Changelog

New features

• Saudi Arabia Proxy Location — Added RESIDENTIAL_SA as a proxy geolocation option for automations targeting Saudi Arabian websites.
• Workflow-Level Error Code Mapping — Workflows now support an error_code_mapping field that all blocks inherit automatically. Block-level mappings still take precedence, giving you a single place to define default error handling without repeating it on every block. (#5506)

Improvements

• Smarter Extraction Cache Keys — The extraction cache now normalizes session-volatile tokens (random UUIDs, CSRF values, nonce query params, timestamps) before hashing cache keys. Semantically identical page visits now share cached extraction results, improving cache hit rates across workflow re-runs. (#5504)
• Reduced Rate-Limit Errors on Extraction Prompts — Extraction prompts are now capped in size before being sent to the LLM. Oversized previous_extracted_information, extracted_information_schema, and element trees are truncated rather than sent whole, significantly reducing 429 RESOURCE_EXHAUSTED errors on high-volume workflows. (#5502)

Bug fixes

• Fixed CSV parsing failing on files whose first header row exceeded ~1 MB, causing valid wide-format CSVs to be rejected with “Could not determine delimiter”. (#5510)
• Fixed a browser automation crash on sites with cross-origin stylesheets (e.g. MUI-based apps). CSS hover styles are now parsed entirely in memory without modifying the DOM, preventing React reconciliation errors that caused the agent to loop until max steps. (#5508)
• Fixed context-window overflow failures being misdiagnosed. Steps that fail due to oversized input are now correctly counted and short-circuit to a clear LLM_ERROR failure reason instead of triggering a redundant LLM-based failure analysis. (#5507)
• Fixed workflow file-download blocks losing their configured download_suffix at task finalization, causing downloaded files to be persisted without the expected filename suffix. (#5499)
• Fixed cached scripts for non-ATS workflows incorrectly generating fill_form() calls, which defeated the performance benefits of caching by delegating back to AI at runtime. (#5497)
• Fixed the pagination navigator on the Workflow Past Runs and Run History pages being right-aligned instead of centered.

New features

• Adaptive Caching (Script Generation) — Skyvern can now learn from repeated workflow runs and generate cached scripts that speed up future executions. When a block runs successfully, Skyvern records a reusable script and replays it on subsequent runs, falling back to the AI agent only if the script fails. (#4908, #4916, #4917, #4920, #4922, #4931)
• Workflow Trigger Block — A new block type that lets one workflow trigger another from within the workflow editor. Chain workflows together as building blocks for complex automations. (#4885)
• Browser-based CLI Signup — New users can sign up for Skyvern Cloud directly from the CLI. Running skyvern signup opens a browser flow that creates your account and stores your API key locally. (#4925)
• Interactive ngrok Tunnel — skyvern browser serve can now create an ngrok tunnel automatically, making it easy to expose your local browser to Skyvern Cloud without manual network configuration. (#4924)
• South Korea Proxy Location — Added RESIDENTIAL_KR as a proxy geolocation option for automations targeting South Korean websites. (#4918)
• Downloaded Files Tab — Browser session detail pages now include a Downloaded Files tab for viewing and accessing files captured during a session. (#4911)
• CDP Proxy Authentication — Remote browser connections now support authenticated proxies via the CDP Fetch.authRequired protocol. (#4936)
• Remote Browser Download Interception — File downloads triggered by XHR requests and other non-navigation events are now captured when using remote CDP browser connections. (#4906, #4921, #4934)

Improvements

• Renamed “Login-Free” to “Saved Profile” — Clearer terminology throughout the UI. Browser session checkbox wording has also been improved. (#4914)
• TOTP Verification Improvements — Better timer handling, expired code retry, 2FA banner suppression, and more reliable goal-text extraction during TOTP flows. (#4860)
• TOTP Webhook includes workflow_permanent_id — Easier to identify which workflow is requesting a 2FA code. (#4871)
• Consolidated Gemini 3 Pro Model Key — VERTEX_GEMINI_3_PRO now auto-resolves to the latest version, so you no longer need to update config when Google releases point versions. (#4926)
• Browser Rotation for Remote Connections — Remote browser environments now support context rotation, improving reliability for long-running automations. (#4929)

Bug fixes

• Fixed a race condition where final task status could be overwritten by a late-arriving non-final status. (#4928)
• Conditional blocks no longer crash when evaluating empty or null parameters. (#4907)
• Browser sessions using FORCE_BROWSER_SESSION no longer time out prematurely. (#4903)
• SVG-based captcha text is now correctly extracted during DOM scraping. (#4880)
• MCP remote tool handlers now route to the correct API environment. (#4869)
• MCP remote auth token comparison no longer fails on encrypted tokens. (#4863)
• skyvern quickstart now handles local PostgreSQL without a pre-existing skyvern role. (#4878)
• WebSocket URLs through ngrok tunnels now rewrite correctly for live browser streaming. (#4927)

New features

• current_date Reserved Parameter — Workflows now have a built-in current_date parameter that resolves to today’s date automatically. (#4854)
• Reserved Parameters in Block Editor — The workflow block parameter picker now shows reserved system parameters, making them discoverable without memorizing names. (#4857)
• Natural Language Loop data_schema — Loop blocks driven by natural language extraction now support a data_schema field for consistent output structure. (#4851)
• Gemini 3.1 Pro and Inception Mercury-2 — Two new LLM engine options for your automations. (#4847)
• MCP Server on API Service — The MCP remote server is now mounted at /mcp on the existing API, eliminating the need for a separate process. (#4843)

Improvements

• Workflow save validation (HTTP 422) now returns clear, field-level error messages. (#4852)
• Browser launch errors now show actionable messages instead of raw exceptions. (#4844)
• Cloud LLM fallback overrides are now properly applied when selecting models. (#4839)

Bug fixes

• Conditional blocks no longer crash when all branches use Jinja templates. (#4849)
• Conditional blocks no longer misinterpret resolved Jinja template variables. (#4841)

New features

• Browser Profile Testing — Test browser profiles directly from the UI to verify that saved login sessions are still valid. Workflows can also run with a saved browser profile, enabling login-free automations that reuse authenticated sessions. (#4818)

Bug fixes

• When both a TOTP credential and a webhook are configured for 2FA, the credential is now correctly prioritized. (#4811)

New features

• Skyvern Skills Package — The new skyvern skill CLI command provides pre-built workflow templates and reference documentation for common automation patterns. (#4817)
• Clear Cached Scripts API — A new endpoint lets you clear cached automation scripts for a workflow when a target website changes and you want to force Skyvern to re-learn. (#4809)

New features

• Automatic 2FA Detection Without TOTP Credentials — Skyvern now detects when a website asks for a 2FA code even without pre-configured TOTP credentials. The system pauses and waits for you to provide the code via the UI or webhook. (#4786)
• Full CLI Parity with MCP — The CLI now supports browser session management, credential management, block operations, and enhanced workflow commands. Everything available through MCP is now accessible from the command line. (#4789, #4792, #4793)
• Claude Opus 4.6 CUA Support — Anthropic’s Claude Opus 4.6 is now available as a Computer Use Agent model for driving browser interactions. (#4780)

Improvements

• The workflow run timeline now properly displays loop iterations and conditional branches, making complex runs easier to debug. (#4782)

Bug fixes

• Fixed conditional blocks evaluating the wrong value after Jinja template rendering. (#4801)
• Fixed MFA resolution priority when both TOTP credentials and webhooks are configured. (#4800)

New features

• Anthropic Claude Opus 4.6 — Added Claude Opus 4.6 as an available LLM engine for web automation tasks. (#4777, #4778)