@tangle-network/agent-eval 0.72.0 → 0.72.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-eval",
-  "version": "0.72.0",
+  "version": "0.72.3",
   "description": "Substrate for self-improving agents: traces, verifiable rewards, preferences, GEPA / reflective mutation, auto-research, replay, sequential anytime-valid stats, and release gates.",
   "homepage": "https://github.com/tangle-network/agent-eval#readme",
   "repository": {
@@ -19,6 +19,11 @@
       "import": "./dist/index.js",
       "default": "./dist/index.js"
     },
+    "./analyst": {
+      "types": "./dist/analyst/index.d.ts",
+      "import": "./dist/analyst/index.js",
+      "default": "./dist/analyst/index.js"
+    },
     "./control": {
       "types": "./dist/control.d.ts",
       "import": "./dist/control.js",
@@ -104,6 +109,11 @@
       "import": "./dist/campaign/index.js",
       "default": "./dist/campaign/index.js"
     },
+    "./workflow": {
+      "types": "./dist/workflow/index.d.ts",
+      "import": "./dist/workflow/index.js",
+      "default": "./dist/workflow/index.js"
+    },
     "./contract": {
       "types": "./dist/contract/index.d.ts",
       "import": "./dist/contract/index.js",
@@ -144,6 +154,19 @@
   "publishConfig": {
     "access": "public"
   },
+  "scripts": {
+    "build": "tsup && pnpm openapi",
+    "dev": "tsup --watch",
+    "prepare": "husky",
+    "prepublishOnly": "pnpm build",
+    "test": "vitest run",
+    "test:watch": "vitest",
+    "typecheck": "tsc --noEmit",
+    "lint": "biome check src",
+    "format": "biome format --write src",
+    "openapi": "node dist/cli.js openapi --out dist/openapi.json",
+    "verify:package": "node scripts/verify-package-exports.mjs"
+  },
   "dependencies": {
     "@asteasolutions/zod-to-openapi": "^8.5.0",
     "@ax-llm/ax": "^19.0.25",
@@ -171,6 +194,16 @@
     "typescript": "^5.7.0",
     "vitest": "^3.0.0"
   },
+  "pnpm": {
+    "minimumReleaseAge": 4320,
+    "minimumReleaseAgeExclude": [
+      "@tangle-network/sandbox"
+    ],
+    "overrides": {
+      "postcss@<8.5.10": "^8.5.10",
+      "ws@>=8.0.0 <8.20.1": "^8.20.1"
+    }
+  },
   "engines": {
     "node": ">=20"
   },
@@ -183,14 +216,5 @@
     ]
   },
   "license": "MIT",
-  "scripts": {
-    "build": "tsup && pnpm openapi",
-    "dev": "tsup --watch",
-    "test": "vitest run",
-    "test:watch": "vitest",
-    "typecheck": "tsc --noEmit",
-    "lint": "biome check src",
-    "format": "biome format --write src",
-    "openapi": "node dist/cli.js openapi --out dist/openapi.json"
-  }
-}
+  "packageManager": "pnpm@10.22.0"
+}

package/dist/agent-profile-DzcPHR1Z.d.ts

@@ -1,114 +0,0 @@
-import { A as AgentEvalError } from './errors-Dwqw-T_m.js';
-import { R as RunRecord } from './run-record-BgTFzO2r.js';
-
-/**
- * Backend-integrity guard: distinguish "agent failed" from "eval ran against
- * a stub / unconfigured backend." Without this guard a canonical eval can
- * silently report `0/N passed` and look like an agent-quality problem when
- * the LLM was never actually called — the failure mode we just hit running
- * the 4-vertical parallel eval (legal-sandbox-stub returned hard-coded 33-104
- * char strings; gtm/creative defaulted to a cli-bridge that wasn't running).
- *
- * The shape:
- *
- *   const report = summarizeBackendIntegrity(records)
- *   assertRealBackend(records)   // throws BackendIntegrityError if 100% stub
- *
- * A record is "stub-mode" if its `tokenUsage.input === 0 && tokenUsage.output === 0`.
- * (`costUsd` alone is unreliable — some backends successfully call LLMs but
- *  don't propagate pricing, producing real tokens with $0 cost.)
- *
- * Verdicts:
- *   - `real`   — at least one record has nonzero token usage
- *   - `stub`   — every record is stub-mode (eval ran blind)
- *   - `mixed`  — some records real, some stub (partial backend failure;
- *                often the 429-cascade or auth-half-failed case)
- */
-
-interface BackendIntegrityReport {
-    /** Total records inspected. */
-    totalRecords: number;
-    /** Records with input=0 AND output=0 (a stub fingerprint). */
-    stubRecords: number;
-    /** Records with nonzero token usage (real LLM activity). */
-    realRecords: number;
-    /** Records where output>0 but costUsd=0 (real LLM, broken cost ledger). */
-    uncostedRecords: number;
-    /** Sum of input tokens across all records. */
-    totalInputTokens: number;
-    /** Sum of output tokens across all records. */
-    totalOutputTokens: number;
-    /** Sum of costUsd across all records. */
-    totalCostUsd: number;
-    /** Worst-case integrity verdict. */
-    verdict: 'real' | 'mixed' | 'stub';
-    /** Human-readable diagnosis suitable for terminal output. */
-    diagnosis: string;
-}
-/**
- * Error thrown when an integrity assertion fails. Caller can pattern-match
- * by `code === 'AGENT_EVAL_BACKEND_STUB'` to differentiate from other
- * errors.
- */
-declare class BackendIntegrityError extends AgentEvalError {
-    readonly report: BackendIntegrityReport;
-    constructor(message: string, report: BackendIntegrityReport);
-}
-/**
- * Inspect a batch of RunRecords and return an integrity report. Pure
- * function — no I/O, no logging. The caller decides what to do with the
- * verdict (print warning, throw, gate CI, etc.).
- */
-declare function summarizeBackendIntegrity(records: ReadonlyArray<RunRecord>): BackendIntegrityReport;
-/**
- * Throw BackendIntegrityError if the verdict is 'stub' — i.e. every record
- * shows zero LLM activity. Non-strict callers can pass `{ allowMixed: false }`
- * to also reject mixed verdicts (recommended for CI gates).
- *
- * Real backends pass through silently.
- */
-declare function assertRealBackend(records: ReadonlyArray<RunRecord>, opts?: {
-    allowMixed?: boolean;
-}): BackendIntegrityReport;
-
-/**
- * @stable
- *
- * AgentProfile — the eval harness's unit of variation.
- *
- * A profile pins everything that changes agent behaviour for a benchmark
- * cell: the model, the active skills, the prompt version, the available
- * tools. Vary the profile — swap a model, add a skill — and re-run the suite
- * to benchmark the change. The scorecard keys a cell on
- * `(scenarioId, profileHash)`, so the model is not a separate axis: it lives
- * inside the profile, and two profiles with the same model but different
- * skills are different cells.
- *
- * `agentProfileHash` is the profile's behaviour identity. Two profiles that
- * produce the same agent behaviour share a hash (and a scorecard cell);
- * reordering `skills` or `tools` does not change it; the human-facing `id`
- * label does not affect it.
- */
-interface AgentProfile {
-    /** Human-facing label, e.g. `sonnet-legal-skills-v3`. Not part of the hash. */
-    id: string;
-    /** Model snapshot id this profile pins, e.g. `claude-sonnet-4-6@2025-04-15`. */
-    model: string;
-    /** Skill ids/versions active in this profile — the primary behaviour lever. */
-    skills?: string[];
-    /** Prompt version identifier. */
-    promptVersion?: string;
-    /** Tool ids available to the agent. */
-    tools?: string[];
-    /** Any other behaviour-bearing knobs that should fingerprint into the hash. */
-    metadata?: Record<string, string | number | boolean>;
-}
-/**
- * Deterministic behaviour identity of a profile — a sha256 over the
- * behaviour-bearing fields. `skills` and `tools` are order-insensitive; the
- * `id` label is excluded. Throws on a profile with no `model` — an unkeyable
- * profile must fail loud rather than collapse into a blank-model cell.
- */
-declare function agentProfileHash(profile: AgentProfile): string;
-
-export { type AgentProfile as A, type BackendIntegrityReport as B, BackendIntegrityError as a, agentProfileHash as b, assertRealBackend as c, summarizeBackendIntegrity as s };