v1.4.1 — Umbrella Release: Per-Model MCP Proxy
What's New
Coordinated umbrella release — all three @tscg/* packages at v1.4.1.
@tscg/mcp-proxy — Per-Model Target Resolution
--target=claude-opus-4-7auto-enables full compression with operator archetype (hungry/robust/sensitive)--target=claude-sonnet-4for Sonnet-optimized compression- Backward compatible:
createProxy({})still works (description-only mode) - 61 tests passing
@tscg/core — Regression Fix
- Fix
hasExplicitPrinciplesregression when custom principles array is provided - 47 tests passing
@tscg/tool-optimizer — Version Sync
- Umbrella version sync to 1.4.1, no functional changes
Benchmark Validation
480-Call MCP Proxy Benchmark (n=40 per cell, 2 seeds)
| Model | Tools | Baseline | TSCG Proxy | Δ Accuracy | Token Savings |
|---|---|---|---|---|---|
| Opus 4.7 | 16 | 70.0% | 75.0% | +5.0pp | 53.1% |
| Opus 4.7 | 43 | 75.0% | 75.0% | ±0.0pp | 55.8% |
| Opus 4.7 | 50 | 77.5% | 77.5% | ±0.0pp | 55.5% |
| Sonnet 4 | 16 | 80.0% | 77.5% | -2.5pp | 53.1% |
| Sonnet 4 | 43 | 85.0% | 82.5% | -2.5pp | 55.8% |
| Sonnet 4 | 50 | 77.5% | 77.5% | ±0.0pp | 55.5% |
Tool-Optimizer E2E (30 calls, Sonnet 4 @ 16 tools)
withTSCG(): 86.7% vs baseline 80.0% (+6.7pp), 36.6% character savings
External Validation (4 independent benchmarks)
- BFCL: 108–181% Accuracy Retention Rate across 3 frontier models
- ToolBench: +5.0pp (75.0% → 80.0%)
- Real MCP Server: 100% syntactic validity, 90–97% server acceptance
- TAB → MCP Transfer: 0.1pp prediction accuracy
Install
npm install @tscg/core@1.4.1
npm install @tscg/mcp-proxy@1.4.1
npm install @tscg/tool-optimizer@1.4.1