@ls-stack/agent-eval 0.36.0 → 0.38.0

package/dist/{runner-BCs5rzej.mjs → runner-B1wfPKNH.mjs}

@@ -1,5 +1,5 @@
-import { n as createRunner } from "./cli-Dg3abrOv.mjs";
-import "./src-DBypR4TV.mjs";
+import { n as createRunner } from "./cli-BUX6tr9J.mjs";
+import "./src-BwJ5tod2.mjs";
 //#region ../../apps/server/src/runner.ts
 let runnerInstance = null;
 function getRunnerInstance() {

package/dist/{runner-znY6PY1M.mjs → runner-CoRf7Vef.mjs}

@@ -1,2 +1,2 @@
-import { n as initRunner, t as getRunnerInstance } from "./runner-BCs5rzej.mjs";
+import { n as initRunner, t as getRunnerInstance } from "./runner-B1wfPKNH.mjs";
 export { getRunnerInstance, initRunner };

package/dist/src-BwJ5tod2.mjs

@@ -0,0 +1,3 @@
+import "./runOrchestration-BhUFWvq9.mjs";
+import "./cli-BUX6tr9J.mjs";
+export {};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ls-stack/agent-eval",
-  "version": "0.36.0",
+  "version": "0.38.0",
   "type": "module",
   "bin": {
     "agent-evals": "./dist/bin.mjs"
@@ -31,9 +31,9 @@
   "devDependencies": {
     "@types/node": "^24.7.2",
     "typescript": "^5.9.2",
+    "@agent-evals/sdk": "0.0.1",
     "@agent-evals/runner": "0.0.1",
-    "@agent-evals/shared": "0.0.1",
-    "@agent-evals/sdk": "0.0.1"
+    "@agent-evals/shared": "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

@@ -27,7 +27,13 @@ display rules), read the TypeScript declarations shipped with the package:
 - Unfiltered `agent-evals run` is disabled by default; use `--eval` or `--case`
   for targeted CLI runs. Set `allowCliRunAll: true` in
   `agent-evals.config.ts` to opt into run-all CLI behavior. The web UI can
-  still run grouped evals and confirms before starting more than five.
+  still run grouped evals and confirms before starting more than five. On a
+  single eval page, the Run chevron can open a picker to run specific authored
+  case ids; those case-picked runs are temporary by default and can be made
+  durable in the modal.
+- `agent-evals run --temporary` persists a run like normal history, but deletes
+  it before the next run starts. Temporary runs appear in `show-runs` and the UI
+  while present; normal runs are never deleted by temporary-run cleanup.
 - `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 UI
   shows a pending reload banner and blocks new runs until the current run
@@ -485,6 +491,10 @@ Mental model:
   `.agent-evals/cache/<owner>.json`; each namespace is capped at 100 entries by
   default. Configure `cache.maxEntriesPerNamespace` for the default cap and
   `cache.maxEntriesByNamespace` for exact namespace-specific caps.
+- Nested cached JSON values at or above roughly 10K JSON characters are stored
+  as content-addressed Brotli blobs under `.agent-evals/cache-blobs/` and
+  referenced from cache JSON by sha256. Identical large payloads share the same
+  blob.
 - Authored raw cache keys are stored for debugging under
   `.agent-evals/cache-debug/<owner>.json`. This folder may include prompts,
   user inputs, or other sensitive data, should be gitignored, and is not needed
@@ -492,11 +502,11 @@ Mental model:
   can be filtered to hits or new entries added by cache misses/refreshes.
   Misses/refreshes with `cache.store: false` are shown as non-stored activity
   without fetch/delete controls.
-- Cached payloads use advanced serialization/deserialization with the Web API
-  plugin set, so return values and recorded SDK effects preserve richer
-  built-ins such as `Date`, `Map`, `Set`, typed arrays, `URL`, `Headers`,
-  `Blob`, and `File` on hits. Undefined values are omitted by default instead
-  of being written to cache files; direct serializer callers can pass
+- Cached payloads use JSON-safe tagged serialization, so return values and
+  recorded SDK effects preserve richer built-ins such as `Date`, `Map`, `Set`,
+  typed arrays, `URL`, `Headers`, `Blob`, and `File` on hits. Undefined values
+  are omitted by default instead of being written to cache files; direct
+  serializer callers can pass
   `{ preserveUndefined: true }` when explicit undefined wrappers are needed.
   Cache keys still use the deterministic key-hashing rules above.
 - Cache mode per run is controlled by CLI flags (see `agent-evals run --help`).
@@ -507,6 +517,8 @@ Run output lives under `.agent-evals/runs/<run-id>/` and cache entries under
 `.agent-evals/cache/<eval-id>.json`. Files in a run directory include run
 metadata, a run summary, per-case results, and per-case trace JSON. Inspect
 these when debugging persisted output, costs, columns, traces, or failures.
+Temporary runs use the same directory layout, but are removed before the next
+run of any kind starts.
 
 Use `agent-evals show-runs` when you need stable file
 paths before reading saved output: