pi-browser-use 0.10.1 → 0.11.1

package/docs/performance.md

@@ -41,6 +41,8 @@ Budgets live in [`performance-budgets.json`](../performance-budgets.json). They
 | Published map files               |             0 |
 | Production dependency closure     |  140 packages |
 
+The published artifact stays byte-lean by design: `tsconfig.json` sets `removeComments` so emitted JavaScript and declarations carry no comment bytes, and build maps never ship. Source commentary lives in the repository, not in the installed package.
+
 ## Reproduce locally and in CI
 
 ```sh
@@ -57,3 +59,15 @@ 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/extension/README.md

@@ -17,6 +17,8 @@ default, override via `tabBridgePort` in `pi-browser-use` settings plus the
 extension's stored `piBridgeUrl`) for open-tab requests, creates each tab
 inactive in the `pi-browser-use` group, keeps the group collapsed, and
 reports back per-token so Pi selects the exact tab — never by URL matching.
+Polling is event-driven plus a one-minute alarm; it does not run a persistent
+timer that can keep Chrome alive after its windows are closed.
 
 ## Permissions
 

package/extension/background.js

@@ -112,7 +112,6 @@ chrome.runtime.onMessage.addListener((message, _sender, sendResponse) => {
  * The bridge URL override lives in local storage (default loopback port).
  */
 const PI_BRIDGE_DEFAULT_URL = 'http://127.0.0.1:31973'
-const PI_BRIDGE_POLL_MS = 2000
 let piBridgeUrlOverride = null
 
 try {
@@ -163,15 +162,11 @@ async function pollBridgeOnce() {
   }
 }
 
-setInterval(() => {
-  void pollBridgeOnce().catch(() => {})
-}, PI_BRIDGE_POLL_MS)
-
 /**
- * MV3 service workers are short-lived: setInterval dies permanently once the
- * worker is evicted, so the fast poll above only covers the awake case. The
- * repeating alarm wakes a suspended worker and guarantees a poll at least
- * every minute (safe floor for the alarm cadence). Re-created on every
+ * MV3 service workers are short-lived, and a persistent timer can keep Chrome
+ * alive after the user closes its windows. Use an alarm instead: it wakes a
+ * suspended worker without holding the browser open, and guarantees a poll at
+ * least every minute (safe floor for the alarm cadence). Re-created on every
  * startup since alarm persistence across restarts is not guaranteed.
  */
 const PI_BRIDGE_ALARM = 'pi-bridge-poll'

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-browser-use",
-  "version": "0.10.1",
+  "version": "0.11.1",
   "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": "0.11.1",
   "description": "Managed persistent, fresh, and existing Chrome sessions with focus-safe browser tools and CLI-first skills.",
   "author": {
     "name": "0xPlayerOne"