four key dimensions:

• Session State Management: Defaults to OFF. Once activated, it remains active for the entire session. Users can trigger a one-time skip (e.g., "skip this time") or a full deactivation via natural language.
• Intent Distillation: When conversations exceed a certain number of turns or context becomes cluttered, it extracts the current true task intent, filters out irrelevant historical discussions, and distills them into a single intent statement.
• Input Compression: Removes duplicate information and decorative filler (e.g., "maybe", "I think"), converting verbose natural language descriptions into high-density structured data like JSON or tables. It performs quality checks to ensure edge cases are not lost.
• Output Constraints: Applies specific token-saving directives based on output types. For example, code outputs skip boilerplate explanations and redundant comments; explanatory outputs are limited to 3 core points; data reports show only the Top 10 with a link to full data.

Boundaries and Safety Constraints

This mechanism is not suitable for all scenarios and must strictly adhere to safety limits:
- No Compression of Sensitive Data: Never remove security-sensitive information (e.g., API keys, credentials) or critical error-handling code to save tokens.
- Avoid Over-compression: Compression must not result in ambiguity. It must be performed while preserving rare scenarios (edge cases) and user preference constraints.
- Mandatory Integrity Validation: Before finalizing the optimization, validation scripts (e.g., validate_completeness.py) must be run to ensure the optimized input still contains all required information to complete the task.
- Exemption for Detail-heavy Tasks: When users require complete API documentation or detailed diagnostic logs, the one-time skip mechanism should be triggered to maintain full output.