@evolvingmachines/sdk 0.0.51 → 0.0.52-project-sable.20260729.8632c49

package/hosted-error-codes.json

@@ -0,0 +1,57 @@
+{
+  "$comment": "The hosted evals API's error-code vocabulary, checked in so two repos that share no test runner can be held to one list. The server owns it: HOSTED_API_ERROR_CODES in the dashboard (lib/evaluations/api-errors.ts) is the only place a code may be added. This file is that array’s shadow, and three tests keep everyone honest — the dashboard regenerates these exact bytes and diffs them, the TypeScript SDK asserts HOSTED_ERROR_CODES equals codes[], and the Python SDK asserts both its tuple and its Literal do. Adding a code therefore means: server array, this file, both SDK lists.",
+  "codes": [
+    "missing_authorization",
+    "invalid_api_key",
+    "credential_service_unavailable",
+    "rate_limited",
+    "insufficient_credits",
+    "invalid_json",
+    "invalid_input",
+    "invalid_limit",
+    "invalid_status",
+    "invalid_cursor",
+    "invalid_after",
+    "invalid_format",
+    "invalid_ids",
+    "invalid_multipart",
+    "idempotency_key_reused",
+    "benchmark_not_found",
+    "benchmark_version_not_found",
+    "benchmark_name_taken",
+    "benchmark_in_use",
+    "benchmark_not_owned",
+    "upstream_not_watchable",
+    "no_active_version",
+    "version_not_ready",
+    "unknown_task_keys",
+    "no_tasks",
+    "custom_harness_not_found",
+    "custom_harness_name_taken",
+    "custom_harness_name_reserved",
+    "custom_harness_invalid_name",
+    "custom_harness_source_required",
+    "custom_harness_source_conflict",
+    "custom_harness_invalid_env",
+    "custom_harness_too_large",
+    "custom_harness_limit_reached",
+    "harness_version_not_found",
+    "job_too_large",
+    "provider_unsupported",
+    "job_not_found",
+    "job_not_terminal",
+    "no_failed_runs",
+    "trial_not_found",
+    "concurrent_update",
+    "regrade_source_ineligible",
+    "no_regradable_runs",
+    "regrade_not_found",
+    "import_not_found",
+    "import_too_large",
+    "invalid_archive",
+    "package_not_retained",
+    "package_corrupt",
+    "package_missing",
+    "internal_error"
+  ]
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@evolvingmachines/sdk",
-  "version": "0.0.51",
+  "version": "0.0.52-project-sable.20260729.8632c49",
   "keywords": [
     "ai",
     "agents",
@@ -23,12 +23,16 @@
     "url": "git+https://github.com/evolving-machines-lab/evolve.git"
   },
   "type": "module",
+  "bin": {
+    "evolve-evals": "dist/hosted/cli.js"
+  },
   "main": "dist/index.cjs",
   "types": "dist/index.d.ts",
   "license": "Apache-2.0",
   "files": [
     "dist",
-    "LICENSE"
+    "LICENSE",
+    "hosted-error-codes.json"
   ],
   "exports": {
     ".": {
@@ -40,10 +44,11 @@
   "scripts": {
     "build": "tsup --minify",
     "dev": "tsup src/index.ts --watch",
-    "type-check": "tsc --noEmit",
+    "type-check": "tsc --noEmit && tsc -p tsconfig.types-test.json",
     "test": "npm run test:unit && npm run test:integration",
     "pretest:unit": "npm run build",
-    "test:unit": "tsx tests/unit/semaphore.test.ts && tsx tests/unit/swarm-concurrency.test.ts && tsx tests/unit/swarm-verify.test.ts && tsx tests/unit/swarm-retry-verify.test.ts && tsx tests/unit/prompt-construction.test.ts && tsx tests/unit/observability-metadata.test.ts && tsx tests/unit/skills-integration.test.ts && tsx tests/unit/auth-config.test.ts && tsx tests/unit/browser-config.test.ts && tsx tests/unit/integrations-config.test.ts && tsx tests/unit/plugins-config.test.ts && tsx tests/unit/session-runtime.test.ts && tsx tests/unit/cost-api.test.ts && tsx tests/unit/storage-config.test.ts && tsx tests/unit/checkpoint-tar.test.ts && tsx tests/unit/checkpoint-errors.test.ts && tsx tests/unit/checkpoint-flows.test.ts && tsx tests/unit/checkpoint-dx.test.ts && tsx tests/unit/checkpoint-edge-cases.test.ts && tsx tests/unit/kimi-parser.test.ts && tsx tests/unit/kimi-mcp.test.ts && tsx tests/unit/opencode-parser.test.ts && tsx tests/unit/droid-parser.test.ts && tsx tests/unit/droid-mcp.test.ts && tsx tests/unit/provider-parity.test.ts && tsx tests/unit/storage-client.test.ts && tsx tests/unit/sessions-client.test.ts",
+    "test:unit": "tsx tests/unit/semaphore.test.ts && tsx tests/unit/swarm-concurrency.test.ts && tsx tests/unit/swarm-verify.test.ts && tsx tests/unit/swarm-retry-verify.test.ts && tsx tests/unit/prompt-construction.test.ts && tsx tests/unit/observability-metadata.test.ts && tsx tests/unit/skills-integration.test.ts && tsx tests/unit/auth-config.test.ts && tsx tests/unit/browser-config.test.ts && tsx tests/unit/integrations-config.test.ts && tsx tests/unit/managed-secrets.test.ts && tsx tests/unit/plugins-config.test.ts && tsx tests/unit/session-runtime.test.ts && tsx tests/unit/cost-api.test.ts && tsx tests/unit/storage-config.test.ts && tsx tests/unit/checkpoint-tar.test.ts && tsx tests/unit/checkpoint-errors.test.ts && tsx tests/unit/checkpoint-flows.test.ts && tsx tests/unit/checkpoint-dx.test.ts && tsx tests/unit/checkpoint-edge-cases.test.ts && tsx tests/unit/codex-parser-errors.test.ts && tsx tests/unit/kimi-parser.test.ts && tsx tests/unit/kimi-mcp.test.ts && tsx tests/unit/opencode-parser.test.ts && tsx tests/unit/droid-parser.test.ts && tsx tests/unit/droid-mcp.test.ts && tsx tests/unit/provider-parity.test.ts && tsx tests/unit/storage-client.test.ts && tsx tests/unit/sessions-client.test.ts && tsx tests/unit/hosted-client.test.ts && tsx tests/unit/hosted-ergonomics.test.ts && tsx tests/unit/hosted-cli.test.ts && tsx tests/unit/hosted-types.test.ts && tsx tests/unit/hosted-error-codes.test.ts && tsx tests/unit/sandbox-artifacts.test.ts && tsx tests/unit/upload-file-from-path.test.ts",
+    "test:unit:upload-from-path": "tsx tests/unit/upload-file-from-path.test.ts",
     "test:unit:parity": "tsx tests/unit/provider-parity.test.ts",
     "test:unit:semaphore": "tsx tests/unit/semaphore.test.ts",
     "test:unit:swarm": "tsx tests/unit/swarm-concurrency.test.ts",
@@ -55,8 +60,10 @@
     "test:unit:auth": "tsx tests/unit/auth-config.test.ts",
     "test:unit:browser": "tsx tests/unit/browser-config.test.ts",
     "test:unit:integrations": "tsx tests/unit/integrations-config.test.ts",
+    "test:unit:managed-secrets": "tsx tests/unit/managed-secrets.test.ts",
     "test:unit:plugins": "tsx tests/unit/plugins-config.test.ts",
     "test:unit:storage-config": "tsx tests/unit/storage-config.test.ts",
+    "test:unit:error-codes": "tsx tests/unit/hosted-error-codes.test.ts",
     "test:unit:checkpoint-tar": "tsx tests/unit/checkpoint-tar.test.ts",
     "test:unit:checkpoint-errors": "tsx tests/unit/checkpoint-errors.test.ts",
     "test:unit:checkpoint-flows": "tsx tests/unit/checkpoint-flows.test.ts",
@@ -64,6 +71,7 @@
     "test:unit:checkpoint-edge-cases": "tsx tests/unit/checkpoint-edge-cases.test.ts",
     "test:unit:storage-client": "tsx tests/unit/storage-client.test.ts",
     "test:unit:sessions-client": "tsx tests/unit/sessions-client.test.ts",
+    "test:unit:hosted-client": "tsx tests/unit/hosted-client.test.ts",
     "test:unit:kimi-parser": "tsx tests/unit/kimi-parser.test.ts",
     "test:unit:kimi-mcp": "tsx tests/unit/kimi-mcp.test.ts",
     "pretest:integration": "npm run build",
@@ -93,13 +101,15 @@
     "test:claude": "tsx tests/integration/01-all-agents-parallel.ts claude",
     "test:codex": "tsx tests/integration/01-all-agents-parallel.ts codex",
     "test:gemini": "tsx tests/integration/01-all-agents-parallel.ts gemini",
-    "test:qwen": "tsx tests/integration/01-all-agents-parallel.ts qwen"
+    "test:qwen": "tsx tests/integration/01-all-agents-parallel.ts qwen",
+    "test:unit:hosted-cli": "tsx tests/unit/hosted-cli.test.ts",
+    "test:unit:sandbox-artifacts": "tsx tests/unit/sandbox-artifacts.test.ts"
   },
   "dependencies": {
     "@agentclientprotocol/sdk": "^0.5.1",
-    "@evolvingmachines/e2b": "^0.0.51",
-    "@evolvingmachines/daytona": "^0.0.51",
-    "@evolvingmachines/modal": "^0.0.51",
+    "@evolvingmachines/e2b": "^0.0.52-project-sable.20260729.8632c49",
+    "@evolvingmachines/daytona": "^0.0.52-project-sable.20260729.8632c49",
+    "@evolvingmachines/modal": "^0.0.52-project-sable.20260729.8632c49",
     "ajv": "^8.17.1",
     "p-map": "^7.0.2",
     "zod": "^3.24.0",