amankansal-lt · Oct 12, 2026 · Oct 6, 2026 · Oct 6, 2026 · Oct 6, 2026 · Oct 6, 2026
diff --git a/.changeset/agent-device-0-21-25.md b/.changeset/agent-device-0-21-25.md
@@ -0,0 +1,6 @@
+---
+"@e2e-dev/mobile": patch
+"@e2e-dev/limrun": patch
+---
+
+Update agent-device to 0.21.25, the first release whose daemon loads WebDriver provider plugins such as `@agent-device/testmu`, which `@e2e-dev/testmu` needs on the same agent-device copy as `@e2e-dev/mobile`. agent-device 0.21.22 loads only plugins that return a native provider runtime, so it refuses that plugin. `@e2e-dev/limrun` pins the same version as `@e2e-dev/mobile`, so its `agent-device` peer moves to 0.21.25 too.
diff --git a/.changeset/browser-install-verdict.md b/.changeset/browser-install-verdict.md
diff --git a/.changeset/cache-empty-typed-value.md b/.changeset/cache-empty-typed-value.md
diff --git a/.changeset/decision-dismiss-keyboard.md b/.changeset/decision-dismiss-keyboard.md
@@ -0,0 +1,5 @@
+---
+"@e2e-dev/decision": patch
+---
+
+The decision executor offers `dismiss_keyboard` while an on-screen keyboard is showing and the engine can dismiss it. Before, a control the keyboard covered after typing (an iOS number pad over a submit button) could not be reached, and the step failed.
diff --git a/.changeset/decision-executor.md b/.changeset/decision-executor.md
diff --git a/.changeset/decision-stall-clock.md b/.changeset/decision-stall-clock.md
@@ -0,0 +1,5 @@
+---
+"@e2e-dev/decision": patch
+---
+
+The decision executor's stall guard now counts an action that errored as no progress, so three actions in a row that failed or changed nothing on screen block the step. It used to look only at whether the page changed, and a ticking timer on screen changes it every time: a tap that failed with `APP_UNREACHABLE` again and again never tripped the guard, and the model could repeat it until it gave up and claimed the step failed. A skipped pick and a rejected `done` or `failed` claim also carry an error in the step's history, so they count toward the three as well.
diff --git a/.changeset/download-full-ledger-pass.md b/.changeset/download-full-ledger-pass.md
diff --git a/.changeset/engine-prepare-headed.md b/.changeset/engine-prepare-headed.md
diff --git a/.changeset/error-default-redactor.md b/.changeset/error-default-redactor.md
diff --git a/.changeset/evaluate-string-expression.md b/.changeset/evaluate-string-expression.md
diff --git a/.changeset/failure-evidence-budget.md b/.changeset/failure-evidence-budget.md
diff --git a/.changeset/forced-conclusion-wrong-tool.md b/.changeset/forced-conclusion-wrong-tool.md
diff --git a/.changeset/foreign-engine-errors.md b/.changeset/foreign-engine-errors.md
diff --git a/.changeset/in-process-stray-rejections.md b/.changeset/in-process-stray-rejections.md
diff --git a/.changeset/limrun-provider.md b/.changeset/limrun-provider.md
@@ -0,0 +1,9 @@
+---
+"@e2e-dev/mobile": patch
+"@e2e-dev/limrun": minor
+---
+
+Add the Limrun device provider with one hosted instance per worker, app
+installation, recording through agent-device, and cleanup on run exit.
+Renew retained driver leases while other targets run. Keep provider-owned
+local daemons isolated from ambient remote-daemon environment settings.
diff --git a/.changeset/mcp-error-redaction.md b/.changeset/mcp-error-redaction.md
diff --git a/.changeset/mcp-headless-default.md b/.changeset/mcp-headless-default.md
diff --git a/.changeset/mcp-user-output-redaction.md b/.changeset/mcp-user-output-redaction.md
diff --git a/.changeset/npm-exec-cli-hint.md b/.changeset/npm-exec-cli-hint.md
diff --git a/.changeset/oxc-typescript-loader.md b/.changeset/oxc-typescript-loader.md
diff --git a/.changeset/plain-string-secret-trace.md b/.changeset/plain-string-secret-trace.md
diff --git a/.changeset/replay-lost-list-once.md b/.changeset/replay-lost-list-once.md
diff --git a/.changeset/retry-crash-attempts.md b/.changeset/retry-crash-attempts.md
diff --git a/.changeset/run-user-output-redaction.md b/.changeset/run-user-output-redaction.md
diff --git a/.changeset/serial-group-verdicts.md b/.changeset/serial-group-verdicts.md
diff --git a/.changeset/strict-mode-reresolve.md b/.changeset/strict-mode-reresolve.md
diff --git a/.changeset/swipe-path-duration.md b/.changeset/swipe-path-duration.md
diff --git a/.changeset/tainted-fallback-policy-denied.md b/.changeset/tainted-fallback-policy-denied.md
diff --git a/.changeset/telemetry-identity.md b/.changeset/telemetry-identity.md
diff --git a/.changeset/testmu-devices.md b/.changeset/testmu-devices.md
@@ -2,4 +2,4 @@
 "@e2e-dev/testmu": minor
 ---
 
-`@e2e-dev/testmu`: `mobile({ device: testmu({ device, osVersion, app }) })` runs a mobile target on TestMu AI (formerly LambdaTest) Android emulators, iOS simulators, and real devices (`deviceType: 'real'`), one per worker slot, through agent-device's `testmu` cloud provider. Each slot leases a device from an agent-device daemon the provider starts for the run under `stateDir` (default `.e2e/testmu`), in a directory per run that it marks as its own and that later runs remove once nothing in it has changed for 24 hours; allocating a lease starts its TestMu AI session, which installs `app` (an `lt://` id, an `https` URL, or a local path), so every slot is billed from the moment it is leased, and releasing the lease ends it. It reads `LT_USERNAME` and `LT_ACCESS_KEY` from the run's environment and fails the lease before starting anything when either is missing, sets up the device with `orientation`, `geoLocation`, `timezone`, `language`, `locale`, and `appiumVersion` when given, groups sessions on the dashboard under `project` (default `e2e`) and `build` (default the run id) and names each slot's session `sessionName` (default `e2e-<run id>-<target>-<slot>`, with `-<slot>` appended to a given name when the target has more than one slot), allocates each lease with agent-device's longest inactivity window and heartbeats it every 2 minutes while the run holds it, so a session that takes up to 10 minutes to start keeps its lease, links TestMu AI's recording of the slot's session as the video of an attempt that records one (the whole session, starting at the session's start time), found through TestMu AI's sessions API by the session's build and name, releases a lease it allocated when the run was interrupted, and releases each lease once. It needs an agent-device that includes the `testmu` provider, which no published release does yet, with `@e2e-dev/mobile`'s agent-device overridden to the same version; `agent-device` and `@e2e-dev/mobile` are its peers.
+`@e2e-dev/testmu`: `mobile({ device: testmu({ device, osVersion, app }) })` runs a mobile target on TestMu AI (formerly LambdaTest) Android emulators, iOS simulators, and real devices (`deviceType: 'real'`), one per worker slot, through agent-device's `@agent-device/testmu` plugin. Each slot leases a device from an agent-device daemon the provider starts for the run under `stateDir` (default `.e2e/testmu`), in a directory per run that it marks as its own and that later runs remove once nothing in it has changed for 24 hours; allocating a lease starts its TestMu AI session, which installs `app` (an `lt://` id, an `https` URL, or a local path), so every slot is billed from the moment it is leased, and releasing the lease ends it. It reads `LT_USERNAME` and `LT_ACCESS_KEY` from the run's environment and fails the lease before starting anything when either is missing, sets up the device with `orientation`, `geoLocation`, `timezone`, `language`, `locale`, and `appiumVersion` when given, groups sessions on the dashboard under `project` (default `e2e`) and `build` (default the run id) and names each slot's session `sessionName` (default `e2e-<run id>-<target>-<slot>`, with `-<slot>` appended to a given name when the target has more than one slot), allocates each lease with agent-device's longest inactivity window and heartbeats it every 2 minutes while the run holds it, so a session that takes up to 10 minutes to start keeps its lease, links TestMu AI's recording of the slot's session as the video of an attempt that records one (the whole session, starting at the session's start time), found through TestMu AI's sessions API by the session's build and name, releases a lease it allocated when the run was interrupted, and releases each lease once. Each lease is allocated with `retainOnClose`, so the engine closing a worker's agent-device session does not end the TestMu AI session; a daemon that does not keep the lease fails it and the provider releases it. It needs agent-device 0.21.25 or later, the same copy `@e2e-dev/mobile` 0.11.1 or later uses, with the plugin added once per machine (`agent-device plugins add @agent-device/testmu`, then `agent-device daemon stop` for a daemon already running); `agent-device` and `@e2e-dev/mobile` are its peers.
diff --git a/.changeset/web-hung-page-timeouts.md b/.changeset/web-hung-page-timeouts.md
diff --git a/.changeset/worker-hang-watchdog.md b/.changeset/worker-hang-watchdog.md
diff --git a/.claude/skills/authoring-docs b/.claude/skills/authoring-docs
@@ -0,0 +1 @@
+../../.dev/skills/authoring-docs
diff --git a/.claude/skills/babysit b/.claude/skills/babysit
@@ -0,0 +1 @@
+../../.dev/skills/babysit
diff --git a/.claude/skills/ship-pr b/.claude/skills/ship-pr
@@ -0,0 +1 @@
+../../.dev/skills/ship-pr
diff --git a/.claude/skills/verify b/.claude/skills/verify
@@ -0,0 +1 @@
+../../.dev/skills/verify
diff --git a/.claude/skills/writing-pr b/.claude/skills/writing-pr
@@ -0,0 +1 @@
+../../.dev/skills/writing-pr