pi-browser-use 0.10.1 → 0.11.0

package/dist/chrome-launcher.js

@@ -179,7 +179,7 @@ class OwnedChromeProcess {
     waitForExit() {
         return this.exitPromise;
     }
-    async shutdown(graceMs = 3_000) {
+    async shutdown(graceMs = 10_000) {
         if (this.exited)
             return;
         this.child.kill('SIGTERM');

package/docs/performance.md

@@ -57,3 +57,15 @@ npm run bench -- --startup-only --json
 ```
 
 `perf:audit` and `bench --json` are suitable for machine capture. Use at least seven cold imports and ten runtime iterations when updating the table. Record Node, OS, architecture, commit, and whether Chrome was headless; do not compare results collected with different modes or iteration counts.
+
+## Task-level evaluation
+
+`npm run eval -- --json` runs deterministic browser tasks against fixed DOM fixtures in a fresh isolated session. Each task attempt gets its own runtime and records objective checks, startup/task/step timings, bounded failure text, evidence-capture error counts, and artifact paths in `eval-results/result.json`. Failed attempts save a screenshot by default; use `--evidence all` to capture evidence after every successful step or `--evidence none` to disable it. Failures are classified as `harness` (the browser environment never became usable: startup failed, no step succeeded, or cleanup broke) or `task` (the scenario regressed), and the summary reports `harnessFailures` separately so a broken browser environment is not mistaken for task regressions. Unknown CLI options are rejected. The report contract is documented in [Eval result schema](eval-results.md), including the `eval-budgets.json` gate that mirrors the Code Foundry eval contract.
+
+```sh
+npm run eval -- --iterations 3 --json
+npm run eval -- --task form-submit,extract-list --evidence all
+npm run eval -- --list
+```
+
+These tasks evaluate runtime and browser-tool behavior, not model quality. Keep them deterministic and network-free for local/CI regression checks. Model-agent evaluations should use the same task IDs and metadata but run in a separate harness, with the model, prompt, reasoning budget, dependency hash, and retry policy recorded. Do not silently truncate normal browser results; bounded evidence is only for judge-facing output.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-browser-use",
-  "version": "0.10.1",
+  "version": "0.11.0",
   "description": "Opinionated browser automation via chrome-devtools-mcp: native Pi extension and portable Agent Plugins 1.0 skills + MCP server.",
   "keywords": [
     "automation",
@@ -54,6 +54,8 @@
     "typecheck": "tsc -p tsconfig.json --noEmit",
     "prepublishOnly": "npm run build",
     "bench": "node scripts/bench.mjs",
+    "eval": "node scripts/eval.mjs",
+    "preeval": "npm run build",
     "test": "node --test test/*.test.mjs",
     "preperf:audit": "npm run build",
     "perf:audit": "node scripts/performance.mjs",
@@ -73,7 +75,7 @@
     "@types/node": "^22.0.0",
     "oxfmt": "^0.66.0",
     "oxlint": "^1.81.0",
-    "typescript": "^5.9.0"
+    "typescript": "^6.0.3"
   },
   "peerDependencies": {
     "@earendil-works/pi-ai": ">=0.80.10",

package/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://agent-plugins.org/schemas/1.0.0/plugin.schema.json",
   "name": "pi-browser-use",
-  "version": "0.10.1",
+  "version": "0.11.0",
   "description": "Managed persistent, fresh, and existing Chrome sessions with focus-safe browser tools and CLI-first skills.",
   "author": {
     "name": "0xPlayerOne"