tscg-openclaw

name	@tscg/openclaw
version	1.4.1
description	TSCG Tool-Schema Compression for OpenClaw agents
author	TSCG Team
license	MIT
tags	["tscg","tool-compression","openclaw","optimization"]
minHostVersion	2026.0.0

Agent Instructions

You are a TSCG-enhanced OpenClaw agent.

Your tool definitions are automatically compressed using TSCG (Tool-Schema Compression Grammar). This reduces token usage by 40-65% with <2pp accuracy impact.

The @tscg/openclaw plugin handles compression transparently via the beforeToolsList hook. If you notice unusual tool behavior, the original uncompressed definitions are always preserved as fallback.

Profile Resolution

Your model is matched to an optimal compression profile using a 4-tier resolution strategy:

1. Cache -- custom profile from self-tuning (highest priority)
2. Static -- built-in profile map for known models
3. Size-heuristic -- parameter-count-based archetype for unknown models
4. Fallback -- conservative SDM-only profile (always safe)

Available Profiles

Model	Archetype
claude-opus	hungry
claude-sonnet	robust
claude-haiku	balanced
gpt-4	sensitive
gpt-5	robust
qwen3 / phi4 / llama3.1 / gemma3 / mistral	small-model
deepseek-v3 / deepseek-r1	conservative

Self-Tuning

For best results, run tscg-openclaw tune --model <your-model> to create a custom profile tailored to your specific model's compression tolerance.

References

Reference documents are available in the references/ directory:

• self-tuning.md -- Guide to quick vs full benchmarks and cost optimization
• benchmarks.md -- 720-call benchmark results and operator impact data
• model-profiles.md -- Static profile map and size heuristic details
• troubleshooting.md -- Common issues and fixes