@halo-sdk/eval 1.0.0

package/README.md

@@ -0,0 +1,30 @@
+# @halo-sdk/eval
+
+Two things, deliberately scoped:
+
+1. **Cache cost benchmark** — the differentiated "evidence" that proves Halo's prefix-cache moat. `benchmarkCache(agent, inputs)` drives an agent through a multi-turn scenario and reports hit rate, token split, estimated spend, and an A–F grade. `compareCache(scenario, a, b)` runs the same scenario through two agents — e.g. to show `SummarizeAppendStrategy` retains hit-rate where naive truncation collapses it.
+
+2. **Behavioral-eval seam** — a thin `runEvalCases(agent, cases)` harness. For real behavioral/quality evaluation (LLM-as-judge, datasets, regression gates), point **promptfoo** or **vitest** at your agent. Halo does not reimplement generic eval.
+
+## Usage
+
+```ts
+import { benchmarkCache, compareCache, runEvalCases } from "@halo-sdk/eval";
+
+const report = await benchmarkCache(agent, [
+  "Summarize the cache design.",
+  "Now expand on breakpoints.",
+  "And how does it compare to OpenAI?",
+]);
+console.log(report.hitRate, report.grade, report.estimatedUsd);
+
+const cmp = await compareCache(
+  scenario,
+  { name: "truncate", agent: a },
+  { name: "summarize", agent: b },
+);
+
+const evalReport = await runEvalCases(agent, [
+  { name: "greets", input: "say hi", assert: (out) => out.toLowerCase().includes("hi") },
+]);
+```

package/dist/index.cjs

@@ -0,0 +1,72 @@
+"use strict";
+var __defProp = Object.defineProperty;
+var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
+var __getOwnPropNames = Object.getOwnPropertyNames;
+var __hasOwnProp = Object.prototype.hasOwnProperty;
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+var __copyProps = (to, from, except, desc) => {
+  if (from && typeof from === "object" || typeof from === "function") {
+    for (let key of __getOwnPropNames(from))
+      if (!__hasOwnProp.call(to, key) && key !== except)
+        __defProp(to, key, { get: () => from[key], enumerable: !(desc = __getOwnPropDesc(from, key)) || desc.enumerable });
+  }
+  return to;
+};
+var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
+
+// src/index.ts
+var index_exports = {};
+__export(index_exports, {
+  benchmarkCache: () => benchmarkCache,
+  compareCache: () => compareCache,
+  runEvalCases: () => runEvalCases
+});
+module.exports = __toCommonJS(index_exports);
+var import_core = require("@halo-sdk/core");
+async function benchmarkCache(agent, inputs, opts) {
+  const now = opts?.now ?? (() => Date.now());
+  const profiler = new import_core.CacheProfiler();
+  const off = agent.observe(profiler.observe);
+  const start = now();
+  try {
+    for (const input of inputs) {
+      await agent.generateText(input);
+    }
+  } finally {
+    off();
+  }
+  return { ...profiler.report(), durationMs: now() - start, inputs: inputs.length };
+}
+async function compareCache(scenario, a, b, opts) {
+  const ra = await benchmarkCache(a.agent, scenario, opts);
+  const rb = await benchmarkCache(b.agent, scenario, opts);
+  return { [a.name]: ra, [b.name]: rb };
+}
+async function runEvalCases(agent, cases) {
+  const results = [];
+  for (const c of cases) {
+    try {
+      const { content } = await agent.generateText(c.input);
+      const passed = await c.assert(content);
+      results.push({ name: c.name, passed, output: content });
+    } catch (err) {
+      results.push({
+        name: c.name,
+        passed: false,
+        output: "",
+        error: err instanceof Error ? err.message : String(err)
+      });
+    }
+  }
+  return { results, passed: results.filter((r) => r.passed).length, total: results.length };
+}
+// Annotate the CommonJS export names for ESM import in node:
+0 && (module.exports = {
+  benchmarkCache,
+  compareCache,
+  runEvalCases
+});
+//# sourceMappingURL=index.cjs.map

package/dist/index.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/index.ts"],"sourcesContent":["import { CacheProfiler, type CacheProfile, type Observer } from \"@halo-sdk/core\";\n\n// ── Cache cost benchmark (the differentiated \"evidence\") ──\n\n/** The minimal agent surface the benchmark drives. `HaloAgent` satisfies it. */\nexport interface BenchmarkableAgent {\n  observe(fn: Observer): () => void;\n  generateText(input: string): Promise<unknown>;\n}\n\nexport interface CacheBenchmarkResult extends CacheProfile {\n  /** Wall-clock duration of the run in ms. */\n  durationMs: number;\n  /** Inputs that were run, in order. */\n  inputs: number;\n}\n\n/**\n * Drive an agent through a multi-turn scenario and measure the prefix-cache\n * moat: hit rate, token split, estimated spend, and an A–F grade. This is the\n * internal *evidence* benchmark — it proves the cost/latency claims, it is not a\n * behavioral quality eval (see {@link runEvalCases} + promptfoo for that).\n */\nexport async function benchmarkCache(\n  agent: BenchmarkableAgent,\n  inputs: string[],\n  opts?: { now?: () => number },\n): Promise<CacheBenchmarkResult> {\n  const now = opts?.now ?? (() => Date.now());\n  const profiler = new CacheProfiler();\n  const off = agent.observe(profiler.observe);\n  const start = now();\n  try {\n    for (const input of inputs) {\n      await agent.generateText(input);\n    }\n  } finally {\n    off();\n  }\n  return { ...profiler.report(), durationMs: now() - start, inputs: inputs.length };\n}\n\n/**\n * Compare two strategies (e.g. truncate vs. summarize-append) by running the\n * same scenario through two agent factories and reporting both cache profiles.\n * Used to demonstrate that cache-preserving context management retains hit-rate\n * where naive truncation collapses it.\n */\nexport async function compareCache(\n  scenario: string[],\n  a: { name: string; agent: BenchmarkableAgent },\n  b: { name: string; agent: BenchmarkableAgent },\n  opts?: { now?: () => number },\n): Promise<{ [name: string]: CacheBenchmarkResult }> {\n  const ra = await benchmarkCache(a.agent, scenario, opts);\n  const rb = await benchmarkCache(b.agent, scenario, opts);\n  return { [a.name]: ra, [b.name]: rb };\n}\n\n// ── Behavioral eval seam ──\n\n/**\n * A behavioral test case. This package ships only a thin harness — for richer\n * behavioral/quality evaluation (LLM-as-judge, datasets, regression gates)\n * point a real tool like **promptfoo** or **vitest** at your agent. Halo's\n * differentiated contribution is the cache benchmark above, not a generic eval.\n */\nexport interface EvalCase {\n  name: string;\n  input: string;\n  /** Return true if the output passes. */\n  assert: (output: string) => boolean | Promise<boolean>;\n}\n\nexport interface EvalCaseResult {\n  name: string;\n  passed: boolean;\n  output: string;\n  error?: string;\n}\n\nexport interface EvalRunReport {\n  results: EvalCaseResult[];\n  passed: number;\n  total: number;\n}\n\n/** Run behavioral cases against an agent and tally pass/fail. */\nexport async function runEvalCases(\n  agent: { generateText(input: string): Promise<{ content: string }> },\n  cases: EvalCase[],\n): Promise<EvalRunReport> {\n  const results: EvalCaseResult[] = [];\n  for (const c of cases) {\n    try {\n      const { content } = await agent.generateText(c.input);\n      const passed = await c.assert(content);\n      results.push({ name: c.name, passed, output: content });\n    } catch (err: unknown) {\n      results.push({\n        name: c.name,\n        passed: false,\n        output: \"\",\n        error: err instanceof Error ? err.message : String(err),\n      });\n    }\n  }\n  return { results, passed: results.filter((r) => r.passed).length, total: results.length };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,kBAAgE;AAuBhE,eAAsB,eACpB,OACA,QACA,MAC+B;AAC/B,QAAM,MAAM,MAAM,QAAQ,MAAM,KAAK,IAAI;AACzC,QAAM,WAAW,IAAI,0BAAc;AACnC,QAAM,MAAM,MAAM,QAAQ,SAAS,OAAO;AAC1C,QAAM,QAAQ,IAAI;AAClB,MAAI;AACF,eAAW,SAAS,QAAQ;AAC1B,YAAM,MAAM,aAAa,KAAK;AAAA,IAChC;AAAA,EACF,UAAE;AACA,QAAI;AAAA,EACN;AACA,SAAO,EAAE,GAAG,SAAS,OAAO,GAAG,YAAY,IAAI,IAAI,OAAO,QAAQ,OAAO,OAAO;AAClF;AAQA,eAAsB,aACpB,UACA,GACA,GACA,MACmD;AACnD,QAAM,KAAK,MAAM,eAAe,EAAE,OAAO,UAAU,IAAI;AACvD,QAAM,KAAK,MAAM,eAAe,EAAE,OAAO,UAAU,IAAI;AACvD,SAAO,EAAE,CAAC,EAAE,IAAI,GAAG,IAAI,CAAC,EAAE,IAAI,GAAG,GAAG;AACtC;AA+BA,eAAsB,aACpB,OACA,OACwB;AACxB,QAAM,UAA4B,CAAC;AACnC,aAAW,KAAK,OAAO;AACrB,QAAI;AACF,YAAM,EAAE,QAAQ,IAAI,MAAM,MAAM,aAAa,EAAE,KAAK;AACpD,YAAM,SAAS,MAAM,EAAE,OAAO,OAAO;AACrC,cAAQ,KAAK,EAAE,MAAM,EAAE,MAAM,QAAQ,QAAQ,QAAQ,CAAC;AAAA,IACxD,SAAS,KAAc;AACrB,cAAQ,KAAK;AAAA,QACX,MAAM,EAAE;AAAA,QACR,QAAQ;AAAA,QACR,QAAQ;AAAA,QACR,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MACxD,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO,EAAE,SAAS,QAAQ,QAAQ,OAAO,CAAC,MAAM,EAAE,MAAM,EAAE,QAAQ,OAAO,QAAQ,OAAO;AAC1F;","names":[]}

package/dist/index.d.ts

@@ -0,0 +1,68 @@
+import { type CacheProfile, type Observer } from "@halo-sdk/core";
+/** The minimal agent surface the benchmark drives. `HaloAgent` satisfies it. */
+export interface BenchmarkableAgent {
+    observe(fn: Observer): () => void;
+    generateText(input: string): Promise<unknown>;
+}
+export interface CacheBenchmarkResult extends CacheProfile {
+    /** Wall-clock duration of the run in ms. */
+    durationMs: number;
+    /** Inputs that were run, in order. */
+    inputs: number;
+}
+/**
+ * Drive an agent through a multi-turn scenario and measure the prefix-cache
+ * moat: hit rate, token split, estimated spend, and an A–F grade. This is the
+ * internal *evidence* benchmark — it proves the cost/latency claims, it is not a
+ * behavioral quality eval (see {@link runEvalCases} + promptfoo for that).
+ */
+export declare function benchmarkCache(agent: BenchmarkableAgent, inputs: string[], opts?: {
+    now?: () => number;
+}): Promise<CacheBenchmarkResult>;
+/**
+ * Compare two strategies (e.g. truncate vs. summarize-append) by running the
+ * same scenario through two agent factories and reporting both cache profiles.
+ * Used to demonstrate that cache-preserving context management retains hit-rate
+ * where naive truncation collapses it.
+ */
+export declare function compareCache(scenario: string[], a: {
+    name: string;
+    agent: BenchmarkableAgent;
+}, b: {
+    name: string;
+    agent: BenchmarkableAgent;
+}, opts?: {
+    now?: () => number;
+}): Promise<{
+    [name: string]: CacheBenchmarkResult;
+}>;
+/**
+ * A behavioral test case. This package ships only a thin harness — for richer
+ * behavioral/quality evaluation (LLM-as-judge, datasets, regression gates)
+ * point a real tool like **promptfoo** or **vitest** at your agent. Halo's
+ * differentiated contribution is the cache benchmark above, not a generic eval.
+ */
+export interface EvalCase {
+    name: string;
+    input: string;
+    /** Return true if the output passes. */
+    assert: (output: string) => boolean | Promise<boolean>;
+}
+export interface EvalCaseResult {
+    name: string;
+    passed: boolean;
+    output: string;
+    error?: string;
+}
+export interface EvalRunReport {
+    results: EvalCaseResult[];
+    passed: number;
+    total: number;
+}
+/** Run behavioral cases against an agent and tally pass/fail. */
+export declare function runEvalCases(agent: {
+    generateText(input: string): Promise<{
+        content: string;
+    }>;
+}, cases: EvalCase[]): Promise<EvalRunReport>;
+//# sourceMappingURL=index.d.ts.map

package/dist/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA,OAAO,EAAiB,KAAK,YAAY,EAAE,KAAK,QAAQ,EAAE,MAAM,gBAAgB,CAAC;AAIjF,gFAAgF;AAChF,MAAM,WAAW,kBAAkB;IACjC,OAAO,CAAC,EAAE,EAAE,QAAQ,GAAG,MAAM,IAAI,CAAC;IAClC,YAAY,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC;CAC/C;AAED,MAAM,WAAW,oBAAqB,SAAQ,YAAY;IACxD,4CAA4C;IAC5C,UAAU,EAAE,MAAM,CAAC;IACnB,sCAAsC;IACtC,MAAM,EAAE,MAAM,CAAC;CAChB;AAED;;;;;GAKG;AACH,wBAAsB,cAAc,CAClC,KAAK,EAAE,kBAAkB,EACzB,MAAM,EAAE,MAAM,EAAE,EAChB,IAAI,CAAC,EAAE;IAAE,GAAG,CAAC,EAAE,MAAM,MAAM,CAAA;CAAE,GAC5B,OAAO,CAAC,oBAAoB,CAAC,CAa/B;AAED;;;;;GAKG;AACH,wBAAsB,YAAY,CAChC,QAAQ,EAAE,MAAM,EAAE,EAClB,CAAC,EAAE;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,kBAAkB,CAAA;CAAE,EAC9C,CAAC,EAAE;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,kBAAkB,CAAA;CAAE,EAC9C,IAAI,CAAC,EAAE;IAAE,GAAG,CAAC,EAAE,MAAM,MAAM,CAAA;CAAE,GAC5B,OAAO,CAAC;IAAE,CAAC,IAAI,EAAE,MAAM,GAAG,oBAAoB,CAAA;CAAE,CAAC,CAInD;AAID;;;;;GAKG;AACH,MAAM,WAAW,QAAQ;IACvB,IAAI,EAAE,MAAM,CAAC;IACb,KAAK,EAAE,MAAM,CAAC;IACd,wCAAwC;IACxC,MAAM,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,OAAO,GAAG,OAAO,CAAC,OAAO,CAAC,CAAC;CACxD;AAED,MAAM,WAAW,cAAc;IAC7B,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE,OAAO,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,aAAa;IAC5B,OAAO,EAAE,cAAc,EAAE,CAAC;IAC1B,MAAM,EAAE,MAAM,CAAC;IACf,KAAK,EAAE,MAAM,CAAC;CACf;AAED,iEAAiE;AACjE,wBAAsB,YAAY,CAChC,KAAK,EAAE;IAAE,YAAY,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAAC;QAAE,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;CAAE,EACpE,KAAK,EAAE,QAAQ,EAAE,GAChB,OAAO,CAAC,aAAa,CAAC,CAiBxB"}

package/dist/index.js

@@ -0,0 +1,45 @@
+// src/index.ts
+import { CacheProfiler } from "@halo-sdk/core";
+async function benchmarkCache(agent, inputs, opts) {
+  const now = opts?.now ?? (() => Date.now());
+  const profiler = new CacheProfiler();
+  const off = agent.observe(profiler.observe);
+  const start = now();
+  try {
+    for (const input of inputs) {
+      await agent.generateText(input);
+    }
+  } finally {
+    off();
+  }
+  return { ...profiler.report(), durationMs: now() - start, inputs: inputs.length };
+}
+async function compareCache(scenario, a, b, opts) {
+  const ra = await benchmarkCache(a.agent, scenario, opts);
+  const rb = await benchmarkCache(b.agent, scenario, opts);
+  return { [a.name]: ra, [b.name]: rb };
+}
+async function runEvalCases(agent, cases) {
+  const results = [];
+  for (const c of cases) {
+    try {
+      const { content } = await agent.generateText(c.input);
+      const passed = await c.assert(content);
+      results.push({ name: c.name, passed, output: content });
+    } catch (err) {
+      results.push({
+        name: c.name,
+        passed: false,
+        output: "",
+        error: err instanceof Error ? err.message : String(err)
+      });
+    }
+  }
+  return { results, passed: results.filter((r) => r.passed).length, total: results.length };
+}
+export {
+  benchmarkCache,
+  compareCache,
+  runEvalCases
+};
+//# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/index.ts"],"sourcesContent":["import { CacheProfiler, type CacheProfile, type Observer } from \"@halo-sdk/core\";\n\n// ── Cache cost benchmark (the differentiated \"evidence\") ──\n\n/** The minimal agent surface the benchmark drives. `HaloAgent` satisfies it. */\nexport interface BenchmarkableAgent {\n  observe(fn: Observer): () => void;\n  generateText(input: string): Promise<unknown>;\n}\n\nexport interface CacheBenchmarkResult extends CacheProfile {\n  /** Wall-clock duration of the run in ms. */\n  durationMs: number;\n  /** Inputs that were run, in order. */\n  inputs: number;\n}\n\n/**\n * Drive an agent through a multi-turn scenario and measure the prefix-cache\n * moat: hit rate, token split, estimated spend, and an A–F grade. This is the\n * internal *evidence* benchmark — it proves the cost/latency claims, it is not a\n * behavioral quality eval (see {@link runEvalCases} + promptfoo for that).\n */\nexport async function benchmarkCache(\n  agent: BenchmarkableAgent,\n  inputs: string[],\n  opts?: { now?: () => number },\n): Promise<CacheBenchmarkResult> {\n  const now = opts?.now ?? (() => Date.now());\n  const profiler = new CacheProfiler();\n  const off = agent.observe(profiler.observe);\n  const start = now();\n  try {\n    for (const input of inputs) {\n      await agent.generateText(input);\n    }\n  } finally {\n    off();\n  }\n  return { ...profiler.report(), durationMs: now() - start, inputs: inputs.length };\n}\n\n/**\n * Compare two strategies (e.g. truncate vs. summarize-append) by running the\n * same scenario through two agent factories and reporting both cache profiles.\n * Used to demonstrate that cache-preserving context management retains hit-rate\n * where naive truncation collapses it.\n */\nexport async function compareCache(\n  scenario: string[],\n  a: { name: string; agent: BenchmarkableAgent },\n  b: { name: string; agent: BenchmarkableAgent },\n  opts?: { now?: () => number },\n): Promise<{ [name: string]: CacheBenchmarkResult }> {\n  const ra = await benchmarkCache(a.agent, scenario, opts);\n  const rb = await benchmarkCache(b.agent, scenario, opts);\n  return { [a.name]: ra, [b.name]: rb };\n}\n\n// ── Behavioral eval seam ──\n\n/**\n * A behavioral test case. This package ships only a thin harness — for richer\n * behavioral/quality evaluation (LLM-as-judge, datasets, regression gates)\n * point a real tool like **promptfoo** or **vitest** at your agent. Halo's\n * differentiated contribution is the cache benchmark above, not a generic eval.\n */\nexport interface EvalCase {\n  name: string;\n  input: string;\n  /** Return true if the output passes. */\n  assert: (output: string) => boolean | Promise<boolean>;\n}\n\nexport interface EvalCaseResult {\n  name: string;\n  passed: boolean;\n  output: string;\n  error?: string;\n}\n\nexport interface EvalRunReport {\n  results: EvalCaseResult[];\n  passed: number;\n  total: number;\n}\n\n/** Run behavioral cases against an agent and tally pass/fail. */\nexport async function runEvalCases(\n  agent: { generateText(input: string): Promise<{ content: string }> },\n  cases: EvalCase[],\n): Promise<EvalRunReport> {\n  const results: EvalCaseResult[] = [];\n  for (const c of cases) {\n    try {\n      const { content } = await agent.generateText(c.input);\n      const passed = await c.assert(content);\n      results.push({ name: c.name, passed, output: content });\n    } catch (err: unknown) {\n      results.push({\n        name: c.name,\n        passed: false,\n        output: \"\",\n        error: err instanceof Error ? err.message : String(err),\n      });\n    }\n  }\n  return { results, passed: results.filter((r) => r.passed).length, total: results.length };\n}\n"],"mappings":";AAAA,SAAS,qBAAuD;AAuBhE,eAAsB,eACpB,OACA,QACA,MAC+B;AAC/B,QAAM,MAAM,MAAM,QAAQ,MAAM,KAAK,IAAI;AACzC,QAAM,WAAW,IAAI,cAAc;AACnC,QAAM,MAAM,MAAM,QAAQ,SAAS,OAAO;AAC1C,QAAM,QAAQ,IAAI;AAClB,MAAI;AACF,eAAW,SAAS,QAAQ;AAC1B,YAAM,MAAM,aAAa,KAAK;AAAA,IAChC;AAAA,EACF,UAAE;AACA,QAAI;AAAA,EACN;AACA,SAAO,EAAE,GAAG,SAAS,OAAO,GAAG,YAAY,IAAI,IAAI,OAAO,QAAQ,OAAO,OAAO;AAClF;AAQA,eAAsB,aACpB,UACA,GACA,GACA,MACmD;AACnD,QAAM,KAAK,MAAM,eAAe,EAAE,OAAO,UAAU,IAAI;AACvD,QAAM,KAAK,MAAM,eAAe,EAAE,OAAO,UAAU,IAAI;AACvD,SAAO,EAAE,CAAC,EAAE,IAAI,GAAG,IAAI,CAAC,EAAE,IAAI,GAAG,GAAG;AACtC;AA+BA,eAAsB,aACpB,OACA,OACwB;AACxB,QAAM,UAA4B,CAAC;AACnC,aAAW,KAAK,OAAO;AACrB,QAAI;AACF,YAAM,EAAE,QAAQ,IAAI,MAAM,MAAM,aAAa,EAAE,KAAK;AACpD,YAAM,SAAS,MAAM,EAAE,OAAO,OAAO;AACrC,cAAQ,KAAK,EAAE,MAAM,EAAE,MAAM,QAAQ,QAAQ,QAAQ,CAAC;AAAA,IACxD,SAAS,KAAc;AACrB,cAAQ,KAAK;AAAA,QACX,MAAM,EAAE;AAAA,QACR,QAAQ;AAAA,QACR,QAAQ;AAAA,QACR,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MACxD,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO,EAAE,SAAS,QAAQ,QAAQ,OAAO,CAAC,MAAM,EAAE,MAAM,EAAE,QAAQ,OAAO,QAAQ,OAAO;AAC1F;","names":[]}

package/package.json

@@ -0,0 +1,50 @@
+{
+  "name": "@halo-sdk/eval",
+  "version": "1.0.0",
+  "description": "Cache cost benchmark + behavioral-eval seam for Halo SDK — measure the prefix-cache moat, plug in promptfoo/vitest for quality",
+  "keywords": [
+    "ai",
+    "benchmark",
+    "eval",
+    "llm",
+    "prefix-cache"
+  ],
+  "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/halo-sdk/halo-ai",
+    "directory": "packages/eval"
+  },
+  "files": [
+    "dist"
+  ],
+  "type": "module",
+  "main": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js",
+      "require": "./dist/index.cjs"
+    }
+  },
+  "publishConfig": {
+    "access": "public"
+  },
+  "devDependencies": {
+    "typescript": "^5.8.0",
+    "vitest": "^3.0.0",
+    "@halo-sdk/core": "1.1.0"
+  },
+  "peerDependencies": {
+    "@halo-sdk/core": ">=1.1.0"
+  },
+  "scripts": {
+    "build": "tsc --build --emitDeclarationOnly && tsup",
+    "dev": "tsup --watch",
+    "clean": "del-cli dist *.tsbuildinfo",
+    "publint": "publint",
+    "test": "vitest run",
+    "test:watch": "vitest"
+  }
+}