releases
4.2.6
2026-10-29
FeelFish Team
4.2.6
- ๐๏ธ Two new Qwen audio models in the toolbox: turn text into narration, or generate voices, sound effects and ambience from a script; describe a voice in words or upload a reference clip
- ๐ง More natural audiobook cut points: file boundaries fall at sentence ends, with no music or effects left at the end of a segment
- ๐๏ธ Right-click a tab to locate its file in the sidebar, which makes long projects easier to navigate
- ๐ง Fixed "Reset and rebuild smart context" hanging on the confirm dialog: it now closes as soon as you confirm, and progress and cancel live in the toolbar
- ๐ Failed media tasks in the status bar are marked as read when opened instead of staying red
- ๐ง Smaller fixes
4.2.5
- โ๏ธ Cloud sync is calmer: failed files show their path and reason one by one, a single failed file no longer aborts the whole round, and path changes no longer raise a false "version conflict"
- ๐๏ธ Archived conversations now sync to the cloud too, so switching devices no longer loses them
- ๐ Drag docx / PDF files straight into the input box; they are converted to text when sent
- ๐ฑ๏ธ Fixed "focused but can't type" after removing an input slot
- ๐ Claiming invite rewards is more stable: no more getting stuck while processing, and duplicate claims are guarded
- ๐ค When the AI occasionally sends malformed tool arguments, the run is no longer aborted and the AI retries on its own
- ๐ง Smaller fixes
4.2.4
- ๐งฉ New "General custom" solution and agent: a blank agent with every tool, covering everything from writing to media tasks
- ๐ผ๏ธ Drag images and text files straight into the input box and send them along with your message
- ๐งฐ Easier custom media endpoints: parameters are checked on save with the reason, config changes apply without a restart, and async endpoints wait for their own results
- ๐ The task board can clear all failed tasks at once
- ๐ More dependable read-aloud: running out of credits offers a top-up entry, "Online model" only lists the TTS you set up yourself, and switching sources no longer keeps the previous error
- ๐ Fixed drafts still being overwritten by the agent after they were cleared, so unsaved text is no longer lost
- ๐ง Smaller fixes
4.2.3
- ๐งฐ Easier custom media-model setup: parameters are checked on save, so a wrong type or placeholder explains itself; the Secretary can also recognize asynchronous endpoints and wait for results
- ๐ Clearer endpoint tests: upstream task ids are tracked with a heads-up about possible charges, and common error codes get plain-language messages
- ๐ Rich-text saving no longer falls back to an empty document: a failed body read is reported instead of writing empty content back
- โ Approval cards show the request about to be sent, long commands scroll within a capped height, and dialogs follow your theme and language
4.2.2
- โ Command approvals can now be dismissed for the current conversation with "Don't remind me again this conversation": the same card no longer pops up repeatedly, and confirmation comes back with your next message
- ๐ Fixed "save seems to do nothing": after a file is changed by another program, saving now clearly reports the conflict and offers a compare entry instead of failing silently
- ๐ง More reliable remote control: expired artifact requests no longer stall sync, and leftover requests finish cleanly
4.2.1
- ๐ Commands no longer fail outright: when the AI thinks a command needs your call, it explains why and asks you to confirm instead of erroring out
- ๐ง The solution/agent picker now shows the full solution description at the top of its second column
4.2.0
- ๐งญ New "Effortless": drafting and review are tuned by us over time, so you get the latest without updating the app, and skills and agents stay freely switchable
- ๐ The Secretary can now operate web pages for you: opening, signing in, filling in, publishing. Control comes back to you when needed, and publishing or paying always asks first
- ๐ More reliable web fetching: pages that block it are reported honestly, and with your consent they can be read from a screenshot instead of feeding garbled text to the AI
- ๐ผ๏ธ Paste images straight into the input box and the AI can write from them; images are compressed into the project automatically
- ๐ New chapter-feedback entry in the status bar, shown once on the first chapter of new projects; your feedback goes straight into improving the solution
- โ Importing local files can remember "don't ask again for similar tasks"
- ๐ง Smaller fixes