v1.2.40

• Actors on puppeteer 25 (ESM-only) are sent to the remote browser when puppeteer is loaded with require() before the SDK on a run without the preload. The SDK only recognised puppeteer's old lib/cjs build, so such a puppeteer.launch() tried to start a local browser; the named launch export (const { launch } = require('puppeteer')) is covered too. puppeteer 24 and older behave as before.

v1.2.39

• A refused charge names the limit that refused it. When a charge is refused or cut short, the platform now says why (HTTP 402 error.data.reason, and limitedBy on a partial acceptance): the run's item limit (maxItems), its maximum total charge (maxTotalChargeUsd) or the account balance. The SDK used to call all of them the spending limit. Now its log lines name the real limit; ChargingManager.getChargeLimitReason() returns 'max-items', 'max-charge' or 'insufficient-funds' and getChargeLimitMessage() the stop message ("Stopped: reached the item limit you set (800 items)", "Stopped: reached the maximum charge you set ($3.00)", "Stopped: your account has no credit left for this run"); and once a limit was reached, a status message (Actor.setStatusMessage, Actor.exit, Actor.fail) whose clause blames the maximum charge gets that clause replaced by the real stop message. Platforms that do not send the reason yet: inferred from the 402 message and ACTOR_MAX_PAID_DATASET_ITEMS. What is charged and saved does not change.

v1.2.38

• A key-value store record write (Actor.setValue, KeyValueStore.setValue) is no longer lost to one failed answer. Each write now sends an Idempotency-Key, and when its answer is lost (a 502/504 page from the edge or load balancer, a dropped connection, a timeout) it is sent again with the same key; the API applies and counts it once. Before, the write was never re-sent, so one such page on the final OUTPUT write failed the whole run. An API that answers a keyed write without X-Scrapely-Idempotency (an older server) gets no re-sends, as before.

v1.2.37

• Opening a named dataset, key-value store or request queue (Actor.openDataset('name') and the like) retries when the API answers 503 storage-name-busy. The API sends it when another request kept creating or renaming a storage of that name for over 15 s; nothing was changed, so the open is simply repeated. It used to fail the open.
• Once Actor.fail() has run, the run stays failed: a later Actor.exit(), or the end of Actor.main(), keeps the failure's exit code and status message. fail() normally ends the process at once, so this only changed runs that called fail({ exit: false }) and then exit() or returned from Actor.main(); they used to exit with 0. A second fail() keeps the first failure. fail({ exitCode: undefined }) exits with 1.

v1.2.36

• Debug captures (Actor.debugCapture, Actor.debugCaptureSync) keep payloads over 256 KiB whole. They are stored gzip-compressed and base64-encoded as { key, payload, encoding: 'gzip-base64', length, truncated } instead of being cut at 262,144 characters; up to 8 Mi characters are kept. A payload that does not compress enough keeps its first 256 KiB, now marked truncated: true with its full length. Uint8Array / Buffer payloads are stored as UTF-8 text. Captures are only active on the platform's own diagnostic runs (SCRAPELY_DEBUG_CAPTURE=1).