@@ -2,6 +2,7 @@
 
 ## Unreleased
 
+- Provider plugins have a new contract and one provider registry. A package exports a factory at `./agent-device` that returns `{ apiVersion: 2, provider, credentials, connection?, setup(host) }`, typed by `AgentDevicePluginFactory` from `agent-device/plugins`; the store calls it with the package's `config.json` options and a JavaScript host can call it directly. Bundled providers, installed plugins, and embedder plugins register in that order through one registry that claims each provider id once: a duplicate id is recorded as failed, and `cloud` and `proxy` stay reserved. A plugin that fails to load, validate, or set up no longer stops the daemon; it is listed as `failed` with a typed `error.reason`, and the other providers still register. The inventory is published in `daemon.json` and shown by the new `agent-device daemon status [--json]`; `plugins list --json` works without a daemon and reports a running daemon's inventory under `daemon`. A client now takes provider credential variable names from that inventory, narrowed per lease backend, and falls back to v1 manifests. Version 1 plugins (`agentDevicePlugin.apiVersion: 1`) keep loading. A Limrun credential fingerprint now hashes the exact variable values, so padding whitespace around a value counts as a change like it does for BrowserStack, and a partial Limrun instance set shows in `daemon status` and as `provider_plugin_failed` in `daemon.log` instead of `provider_runtime_skipped`.
 - Add `uninstall <app>` and `client.apps.uninstall({ app })` to remove an installed app on adb-backed Android targets, iOS simulators, and CoreDevice iOS devices. Both report `wasInstalled`. A package name or bundle id that is not installed succeeds with `wasInstalled: false`. Other targets, Appium and Limrun sessions included, refuse it before anything runs.
 - Add `settings mobile-data on|off` for adb-backed Android targets. It switches mobile data through `svc data` and reports the `mobileData` state stored in the global `mobile_data` row; multi-SIM read-back is not covered yet. Every other target refuses it before anything runs, and `--app` is refused because the switch is device-wide.
 - On iOS, a `fill` that runs under a degraded XCTest snapshot channel now aims at a point, including ref, selector and replay fills. When the runner identifies no text input under that point, it refuses with `NO_TEXT_INPUT_AT_POINT` (`error.details.reason: no_text_input_at_point`, `dispatched: no`) before dispatching its focus tap, and no field changes; when the input lookup itself cannot answer, it refuses with `TEXT_INPUT_PROBE_UNAVAILABLE` (`error.details.reason: text_input_probe_unavailable`). Previously such a fill cleared and typed through the app-level first responder, so it could rewrite a field the caller never named, or report a clear failure after emptying it. A ref or selector fill whose rect centre misses the real field (an outer container, a non-input node) now needs a point inside the input's own rect.
 
@@ -0,0 +1,66 @@
+# 0034. Provider plugin registry
+
+Status: accepted. Umbrella #3747; contract and registry #3755.
+
+## Rules at a glance
+
+- One contract. A provider plugin is a factory that returns `{ apiVersion: 2, provider, credentials,
+  connection?, setup(host) }`. `setup` returns `{ runtime, platformModule, dispose? }` or
+  `{ webDriver, dispose? }`. A `{ webDriver }` result stays a data descriptor the host interprets.
+- One registry. Sources compose in the order built-in, store, embedder. Built-in providers use the
+  same shape as an installed plugin, and one function validates what every factory returns.
+- Plugins import only types from agent-device. The type-only contract is `agent-device/plugins`;
+  there is no runtime helper, and the contract does not depend on a plugin's module URL.
+- The first registration of a provider id claims it, even when that registration fails. Later ones
+  are recorded as failed. `cloud` and `proxy` are routing identities no plugin can claim.
+- A plugin that fails to enumerate, load, validate or set up is recorded and skipped. The daemon
+  starts with the rest. The inventory lists every plugin as `active` or `failed`, with a typed
+  `error.reason`, and is published in `daemon.json` and by `agent-device daemon status`.
+- The plugin set is fixed for a daemon's lifetime. A plugin change takes effect on the next daemon.
+- v1 plugins (`agentDevicePlugin.apiVersion: 1`) without a `./agent-device` export are dispatched by
+  their manifest before any code runs and load through an adapter until 0.22 removes them (#3759).
+  A package that keeps the manifest and also exports `./agent-device` loads as v2; hosts without
+  v2 still read the manifest.
+
+## Context
+
+Three things loaded providers: a hand-kept `RESERVED_PLUGIN_PROVIDERS` list, a second list of
+providers the daemon must have a runtime for, and a store that treated one bad plugin as a daemon
+startup failure. Each provider kind also needed its own path for credential fingerprints and for
+`connect` capabilities, which read manifests the bundled providers do not have. An embedding host
+had to fake the store to load a plugin from its own dependencies (#3747).
+
+## Decisions
+
+- **A factory takes options.** The store calls it with `config.json` `plugins[name].options`; a JS
+  host calls it directly with its own, such as credentials. A plugin that takes credentials as
+  options declares no environment variables for them, so they stay out of the environment
+  fingerprint.
+- **A package exports its entry as `./agent-device`.** The host resolves it with import
+  conditions and falls back to the package root only when the subpath is absent. A subpath whose
+  import or setup throws never falls back. A v2 package needs no manifest.
+- **The registry owns identity, not the retired lists.** Duplicate ids and bundled ids are one
+  question answered by claim order. Credential variables, per-backend narrowing and connect
+  capabilities are read from the registry's plugins, not from manifests. A v1 manifest still
+  supplies them for a v1 plugin.
+- **Failure is data.** A failed plugin is an inventory entry, not an exception. Behavior keys on
+  `error.reason`, never on the message.
+- **A client reads a daemon's inventory from `daemon.json`.** The client credential fingerprint
+  takes variable names from it and falls back to v1 manifests. A v2 plugin declares its connect
+  capabilities in code, so a client loads the store, without setting up a runtime, before it reads
+  them.
+
+## Daemon ownership
+
+The target is that starting or replacing a daemon is an explicit act, not a side effect of a client
+call. Today a client call can still start a daemon; #3754 makes that explicit for the client
+lifecycle, and #3757 adds `agent-device/host`, where an embedder owns its daemon and passes its
+plugins as the third source.
+
+## Alternatives refused
+
+- **Keep the manifest as the contract.** It cannot describe options-dependent credentials, and a
+  bundled provider has no package to hold one.
+- **A runtime helper such as `definePlugin`.** It would make plugins depend on a runtime export of
+  agent-device and its version, which is what a type-only contract avoids.
+- **Fail the daemon on a bad plugin.** One broken optional package then takes every provider down.
@@ -35,6 +35,7 @@
 | [0031 macOS Native App Backend](0031-macos-native-app-backend.md) | `AGENT_DEVICE_MACOS_APP_BACKEND`, driving macOS app sessions without XCTest Automation Mode, why pointer actions are accessibility actions only, and the ghost cursor |
 | [0032 Layering Graph Engine](0032-layering-graph-engine.md) | evaluating replacement of `scripts/layering/` graph rules with dependency-cruiser or fallow `boundaries`, why graph-based layering checks share one replaceable import model, the measured edge-set and parity gaps, and the revisit triggers |
 | [0033 Collocation decision table](0033-collocation-decision-table.md) | moving or adding a file listed in umbrella #3276 §5, re-proposing a move that table refused, reading a zone's collocation decisions |
+| [0034 Provider plugin registry](0034-provider-plugin-registry.md) | provider plugins, the `agent-device/plugins` contract, plugin sources and order, duplicate provider ids, plugin failure isolation, the daemon plugin inventory and `daemon status` |
 
 ADRs record *why*; the registries and gates they describe are the living source of truth — when
 prose and a registry disagree, the registry wins and the ADR needs a follow-up.
 
@@ -85,7 +85,7 @@ Entry points for integrations:
 - `agent-device/plugins/webdriver`
   - experimental `WebDriverPluginOptions` for providers using the shared engine.
 - `agent-device/plugins`
-  - experimental factory context: `ProviderPluginHost`; see [provider plugins](./plugins.md).
+  - experimental plugin contract: `AgentDevicePluginFactory` and `ProviderPluginHost`; see [provider plugins](./plugins.md).
 
 ## Android ADB providers
 
 
@@ -29,7 +29,7 @@ An agent in a sandbox can drive an instance that someone else created, such as t
 | iOS | `LIM_IOS_INSTANCE_URL` (`status.apiUrl`), `LIM_IOS_INSTANCE_TOKEN` (`status.token`) |
 | Android | `LIM_ANDROID_INSTANCE_URL` (`status.apiUrl`), `LIM_ANDROID_INSTANCE_TOKEN` (`status.token`), `LIM_ANDROID_INSTANCE_ADB_URL` (`status.adbWebSocketUrl`) |
 
-The `lim` CLI reads the same variables, so one set works for both tools. Set all of a platform's variables or none. With a partial set, `connect` fails, and the daemon disables Limrun and records `provider_runtime_skipped` in `daemon.log`.
+The `lim` CLI reads the same variables, so one set works for both tools. Set all of a platform's variables or none. With a partial set, `connect` fails, and the daemon registers without Limrun, lists it as `failed` in `agent-device daemon status`, and records `provider_plugin_failed` in `daemon.log`.
 
 ```bash
 export LIM_IOS_INSTANCE_URL=...