Changelog

Version 0.82.7 - November 04, 2026

Improvements

• Optimized browser text wait operations in Autoworklet to query the extension in 50-second intervals for improved stability.
• Adjusted Ollama GLM-4 OCR configuration to prevent unintended offload behavior.

Fixes

• Resolved an issue where a corrupt GGUF header disrupted the GLM-OCR model finder.
• Fixed GLM-OCR model naming compatibility for environments running Ollama 0.30 and newer.
• Prevented text inside script and style tags from matching wait conditions in the browser extension, and stopped generating year-based IDs.
• Ensured Autoworklet avoids storing generated IDs without verifiable names or text as locators.
• Corrected desktop input typing to check native window focus rather than browser focus.
• Fixed UI control wait conditions to properly reject unreadable controls and re-evaluate control state after waiting.

Version 0.82.6 - November 03, 2026

Fixes

• Resolved an issue in Autoworklet's OCR where Ollama processes running the glm-ocr build failed to terminate properly.
• Fixed an issue in Autoworklet's OCR where legitimate repeated text was prematurely truncated by n-gram stop sequences in glm_goollama.

Version 0.82.5 - November 03, 2026

Improvements

• Added the ability in Autoworklet to learn deterministic row paths from agent interactions and replay them without invoking model calls.
• Enhanced data validation to detect misread values, renamed labels, and extraction from mismatched fields.
• Added support for waiting on enabled controls via ui.wait(enabled=True) to prevent interacting with disabled interface elements.
• Updated list lookup tasks with automated input normalization, field verification, and sample auditing.
• Introduced validation warnings in the Autoworklet CLI when publishing worklets that contain machine-specific local paths.
• Enabled automatic verification and single-retry fallback logic for read-only agent queries.

Fixes

• Fixed browser wait timeout parsing for millisecond values and extended supported wait times up to 15 minutes.
• Prevented unstable generated IDs and hashed CSS classes from being used in element locators.
• Improved placeholder detection so entries like 'N/A' or dashes are treated as empty without misidentifying valid text entries.
• Handled blank and floating-point row inputs gracefully during replays without halting the run.
• Corrected desktop UI state tracking to inspect control enablement, capture multi-word values accurately, and prevent reading sensitive password fields.
• Resolved startup race conditions during cold browser launches to ensure the session waits for the extension to connect.
• Restricted write-task retries to prevent unverified agent answers from being returned.

Version 0.82.4 - November 02, 2026

Improvements

• Updated Autoworklet API documentation to clarify cloud run execution environments, read-only file behavior, and spawn_worklet callback handling.

Fixes

• Resolved an issue in Autoworklet runner ensuring executions and callbacks operate correctly within their designated output directory when worklet files are read-only.

Version 0.82.3 - November 02, 2026

Improvements

• Added the --transcript flag to awl agent to securely record turn-by-turn executions while masking sensitive data.
• Enhanced browser_route to support fulfilling requests with custom response headers to enable route redirections.
• Improved native UI inspection performance with budgeted walks and targeted region queries for dense windows.
• Added repeat_safe metadata to ActionResult and running statement tracking to improve reliability in self-healing workflows.
• Added row-lookup scaffolding and improved fallback handling for agent list-input tasks.

Fixes

• Fixed keyboard chord handling and key-release behavior across platforms to prevent stuck keys and X server disconnections.
• Hardened transcript redaction to ensure masked secrets remain protected even after variable scope changes or interrupts.
• Resolved issues where Autoworklet dashboard actions and deployment views could unintentionally spawn host daemons.
• Added safety confirmation prompts to awl learn, from-video, and from-audio commands before executing on the OS target.
• Prevented self-healing first aid routines from re-executing statements that have already performed actions.
• Fixed file upload target paths in awl create to correctly place files into the worklet managed directory.
• Fixed API schema validation and error reporting in type_with_prompt when communicating through proxies.
• Ensured launch-key files are cleanly deleted when the dashboard service terminates.

Version 0.82.2 - November 01, 2026

Fixes

• Resolved an issue in Autoworklet where the agent could navigate away before capturing required evidence screenshots.

Version 0.82.1 - November 01, 2026

Improvements

• Enabled verification by default across all targets in the Autoworklet agent CLI, retiring the AWL_AGENT_VERIFY_REMOTE flag.
• Enhanced the Autoworklet agent judge to evaluate returned read data and files written by worklets.
• Added support in Autoworklet agent for directing native application menus via exact live paths.
• Added the ability to cancel running intake jobs directly from the Workbench Intake pane.
• Updated explore and validate jobs in the dashboard to return job deadlines upon initiation.

Fixes

• Strengthened secret masking in Autoworklet across browser reads, logs, error reports, and batch workbook outputs to redact credentials, keys, and card numbers.
• Prevented window management commands from mistakenly targeting panels, desktops, or unrelated background windows.
• Improved window resolution and dialog handling during focus, resize, and move operations.
• Fixed the Autoworklet agent judge to treat printed script lines as unverified claims and challenge premature completion signals.
• Resolved OCR backend failure handling to prevent silent blank reads and ensure proper model retries and cache keying.
• Fixed UI element detection in the move-to target dialog.
• Hardened dashboard and Workbench error handling to properly report crashed worker runs without leaking raw stack traces.

Version 0.82.0 - October 29, 2026

Improvements

• Added native OS keyboard fallback to handle JavaScript dialogs when browser debugging is restricted.
• Updated the Autoworklet browser extension to observe dialogs and attach to tabs from initial creation.
• Added the --no-cull-viewport option to browser snapshots in the CLI to preserve controls located outside the visible viewport.
• Standardized user interaction commands (hover, click, fill) across the API and CLI to return structured ActionResult payloads.
• Optimized video recording by encoding H.264 video streams on the fly to prevent delays when stopping recordings.
• Enhanced data privacy by masking sensitive and secret input fields in screenshots, logs, and vision model prompts.
• Strengthened platform security by immediately rejecting revoked authentication tokens.

Fixes

• Resolved Autoworklet Workbench display and lock issues when sandbox starts exceed their timeout or encounter provider errors.
• Fixed link resolution in browser_extract_links to strictly adhere to HTML base URL rules, correctly resolving protocol-relative and relative links.
• Corrected file handling for remote captures and Docker screenshots so saved images and videos land at expected host paths.
• Fixed GUI layout truncation affecting wrapped element cards and item rows across library, macro, and crop lists.
• Corrected coordinate calculations for clicks and inputs within focused subframes to map accurately to top-page viewport pixels.
• Prevented stalled automation runs by generating synthetic release events when native keyup or mouseup actions are refused.
• Fixed assertion utilities (assert_image and assert_screen) to clean up temporary comparison files and avoid overwriting baseline images.
• Updated CLI commands (verify-state, assert-screen, wait-settle) to exit with code 1 upon failed checks.

Version 0.80.1 - October 27, 2026

Improvements

• Added first-class cross-origin iframe support in Autoworklet, enabling trusted click and typing interactions mapped directly to the top-level viewport via CDP.
• Introduced domain-scoped sensitive data handling to ensure secrets are strictly confined to authorized destination hosts and fail closed off-host.
• Enabled trusted CDP input by default for click, type, and hover interactions across browser automations.
• Added the opt-in browser_batch command to streamline multi-step browser actions with strict deadline management.
• Optimized DOM snapshot performance by viewport-culling snapshots and automatically redacting sensitive input values.
• Enhanced JPEG screenshot capturing with optimized payload budgeting, reducing capture attempts for oversized images.

Fixes

• Improved sensitive data redaction across Autoworklet logs and CLI outputs, correctly masking repr representations, URL-encoded secrets, and edge-case character sets.
• Fixed execution handling in run_js to report an explicit failure when a script does not execute instead of returning a navigated state.
• Ensured browser batch operations strictly respect execution deadlines, properly clean up held input presses, and report timeout states accurately.
• Resolved dialog handling hangs and improved navigation detection to accurately report modal dialogs and navigation events during running actions.
• Addressed input release retries to maintain correct mouse button states during trusted interaction failures.
• Ensured Autoworklet overlay UI is properly hidden inside same-origin nested frames.
• Preserved backward compatibility for multi-click and right-click actions across supported extension versions.

Version 0.80.0 - October 22, 2026

Improvements

• Added error origin tracking to failed run documents in the Autoworklet API for improved diagnostics.

Fixes

• Resolved selector parsing issues with pseudo-class arguments by properly handling escaped quotes, apostrophes, and nested parentheses.
• Improved error attribution in the Autoworklet API to accurately distinguish sandbox script errors from internal engine errors.
• Fixed remote host execution handling by correctly inlining callback defaults and calculating UTF-8 column offsets via AST parsing.
• Ensured recorder configuration defaults are preserved when encountering mismatched data types in configuration files.
• Corrected wait backoff logic to trigger only when stillness is verified across all screen checks.
• Ensured chat interface cards render inline reliably after headless output completion.