Skip to content

TSCG v1.4.3 - Hotfix Release (Dependency Refs)

Latest
Latest

Choose a tag to compare

View all tags
@SKZL-AI SKZL-AI released this 26 Apr 15:58
· 8 commits to main since this release
v1.4.3
734468a

What

v1.4.2 was published to npm with workspace-style dependency references (workspace:* and file:../core) that work in monorepo development but break npm install for end users. v1.4.3 fixes the dependency metadata across all 4 packages.

All v1.4.2 features, logic, and benchmark-validated behavior are unchanged.

Empirical Validation: 20,000+ API Calls Across 13+ LLMs

TSCG is backed by one of the largest independent tool-schema compression benchmarks in the ecosystem:

  • 20,000+ real API calls across controlled A/B experiments
  • 13+ language models tested: Claude Opus 4.7, Claude Sonnet 4, GPT-4o, GPT-4o-mini, GPT-5.2, GPT-5.4, GPT-5.5, Qwen3-32B, Qwen3-14B, Phi-4-mini, Llama 3.1 8B, Gemma 3 12B, DeepSeek-v3
  • 3 frontier API providers: Anthropic, OpenAI, Ollama (local)
  • 7 sub-15B open-weight models validated for local deployment
  • Per-operator isolation sweeps: 9-condition leave-one-in probes identifying per-model operator sensitivity
  • Combination-effect detection: discovery of super-additive negative interactions (Scenario B) in GPT-5.5
  • External validation: BFCL benchmark (108-181% accuracy retention), ToolBench, API-Bank

Key findings:

  • 45-62% token savings with accuracy retention or improvement across model families
  • Per-model operator sensitivity is real: GPT-5.4 is config-robust (CFO +15pp), GPT-5.5 is combination-fragile
  • Small models (4B-14B) benefit from compression even more than frontier models

See findings/ for the complete empirical research library.

Impact

If you installed any @tscg/* package at v1.4.2, you may have hit errors like:

  • npm ERR! Could not resolve workspace:* dependency
  • npm ERR! ENOENT file:../core

v1.4.3 resolves this:

npm install @tscg/openclaw@1.4.3

What Changed (v1.4.3 vs v1.4.2)

Metadata-only fix:

  • @tscg/openclaw: dep ref file:../core -> ^1.4.3
  • @tscg/mcp-proxy: dev-dep ref workspace:* -> ^1.4.3
  • @tscg/tool-optimizer: dev-dep ref workspace:* -> ^1.4.3
  • All packages: peer-dep refs bumped -> ^1.4.3

No source code changes. No behavioral differences from v1.4.2.

v1.4.2 Features (All Included in v1.4.3)

  • Per-operator adaptive sweep (tune --sweep) -- 9-condition isolation testing (180 calls, ~USD 1)
  • Combination-effect detection (Scenario B) -- automatic fallback for fragile model configurations
  • selectOptimalProfile() -- derives optimal operator set from per-operator deltas with classification thresholds
  • mcp-proxy cacheReader integration -- adaptive profiles from openclaw sweep override static profiles at runtime
  • GPT-5.4 + GPT-5.5 static profiles in mcp-proxy, derived from 440-call benchmark
  • 459+ tests across all packages

Deprecated

v1.4.2 on npm has been deprecated. Users should upgrade to v1.4.3.

Assets 2
Loading