@dzhechkov/skills-bto 1.3.0 → 1.3.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@dzhechkov/skills-bto",
-  "version": "1.3.0",
+  "version": "1.3.1",
   "description": "Build-Benchmark-Test-Optimize skill pack for Claude Code — deterministic benchmarking, quality gates, witness chain, judge attestation, and optimization",
   "main": "src/cli.js",
   "bin": {
@@ -11,10 +11,6 @@
     "src/",
     "templates/"
   ],
-  "scripts": {
-    "test": "node bin/cli.js doctor",
-    "prepublishOnly": "node scripts/sync-templates.js"
-  },
   "keywords": [
     "claude",
     "claude-code",
@@ -35,9 +31,10 @@
   },
   "repository": {
     "type": "git",
-    "url": "https://github.com/dzhechko/product-keysarium-2026"
+    "url": "https://github.com/djd1m/dz-harness-hub.git",
+    "directory": "packages/@dzhechkov/skills-bto"
   },
-  "homepage": "https://github.com/dzhechko/product-keysarium-2026#readme",
+  "homepage": "https://github.com/djd1m/dz-harness-hub/tree/main/packages/@dzhechkov/skills-bto#readme",
   "bugs": {
     "url": "https://github.com/dzhechko/product-keysarium-2026/issues"
   },
@@ -48,5 +45,8 @@
     "@dzhechkov/keysarium-core": {
       "optional": true
     }
+  },
+  "publishConfig": {
+    "access": "public"
   }
 }

package/templates/.claude/skills/bto/modules/benchmark.md

@@ -445,8 +445,74 @@ PERFORMANCE SCORE: X.XX
 
 ---
 
+## Layer B4: Cost Efficiency (advisory axis)
+
+> Inspired by Princeton HAL harness — cost is tracked as a **separate first-class
+> axis**, NOT blended into the quality score. A skill is never blocked for being
+> expensive; instead the cost↔quality tradeoff is surfaced so the operator can
+> decide. Two skills with equal quality but a 10x cost difference must be
+> distinguishable at a glance.
+
+### Static estimate (pre-run)
+
+Estimate the skill's cost band from its declared model usage *before* any LLM
+call is made, by counting how many times the artifact routes work to each tier:
+
+```
+relative_units = haiku_calls * 1 + sonnet_calls * 4 + opus_calls * 20
+
+Band:
+  free    : relative_units == 0   (deterministic, no LLM)
+  low     : 1  - 8
+  medium  : 9  - 40
+  high     : > 40
+```
+
+> The deterministic implementation lives in `@dzhechkov/harness-core`
+> (`estimateSkillCost()`), surfaced on every `BenchmarkScore.cost`. Prefer it to
+> hand-counting when benchmarking via the CLI.
+
+### Measured cost (post-run, when TEST has executed)
+
+When TEST has run real judges, aggregate actual token usage into a dollar cost
+using the per-token pricing table and provider-normalized usage:
+
+```
+invocation_cost_usd = Σ over calls [ normalizeUsage(call) priced via MODEL_PRICES ]
+cost_efficiency      = quality_score / invocation_cost_usd     (quality per $)
+```
+
+`normalizeUsage()` reconciles OpenAI / Anthropic / Bedrock token shapes and
+applies cached-read discounts (~10% of prompt price). Both helpers are exported
+from `@dzhechkov/harness-core` (`invocationCost`, `costEfficiency`).
+
+### Layer B4 Output
+
+```
+LAYER B4 -- Cost Efficiency (advisory)
+  Static band:        medium  (relative_units = 23)
+  Tier calls:         haiku x3, sonnet x5, opus x0
+  Measured cost:      $0.042 / invocation   (if TEST ran)
+  Cost-efficiency:    202 quality-points/$  (8.5 / $0.042)
+  Verdict:            efficient — no opus on the hot path
+```
+
+### Cost Gate (advisory only — never blocks)
+
+| Band | Advisory |
+|------|----------|
+| free / low | ✅ Lean. No action. |
+| medium | ℹ️ Acceptable. Confirm opus calls (if any) are reserved for disagreement/synthesis. |
+| high | ⚠️ Flag for OPTIMIZE: can any opus call drop to sonnet, or any layer merge, without quality loss? |
+
+---
+
 ## Overall Benchmark Score
 
+> Cost (B4) is reported alongside but is **not** part of `BENCHMARK_SCORE` —
+> quality and cost are orthogonal axes (HAL principle). Quality gates the
+> pipeline; cost informs the operator.
+
 ```
 BENCHMARK_SCORE = B0_golden * 0.30 + B1_tests * 0.35 + B2_consistency * 0.15 + B3_performance * 0.20
 ```