@usestatemachines/sdk 0.3.0 → 0.5.0

package/README.md

@@ -1,7 +1,7 @@
 # @usestatemachines/sdk
 
 TypeScript SDK for [State](https://usestatemachines.com): start disposable environments of replica
-apps, connect to them, snapshot their data, and stop them.
+apps, connect to them, snapshot their data, suspend and resume them, and stop them.
 
 - Node.js 20.20 or later (22 or 24 recommended), and Cloudflare Workers. ESM only.
 - Retries transient failures safely, including `create`, which is idempotent.
@@ -34,12 +34,13 @@ It takes `{ readyTimeoutMs, signal }`: the wait gives up after `readyTimeoutMs`
 minutes by default). A string app is shorthand for `{ key: s, appId: s }`; nothing else is inferred.
 Salesforce offers several APIs, so pass `apiId`.
 
-If the wait fails after the environment was created, `start()` stops the environment, then
-rethrows the original error. The SDK's errors (`EnvironmentFailedError`, `TimeoutError` or a
-`StateMachinesError`; check with `instanceof`) carry `environmentId` and `stopped`. `stopped` is
-`true` when the stop request succeeded. When it's `false`, the stop wasn't confirmed:
-`stopError` says why, and you should call `sm.environments.stop(err.environmentId)`, which is safe
-to repeat. The exception is your own key, below.
+`env.stop()` ends the environment and discards its data. If the wait fails after the environment
+was created, `start()` stops the environment, then rethrows the original error. The SDK's errors
+(`EnvironmentFailedError`, `TimeoutError` or a `StateMachinesError`; check with `instanceof`)
+carry `environmentId` and `stopped`. `stopped` is `true` when the stop request succeeded. When
+it's `false`, the stop wasn't confirmed: `stopError` says why, and you should call
+`sm.environments.stop(err.environmentId)`, which is safe to repeat. The exception is your own key,
+below.
 
 When you abort the `signal` during the wait, `start()` still stops the environment. If the stop
 succeeded, it rejects with your `signal.reason`, unchanged. If not, it rejects with a
@@ -84,6 +85,35 @@ chosen actor's permissions. Without an `actorId`, `connect()` uses the environme
 default connection, if it has one, or else the first actor the app lists, the administrator, so
 pick a less-privileged actor to test permission behaviour.
 
+## Suspending and resuming
+
+Suspend an environment to end its sandboxes and keep its data, then resume it later on the same
+data:
+
+```ts
+await env.suspend();
+await env.waitUntilSuspended();
+// Later, from the same handle or with sm.environments.resume(env.id):
+await env.resume();
+await env.waitUntilReady();
+```
+
+- `env.suspend()` is `sm.environments.suspend(id)`. The environment reads `suspending`, then
+  `suspended`, its apps read `suspended` and `suspendedAt` is set. Only a running environment
+  whose apps are all ready can be suspended; anything else answers `409 not-ready`.
+- While suspended, the lifetime is paused: `expiresAt` moves out by the suspended time when it
+  resumes, and the environment doesn't count toward the live-environment cap.
+- `resume()` starts it again; it answers `409 not-suspended` unless the environment is
+  `suspended` or `failed`, and `429 environment-limit` when the cap is reached. `waitUntilSuspended()` resolves
+  on `suspended` and throws `EnvironmentFailedError` if the environment ends instead.
+- Suspending isn't cleanup: a suspended environment keeps its data until you stop it.
+  `stop()` interrupts ongoing work, including suspension and snapshot capture.
+
+Lifecycle commands require orchestration availability. `stop`, `suspend` and `resume` each
+send a fresh idempotency key and reuse it across automatic retries. You can pass your own
+`{ idempotencyKey }`. Failures carry that key so you can retry the same command after a timeout;
+a timeout does not establish whether the command was accepted.
+
 ## Full control
 
 Use `create`, `waitUntilReady` and `connect` directly when you want to run each step yourself, or
@@ -116,18 +146,19 @@ try {
 
 1. `create(input)` returns one `Environment` as soon as the API accepts it, before the apps boot.
 2. `waitUntilReady(id)` polls until it is `running`. It throws `EnvironmentFailedError` if the
-   environment becomes `failed`, `stopping` or `stopped`, and `TimeoutError` after `timeoutMs`
-   (default 10 minutes). It **never stops the environment**, even when it throws.
+   environment becomes `failed`, `suspending`, `suspended`, `stopping` or `stopped`, and
+   `TimeoutError` after `timeoutMs` (default 10 minutes). It **never stops the environment**,
+   even when it throws.
 3. `connect(id, appKey, { apiId, actorId })` returns a `Connection` to one app (see
    [Connections](#connections)).
 4. `stop(id)` returns as soon as the API accepts it, and a second call also resolves.
-   `waitUntilStopped(id)` resolves only when the environment is `stopped` or `failed` **and**
+   `waitUntilStopped(id)` resolves only when the environment is `stopped` **and**
    every app's `disposal` is `complete`. If any disposal is `failed` or `unknown`, it throws
    `DisposalIncompleteError` with the environment.
 
-`lifetimeMinutes` counts from when the environment starts, so boot time uses part of it. When it
-runs out, the environment stops with `stopReason: 'expired'`. The default and maximum are in
-`(await sm.credential.get()).limits.environmentLifetime`.
+`lifetimeMinutes` counts from when the environment starts, so boot time uses part of it; time spent
+suspended doesn't count. When it runs out, the environment stops with `stopReason: 'expired'`. The
+default and maximum are in `(await sm.credential.get()).limits.environmentLifetime`.
 
 Both waits accept `{ timeoutMs, signal }`. Aborting the `signal` rejects the wait with
 `signal.reason`. They poll at growing intervals of about 1 to 5 seconds, with random jitter, and
@@ -169,8 +200,10 @@ Pass `snapshotId` instead of `apps` to `start()` or `create()`. Each app keeps t
 captured; `snapshotAppInstanceKeys` restores only some of them. The new environment's
 `sourceEnvironmentSnapshotId` and each app's `sourceAppInstanceSnapshotId` name the source.
 
-- A snapshot is taken per app instance (`consistency: 'per-app-instance'`), not at one instant
-  across apps. It can only be taken while the environment is `running` and each app is `ready`.
+- Every app is paused while the snapshot is cut, so a multi-app snapshot records one instant
+  across apps (`consistency: 'environment'`). A single app, or a capture whose pause was refused,
+  is `'app_instance'`. It can only be taken while the environment is `running` and each app is
+  `ready`.
 - `readyTimeoutMs` bounds the wait for `ready` (10 minutes by default), as in `start()`. While the
   API is busy (429 `snapshot-limit`), the capture is retried with the same idempotency key for up
   to 60 seconds.
@@ -241,8 +274,8 @@ attempts:
 | --- | --- | --- |
 | `502`, `503`, `504` | Yes | The API is saturated (`503`), or Cloudflare can't reach it for a moment (`502`, `504`). |
 | No response: connection error or attempt timeout | Yes | The request may never have arrived. |
-| `429` (`environment-limit`) | No | A quota on live environments. Stop one, then create again. |
-| `429` (`snapshot-limit`) | Only by `snapshots.capture()` and `env.snapshot()`, for up to 60 seconds | The API is running as many captures as it can. |
+| `429` (`environment-limit`) | No | A quota on live environments. Suspend or stop one, then create again. |
+| `429` (`snapshot-limit`) | Only by `snapshots.capture()` and `env.snapshot()`, for up to 60 seconds | Sent by older APIs when too many captures ran. The current API answers `409 capture-in-progress` instead. |
 | Any other status: other `4xx`, and other `5xx` | No | Not retried by the SDK. |
 
 - A `Retry-After` header is honoured, up to 5 seconds.
@@ -334,7 +367,8 @@ created also carry `snapshotId` and `idempotencyKey`.
 
 | Code | Status | Meaning |
 | --- | --- | --- |
-| `snapshot-limit` | 429 | The API is running as many captures as it can. `capture` retries it for up to 60 seconds; `create` doesn't. |
+| `capture-in-progress` | 409 | This environment is already capturing a snapshot; one runs at a time. Wait for it (`waitUntilReady`), then capture again. |
+| `snapshot-limit` | 429 | Older APIs only: too many captures were running. `capture` retries it for up to 60 seconds; `create` doesn't. |
 | `not-ready` | 409 | Capture: the environment isn't `running` or an app isn't `ready`. Restore: the snapshot isn't `ready` yet. Delete: the snapshot is still `capturing`. |
 | `version-conflict` | 409 | `rename` sent a stale `expectedVersion`, or the snapshot was deleted. `get` it: if `deletedAt` is set it's gone, otherwise retry with its `version`. |
 | `release-unavailable` | 422 | Restore: no release on this deployment matches the snapshot's apps. |
@@ -346,7 +380,7 @@ try {
   await sm.environments.create(input);
 } catch (error) {
   if (error instanceof StateMachinesError && error.code === ErrorCode.EnvironmentLimit) {
-    console.error('Too many live environments; stop one first.');
+    console.error('Too many live environments; suspend or stop one first.');
   } else {
     throw error;
   }
@@ -433,6 +467,30 @@ const response = await fetch(new URL('services/data/v67.0/limits', salesforce.ba
 });
 ```
 
+## Recorded requests
+
+Request history is best effort. Missing records do not mean that no requests ran.
+Stored records remain readable after an environment stops.
+
+```ts
+const stats = await sm.environments.getRequestStats(env.id, {
+  from: new Date(Date.now() - 60 * 60_000),
+  to: new Date(),
+});
+console.log(stats.count, stats.p95Ms);
+const page = await sm.environments.listRequests(env.id, { appKey: 'salesforce', limit: 20 });
+const first = page.data[0];
+if (first) {
+  const detail = await sm.environments.getRequest(env.id, first.id);
+  console.log(detail.status, detail.responseBody.preview);
+  const body = await sm.environments.getRequestPayload(env.id, first.id, 'response');
+  const bytes = await new Response(body).arrayBuffer();
+}
+```
+
+Payload coverage reports credential redaction and omissions. The detail preview
+is bounded; the payload method returns the full stored body as a byte stream.
+
 ## Supported runtimes
 
 - **Node.js 20.20 or later**, with 22 or 24 recommended: Node 20 is past its end of life, and
@@ -449,6 +507,14 @@ TypeScript consumers need `lib: ["DOM"]` or `@types/node`, for the `fetch`, `Req
   version.
 - The API may add error codes and enum values, such as a new status or stop reason. Enums are
   open: handle values you don't recognise.
+- 0.5.0 removes database history and fork metadata. See [What's new in 0.5.0](#whats-new-in-050).
+- 0.4.0 adds suspend and resume and keeps every 0.3.0 call: `stop` still ends the environment.
+  `Environment` gains `suspendedAt` and `sourceEnvironmentId`, so object literals typed as
+  `Environment` need them. A snapshot's `consistency` is now `'environment'`, `'app_instance'`
+  or `null` while it isn't ready (it was always `'per-app-instance'`). See
+  [What's new in 0.4.0](#whats-new-in-040).
+- Upgrade from 0.3.0 to capture snapshots: the API now answers `consistency: null` while a
+  capture runs, which 0.3.0 rejects as an invalid response even though the capture started.
 - 0.3.0 adds the snapshot APIs and keeps every 0.2.0 call. An input with neither `apps` nor
   `snapshotId` now fails with `Pass apps or snapshotId.` `Environment` and its apps gain
   `sourceEnvironmentSnapshotId` and `sourceAppInstanceSnapshotId` (`null` unless restored), so
@@ -472,6 +538,43 @@ TypeScript consumers need `lib: ["DOM"]` or `@types/node`, for the `fetch`, `Req
   releases don't declare a default actor yet.
 - `await using` isn't supported. Stop the environment in `finally`.
 
+## What's new in 0.5.0
+
+### Added
+
+- `sm.environments.listRequests`, `getRequest`, `getRequestPayload` and `getRequestStats`
+  read best-effort gateway request history, including after the environment stops.
+- A failed environment retains its application data and can be resumed or stopped.
+  Stop takes priority over startup, recovery and snapshot capture.
+
+### Removed
+
+- The generated `listEnvironmentHistory`, `getEnvironmentPosition` and
+  `getEnvironmentExchange` operations and their history/position/exchange models.
+  Use the request-history methods above for recorded replica traffic.
+- `EnvironmentApp.forkedAt`. Database object history and point-in-time forks are no longer
+  part of the API.
+
+`waitUntilStopped()` now requires `stopped` and completed disposal. A retained `failed`
+environment is not stopped. Lifecycle commands require Temporal availability; existing
+replica traffic does not.
+
+## What's new in 0.4.0
+
+- `sm.environments.suspend(id)` ends an environment's sandboxes and keeps its data;
+  `sm.environments.resume(id)` runs it again on that data, and
+  `sm.environments.waitUntilSuspended(id)` waits for the suspension (see
+  [Suspending and resuming](#suspending-and-resuming)).
+- `RunningEnvironment` gains `suspend()`, `resume()`, `waitUntilSuspended()` and
+  `waitUntilReady()`.
+- `stop` is unchanged: it ends the environment and discards its data, suspended or not.
+- New statuses: `suspending` and `suspended` for environments, `suspended` and `recovering` for
+  apps. New fields: `suspendedAt` and `sourceEnvironmentId` on an environment, `forkedAt` on an
+  app. `waitUntilReady()` throws `EnvironmentFailedError` on `suspending` and `suspended`.
+- Snapshots: `consistency` is `'environment'` (every app paused together), `'app_instance'` (each
+  app captured on its own) or `null` until ready. One capture runs per environment at a time; a
+  second gets `409 capture-in-progress`.
+
 ## What's new in 0.3.0
 
 - `sm.snapshots`: `capture`, `create`, `waitUntilReady`, `get`, `list`, `rename` and `delete`

package/agents/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: state-machines
-description: Use when an agent or test needs a disposable replica of a SaaS app such as Salesforce, to run integration tests or try API calls against a sandboxed copy instead of a real org. Covers the @usestatemachines/sdk TypeScript SDK - starting environments, choosing an actor, snapshots, cleanup, and errors such as environment-limit and release-unavailable. Also use when the user mentions State Machines, STATE_API_KEY or @usestatemachines/sdk.
+description: Use when an agent or test needs a disposable replica of a SaaS app such as Salesforce, to run integration tests or try API calls against a sandboxed copy instead of a real org. Covers the @usestatemachines/sdk TypeScript SDK - starting environments, choosing an actor, snapshots, suspending and resuming, cleanup, and errors such as environment-limit and release-unavailable. Also use when the user mentions State Machines, STATE_API_KEY or @usestatemachines/sdk.
 ---
 
 Read `node_modules/@usestatemachines/sdk/README.md` before relying on memory; it matches the installed version. If the package is missing, run `npm install @usestatemachines/sdk`.
@@ -12,7 +12,7 @@ Read `node_modules/@usestatemachines/sdk/README.md` before relying on memory; it
 3. Open a `try` right after `start()` returns, and call `env.stop()` in its `finally`.
 4. Connect through one API profile. Salesforce offers several and has no default, so pass `apiId`: `env.connect('salesforce', { apiId: 'rest' })`. This acts as the environment's declared default connection's actor, or else the app's first actor, the administrator. To act with another user's permissions, choose one from `await env.actors('salesforce')` and pass its `id` as `actorId`. Salesforce actor IDs are its user IDs, such as `005…`, never a role name, so read them instead of hard-coding one.
 5. Call the app with the connection's `fetch`. A leading `/` means the root of `baseUrl`.
-6. In `finally`, `await env.stop()`. In CI, then `await env.waitUntilStopped()`.
+6. In `finally`, `await env.stop()`; it ends the environment and discards its data. In CI, then `await env.waitUntilStopped()`.
 
 The whole flow, as the package ships it. It is ESM with top-level `await`: save it as `state-quickstart.mts` and run `npx tsx state-quickstart.mts`.
 
@@ -34,9 +34,10 @@ try {
 ## Rules
 
 - Always stop in `finally`. `waitUntilStopped` confirms every app was disposed; use it in CI so a run fails when cleanup can't be confirmed.
+- `env.suspend()` ends the sandboxes and keeps the data, with the lifetime paused; it is not cleanup. Await `env.waitUntilSuspended()` before `env.resume()`, then `env.waitUntilReady()` before connecting again. A suspended environment still needs `env.stop()`.
 - To keep a failed environment for inspection, use the full-control path: `sm.environments.create`, then `sm.environments.waitUntilReady(environment.id)`, then `sm.environments.connect(environment.id, key, options)`. Nothing on that path stops the environment for you.
 - Lifetime counts from the start, including boot. It defaults to 15 minutes, or the deployment maximum if that is lower: `(await sm.credential.get()).limits.environmentLifetime`. Set `lifetimeMinutes` in `start` or `create` for longer runs.
-- Live environments are capped per organization: `limits.liveEnvironments` from `sm.credential.get()`. Stop the ones you no longer need.
+- Live environments are capped per organization: `limits.liveEnvironments` from `sm.credential.get()`. Stop the ones you no longer need; suspended ones don't count until they resume.
 - Never print, log or commit `STATE_API_KEY`, tokens, or `headers()` or `actorToken()` output.
 - Salesforce's SOAP profile (`soap-partner`) takes the actor token in the envelope, not a header: put `connection.actorToken()` in `SessionHeader/sessionId`. For it, `headers()` carries only `X-State-Environment-Token`.
 - `connection.fetch` never sends your API key, never follows redirects and only reaches URLs under `baseUrl`.
@@ -56,18 +57,20 @@ The SDK already retries `502`, `503`, `504` and connection errors with one idemp
 | --- | --- | --- |
 | `release-unavailable`, 422 | Unknown or retired `appId` or `appReleaseId`. A string app is used as the `appId`. | Check the IDs against `sm.catalog.get()`. Omit `appReleaseId` for the default release. |
 | `release-unavailable`, 503, or a failed app with that `errorCode` | This deployment doesn't serve the release. The SDK retries a 503 within a 10-minute budget. | Omit `appReleaseId`. If it persists, report the `requestId`. Pass a `signal` to bound the wait. |
-| `environment-limit`, 429 | The organization is at its live-environment cap. Not retried. | `sm.environments.list({ status: 'live' })`, stop one you no longer need, then start again. |
+| `environment-limit`, 429 | The organization is at its live-environment cap. Not retried. | `sm.environments.list({ status: 'live' })`, stop or suspend one you no longer need, then start again. |
+| `not-suspended` or `suspend-in-progress`, 409 | `resume()` before the environment finished suspending, or `stop()` while it suspends. | Await `env.waitUntilSuspended()` first. |
 | `lifetime-too-long`, 422 | `lifetimeMinutes` is above the maximum. | Use at most `limits.environmentLifetime.maximumMinutes`. |
 | `TypeError: Unknown actor <id> for <appKey>. Choose one of: <ids>.` | The `actorId` isn't one of the app's actors. No token was revealed. | Pass an `id` from `await env.actors(appKey)`, or omit `actorId` to act as the default actor. |
 | `TypeError: Choose one API for <appKey>: <ids>.` | The app has several API profiles and `apiId` is missing or unknown. Salesforce always needs one. | Pass `apiId` from the listed IDs, such as `rest`. |
 | `start()` rejects with `EnvironmentFailedError` or `TimeoutError` | The environment failed, or wasn't `running` within `readyTimeoutMs` (10 minutes by default). If `err.stopped` is `true`, `start()` already stopped it. | Read `err.environment` (`null` on a `TimeoutError` that saw no read): its `stopReason` and each app's `errorCode`. If `err.stopped` is `false`, read `err.stopError` and call `sm.environments.stop(err.environmentId)`. To keep a failed environment, use the full-control path. |
 | `StateMachinesError` (`status` null) for invalid `apps`: `Pass at least one app.`, `App key "<key>" is used twice.` or `"<s>" can't be used as an app key; pass { key, appId } instead.` | Empty list, duplicate key, or a string that isn't a valid key. No request was sent. | Pass one to ten apps with unique keys; use `{ key, appId }` when the ID isn't a valid key. |
 | `TypeError: App ... is not ready to connect` | On the full-control path, connect() ran before the app was `ready`. | Await `waitUntilReady` first, or use `start()`. |
-| `connection-unavailable`, 409 | The app stopped, or the environment is no longer `running`. | Check `sm.environments.get(env.id)`; if it stopped or expired, start a new one. |
-| `EnvironmentFailedError` from `waitUntilReady` | On the full-control path, the environment became `failed`, `stopping` or `stopped`. | Read `err.environment.stopReason` and each app's `errorCode`. Still stop it in `finally`. |
+| `connection-unavailable`, 409 | The app stopped, or the environment is no longer `running`. | Check `sm.environments.get(env.id)`; if it is suspended, resume it; if it stopped or expired, start a new one. |
+| `EnvironmentFailedError` from `waitUntilReady` | On the full-control path, the environment became `failed`, `suspending`, `suspended`, `stopping` or `stopped`. | Read `err.environment.status`, its `stopReason` and each app's `errorCode`. Still stop it in `finally`. |
 | A vendor client such as jsforce is rejected by the gateway | The gateway needs `X-State-Environment-Token` on every request; jsforce only sends it with custom headers configured. | Prefer `connection.fetch`. If you pass `headers()` to another client, send them only to `baseUrl` and disable redirects. |
 | `not-ready`, 409 | Capture: the environment isn't `running` or an app isn't `ready`. Restore: the snapshot isn't `ready` yet. Delete: the snapshot is still `capturing`. | Capture from the handle `start()` returns. Before a restore or delete, `await sm.snapshots.waitUntilReady(snapshotId)`. |
-| `snapshot-limit`, 429 | The API is running as many captures as it can. `capture` and `env.snapshot()` already retried it for 60 seconds. | Wait, then capture again. |
+| `capture-in-progress`, 409 | This environment is already capturing; one capture runs at a time. | `await sm.snapshots.waitUntilReady(<the running snapshot's id>)`, then capture again. |
+| `snapshot-limit`, 429 | Older APIs only: too many captures were running. `capture` and `env.snapshot()` already retried it for 60 seconds. | Wait, then capture again. |
 | `SnapshotFailedError` | The capture failed. `err.snapshot.errorCode` is `capture-failed`, or `capture-interrupted` when the API restarted mid-capture. The snapshot is kept. | Capture again with a new idempotency key (the default); the old key returns the same failed snapshot. |
 | `SnapshotTimeoutError` | The snapshot wasn't `ready` within `readyTimeoutMs`. The capture continues on the server. | `await sm.snapshots.waitUntilReady(err.snapshotId)`. |
 | `forbidden`, 403, from `sm.snapshots.delete` | The key lacks `snapshots:delete`. Keys created by members never hold it. | Use a key whose creator holds `snapshots:delete`, or keep the snapshot. |

package/agents/index.js

@@ -1,22 +1,22 @@
 // Generated by scripts/build-agents.mjs from templates/ and examples/quickstart.ts.
 
 export const prompt =
-  "Set up State Machines so this project can run code against a disposable replica of a SaaS app, such as Salesforce. Follow these steps in order.\n\n1. Install the SDK: `npm install @usestatemachines/sdk`. It is ESM only and runs on Node.js 20.20 or later and on Cloudflare Workers, not in browsers.\n2. Check that `STATE_API_KEY` is set without printing it: `test -n \"$STATE_API_KEY\"`. If it is missing, stop and ask me to set it; the SDK doesn't read `.env` files. In Cloudflare Workers, pass the key as `new StateMachines({ apiKey })`.\n3. Save this quickstart unchanged as `state-quickstart.mts` and run `npx tsx state-quickstart.mts` to confirm the setup. It prints `200` and the org's limits:\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n4. Adapt it to my task. Log `await sm.catalog.get()` to find each app's ID and the actors and API profiles its release offers; don't guess IDs. Start only the apps the task needs; a string such as `'salesforce'` means `{ key: 'salesforce', appId: 'salesforce' }`. Salesforce always needs an `apiId`. `connect()` acts as the declared default connection's actor, or else the app's first actor, the administrator; to act as a specific user, pass an `actorId` from `await env.actors('salesforce')`. To reuse seeded data, `await env.snapshot()` once, then `sm.environments.start({ snapshotId })`.\n5. Always call `env.stop()` in a `finally`. In CI, then await `env.waitUntilStopped()` so the run fails if cleanup isn't confirmed. If `start()` fails after creating the environment, it stops it; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`. If the create itself fails, call `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same `input`; never retry without the key, which can start a second environment.\n6. Never print, log or commit keys or tokens, including `connection.headers()` and `connection.actorToken()`, which return live tokens.\n\nRead more: `node_modules/@usestatemachines/sdk/README.md`, https://state-web.usestatemachines.workers.dev/llms.txt and https://www.npmjs.com/package/@usestatemachines/sdk\n";
+  "Set up State Machines so this project can run code against a disposable replica of a SaaS app, such as Salesforce. Follow these steps in order.\n\n1. Install the SDK: `npm install @usestatemachines/sdk`. It is ESM only and runs on Node.js 20.20 or later and on Cloudflare Workers, not in browsers.\n2. Check that `STATE_API_KEY` is set without printing it: `test -n \"$STATE_API_KEY\"`. If it is missing, stop and ask me to set it; the SDK doesn't read `.env` files. In Cloudflare Workers, pass the key as `new StateMachines({ apiKey })`.\n3. Save this quickstart unchanged as `state-quickstart.mts` and run `npx tsx state-quickstart.mts` to confirm the setup. It prints `200` and the org's limits:\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n4. Adapt it to my task. Log `await sm.catalog.get()` to find each app's ID and the actors and API profiles its release offers; don't guess IDs. Start only the apps the task needs; a string such as `'salesforce'` means `{ key: 'salesforce', appId: 'salesforce' }`. Salesforce always needs an `apiId`. `connect()` acts as the declared default connection's actor, or else the app's first actor, the administrator; to act as a specific user, pass an `actorId` from `await env.actors('salesforce')`. To reuse seeded data, `await env.snapshot()` once, then `sm.environments.start({ snapshotId })`.\n5. Always call `env.stop()` in a `finally`; it ends the environment and discards its data. In CI, then await `env.waitUntilStopped()` so the run fails if cleanup isn't confirmed. If `start()` fails after creating the environment, it stops it; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`. If the create itself fails, call `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same `input`; never retry without the key, which can start a second environment.\n6. Never print, log or commit keys or tokens, including `connection.headers()` and `connection.actorToken()`, which return live tokens.\n\nRead more: `node_modules/@usestatemachines/sdk/README.md`, https://state-web.usestatemachines.workers.dev/llms.txt and https://www.npmjs.com/package/@usestatemachines/sdk\n";
 
 export const llmsTxt =
-  "# State Machines\n\n> State Machines runs disposable environments of replica SaaS apps, such as Salesforce, for tests and coding agents. Start, connect to and stop them with the TypeScript SDK `@usestatemachines/sdk`.\n\nInstall with `npm install @usestatemachines/sdk`. It is ESM only and runs on Node.js 20.20 or later (22 or 24 recommended) and Cloudflare Workers, not browsers. TypeScript needs `lib: [\"DOM\"]` or `@types/node`. `new StateMachines()` reads the API key from `STATE_API_KEY` and doesn't read `.env` files; in Cloudflare Workers, pass `apiKey` explicitly. Never print the key. The installed `node_modules/@usestatemachines/sdk/README.md` matches your version: read it before relying on memory.\n\n## Concepts\n\n- Environment: one disposable run of one or more apps. `const env = await sm.environments.start({ apps: ['salesforce'] })` creates it, waits until it is `running` and returns a `RunningEnvironment`. If the environment fails or the wait times out or is aborted, `start()` stops the environment and rethrows the original error; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`. Always call `env.stop()` in a `finally`.\n- Apps: a string is shorthand for `{ key: s, appId: s }`, nothing more. Use the object form, such as `{ key: 'crm', appId: 'salesforce' }`, for your own key or two instances of one app. `appId` names the app in the catalog (`sm.catalog.get()`); `key` is your name for it in this environment, and later calls take the key.\n- Actor: one of the app's users; requests run with that user's permissions. Salesforce actor IDs are its user IDs, such as `005…`, never a role name. `await env.actors('salesforce')` lists them in the release's order. connect() acts as the environment's declared default connection's actor, or else the first actor, the administrator, unless you pass `actorId`; pick a less-privileged actor to test permissions.\n- API profile: one API of the app, with its path and versions. Salesforce offers several (`rest`, `soap-partner`, `bulk2`, `oauth`) and has no default, so always pass `apiId`: `env.connect('salesforce', { apiId: 'rest' })`.\n- Connection: `connection.fetch` never sends your API key, never follows redirects and never retries. `connection.headers()` returns live tokens: treat them as secrets. Salesforce's SOAP profile (`soap-partner`) takes the actor token in the envelope instead: put `connection.actorToken()`, also a secret, in `SessionHeader/sessionId`; `headers()` then carries only `X-State-Environment-Token`.\n- Full control: `sm.environments.create(input)` returns one `Environment` before the apps boot, `sm.environments.waitUntilReady(environment.id)` waits, and `sm.environments.connect(environment.id, key, options)` connects. Nothing on this path stops the environment for you; use it to keep a failed environment for inspection.\n- Snapshot: the saved data of a running environment's apps, taken per app instance. `const snapshot = await env.snapshot({ name: 'seeded' })` captures it and waits until it is `ready`. `sm.environments.start({ snapshotId: snapshot.id })` starts a new environment from it, in place of `apps`; each app keeps its key. Capturing needs `snapshots:create` and `snapshots:read`, restoring needs `snapshots:read`, `rename` needs `snapshots:create`, and `sm.snapshots.delete(id)` needs `snapshots:delete`, which keys created by members don't hold. Snapshots are never deleted automatically, not even failed or timed-out ones: `SnapshotFailedError` and `SnapshotTimeoutError` carry `err.snapshotId`. `sm.snapshots` also has `create`, `waitUntilReady`, `get`, `list` and `rename`.\n- Vendor clients: the gateway needs `X-State-Environment-Token` on every request. A client such as jsforce only sends it if you configure custom headers. Prefer `connection.fetch`; if you pass `headers()` to another client, send them only to `connection.baseUrl` and disable redirects.\n- Pagination: `environments.list`, `environments.events`, `workspaces.events` and `snapshots.list` return one page, `{ data, nextCursor }`. Pass `nextCursor` back as `cursor` until it is `null`.\n\n## Statuses\n\nEnums are open: handle values you don't recognise.\n\n- Environment `status`: `queued`, `starting`, `running`, `stopping`, `stopped`, `failed`. `stopReason`: `requested`, `expired`, `app_failed`, `owner_lost`.\n- App `status`: `pending`, `starting`, `ready`, `stopping`, `stopped`, `failed`. A failed app carries `errorCode`.\n- App `disposal`: `pending`, `complete`, `failed`, `unknown`.\n- Snapshot `status`: `capturing`, `ready`, `failed`. A failed snapshot carries `errorCode`: `capture-failed`, or `capture-interrupted` when the API restarted mid-capture.\n- `start()` and `waitUntilReady` throw `EnvironmentFailedError` if the environment becomes `failed`, `stopping` or `stopped`, and `TimeoutError` after 10 minutes by default (`readyTimeoutMs` for `start()`, `timeoutMs` for `waitUntilReady`). `start()` then stops the environment; `waitUntilReady` never does.\n- `waitUntilStopped` resolves when the environment is `stopped` or `failed` and every app's disposal is `complete`. It throws `DisposalIncompleteError` if a disposal is `failed` or `unknown`, and `TimeoutError` if disposal stays `pending`.\n\n## Limits\n\n- Live environments are capped per organization: `(await sm.credential.get()).limits.liveEnvironments`.\n- `lifetimeMinutes` defaults to 15 minutes, or the deployment maximum if that is lower. Both are in `(await sm.credential.get()).limits.environmentLifetime`. Lifetime counts from the start, so boot uses part of it; at the end the environment stops with `stopReason: 'expired'`.\n- Always call `env.stop()` in a `finally`. In CI, then await `env.waitUntilStopped()`.\n\n## Errors and retries\n\nA failed API call throws `StateMachinesError` with `status`, `code`, `requestId` and `retryAfter`. The SDK retries `502`, `503`, `504`, connection errors and attempt timeouts itself, within a 10-minute retry budget (the last attempt can end later); `create` retries reuse one idempotency key. It retries nothing else: not `429`, other `4xx` or other `5xx`. The one exception is snapshot capture, which retries `429` `snapshot-limit` for up to 60 seconds with one idempotency key. `err.stopped === false` with `err.stopError` means the stop wasn't confirmed; `sm.environments.stop(err.environmentId)` is safe to repeat. Without `stopError`, you passed your own `idempotencyKey`, so `start()` left the environment running because another `start()` with that key may be using it; stop it only once nothing else needs it. Every error from `create`, including the create step of `start()`, carries `err.idempotencyKey`: `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same input returns the environment the first call created. Invalid `apps` fail before any request as a `StateMachinesError` (`status` null, `cause` a `TypeError`; don't replay it) with one of: `Pass at least one app.`, `App key \"<key>\" is used twice.` or `\"<s>\" can't be used as an app key; pass { key, appId } instead.`\n\n| Code | Status | What to do |\n| --- | --- | --- |\n| `environment-limit` | 429 | Not retried. Stop a live environment (`sm.environments.list({ status: 'live' })`), then start again. |\n| `lifetime-too-long` | 422 | Lower `lifetimeMinutes` to at most `limits.environmentLifetime.maximumMinutes`. |\n| `release-unavailable` | 422 | Unknown or retired app or release. Check `appId` and `appReleaseId` against `sm.catalog.get()`. |\n| `release-unavailable` | 503 | This deployment doesn't serve the release. Retried within the budget; omit `appReleaseId` to use the default release, or contact support with the `requestId`. A failed app can carry it as `errorCode`. |\n| `connection-unavailable` | 409 | The app stopped or the environment is no longer `running`. If it stopped or expired, start a new one. (Before revealing any token, connect() throws a `TypeError` for an app that isn't ready, an unknown actor, or a missing `apiId`.) |\n| `idempotency-conflict` | 409 | The idempotency key was used with a different input. Replay with the exact same input. |\n| `snapshot-limit` | 429 | Too many captures are running. `capture` and `env.snapshot()` retry it for up to 60 seconds; if it still fails, wait and capture again. |\n| `not-ready` | 409 | Capture: the environment isn't `running` or an app isn't `ready`. Restore: the snapshot isn't `ready` yet; `await sm.snapshots.waitUntilReady(id)`. Delete: the snapshot is still `capturing`. |\n| `version-conflict` | 409 | `rename` sent a stale `expectedVersion`, or the snapshot was deleted. `get` it: if `deletedAt` is set it's gone, otherwise retry with its `version`. |\n| `release-unavailable` | 422 | On a restore: no release on this deployment matches the snapshot's apps. Start from `apps`, seed again and capture a new snapshot. |\n| `request-timeout` | 503 | Retried automatically. If `create` still fails, replay it with `err.idempotencyKey`. |\n\n## Quickstart\n\nSave it unchanged as `state-quickstart.mts` and run `npx tsx state-quickstart.mts`.\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n## Docs\n\n- [npm package](https://www.npmjs.com/package/@usestatemachines/sdk): install, README and versions\n- [Agent skill](https://state-web.usestatemachines.workers.dev/skill.md): workflow, rules and troubleshooting for coding agents\n- [OpenAPI document](https://cdn.jsdelivr.net/npm/@usestatemachines/sdk/openapi.json): the HTTP API the SDK calls\n- [State Machines](https://usestatemachines.com): product site\n";
+  "# State Machines\n\n> State Machines runs disposable environments of replica SaaS apps, such as Salesforce, for tests and coding agents. Start, connect to, suspend, resume and stop them with the TypeScript SDK `@usestatemachines/sdk`.\n\nInstall with `npm install @usestatemachines/sdk`. It is ESM only and runs on Node.js 20.20 or later (22 or 24 recommended) and Cloudflare Workers, not browsers. TypeScript needs `lib: [\"DOM\"]` or `@types/node`. `new StateMachines()` reads the API key from `STATE_API_KEY` and doesn't read `.env` files; in Cloudflare Workers, pass `apiKey` explicitly. Never print the key. The installed `node_modules/@usestatemachines/sdk/README.md` matches your version: read it before relying on memory.\n\n## Concepts\n\n- Environment: one disposable run of one or more apps. `const env = await sm.environments.start({ apps: ['salesforce'] })` creates it, waits until it is `running` and returns a `RunningEnvironment`. If the environment fails or the wait times out or is aborted, `start()` stops the environment and rethrows the original error; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`. Always call `env.stop()` in a `finally`.\n- Suspend and resume: `env.suspend()` ends the sandboxes and keeps the data; the lifetime pauses. Await `env.waitUntilSuspended()`, later `env.resume()` and `env.waitUntilReady()` to run again on the same data. `env.stop()` ends the environment and discards its data; a suspended environment still needs it.\n- Apps: a string is shorthand for `{ key: s, appId: s }`, nothing more. Use the object form, such as `{ key: 'crm', appId: 'salesforce' }`, for your own key or two instances of one app. `appId` names the app in the catalog (`sm.catalog.get()`); `key` is your name for it in this environment, and later calls take the key.\n- Actor: one of the app's users; requests run with that user's permissions. Salesforce actor IDs are its user IDs, such as `005…`, never a role name. `await env.actors('salesforce')` lists them in the release's order. connect() acts as the environment's declared default connection's actor, or else the first actor, the administrator, unless you pass `actorId`; pick a less-privileged actor to test permissions.\n- API profile: one API of the app, with its path and versions. Salesforce offers several (`rest`, `soap-partner`, `bulk2`, `oauth`) and has no default, so always pass `apiId`: `env.connect('salesforce', { apiId: 'rest' })`.\n- Connection: `connection.fetch` never sends your API key, never follows redirects and never retries. `connection.headers()` returns live tokens: treat them as secrets. Salesforce's SOAP profile (`soap-partner`) takes the actor token in the envelope instead: put `connection.actorToken()`, also a secret, in `SessionHeader/sessionId`; `headers()` then carries only `X-State-Environment-Token`.\n- Full control: `sm.environments.create(input)` returns one `Environment` before the apps boot, `sm.environments.waitUntilReady(environment.id)` waits, and `sm.environments.connect(environment.id, key, options)` connects. Nothing on this path stops the environment for you; use it to keep a failed environment for inspection.\n- Snapshot: the saved data of a running environment's apps. Every app is paused while it is cut, so a multi-app snapshot is one instant (`consistency: 'environment'`); a single app is `'app_instance'`. `const snapshot = await env.snapshot({ name: 'seeded' })` captures it and waits until it is `ready`. `sm.environments.start({ snapshotId: snapshot.id })` starts a new environment from it, in place of `apps`; each app keeps its key. Capturing needs `snapshots:create` and `snapshots:read`, restoring needs `snapshots:read`, `rename` needs `snapshots:create`, and `sm.snapshots.delete(id)` needs `snapshots:delete`, which keys created by members don't hold. Snapshots are never deleted automatically, not even failed or timed-out ones: `SnapshotFailedError` and `SnapshotTimeoutError` carry `err.snapshotId`. `sm.snapshots` also has `create`, `waitUntilReady`, `get`, `list` and `rename`.\n- Recorded requests: `sm.environments.listRequests(env.id)` pages the calls made to the environment's apps (method, path, status, duration); `getRequest(env.id, id)` adds headers and bodies. Recording is best effort and SOAP bodies are not stored.\n- Vendor clients: the gateway needs `X-State-Environment-Token` on every request. A client such as jsforce only sends it if you configure custom headers. Prefer `connection.fetch`; if you pass `headers()` to another client, send them only to `connection.baseUrl` and disable redirects.\n- Pagination: `environments.list`, `environments.events`, `workspaces.events` and `snapshots.list` return one page, `{ data, nextCursor }`. Pass `nextCursor` back as `cursor` until it is `null`.\n\n## Statuses\n\nEnums are open: handle values you don't recognise.\n\n- Environment `status`: `queued`, `starting`, `running`, `suspending`, `suspended`, `stopping`, `stopped`, `failed`. `stopReason`: `requested`, `expired`, `app_failed`, `owner_lost`. A suspended environment keeps its data and carries `suspendedAt`.\n- App `status`: `pending`, `starting`, `ready`, `recovering`, `suspended`, `stopping`, `stopped`, `failed`. A failed app carries `errorCode`.\n- App `disposal`: `pending`, `complete`, `failed`, `unknown`.\n- Snapshot `status`: `capturing`, `ready`, `failed`. A failed snapshot carries `errorCode`: `capture-failed`, `pause-failed` or `pause-lost` when the apps could not be held paused for the capture, or `capture-interrupted` when the API restarted mid-capture.\n- `start()` and `waitUntilReady` throw `EnvironmentFailedError` if the environment becomes `failed`, `suspending`, `suspended`, `stopping` or `stopped`, and `TimeoutError` after 10 minutes by default (`readyTimeoutMs` for `start()`, `timeoutMs` for `waitUntilReady`). `start()` then stops the environment; `waitUntilReady` never does.\n- `waitUntilSuspended` resolves when the environment is `suspended`, and throws `EnvironmentFailedError` if it ends instead.\n- `waitUntilStopped` resolves when the environment is `stopped` and every app's disposal is `complete`. A `failed` environment keeps its data for `resume()` and is not stopped until you call `stop()`. It throws `DisposalIncompleteError` if a disposal is `failed` or `unknown`, and `TimeoutError` if disposal stays `pending`.\n\n## Limits\n\n- Live environments are capped per organization: `(await sm.credential.get()).limits.liveEnvironments`.\n- `lifetimeMinutes` defaults to 15 minutes, or the deployment maximum if that is lower. Both are in `(await sm.credential.get()).limits.environmentLifetime`. Lifetime counts from the start, so boot uses part of it, and pauses while suspended; at the end the environment stops with `stopReason: 'expired'`.\n- A suspended environment doesn't count toward the live cap; resuming it does.\n- Always call `env.stop()` in a `finally`. In CI, then await `env.waitUntilStopped()`.\n\n## Errors and retries\n\nA failed API call throws `StateMachinesError` with `status`, `code`, `requestId` and `retryAfter`. The SDK retries `502`, `503`, `504`, connection errors and attempt timeouts itself, within a 10-minute retry budget (the last attempt can end later); `create` retries reuse one idempotency key. It retries nothing else: not `429`, other `4xx` or other `5xx`. The one exception is snapshot capture, which retries `429` `snapshot-limit` for up to 60 seconds with one idempotency key. `err.stopped === false` with `err.stopError` means the stop wasn't confirmed; `sm.environments.stop(err.environmentId)` is safe to repeat. Without `stopError`, you passed your own `idempotencyKey`, so `start()` left the environment running because another `start()` with that key may be using it; stop it only once nothing else needs it. Every error from `create`, including the create step of `start()`, carries `err.idempotencyKey`: `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same input returns the environment the first call created. Invalid `apps` fail before any request as a `StateMachinesError` (`status` null, `cause` a `TypeError`; don't replay it) with one of: `Pass at least one app.`, `App key \"<key>\" is used twice.` or `\"<s>\" can't be used as an app key; pass { key, appId } instead.`\n\n| Code | Status | What to do |\n| --- | --- | --- |\n| `environment-limit` | 429 | Not retried. Suspend or stop a live environment (`sm.environments.list({ status: 'live' })`), then start or resume again. |\n| `not-ready` | 409 | Only a running environment whose apps are all ready can be suspended. Wait for it, or stop it. |\n| `not-suspended` | 409 | Only a suspended environment can resume. After `suspend()`, await `waitUntilSuspended()` first. |\n| `suspend-in-progress` | 409 | The environment is suspending. Await `waitUntilSuspended()`, then stop it. |\n| `lifetime-too-long` | 422 | Lower `lifetimeMinutes` to at most `limits.environmentLifetime.maximumMinutes`. |\n| `release-unavailable` | 422 | Unknown or retired app or release. Check `appId` and `appReleaseId` against `sm.catalog.get()`. |\n| `release-unavailable` | 503 | This deployment doesn't serve the release. Retried within the budget; omit `appReleaseId` to use the default release, or contact support with the `requestId`. A failed app can carry it as `errorCode`. |\n| `connection-unavailable` | 409 | The app stopped or the environment is no longer `running`. If it is suspended, resume it; if it stopped or expired, start a new one. (Before revealing any token, connect() throws a `TypeError` for an app that isn't ready, an unknown actor, or a missing `apiId`.) |\n| `idempotency-conflict` | 409 | The idempotency key was used with a different input. Replay with the exact same input. |\n| `capture-in-progress` | 409 | This environment is already capturing a snapshot; one runs at a time. Wait for it, then capture again. |\n| `snapshot-limit` | 429 | Older APIs only. `capture` and `env.snapshot()` retry it for up to 60 seconds; if it still fails, wait and capture again. |\n| `not-ready` | 409 | Capture: the environment isn't `running` or an app isn't `ready`. Restore: the snapshot isn't `ready` yet; `await sm.snapshots.waitUntilReady(id)`. Delete: the snapshot is still `capturing`. |\n| `version-conflict` | 409 | `rename` sent a stale `expectedVersion`, or the snapshot was deleted. `get` it: if `deletedAt` is set it's gone, otherwise retry with its `version`. |\n| `release-unavailable` | 422 | On a restore: no release on this deployment matches the snapshot's apps. Start from `apps`, seed again and capture a new snapshot. |\n| `request-timeout` | 503 | Retried automatically. If `create` still fails, replay it with `err.idempotencyKey`. |\n\n## Quickstart\n\nSave it unchanged as `state-quickstart.mts` and run `npx tsx state-quickstart.mts`.\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n## Docs\n\n- [npm package](https://www.npmjs.com/package/@usestatemachines/sdk): install, README and versions\n- [Agent skill](https://state-web.usestatemachines.workers.dev/skill.md): workflow, rules and troubleshooting for coding agents\n- [OpenAPI document](https://cdn.jsdelivr.net/npm/@usestatemachines/sdk/openapi.json): the HTTP API the SDK calls\n- [State Machines](https://usestatemachines.com): product site\n";
 
 export const skill =
-  "---\nname: state-machines\ndescription: Use when an agent or test needs a disposable replica of a SaaS app such as Salesforce, to run integration tests or try API calls against a sandboxed copy instead of a real org. Covers the @usestatemachines/sdk TypeScript SDK - starting environments, choosing an actor, snapshots, cleanup, and errors such as environment-limit and release-unavailable. Also use when the user mentions State Machines, STATE_API_KEY or @usestatemachines/sdk.\n---\n\nRead `node_modules/@usestatemachines/sdk/README.md` before relying on memory; it matches the installed version. If the package is missing, run `npm install @usestatemachines/sdk`.\n\n## Workflow\n\n1. Check that `STATE_API_KEY` is set without printing it (`test -n \"$STATE_API_KEY\"`). If it isn't, ask the user to set it; the SDK doesn't read `.env` files.\n2. Start the environment with only the apps the task needs: `const env = await sm.environments.start({ apps: ['salesforce'] })`. Log `await sm.catalog.get()` for each app's ID and the actors and API profiles its release offers; don't guess IDs. A string app means `{ key: s, appId: s }`; use `{ key, appId }` for your own key. `start()` waits until the environment is `running`. If that fails, it stops the environment and rethrows; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`.\n3. Open a `try` right after `start()` returns, and call `env.stop()` in its `finally`.\n4. Connect through one API profile. Salesforce offers several and has no default, so pass `apiId`: `env.connect('salesforce', { apiId: 'rest' })`. This acts as the environment's declared default connection's actor, or else the app's first actor, the administrator. To act with another user's permissions, choose one from `await env.actors('salesforce')` and pass its `id` as `actorId`. Salesforce actor IDs are its user IDs, such as `005…`, never a role name, so read them instead of hard-coding one.\n5. Call the app with the connection's `fetch`. A leading `/` means the root of `baseUrl`.\n6. In `finally`, `await env.stop()`. In CI, then `await env.waitUntilStopped()`.\n\nThe whole flow, as the package ships it. It is ESM with top-level `await`: save it as `state-quickstart.mts` and run `npx tsx state-quickstart.mts`.\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n## Rules\n\n- Always stop in `finally`. `waitUntilStopped` confirms every app was disposed; use it in CI so a run fails when cleanup can't be confirmed.\n- To keep a failed environment for inspection, use the full-control path: `sm.environments.create`, then `sm.environments.waitUntilReady(environment.id)`, then `sm.environments.connect(environment.id, key, options)`. Nothing on that path stops the environment for you.\n- Lifetime counts from the start, including boot. It defaults to 15 minutes, or the deployment maximum if that is lower: `(await sm.credential.get()).limits.environmentLifetime`. Set `lifetimeMinutes` in `start` or `create` for longer runs.\n- Live environments are capped per organization: `limits.liveEnvironments` from `sm.credential.get()`. Stop the ones you no longer need.\n- Never print, log or commit `STATE_API_KEY`, tokens, or `headers()` or `actorToken()` output.\n- Salesforce's SOAP profile (`soap-partner`) takes the actor token in the envelope, not a header: put `connection.actorToken()` in `SessionHeader/sessionId`. For it, `headers()` carries only `X-State-Environment-Token`.\n- `connection.fetch` never sends your API key, never follows redirects and only reaches URLs under `baseUrl`.\n- Runtimes: Node.js 20.20 or later (22 or 24 recommended), and Cloudflare Workers. ESM only. TypeScript needs `lib: [\"DOM\"]` or `@types/node`. Never ship an API key to a browser.\n- In Cloudflare Workers, pass the key explicitly: `new StateMachines({ apiKey: env.STATE_API_KEY })`, and bound calls with `{ signal: AbortSignal.timeout(20_000) }`, since retries can take minutes.\n- `environments.list`, `environments.events`, `workspaces.events` and `snapshots.list` return one page, `{ data, nextCursor }`. Pass `nextCursor` back as `cursor` until it is `null`.\n- To reuse seeded data, `await env.snapshot({ name })` once, then `sm.environments.start({ snapshotId })` in place of `apps`. Capturing needs `snapshots:create` and `snapshots:read`. Snapshots are never deleted automatically, not even failed ones.\n- Check errors with `instanceof` (`StateMachinesError`, `EnvironmentFailedError`, `DisposalIncompleteError`, `TimeoutError`, `SnapshotFailedError`, `SnapshotTimeoutError`), not by `name`.\n\n## Recovering a failed create\n\nThe SDK already retries `502`, `503`, `504` and connection errors with one idempotency key. If `start()` or `create` still throws with `err.idempotencyKey`, the create failed: don't call it again as a new request, which can start a second environment. Call `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same `input`. Don't replay when `err.status` is `null` and `err.cause` is a `TypeError`: the input was invalid and nothing was sent; fix the input instead. A key only recovers a failed create: don't reuse the key of a `start()` whose wait failed, because the replay returns that stopped environment. With your own key, `start()` doesn't stop the environment on failure, because another `start()` with that key may be using it: the error carries `err.environmentId`, `err.stopped` is `false` and there is no `err.stopError`. Stop it only once nothing else needs it. It returns the environment the first call created, or creates it if that call never arrived. A different input fails with `idempotency-conflict`. Keys are kept for a day.\n\n## Troubleshooting\n\n| Symptom | Cause | Fix |\n| --- | --- | --- |\n| `release-unavailable`, 422 | Unknown or retired `appId` or `appReleaseId`. A string app is used as the `appId`. | Check the IDs against `sm.catalog.get()`. Omit `appReleaseId` for the default release. |\n| `release-unavailable`, 503, or a failed app with that `errorCode` | This deployment doesn't serve the release. The SDK retries a 503 within a 10-minute budget. | Omit `appReleaseId`. If it persists, report the `requestId`. Pass a `signal` to bound the wait. |\n| `environment-limit`, 429 | The organization is at its live-environment cap. Not retried. | `sm.environments.list({ status: 'live' })`, stop one you no longer need, then start again. |\n| `lifetime-too-long`, 422 | `lifetimeMinutes` is above the maximum. | Use at most `limits.environmentLifetime.maximumMinutes`. |\n| `TypeError: Unknown actor <id> for <appKey>. Choose one of: <ids>.` | The `actorId` isn't one of the app's actors. No token was revealed. | Pass an `id` from `await env.actors(appKey)`, or omit `actorId` to act as the default actor. |\n| `TypeError: Choose one API for <appKey>: <ids>.` | The app has several API profiles and `apiId` is missing or unknown. Salesforce always needs one. | Pass `apiId` from the listed IDs, such as `rest`. |\n| `start()` rejects with `EnvironmentFailedError` or `TimeoutError` | The environment failed, or wasn't `running` within `readyTimeoutMs` (10 minutes by default). If `err.stopped` is `true`, `start()` already stopped it. | Read `err.environment` (`null` on a `TimeoutError` that saw no read): its `stopReason` and each app's `errorCode`. If `err.stopped` is `false`, read `err.stopError` and call `sm.environments.stop(err.environmentId)`. To keep a failed environment, use the full-control path. |\n| `StateMachinesError` (`status` null) for invalid `apps`: `Pass at least one app.`, `App key \"<key>\" is used twice.` or `\"<s>\" can't be used as an app key; pass { key, appId } instead.` | Empty list, duplicate key, or a string that isn't a valid key. No request was sent. | Pass one to ten apps with unique keys; use `{ key, appId }` when the ID isn't a valid key. |\n| `TypeError: App ... is not ready to connect` | On the full-control path, connect() ran before the app was `ready`. | Await `waitUntilReady` first, or use `start()`. |\n| `connection-unavailable`, 409 | The app stopped, or the environment is no longer `running`. | Check `sm.environments.get(env.id)`; if it stopped or expired, start a new one. |\n| `EnvironmentFailedError` from `waitUntilReady` | On the full-control path, the environment became `failed`, `stopping` or `stopped`. | Read `err.environment.stopReason` and each app's `errorCode`. Still stop it in `finally`. |\n| A vendor client such as jsforce is rejected by the gateway | The gateway needs `X-State-Environment-Token` on every request; jsforce only sends it with custom headers configured. | Prefer `connection.fetch`. If you pass `headers()` to another client, send them only to `baseUrl` and disable redirects. |\n| `not-ready`, 409 | Capture: the environment isn't `running` or an app isn't `ready`. Restore: the snapshot isn't `ready` yet. Delete: the snapshot is still `capturing`. | Capture from the handle `start()` returns. Before a restore or delete, `await sm.snapshots.waitUntilReady(snapshotId)`. |\n| `snapshot-limit`, 429 | The API is running as many captures as it can. `capture` and `env.snapshot()` already retried it for 60 seconds. | Wait, then capture again. |\n| `SnapshotFailedError` | The capture failed. `err.snapshot.errorCode` is `capture-failed`, or `capture-interrupted` when the API restarted mid-capture. The snapshot is kept. | Capture again with a new idempotency key (the default); the old key returns the same failed snapshot. |\n| `SnapshotTimeoutError` | The snapshot wasn't `ready` within `readyTimeoutMs`. The capture continues on the server. | `await sm.snapshots.waitUntilReady(err.snapshotId)`. |\n| `forbidden`, 403, from `sm.snapshots.delete` | The key lacks `snapshots:delete`. Keys created by members never hold it. | Use a key whose creator holds `snapshots:delete`, or keep the snapshot. |\n| `release-unavailable`, 422, on a restore | No release on this deployment matches the snapshot's apps. | Start from `apps`, seed again and capture a new snapshot. |\n| `TypeError: Pass apiKey or set STATE_API_KEY.` | No key in the environment. | Ask the user to set `STATE_API_KEY`; in Workers, pass `apiKey`. |\n\nMore: https://state-web.usestatemachines.workers.dev/llms.txt and https://www.npmjs.com/package/@usestatemachines/sdk\n";
+  "---\nname: state-machines\ndescription: Use when an agent or test needs a disposable replica of a SaaS app such as Salesforce, to run integration tests or try API calls against a sandboxed copy instead of a real org. Covers the @usestatemachines/sdk TypeScript SDK - starting environments, choosing an actor, snapshots, suspending and resuming, cleanup, and errors such as environment-limit and release-unavailable. Also use when the user mentions State Machines, STATE_API_KEY or @usestatemachines/sdk.\n---\n\nRead `node_modules/@usestatemachines/sdk/README.md` before relying on memory; it matches the installed version. If the package is missing, run `npm install @usestatemachines/sdk`.\n\n## Workflow\n\n1. Check that `STATE_API_KEY` is set without printing it (`test -n \"$STATE_API_KEY\"`). If it isn't, ask the user to set it; the SDK doesn't read `.env` files.\n2. Start the environment with only the apps the task needs: `const env = await sm.environments.start({ apps: ['salesforce'] })`. Log `await sm.catalog.get()` for each app's ID and the actors and API profiles its release offers; don't guess IDs. A string app means `{ key: s, appId: s }`; use `{ key, appId }` for your own key. `start()` waits until the environment is `running`. If that fails, it stops the environment and rethrows; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`.\n3. Open a `try` right after `start()` returns, and call `env.stop()` in its `finally`.\n4. Connect through one API profile. Salesforce offers several and has no default, so pass `apiId`: `env.connect('salesforce', { apiId: 'rest' })`. This acts as the environment's declared default connection's actor, or else the app's first actor, the administrator. To act with another user's permissions, choose one from `await env.actors('salesforce')` and pass its `id` as `actorId`. Salesforce actor IDs are its user IDs, such as `005…`, never a role name, so read them instead of hard-coding one.\n5. Call the app with the connection's `fetch`. A leading `/` means the root of `baseUrl`.\n6. In `finally`, `await env.stop()`; it ends the environment and discards its data. In CI, then `await env.waitUntilStopped()`.\n\nThe whole flow, as the package ships it. It is ESM with top-level `await`: save it as `state-quickstart.mts` and run `npx tsx state-quickstart.mts`.\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n## Rules\n\n- Always stop in `finally`. `waitUntilStopped` confirms every app was disposed; use it in CI so a run fails when cleanup can't be confirmed.\n- `env.suspend()` ends the sandboxes and keeps the data, with the lifetime paused; it is not cleanup. Await `env.waitUntilSuspended()` before `env.resume()`, then `env.waitUntilReady()` before connecting again. A suspended environment still needs `env.stop()`.\n- To keep a failed environment for inspection, use the full-control path: `sm.environments.create`, then `sm.environments.waitUntilReady(environment.id)`, then `sm.environments.connect(environment.id, key, options)`. Nothing on that path stops the environment for you.\n- Lifetime counts from the start, including boot. It defaults to 15 minutes, or the deployment maximum if that is lower: `(await sm.credential.get()).limits.environmentLifetime`. Set `lifetimeMinutes` in `start` or `create` for longer runs.\n- Live environments are capped per organization: `limits.liveEnvironments` from `sm.credential.get()`. Stop the ones you no longer need; suspended ones don't count until they resume.\n- Never print, log or commit `STATE_API_KEY`, tokens, or `headers()` or `actorToken()` output.\n- Salesforce's SOAP profile (`soap-partner`) takes the actor token in the envelope, not a header: put `connection.actorToken()` in `SessionHeader/sessionId`. For it, `headers()` carries only `X-State-Environment-Token`.\n- `connection.fetch` never sends your API key, never follows redirects and only reaches URLs under `baseUrl`.\n- Runtimes: Node.js 20.20 or later (22 or 24 recommended), and Cloudflare Workers. ESM only. TypeScript needs `lib: [\"DOM\"]` or `@types/node`. Never ship an API key to a browser.\n- In Cloudflare Workers, pass the key explicitly: `new StateMachines({ apiKey: env.STATE_API_KEY })`, and bound calls with `{ signal: AbortSignal.timeout(20_000) }`, since retries can take minutes.\n- `environments.list`, `environments.events`, `workspaces.events` and `snapshots.list` return one page, `{ data, nextCursor }`. Pass `nextCursor` back as `cursor` until it is `null`.\n- To reuse seeded data, `await env.snapshot({ name })` once, then `sm.environments.start({ snapshotId })` in place of `apps`. Capturing needs `snapshots:create` and `snapshots:read`. Snapshots are never deleted automatically, not even failed ones.\n- Check errors with `instanceof` (`StateMachinesError`, `EnvironmentFailedError`, `DisposalIncompleteError`, `TimeoutError`, `SnapshotFailedError`, `SnapshotTimeoutError`), not by `name`.\n\n## Recovering a failed create\n\nThe SDK already retries `502`, `503`, `504` and connection errors with one idempotency key. If `start()` or `create` still throws with `err.idempotencyKey`, the create failed: don't call it again as a new request, which can start a second environment. Call `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same `input`. Don't replay when `err.status` is `null` and `err.cause` is a `TypeError`: the input was invalid and nothing was sent; fix the input instead. A key only recovers a failed create: don't reuse the key of a `start()` whose wait failed, because the replay returns that stopped environment. With your own key, `start()` doesn't stop the environment on failure, because another `start()` with that key may be using it: the error carries `err.environmentId`, `err.stopped` is `false` and there is no `err.stopError`. Stop it only once nothing else needs it. It returns the environment the first call created, or creates it if that call never arrived. A different input fails with `idempotency-conflict`. Keys are kept for a day.\n\n## Troubleshooting\n\n| Symptom | Cause | Fix |\n| --- | --- | --- |\n| `release-unavailable`, 422 | Unknown or retired `appId` or `appReleaseId`. A string app is used as the `appId`. | Check the IDs against `sm.catalog.get()`. Omit `appReleaseId` for the default release. |\n| `release-unavailable`, 503, or a failed app with that `errorCode` | This deployment doesn't serve the release. The SDK retries a 503 within a 10-minute budget. | Omit `appReleaseId`. If it persists, report the `requestId`. Pass a `signal` to bound the wait. |\n| `environment-limit`, 429 | The organization is at its live-environment cap. Not retried. | `sm.environments.list({ status: 'live' })`, stop or suspend one you no longer need, then start again. |\n| `not-suspended` or `suspend-in-progress`, 409 | `resume()` before the environment finished suspending, or `stop()` while it suspends. | Await `env.waitUntilSuspended()` first. |\n| `lifetime-too-long`, 422 | `lifetimeMinutes` is above the maximum. | Use at most `limits.environmentLifetime.maximumMinutes`. |\n| `TypeError: Unknown actor <id> for <appKey>. Choose one of: <ids>.` | The `actorId` isn't one of the app's actors. No token was revealed. | Pass an `id` from `await env.actors(appKey)`, or omit `actorId` to act as the default actor. |\n| `TypeError: Choose one API for <appKey>: <ids>.` | The app has several API profiles and `apiId` is missing or unknown. Salesforce always needs one. | Pass `apiId` from the listed IDs, such as `rest`. |\n| `start()` rejects with `EnvironmentFailedError` or `TimeoutError` | The environment failed, or wasn't `running` within `readyTimeoutMs` (10 minutes by default). If `err.stopped` is `true`, `start()` already stopped it. | Read `err.environment` (`null` on a `TimeoutError` that saw no read): its `stopReason` and each app's `errorCode`. If `err.stopped` is `false`, read `err.stopError` and call `sm.environments.stop(err.environmentId)`. To keep a failed environment, use the full-control path. |\n| `StateMachinesError` (`status` null) for invalid `apps`: `Pass at least one app.`, `App key \"<key>\" is used twice.` or `\"<s>\" can't be used as an app key; pass { key, appId } instead.` | Empty list, duplicate key, or a string that isn't a valid key. No request was sent. | Pass one to ten apps with unique keys; use `{ key, appId }` when the ID isn't a valid key. |\n| `TypeError: App ... is not ready to connect` | On the full-control path, connect() ran before the app was `ready`. | Await `waitUntilReady` first, or use `start()`. |\n| `connection-unavailable`, 409 | The app stopped, or the environment is no longer `running`. | Check `sm.environments.get(env.id)`; if it is suspended, resume it; if it stopped or expired, start a new one. |\n| `EnvironmentFailedError` from `waitUntilReady` | On the full-control path, the environment became `failed`, `suspending`, `suspended`, `stopping` or `stopped`. | Read `err.environment.status`, its `stopReason` and each app's `errorCode`. Still stop it in `finally`. |\n| A vendor client such as jsforce is rejected by the gateway | The gateway needs `X-State-Environment-Token` on every request; jsforce only sends it with custom headers configured. | Prefer `connection.fetch`. If you pass `headers()` to another client, send them only to `baseUrl` and disable redirects. |\n| `not-ready`, 409 | Capture: the environment isn't `running` or an app isn't `ready`. Restore: the snapshot isn't `ready` yet. Delete: the snapshot is still `capturing`. | Capture from the handle `start()` returns. Before a restore or delete, `await sm.snapshots.waitUntilReady(snapshotId)`. |\n| `capture-in-progress`, 409 | This environment is already capturing; one capture runs at a time. | `await sm.snapshots.waitUntilReady(<the running snapshot's id>)`, then capture again. |\n| `snapshot-limit`, 429 | Older APIs only: too many captures were running. `capture` and `env.snapshot()` already retried it for 60 seconds. | Wait, then capture again. |\n| `SnapshotFailedError` | The capture failed. `err.snapshot.errorCode` is `capture-failed`, or `capture-interrupted` when the API restarted mid-capture. The snapshot is kept. | Capture again with a new idempotency key (the default); the old key returns the same failed snapshot. |\n| `SnapshotTimeoutError` | The snapshot wasn't `ready` within `readyTimeoutMs`. The capture continues on the server. | `await sm.snapshots.waitUntilReady(err.snapshotId)`. |\n| `forbidden`, 403, from `sm.snapshots.delete` | The key lacks `snapshots:delete`. Keys created by members never hold it. | Use a key whose creator holds `snapshots:delete`, or keep the snapshot. |\n| `release-unavailable`, 422, on a restore | No release on this deployment matches the snapshot's apps. | Start from `apps`, seed again and capture a new snapshot. |\n| `TypeError: Pass apiKey or set STATE_API_KEY.` | No key in the environment. | Ask the user to set `STATE_API_KEY`; in Workers, pass `apiKey`. |\n\nMore: https://state-web.usestatemachines.workers.dev/llms.txt and https://www.npmjs.com/package/@usestatemachines/sdk\n";
 
 export const quickstart =
   "import { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n";
 
 export const templates = Object.freeze({
   prompt:
-    "Set up State Machines so this project can run code against a disposable replica of a SaaS app, such as Salesforce. Follow these steps in order.\n\n1. Install the SDK: `npm install @usestatemachines/sdk`. It is ESM only and runs on Node.js 20.20 or later and on Cloudflare Workers, not in browsers.\n2. Check that `STATE_API_KEY` is set without printing it: `test -n \"$STATE_API_KEY\"`. If it is missing, stop and ask me to set it; the SDK doesn't read `.env` files. In Cloudflare Workers, pass the key as `new StateMachines({ apiKey })`.\n3. Save this quickstart unchanged as `state-quickstart.mts` and run `npx tsx state-quickstart.mts` to confirm the setup. It prints `200` and the org's limits:\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n4. Adapt it to my task. Log `await sm.catalog.get()` to find each app's ID and the actors and API profiles its release offers; don't guess IDs. Start only the apps the task needs; a string such as `'salesforce'` means `{ key: 'salesforce', appId: 'salesforce' }`. Salesforce always needs an `apiId`. `connect()` acts as the declared default connection's actor, or else the app's first actor, the administrator; to act as a specific user, pass an `actorId` from `await env.actors('salesforce')`. To reuse seeded data, `await env.snapshot()` once, then `sm.environments.start({ snapshotId })`.\n5. Always call `env.stop()` in a `finally`. In CI, then await `env.waitUntilStopped()` so the run fails if cleanup isn't confirmed. If `start()` fails after creating the environment, it stops it; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`. If the create itself fails, call `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same `input`; never retry without the key, which can start a second environment.\n6. Never print, log or commit keys or tokens, including `connection.headers()` and `connection.actorToken()`, which return live tokens.\n\nRead more: `node_modules/@usestatemachines/sdk/README.md`, {{origin}}/llms.txt and https://www.npmjs.com/package/@usestatemachines/sdk\n",
+    "Set up State Machines so this project can run code against a disposable replica of a SaaS app, such as Salesforce. Follow these steps in order.\n\n1. Install the SDK: `npm install @usestatemachines/sdk`. It is ESM only and runs on Node.js 20.20 or later and on Cloudflare Workers, not in browsers.\n2. Check that `STATE_API_KEY` is set without printing it: `test -n \"$STATE_API_KEY\"`. If it is missing, stop and ask me to set it; the SDK doesn't read `.env` files. In Cloudflare Workers, pass the key as `new StateMachines({ apiKey })`.\n3. Save this quickstart unchanged as `state-quickstart.mts` and run `npx tsx state-quickstart.mts` to confirm the setup. It prints `200` and the org's limits:\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n4. Adapt it to my task. Log `await sm.catalog.get()` to find each app's ID and the actors and API profiles its release offers; don't guess IDs. Start only the apps the task needs; a string such as `'salesforce'` means `{ key: 'salesforce', appId: 'salesforce' }`. Salesforce always needs an `apiId`. `connect()` acts as the declared default connection's actor, or else the app's first actor, the administrator; to act as a specific user, pass an `actorId` from `await env.actors('salesforce')`. To reuse seeded data, `await env.snapshot()` once, then `sm.environments.start({ snapshotId })`.\n5. Always call `env.stop()` in a `finally`; it ends the environment and discards its data. In CI, then await `env.waitUntilStopped()` so the run fails if cleanup isn't confirmed. If `start()` fails after creating the environment, it stops it; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`. If the create itself fails, call `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same `input`; never retry without the key, which can start a second environment.\n6. Never print, log or commit keys or tokens, including `connection.headers()` and `connection.actorToken()`, which return live tokens.\n\nRead more: `node_modules/@usestatemachines/sdk/README.md`, {{origin}}/llms.txt and https://www.npmjs.com/package/@usestatemachines/sdk\n",
   llmsTxt:
-    "# State Machines\n\n> State Machines runs disposable environments of replica SaaS apps, such as Salesforce, for tests and coding agents. Start, connect to and stop them with the TypeScript SDK `@usestatemachines/sdk`.\n\nInstall with `npm install @usestatemachines/sdk`. It is ESM only and runs on Node.js 20.20 or later (22 or 24 recommended) and Cloudflare Workers, not browsers. TypeScript needs `lib: [\"DOM\"]` or `@types/node`. `new StateMachines()` reads the API key from `STATE_API_KEY` and doesn't read `.env` files; in Cloudflare Workers, pass `apiKey` explicitly. Never print the key. The installed `node_modules/@usestatemachines/sdk/README.md` matches your version: read it before relying on memory.\n\n## Concepts\n\n- Environment: one disposable run of one or more apps. `const env = await sm.environments.start({ apps: ['salesforce'] })` creates it, waits until it is `running` and returns a `RunningEnvironment`. If the environment fails or the wait times out or is aborted, `start()` stops the environment and rethrows the original error; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`. Always call `env.stop()` in a `finally`.\n- Apps: a string is shorthand for `{ key: s, appId: s }`, nothing more. Use the object form, such as `{ key: 'crm', appId: 'salesforce' }`, for your own key or two instances of one app. `appId` names the app in the catalog (`sm.catalog.get()`); `key` is your name for it in this environment, and later calls take the key.\n- Actor: one of the app's users; requests run with that user's permissions. Salesforce actor IDs are its user IDs, such as `005…`, never a role name. `await env.actors('salesforce')` lists them in the release's order. connect() acts as the environment's declared default connection's actor, or else the first actor, the administrator, unless you pass `actorId`; pick a less-privileged actor to test permissions.\n- API profile: one API of the app, with its path and versions. Salesforce offers several (`rest`, `soap-partner`, `bulk2`, `oauth`) and has no default, so always pass `apiId`: `env.connect('salesforce', { apiId: 'rest' })`.\n- Connection: `connection.fetch` never sends your API key, never follows redirects and never retries. `connection.headers()` returns live tokens: treat them as secrets. Salesforce's SOAP profile (`soap-partner`) takes the actor token in the envelope instead: put `connection.actorToken()`, also a secret, in `SessionHeader/sessionId`; `headers()` then carries only `X-State-Environment-Token`.\n- Full control: `sm.environments.create(input)` returns one `Environment` before the apps boot, `sm.environments.waitUntilReady(environment.id)` waits, and `sm.environments.connect(environment.id, key, options)` connects. Nothing on this path stops the environment for you; use it to keep a failed environment for inspection.\n- Snapshot: the saved data of a running environment's apps, taken per app instance. `const snapshot = await env.snapshot({ name: 'seeded' })` captures it and waits until it is `ready`. `sm.environments.start({ snapshotId: snapshot.id })` starts a new environment from it, in place of `apps`; each app keeps its key. Capturing needs `snapshots:create` and `snapshots:read`, restoring needs `snapshots:read`, `rename` needs `snapshots:create`, and `sm.snapshots.delete(id)` needs `snapshots:delete`, which keys created by members don't hold. Snapshots are never deleted automatically, not even failed or timed-out ones: `SnapshotFailedError` and `SnapshotTimeoutError` carry `err.snapshotId`. `sm.snapshots` also has `create`, `waitUntilReady`, `get`, `list` and `rename`.\n- Vendor clients: the gateway needs `X-State-Environment-Token` on every request. A client such as jsforce only sends it if you configure custom headers. Prefer `connection.fetch`; if you pass `headers()` to another client, send them only to `connection.baseUrl` and disable redirects.\n- Pagination: `environments.list`, `environments.events`, `workspaces.events` and `snapshots.list` return one page, `{ data, nextCursor }`. Pass `nextCursor` back as `cursor` until it is `null`.\n\n## Statuses\n\nEnums are open: handle values you don't recognise.\n\n- Environment `status`: `queued`, `starting`, `running`, `stopping`, `stopped`, `failed`. `stopReason`: `requested`, `expired`, `app_failed`, `owner_lost`.\n- App `status`: `pending`, `starting`, `ready`, `stopping`, `stopped`, `failed`. A failed app carries `errorCode`.\n- App `disposal`: `pending`, `complete`, `failed`, `unknown`.\n- Snapshot `status`: `capturing`, `ready`, `failed`. A failed snapshot carries `errorCode`: `capture-failed`, or `capture-interrupted` when the API restarted mid-capture.\n- `start()` and `waitUntilReady` throw `EnvironmentFailedError` if the environment becomes `failed`, `stopping` or `stopped`, and `TimeoutError` after 10 minutes by default (`readyTimeoutMs` for `start()`, `timeoutMs` for `waitUntilReady`). `start()` then stops the environment; `waitUntilReady` never does.\n- `waitUntilStopped` resolves when the environment is `stopped` or `failed` and every app's disposal is `complete`. It throws `DisposalIncompleteError` if a disposal is `failed` or `unknown`, and `TimeoutError` if disposal stays `pending`.\n\n## Limits\n\n- Live environments are capped per organization: `(await sm.credential.get()).limits.liveEnvironments`.\n- `lifetimeMinutes` defaults to 15 minutes, or the deployment maximum if that is lower. Both are in `(await sm.credential.get()).limits.environmentLifetime`. Lifetime counts from the start, so boot uses part of it; at the end the environment stops with `stopReason: 'expired'`.\n- Always call `env.stop()` in a `finally`. In CI, then await `env.waitUntilStopped()`.\n\n## Errors and retries\n\nA failed API call throws `StateMachinesError` with `status`, `code`, `requestId` and `retryAfter`. The SDK retries `502`, `503`, `504`, connection errors and attempt timeouts itself, within a 10-minute retry budget (the last attempt can end later); `create` retries reuse one idempotency key. It retries nothing else: not `429`, other `4xx` or other `5xx`. The one exception is snapshot capture, which retries `429` `snapshot-limit` for up to 60 seconds with one idempotency key. `err.stopped === false` with `err.stopError` means the stop wasn't confirmed; `sm.environments.stop(err.environmentId)` is safe to repeat. Without `stopError`, you passed your own `idempotencyKey`, so `start()` left the environment running because another `start()` with that key may be using it; stop it only once nothing else needs it. Every error from `create`, including the create step of `start()`, carries `err.idempotencyKey`: `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same input returns the environment the first call created. Invalid `apps` fail before any request as a `StateMachinesError` (`status` null, `cause` a `TypeError`; don't replay it) with one of: `Pass at least one app.`, `App key \"<key>\" is used twice.` or `\"<s>\" can't be used as an app key; pass { key, appId } instead.`\n\n| Code | Status | What to do |\n| --- | --- | --- |\n| `environment-limit` | 429 | Not retried. Stop a live environment (`sm.environments.list({ status: 'live' })`), then start again. |\n| `lifetime-too-long` | 422 | Lower `lifetimeMinutes` to at most `limits.environmentLifetime.maximumMinutes`. |\n| `release-unavailable` | 422 | Unknown or retired app or release. Check `appId` and `appReleaseId` against `sm.catalog.get()`. |\n| `release-unavailable` | 503 | This deployment doesn't serve the release. Retried within the budget; omit `appReleaseId` to use the default release, or contact support with the `requestId`. A failed app can carry it as `errorCode`. |\n| `connection-unavailable` | 409 | The app stopped or the environment is no longer `running`. If it stopped or expired, start a new one. (Before revealing any token, connect() throws a `TypeError` for an app that isn't ready, an unknown actor, or a missing `apiId`.) |\n| `idempotency-conflict` | 409 | The idempotency key was used with a different input. Replay with the exact same input. |\n| `snapshot-limit` | 429 | Too many captures are running. `capture` and `env.snapshot()` retry it for up to 60 seconds; if it still fails, wait and capture again. |\n| `not-ready` | 409 | Capture: the environment isn't `running` or an app isn't `ready`. Restore: the snapshot isn't `ready` yet; `await sm.snapshots.waitUntilReady(id)`. Delete: the snapshot is still `capturing`. |\n| `version-conflict` | 409 | `rename` sent a stale `expectedVersion`, or the snapshot was deleted. `get` it: if `deletedAt` is set it's gone, otherwise retry with its `version`. |\n| `release-unavailable` | 422 | On a restore: no release on this deployment matches the snapshot's apps. Start from `apps`, seed again and capture a new snapshot. |\n| `request-timeout` | 503 | Retried automatically. If `create` still fails, replay it with `err.idempotencyKey`. |\n\n## Quickstart\n\nSave it unchanged as `state-quickstart.mts` and run `npx tsx state-quickstart.mts`.\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n## Docs\n\n- [npm package](https://www.npmjs.com/package/@usestatemachines/sdk): install, README and versions\n- [Agent skill]({{origin}}/skill.md): workflow, rules and troubleshooting for coding agents\n- [OpenAPI document](https://cdn.jsdelivr.net/npm/@usestatemachines/sdk/openapi.json): the HTTP API the SDK calls\n- [State Machines](https://usestatemachines.com): product site\n",
+    "# State Machines\n\n> State Machines runs disposable environments of replica SaaS apps, such as Salesforce, for tests and coding agents. Start, connect to, suspend, resume and stop them with the TypeScript SDK `@usestatemachines/sdk`.\n\nInstall with `npm install @usestatemachines/sdk`. It is ESM only and runs on Node.js 20.20 or later (22 or 24 recommended) and Cloudflare Workers, not browsers. TypeScript needs `lib: [\"DOM\"]` or `@types/node`. `new StateMachines()` reads the API key from `STATE_API_KEY` and doesn't read `.env` files; in Cloudflare Workers, pass `apiKey` explicitly. Never print the key. The installed `node_modules/@usestatemachines/sdk/README.md` matches your version: read it before relying on memory.\n\n## Concepts\n\n- Environment: one disposable run of one or more apps. `const env = await sm.environments.start({ apps: ['salesforce'] })` creates it, waits until it is `running` and returns a `RunningEnvironment`. If the environment fails or the wait times out or is aborted, `start()` stops the environment and rethrows the original error; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`. Always call `env.stop()` in a `finally`.\n- Suspend and resume: `env.suspend()` ends the sandboxes and keeps the data; the lifetime pauses. Await `env.waitUntilSuspended()`, later `env.resume()` and `env.waitUntilReady()` to run again on the same data. `env.stop()` ends the environment and discards its data; a suspended environment still needs it.\n- Apps: a string is shorthand for `{ key: s, appId: s }`, nothing more. Use the object form, such as `{ key: 'crm', appId: 'salesforce' }`, for your own key or two instances of one app. `appId` names the app in the catalog (`sm.catalog.get()`); `key` is your name for it in this environment, and later calls take the key.\n- Actor: one of the app's users; requests run with that user's permissions. Salesforce actor IDs are its user IDs, such as `005…`, never a role name. `await env.actors('salesforce')` lists them in the release's order. connect() acts as the environment's declared default connection's actor, or else the first actor, the administrator, unless you pass `actorId`; pick a less-privileged actor to test permissions.\n- API profile: one API of the app, with its path and versions. Salesforce offers several (`rest`, `soap-partner`, `bulk2`, `oauth`) and has no default, so always pass `apiId`: `env.connect('salesforce', { apiId: 'rest' })`.\n- Connection: `connection.fetch` never sends your API key, never follows redirects and never retries. `connection.headers()` returns live tokens: treat them as secrets. Salesforce's SOAP profile (`soap-partner`) takes the actor token in the envelope instead: put `connection.actorToken()`, also a secret, in `SessionHeader/sessionId`; `headers()` then carries only `X-State-Environment-Token`.\n- Full control: `sm.environments.create(input)` returns one `Environment` before the apps boot, `sm.environments.waitUntilReady(environment.id)` waits, and `sm.environments.connect(environment.id, key, options)` connects. Nothing on this path stops the environment for you; use it to keep a failed environment for inspection.\n- Snapshot: the saved data of a running environment's apps. Every app is paused while it is cut, so a multi-app snapshot is one instant (`consistency: 'environment'`); a single app is `'app_instance'`. `const snapshot = await env.snapshot({ name: 'seeded' })` captures it and waits until it is `ready`. `sm.environments.start({ snapshotId: snapshot.id })` starts a new environment from it, in place of `apps`; each app keeps its key. Capturing needs `snapshots:create` and `snapshots:read`, restoring needs `snapshots:read`, `rename` needs `snapshots:create`, and `sm.snapshots.delete(id)` needs `snapshots:delete`, which keys created by members don't hold. Snapshots are never deleted automatically, not even failed or timed-out ones: `SnapshotFailedError` and `SnapshotTimeoutError` carry `err.snapshotId`. `sm.snapshots` also has `create`, `waitUntilReady`, `get`, `list` and `rename`.\n- Recorded requests: `sm.environments.listRequests(env.id)` pages the calls made to the environment's apps (method, path, status, duration); `getRequest(env.id, id)` adds headers and bodies. Recording is best effort and SOAP bodies are not stored.\n- Vendor clients: the gateway needs `X-State-Environment-Token` on every request. A client such as jsforce only sends it if you configure custom headers. Prefer `connection.fetch`; if you pass `headers()` to another client, send them only to `connection.baseUrl` and disable redirects.\n- Pagination: `environments.list`, `environments.events`, `workspaces.events` and `snapshots.list` return one page, `{ data, nextCursor }`. Pass `nextCursor` back as `cursor` until it is `null`.\n\n## Statuses\n\nEnums are open: handle values you don't recognise.\n\n- Environment `status`: `queued`, `starting`, `running`, `suspending`, `suspended`, `stopping`, `stopped`, `failed`. `stopReason`: `requested`, `expired`, `app_failed`, `owner_lost`. A suspended environment keeps its data and carries `suspendedAt`.\n- App `status`: `pending`, `starting`, `ready`, `recovering`, `suspended`, `stopping`, `stopped`, `failed`. A failed app carries `errorCode`.\n- App `disposal`: `pending`, `complete`, `failed`, `unknown`.\n- Snapshot `status`: `capturing`, `ready`, `failed`. A failed snapshot carries `errorCode`: `capture-failed`, `pause-failed` or `pause-lost` when the apps could not be held paused for the capture, or `capture-interrupted` when the API restarted mid-capture.\n- `start()` and `waitUntilReady` throw `EnvironmentFailedError` if the environment becomes `failed`, `suspending`, `suspended`, `stopping` or `stopped`, and `TimeoutError` after 10 minutes by default (`readyTimeoutMs` for `start()`, `timeoutMs` for `waitUntilReady`). `start()` then stops the environment; `waitUntilReady` never does.\n- `waitUntilSuspended` resolves when the environment is `suspended`, and throws `EnvironmentFailedError` if it ends instead.\n- `waitUntilStopped` resolves when the environment is `stopped` and every app's disposal is `complete`. A `failed` environment keeps its data for `resume()` and is not stopped until you call `stop()`. It throws `DisposalIncompleteError` if a disposal is `failed` or `unknown`, and `TimeoutError` if disposal stays `pending`.\n\n## Limits\n\n- Live environments are capped per organization: `(await sm.credential.get()).limits.liveEnvironments`.\n- `lifetimeMinutes` defaults to 15 minutes, or the deployment maximum if that is lower. Both are in `(await sm.credential.get()).limits.environmentLifetime`. Lifetime counts from the start, so boot uses part of it, and pauses while suspended; at the end the environment stops with `stopReason: 'expired'`.\n- A suspended environment doesn't count toward the live cap; resuming it does.\n- Always call `env.stop()` in a `finally`. In CI, then await `env.waitUntilStopped()`.\n\n## Errors and retries\n\nA failed API call throws `StateMachinesError` with `status`, `code`, `requestId` and `retryAfter`. The SDK retries `502`, `503`, `504`, connection errors and attempt timeouts itself, within a 10-minute retry budget (the last attempt can end later); `create` retries reuse one idempotency key. It retries nothing else: not `429`, other `4xx` or other `5xx`. The one exception is snapshot capture, which retries `429` `snapshot-limit` for up to 60 seconds with one idempotency key. `err.stopped === false` with `err.stopError` means the stop wasn't confirmed; `sm.environments.stop(err.environmentId)` is safe to repeat. Without `stopError`, you passed your own `idempotencyKey`, so `start()` left the environment running because another `start()` with that key may be using it; stop it only once nothing else needs it. Every error from `create`, including the create step of `start()`, carries `err.idempotencyKey`: `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same input returns the environment the first call created. Invalid `apps` fail before any request as a `StateMachinesError` (`status` null, `cause` a `TypeError`; don't replay it) with one of: `Pass at least one app.`, `App key \"<key>\" is used twice.` or `\"<s>\" can't be used as an app key; pass { key, appId } instead.`\n\n| Code | Status | What to do |\n| --- | --- | --- |\n| `environment-limit` | 429 | Not retried. Suspend or stop a live environment (`sm.environments.list({ status: 'live' })`), then start or resume again. |\n| `not-ready` | 409 | Only a running environment whose apps are all ready can be suspended. Wait for it, or stop it. |\n| `not-suspended` | 409 | Only a suspended environment can resume. After `suspend()`, await `waitUntilSuspended()` first. |\n| `suspend-in-progress` | 409 | The environment is suspending. Await `waitUntilSuspended()`, then stop it. |\n| `lifetime-too-long` | 422 | Lower `lifetimeMinutes` to at most `limits.environmentLifetime.maximumMinutes`. |\n| `release-unavailable` | 422 | Unknown or retired app or release. Check `appId` and `appReleaseId` against `sm.catalog.get()`. |\n| `release-unavailable` | 503 | This deployment doesn't serve the release. Retried within the budget; omit `appReleaseId` to use the default release, or contact support with the `requestId`. A failed app can carry it as `errorCode`. |\n| `connection-unavailable` | 409 | The app stopped or the environment is no longer `running`. If it is suspended, resume it; if it stopped or expired, start a new one. (Before revealing any token, connect() throws a `TypeError` for an app that isn't ready, an unknown actor, or a missing `apiId`.) |\n| `idempotency-conflict` | 409 | The idempotency key was used with a different input. Replay with the exact same input. |\n| `capture-in-progress` | 409 | This environment is already capturing a snapshot; one runs at a time. Wait for it, then capture again. |\n| `snapshot-limit` | 429 | Older APIs only. `capture` and `env.snapshot()` retry it for up to 60 seconds; if it still fails, wait and capture again. |\n| `not-ready` | 409 | Capture: the environment isn't `running` or an app isn't `ready`. Restore: the snapshot isn't `ready` yet; `await sm.snapshots.waitUntilReady(id)`. Delete: the snapshot is still `capturing`. |\n| `version-conflict` | 409 | `rename` sent a stale `expectedVersion`, or the snapshot was deleted. `get` it: if `deletedAt` is set it's gone, otherwise retry with its `version`. |\n| `release-unavailable` | 422 | On a restore: no release on this deployment matches the snapshot's apps. Start from `apps`, seed again and capture a new snapshot. |\n| `request-timeout` | 503 | Retried automatically. If `create` still fails, replay it with `err.idempotencyKey`. |\n\n## Quickstart\n\nSave it unchanged as `state-quickstart.mts` and run `npx tsx state-quickstart.mts`.\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n## Docs\n\n- [npm package](https://www.npmjs.com/package/@usestatemachines/sdk): install, README and versions\n- [Agent skill]({{origin}}/skill.md): workflow, rules and troubleshooting for coding agents\n- [OpenAPI document](https://cdn.jsdelivr.net/npm/@usestatemachines/sdk/openapi.json): the HTTP API the SDK calls\n- [State Machines](https://usestatemachines.com): product site\n",
   skill:
-    "---\nname: state-machines\ndescription: Use when an agent or test needs a disposable replica of a SaaS app such as Salesforce, to run integration tests or try API calls against a sandboxed copy instead of a real org. Covers the @usestatemachines/sdk TypeScript SDK - starting environments, choosing an actor, snapshots, cleanup, and errors such as environment-limit and release-unavailable. Also use when the user mentions State Machines, STATE_API_KEY or @usestatemachines/sdk.\n---\n\nRead `node_modules/@usestatemachines/sdk/README.md` before relying on memory; it matches the installed version. If the package is missing, run `npm install @usestatemachines/sdk`.\n\n## Workflow\n\n1. Check that `STATE_API_KEY` is set without printing it (`test -n \"$STATE_API_KEY\"`). If it isn't, ask the user to set it; the SDK doesn't read `.env` files.\n2. Start the environment with only the apps the task needs: `const env = await sm.environments.start({ apps: ['salesforce'] })`. Log `await sm.catalog.get()` for each app's ID and the actors and API profiles its release offers; don't guess IDs. A string app means `{ key: s, appId: s }`; use `{ key, appId }` for your own key. `start()` waits until the environment is `running`. If that fails, it stops the environment and rethrows; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`.\n3. Open a `try` right after `start()` returns, and call `env.stop()` in its `finally`.\n4. Connect through one API profile. Salesforce offers several and has no default, so pass `apiId`: `env.connect('salesforce', { apiId: 'rest' })`. This acts as the environment's declared default connection's actor, or else the app's first actor, the administrator. To act with another user's permissions, choose one from `await env.actors('salesforce')` and pass its `id` as `actorId`. Salesforce actor IDs are its user IDs, such as `005…`, never a role name, so read them instead of hard-coding one.\n5. Call the app with the connection's `fetch`. A leading `/` means the root of `baseUrl`.\n6. In `finally`, `await env.stop()`. In CI, then `await env.waitUntilStopped()`.\n\nThe whole flow, as the package ships it. It is ESM with top-level `await`: save it as `state-quickstart.mts` and run `npx tsx state-quickstart.mts`.\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n## Rules\n\n- Always stop in `finally`. `waitUntilStopped` confirms every app was disposed; use it in CI so a run fails when cleanup can't be confirmed.\n- To keep a failed environment for inspection, use the full-control path: `sm.environments.create`, then `sm.environments.waitUntilReady(environment.id)`, then `sm.environments.connect(environment.id, key, options)`. Nothing on that path stops the environment for you.\n- Lifetime counts from the start, including boot. It defaults to 15 minutes, or the deployment maximum if that is lower: `(await sm.credential.get()).limits.environmentLifetime`. Set `lifetimeMinutes` in `start` or `create` for longer runs.\n- Live environments are capped per organization: `limits.liveEnvironments` from `sm.credential.get()`. Stop the ones you no longer need.\n- Never print, log or commit `STATE_API_KEY`, tokens, or `headers()` or `actorToken()` output.\n- Salesforce's SOAP profile (`soap-partner`) takes the actor token in the envelope, not a header: put `connection.actorToken()` in `SessionHeader/sessionId`. For it, `headers()` carries only `X-State-Environment-Token`.\n- `connection.fetch` never sends your API key, never follows redirects and only reaches URLs under `baseUrl`.\n- Runtimes: Node.js 20.20 or later (22 or 24 recommended), and Cloudflare Workers. ESM only. TypeScript needs `lib: [\"DOM\"]` or `@types/node`. Never ship an API key to a browser.\n- In Cloudflare Workers, pass the key explicitly: `new StateMachines({ apiKey: env.STATE_API_KEY })`, and bound calls with `{ signal: AbortSignal.timeout(20_000) }`, since retries can take minutes.\n- `environments.list`, `environments.events`, `workspaces.events` and `snapshots.list` return one page, `{ data, nextCursor }`. Pass `nextCursor` back as `cursor` until it is `null`.\n- To reuse seeded data, `await env.snapshot({ name })` once, then `sm.environments.start({ snapshotId })` in place of `apps`. Capturing needs `snapshots:create` and `snapshots:read`. Snapshots are never deleted automatically, not even failed ones.\n- Check errors with `instanceof` (`StateMachinesError`, `EnvironmentFailedError`, `DisposalIncompleteError`, `TimeoutError`, `SnapshotFailedError`, `SnapshotTimeoutError`), not by `name`.\n\n## Recovering a failed create\n\nThe SDK already retries `502`, `503`, `504` and connection errors with one idempotency key. If `start()` or `create` still throws with `err.idempotencyKey`, the create failed: don't call it again as a new request, which can start a second environment. Call `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same `input`. Don't replay when `err.status` is `null` and `err.cause` is a `TypeError`: the input was invalid and nothing was sent; fix the input instead. A key only recovers a failed create: don't reuse the key of a `start()` whose wait failed, because the replay returns that stopped environment. With your own key, `start()` doesn't stop the environment on failure, because another `start()` with that key may be using it: the error carries `err.environmentId`, `err.stopped` is `false` and there is no `err.stopError`. Stop it only once nothing else needs it. It returns the environment the first call created, or creates it if that call never arrived. A different input fails with `idempotency-conflict`. Keys are kept for a day.\n\n## Troubleshooting\n\n| Symptom | Cause | Fix |\n| --- | --- | --- |\n| `release-unavailable`, 422 | Unknown or retired `appId` or `appReleaseId`. A string app is used as the `appId`. | Check the IDs against `sm.catalog.get()`. Omit `appReleaseId` for the default release. |\n| `release-unavailable`, 503, or a failed app with that `errorCode` | This deployment doesn't serve the release. The SDK retries a 503 within a 10-minute budget. | Omit `appReleaseId`. If it persists, report the `requestId`. Pass a `signal` to bound the wait. |\n| `environment-limit`, 429 | The organization is at its live-environment cap. Not retried. | `sm.environments.list({ status: 'live' })`, stop one you no longer need, then start again. |\n| `lifetime-too-long`, 422 | `lifetimeMinutes` is above the maximum. | Use at most `limits.environmentLifetime.maximumMinutes`. |\n| `TypeError: Unknown actor <id> for <appKey>. Choose one of: <ids>.` | The `actorId` isn't one of the app's actors. No token was revealed. | Pass an `id` from `await env.actors(appKey)`, or omit `actorId` to act as the default actor. |\n| `TypeError: Choose one API for <appKey>: <ids>.` | The app has several API profiles and `apiId` is missing or unknown. Salesforce always needs one. | Pass `apiId` from the listed IDs, such as `rest`. |\n| `start()` rejects with `EnvironmentFailedError` or `TimeoutError` | The environment failed, or wasn't `running` within `readyTimeoutMs` (10 minutes by default). If `err.stopped` is `true`, `start()` already stopped it. | Read `err.environment` (`null` on a `TimeoutError` that saw no read): its `stopReason` and each app's `errorCode`. If `err.stopped` is `false`, read `err.stopError` and call `sm.environments.stop(err.environmentId)`. To keep a failed environment, use the full-control path. |\n| `StateMachinesError` (`status` null) for invalid `apps`: `Pass at least one app.`, `App key \"<key>\" is used twice.` or `\"<s>\" can't be used as an app key; pass { key, appId } instead.` | Empty list, duplicate key, or a string that isn't a valid key. No request was sent. | Pass one to ten apps with unique keys; use `{ key, appId }` when the ID isn't a valid key. |\n| `TypeError: App ... is not ready to connect` | On the full-control path, connect() ran before the app was `ready`. | Await `waitUntilReady` first, or use `start()`. |\n| `connection-unavailable`, 409 | The app stopped, or the environment is no longer `running`. | Check `sm.environments.get(env.id)`; if it stopped or expired, start a new one. |\n| `EnvironmentFailedError` from `waitUntilReady` | On the full-control path, the environment became `failed`, `stopping` or `stopped`. | Read `err.environment.stopReason` and each app's `errorCode`. Still stop it in `finally`. |\n| A vendor client such as jsforce is rejected by the gateway | The gateway needs `X-State-Environment-Token` on every request; jsforce only sends it with custom headers configured. | Prefer `connection.fetch`. If you pass `headers()` to another client, send them only to `baseUrl` and disable redirects. |\n| `not-ready`, 409 | Capture: the environment isn't `running` or an app isn't `ready`. Restore: the snapshot isn't `ready` yet. Delete: the snapshot is still `capturing`. | Capture from the handle `start()` returns. Before a restore or delete, `await sm.snapshots.waitUntilReady(snapshotId)`. |\n| `snapshot-limit`, 429 | The API is running as many captures as it can. `capture` and `env.snapshot()` already retried it for 60 seconds. | Wait, then capture again. |\n| `SnapshotFailedError` | The capture failed. `err.snapshot.errorCode` is `capture-failed`, or `capture-interrupted` when the API restarted mid-capture. The snapshot is kept. | Capture again with a new idempotency key (the default); the old key returns the same failed snapshot. |\n| `SnapshotTimeoutError` | The snapshot wasn't `ready` within `readyTimeoutMs`. The capture continues on the server. | `await sm.snapshots.waitUntilReady(err.snapshotId)`. |\n| `forbidden`, 403, from `sm.snapshots.delete` | The key lacks `snapshots:delete`. Keys created by members never hold it. | Use a key whose creator holds `snapshots:delete`, or keep the snapshot. |\n| `release-unavailable`, 422, on a restore | No release on this deployment matches the snapshot's apps. | Start from `apps`, seed again and capture a new snapshot. |\n| `TypeError: Pass apiKey or set STATE_API_KEY.` | No key in the environment. | Ask the user to set `STATE_API_KEY`; in Workers, pass `apiKey`. |\n\nMore: {{origin}}/llms.txt and https://www.npmjs.com/package/@usestatemachines/sdk\n",
+    "---\nname: state-machines\ndescription: Use when an agent or test needs a disposable replica of a SaaS app such as Salesforce, to run integration tests or try API calls against a sandboxed copy instead of a real org. Covers the @usestatemachines/sdk TypeScript SDK - starting environments, choosing an actor, snapshots, suspending and resuming, cleanup, and errors such as environment-limit and release-unavailable. Also use when the user mentions State Machines, STATE_API_KEY or @usestatemachines/sdk.\n---\n\nRead `node_modules/@usestatemachines/sdk/README.md` before relying on memory; it matches the installed version. If the package is missing, run `npm install @usestatemachines/sdk`.\n\n## Workflow\n\n1. Check that `STATE_API_KEY` is set without printing it (`test -n \"$STATE_API_KEY\"`). If it isn't, ask the user to set it; the SDK doesn't read `.env` files.\n2. Start the environment with only the apps the task needs: `const env = await sm.environments.start({ apps: ['salesforce'] })`. Log `await sm.catalog.get()` for each app's ID and the actors and API profiles its release offers; don't guess IDs. A string app means `{ key: s, appId: s }`; use `{ key, appId }` for your own key. `start()` waits until the environment is `running`. If that fails, it stops the environment and rethrows; an `EnvironmentFailedError` or `TimeoutError` carries `err.environmentId` and `err.stopped`.\n3. Open a `try` right after `start()` returns, and call `env.stop()` in its `finally`.\n4. Connect through one API profile. Salesforce offers several and has no default, so pass `apiId`: `env.connect('salesforce', { apiId: 'rest' })`. This acts as the environment's declared default connection's actor, or else the app's first actor, the administrator. To act with another user's permissions, choose one from `await env.actors('salesforce')` and pass its `id` as `actorId`. Salesforce actor IDs are its user IDs, such as `005…`, never a role name, so read them instead of hard-coding one.\n5. Call the app with the connection's `fetch`. A leading `/` means the root of `baseUrl`.\n6. In `finally`, `await env.stop()`; it ends the environment and discards its data. In CI, then `await env.waitUntilStopped()`.\n\nThe whole flow, as the package ships it. It is ESM with top-level `await`: save it as `state-quickstart.mts` and run `npx tsx state-quickstart.mts`.\n\n```ts\nimport { StateMachines } from '@usestatemachines/sdk';\n\nconst sm = new StateMachines(); // reads STATE_API_KEY\nconst env = await sm.environments.start({ apps: ['salesforce'] });\n\ntry {\n  const salesforce = await env.connect('salesforce', { apiId: 'rest' });\n  const res = await salesforce.fetch('/services/data/v67.0/limits');\n  console.log(res.status, await res.json());\n} finally {\n  await env.stop();\n}\n```\n\n## Rules\n\n- Always stop in `finally`. `waitUntilStopped` confirms every app was disposed; use it in CI so a run fails when cleanup can't be confirmed.\n- `env.suspend()` ends the sandboxes and keeps the data, with the lifetime paused; it is not cleanup. Await `env.waitUntilSuspended()` before `env.resume()`, then `env.waitUntilReady()` before connecting again. A suspended environment still needs `env.stop()`.\n- To keep a failed environment for inspection, use the full-control path: `sm.environments.create`, then `sm.environments.waitUntilReady(environment.id)`, then `sm.environments.connect(environment.id, key, options)`. Nothing on that path stops the environment for you.\n- Lifetime counts from the start, including boot. It defaults to 15 minutes, or the deployment maximum if that is lower: `(await sm.credential.get()).limits.environmentLifetime`. Set `lifetimeMinutes` in `start` or `create` for longer runs.\n- Live environments are capped per organization: `limits.liveEnvironments` from `sm.credential.get()`. Stop the ones you no longer need; suspended ones don't count until they resume.\n- Never print, log or commit `STATE_API_KEY`, tokens, or `headers()` or `actorToken()` output.\n- Salesforce's SOAP profile (`soap-partner`) takes the actor token in the envelope, not a header: put `connection.actorToken()` in `SessionHeader/sessionId`. For it, `headers()` carries only `X-State-Environment-Token`.\n- `connection.fetch` never sends your API key, never follows redirects and only reaches URLs under `baseUrl`.\n- Runtimes: Node.js 20.20 or later (22 or 24 recommended), and Cloudflare Workers. ESM only. TypeScript needs `lib: [\"DOM\"]` or `@types/node`. Never ship an API key to a browser.\n- In Cloudflare Workers, pass the key explicitly: `new StateMachines({ apiKey: env.STATE_API_KEY })`, and bound calls with `{ signal: AbortSignal.timeout(20_000) }`, since retries can take minutes.\n- `environments.list`, `environments.events`, `workspaces.events` and `snapshots.list` return one page, `{ data, nextCursor }`. Pass `nextCursor` back as `cursor` until it is `null`.\n- To reuse seeded data, `await env.snapshot({ name })` once, then `sm.environments.start({ snapshotId })` in place of `apps`. Capturing needs `snapshots:create` and `snapshots:read`. Snapshots are never deleted automatically, not even failed ones.\n- Check errors with `instanceof` (`StateMachinesError`, `EnvironmentFailedError`, `DisposalIncompleteError`, `TimeoutError`, `SnapshotFailedError`, `SnapshotTimeoutError`), not by `name`.\n\n## Recovering a failed create\n\nThe SDK already retries `502`, `503`, `504` and connection errors with one idempotency key. If `start()` or `create` still throws with `err.idempotencyKey`, the create failed: don't call it again as a new request, which can start a second environment. Call `sm.environments.start(input, { idempotencyKey: err.idempotencyKey })` with the same `input`. Don't replay when `err.status` is `null` and `err.cause` is a `TypeError`: the input was invalid and nothing was sent; fix the input instead. A key only recovers a failed create: don't reuse the key of a `start()` whose wait failed, because the replay returns that stopped environment. With your own key, `start()` doesn't stop the environment on failure, because another `start()` with that key may be using it: the error carries `err.environmentId`, `err.stopped` is `false` and there is no `err.stopError`. Stop it only once nothing else needs it. It returns the environment the first call created, or creates it if that call never arrived. A different input fails with `idempotency-conflict`. Keys are kept for a day.\n\n## Troubleshooting\n\n| Symptom | Cause | Fix |\n| --- | --- | --- |\n| `release-unavailable`, 422 | Unknown or retired `appId` or `appReleaseId`. A string app is used as the `appId`. | Check the IDs against `sm.catalog.get()`. Omit `appReleaseId` for the default release. |\n| `release-unavailable`, 503, or a failed app with that `errorCode` | This deployment doesn't serve the release. The SDK retries a 503 within a 10-minute budget. | Omit `appReleaseId`. If it persists, report the `requestId`. Pass a `signal` to bound the wait. |\n| `environment-limit`, 429 | The organization is at its live-environment cap. Not retried. | `sm.environments.list({ status: 'live' })`, stop or suspend one you no longer need, then start again. |\n| `not-suspended` or `suspend-in-progress`, 409 | `resume()` before the environment finished suspending, or `stop()` while it suspends. | Await `env.waitUntilSuspended()` first. |\n| `lifetime-too-long`, 422 | `lifetimeMinutes` is above the maximum. | Use at most `limits.environmentLifetime.maximumMinutes`. |\n| `TypeError: Unknown actor <id> for <appKey>. Choose one of: <ids>.` | The `actorId` isn't one of the app's actors. No token was revealed. | Pass an `id` from `await env.actors(appKey)`, or omit `actorId` to act as the default actor. |\n| `TypeError: Choose one API for <appKey>: <ids>.` | The app has several API profiles and `apiId` is missing or unknown. Salesforce always needs one. | Pass `apiId` from the listed IDs, such as `rest`. |\n| `start()` rejects with `EnvironmentFailedError` or `TimeoutError` | The environment failed, or wasn't `running` within `readyTimeoutMs` (10 minutes by default). If `err.stopped` is `true`, `start()` already stopped it. | Read `err.environment` (`null` on a `TimeoutError` that saw no read): its `stopReason` and each app's `errorCode`. If `err.stopped` is `false`, read `err.stopError` and call `sm.environments.stop(err.environmentId)`. To keep a failed environment, use the full-control path. |\n| `StateMachinesError` (`status` null) for invalid `apps`: `Pass at least one app.`, `App key \"<key>\" is used twice.` or `\"<s>\" can't be used as an app key; pass { key, appId } instead.` | Empty list, duplicate key, or a string that isn't a valid key. No request was sent. | Pass one to ten apps with unique keys; use `{ key, appId }` when the ID isn't a valid key. |\n| `TypeError: App ... is not ready to connect` | On the full-control path, connect() ran before the app was `ready`. | Await `waitUntilReady` first, or use `start()`. |\n| `connection-unavailable`, 409 | The app stopped, or the environment is no longer `running`. | Check `sm.environments.get(env.id)`; if it is suspended, resume it; if it stopped or expired, start a new one. |\n| `EnvironmentFailedError` from `waitUntilReady` | On the full-control path, the environment became `failed`, `suspending`, `suspended`, `stopping` or `stopped`. | Read `err.environment.status`, its `stopReason` and each app's `errorCode`. Still stop it in `finally`. |\n| A vendor client such as jsforce is rejected by the gateway | The gateway needs `X-State-Environment-Token` on every request; jsforce only sends it with custom headers configured. | Prefer `connection.fetch`. If you pass `headers()` to another client, send them only to `baseUrl` and disable redirects. |\n| `not-ready`, 409 | Capture: the environment isn't `running` or an app isn't `ready`. Restore: the snapshot isn't `ready` yet. Delete: the snapshot is still `capturing`. | Capture from the handle `start()` returns. Before a restore or delete, `await sm.snapshots.waitUntilReady(snapshotId)`. |\n| `capture-in-progress`, 409 | This environment is already capturing; one capture runs at a time. | `await sm.snapshots.waitUntilReady(<the running snapshot's id>)`, then capture again. |\n| `snapshot-limit`, 429 | Older APIs only: too many captures were running. `capture` and `env.snapshot()` already retried it for 60 seconds. | Wait, then capture again. |\n| `SnapshotFailedError` | The capture failed. `err.snapshot.errorCode` is `capture-failed`, or `capture-interrupted` when the API restarted mid-capture. The snapshot is kept. | Capture again with a new idempotency key (the default); the old key returns the same failed snapshot. |\n| `SnapshotTimeoutError` | The snapshot wasn't `ready` within `readyTimeoutMs`. The capture continues on the server. | `await sm.snapshots.waitUntilReady(err.snapshotId)`. |\n| `forbidden`, 403, from `sm.snapshots.delete` | The key lacks `snapshots:delete`. Keys created by members never hold it. | Use a key whose creator holds `snapshots:delete`, or keep the snapshot. |\n| `release-unavailable`, 422, on a restore | No release on this deployment matches the snapshot's apps. | Start from `apps`, seed again and capture a new snapshot. |\n| `TypeError: Pass apiKey or set STATE_API_KEY.` | No key in the environment. | Ask the user to set `STATE_API_KEY`; in Workers, pass `apiKey`. |\n\nMore: {{origin}}/llms.txt and https://www.npmjs.com/package/@usestatemachines/sdk\n",
 });