Long tasks that finish, Custom skills and memory that lasts

Write your own skills from HydraAI Client Settings → Skills. Private to your workspace, offered in chat alongside the built-in ones.
Spreadsheets are read in full. Uploads reached Hydra as the first ten rows only; now every sheet and every row.
Hydra remembers what it has done. In long builds it keeps a list of records created and files read, so it stops re-creating or re-asking after older messages are trimmed.
Ask and Plan never change your data. Writes happen only in Fast mode or while a plan is being built.
Mode switches take effect straight away on every model, including after Regenerate.
More models work, including ones that always reason before answering and the newest Claude models.
Drafts are kept per chat. Unsent text survives New Chat, switching chats and reloads, and never lands in the wrong conversation.
Built-in Frappe guides check your site first. Eight guides, including Data Import, now inspect the live site instead of assuming a version.
Long tasks finish. Replies no longer stop with "Could not reach the server" when a step goes quiet for a couple of minutes.
Plans you can follow and steer. Short plan (title, summary, tasks), a side panel with status badge, task progress, a Modify plan box and one Proceed button. Every build now ends Completed, Partial or Failed, never stuck on Executing.
Easier-to-read chat window with even spacing and slightly larger text.

One failed tool call no longer ends the whole reply.
"This conversation has no history to respond to" after an earlier failed message is gone.
Planning tools no longer appear in Ask and Fast chats.
Interim plan updates are accepted while drafting; the "cannot run while planning" warning now says it will run on Proceed.
Older plan versions collapse to a single row after a re-plan.
Long skill playbooks arrived truncated at 3,000 characters; they now arrive whole.
Hydra no longer declares an attached file "doesn't exist" after searching by a guessed name.
Very large attachments no longer push a request over the model's limit.
Auto-written commit messages work again.
The most common sandbox failure (code starting with import) is gone.
The settings gear is shown only to users who can open the page.
Card numbers and other sensitive values are masked everywhere in the audit log.

A friendly “We’ll be right back” page while a site restarts or redeploys
Sites switch into maintenance mode while apps install and the database migrates
Deployments recover a stopped site instead of being blocked
Failure emails now arrive in cases that used to stay silent, such as a static site build that times out

A push that did not land is never reported as a success. Rejected pushes fail the log, and "Verified" checks the remote.
Export Preview shows exactly what the push will contain.
Workflows reach your other sites as fixtures.
Pushing again to an open pull request branch builds on that branch instead of conflicting.
The Create PR dialog is pre-filled with the AI-written description.

A release tag already in use is caught before anything is pushed.
Untouched Workflows, Notifications and custom DocTypes no longer show as changed.
A later push no longer restores edits you had discarded.
Your git identity is used for the push without being written into the app's repository settings.
A failed Initialize Git Repository cleans up and can be retried, with a clear message when the base branch is missing.
The Failed filter in Migration Log works.
Labels and messages match the buttons, fields and settings that exist.

Long tasks that finish, Custom skills and memory that lasts ​