@ls-stack/agent-eval 0.60.0 → 0.60.1

package/dist/index.mjs

@@ -1,4 +1,4 @@
 import { $ as setScopeCacheContext, A as repoFile, B as evalTime, C as evalTracer, D as deserializeCacheValue, E as deserializeCacheRecording, F as EvalRuntimeUsageError, Ft as getEvalRegistry, H as getEvalCaseInput, I as appendToEvalOutput, J as runInEvalRuntimeScope, K as mergeEvalOutput, M as readManualInputFile, N as evalExpect, O as serializeCacheRecording, P as EvalAssertionError, Q as setEvalOutput, R as evalAssert, S as evalSpan, T as hashCacheKeySync, U as incrementEvalOutput, V as getCurrentScope, W as isInEvalScope, X as runInExistingEvalScope, Y as runInEvalScope, at as extractApiCalls, b as buildTraceTree, ct as simulateTokenAllocation, et as startEvalBackgroundJob, it as extractCacheHits, j as manualInputFileValueSchema, k as serializeCacheValue, ot as extractLlmCalls, q as nextEvalId, rt as extractCacheEntries, st as simulateLlmCallCost, ut as getNestedAttribute, w as hashCacheKey, x as captureEvalSpanError, z as evalLog } from "./runExecution-CjWJUUZ5.mjs";
-import { a as materializeManualInputFiles, i as isManualInputFileValue, n as createRunner, o as stageManualInputFile, r as cleanupStagedManualInputFiles, s as stageManualInputFileFromPath, t as runCli } from "./cli-FOyPC8UD.mjs";
-import { n as matchesEvalTags, t as defineEval } from "./src-p-GRSVDb.mjs";
+import { a as materializeManualInputFiles, i as isManualInputFileValue, n as createRunner, o as stageManualInputFile, r as cleanupStagedManualInputFiles, s as stageManualInputFileFromPath, t as runCli } from "./cli-CVBSlTD8.mjs";
+import { n as matchesEvalTags, t as defineEval } from "./src-DjOTPnDz.mjs";
 export { EvalAssertionError, EvalRuntimeUsageError, appendToEvalOutput, buildTraceTree, captureEvalSpanError, cleanupStagedManualInputFiles, createRunner, defineEval, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalExpect, evalLog, evalSpan, evalTime, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, getCurrentScope, getEvalCaseInput, getEvalRegistry, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, isManualInputFileValue, manualInputFileValueSchema, matchesEvalTags, materializeManualInputFiles, mergeEvalOutput, nextEvalId, readManualInputFile, repoFile, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, serializeCacheRecording, serializeCacheValue, setEvalOutput, setScopeCacheContext, simulateLlmCallCost, simulateTokenAllocation, stageManualInputFile, stageManualInputFileFromPath, startEvalBackgroundJob };

package/dist/{runner-CIxj7jYj.mjs → runner-Cu1CQPTB.mjs}

@@ -1,2 +1,2 @@
-import { n as initRunner, t as getRunnerInstance } from "./runner-Dv5cseOt.mjs";
+import { n as initRunner, t as getRunnerInstance } from "./runner-DzDRasWV.mjs";
 export { getRunnerInstance, initRunner };

package/dist/runner-DzDRasWV.mjs

@@ -0,0 +1,15 @@
+import { n as createRunner } from "./cli-CVBSlTD8.mjs";
+import "./src-DjOTPnDz.mjs";
+//#region ../../apps/server/src/runner.ts
+let runnerInstance = null;
+function getRunnerInstance({ loadEnv = true } = {}) {
+	if (!runnerInstance) runnerInstance = createRunner({ loadEnv });
+	return runnerInstance;
+}
+async function initRunner(options = {}) {
+	const runner = getRunnerInstance(options);
+	await runner.init();
+	return runner;
+}
+//#endregion
+export { initRunner as n, getRunnerInstance as t };

package/dist/{src-p-GRSVDb.mjs → src-DjOTPnDz.mjs}

@@ -1,5 +1,5 @@
 import { G as matchesEvalTags$1, Pt as defineEval$1 } from "./runExecution-CjWJUUZ5.mjs";
-import "./cli-FOyPC8UD.mjs";
+import "./cli-CVBSlTD8.mjs";
 //#region src/index.ts
 /** Register an eval definition with typed tag support. */
 function defineEval(definition) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ls-stack/agent-eval",
-  "version": "0.60.0",
+  "version": "0.60.1",
   "type": "module",
   "bin": {
     "agent-evals": "./dist/bin.mjs"
@@ -32,9 +32,9 @@
   "devDependencies": {
     "@types/node": "^24.7.2",
     "typescript": "^5.9.2",
-    "@agent-evals/runner": "0.0.1",
+    "@agent-evals/sdk": "0.0.1",
     "@agent-evals/shared": "0.0.1",
-    "@agent-evals/sdk": "0.0.1"
+    "@agent-evals/runner": "0.0.1"
   },
   "scripts": {
     "build": "pnpm --filter @agent-evals/web build && pnpm --filter @agent-evals/shared build && pnpm --filter @agent-evals/sdk build && pnpm --filter @agent-evals/runner build && tsdown --filter cli-js && tsdown --filter cli-types",

package/skills/agent-eval/SKILL.md

@@ -34,9 +34,10 @@ display rules), read the TypeScript declarations shipped with the package:
   it before the next run starts. Temporary runs appear in `show-runs` while
   present; normal runs are never deleted by temporary-run cleanup. In the app,
   the run drawer can promote a temporary run to durable history.
-- `agent-evals app` watches `agent-evals.config.ts` and reloads config in
-  place when the runner is idle. If config changes during an active run, the
-  reload applies after the current run reaches a terminal state.
+- `agent-evals app` watches `agent-evals.config.ts` and the workspace `.env`
+  and reloads them in place when the runner is idle. If config or `.env`
+  changes during an active run, the reload applies after the current run
+  reaches a terminal state.
 - App-triggered runs log the queued target evals, resolved case concurrency,
   each case start for evals that are actually running, and the terminal run
   summary in the server terminal.

package/dist/runner-Dv5cseOt.mjs

@@ -1,15 +0,0 @@
-import { n as createRunner } from "./cli-FOyPC8UD.mjs";
-import "./src-p-GRSVDb.mjs";
-//#region ../../apps/server/src/runner.ts
-let runnerInstance = null;
-function getRunnerInstance() {
-	if (!runnerInstance) runnerInstance = createRunner();
-	return runnerInstance;
-}
-async function initRunner() {
-	const runner = getRunnerInstance();
-	await runner.init();
-	return runner;
-}
-//#endregion
-export { initRunner as n, getRunnerInstance as t };