v4.100.254

Settings

• Notice wording and the copy bar moved to the provider settings. They
  are part of how a model handles notices, so they are now set per provider,
  in the Notices section below the switch that turns notices on, with Load
  default and Reset, and only shown while notices are on. They are gone from
  Settings > Features. Wording set there in 4.100.253 is not carried over:
  set it again on each provider that uses it.

Fixes

• repeat_last_n keeps -1 on Ollama and opencoti. The field accepted -1
  and showed no error, but the value was never saved, and a -1 saved earlier
  was dropped as soon as any other sampling field was changed. It is now
  saved and sent. Upstream llama.cpp still refuses -1, as before, because
  that server rejects it.
• Sampling fields save what they accept. An integer field such as
  top_k or seed took a decimal with no error and saved it cut to a whole
  number; it now says the value must be a whole number. On llama.cpp, a
  n_keep below -1, or a repeat_last_n, n_keep or num_predict above
  2,147,483,647, made the server refuse every request; the panel now refuses
  them instead.

Agents

• Edits to a line you just wrote are no longer refused. After the model
  added text to the end of a line it had read, the editor still treated the
  added characters as unseen and refused further edits reaching them, as
  "past what you have seen".

• No "you just changed" note after a refused edit. A refused edit was
  followed by a note saying the file had been changed.

• Harness notes no longer read as you speaking. The task checklist and
  the reminders Cerebriline adds while a model works (an edit with no
  checklist, a repeated passage, a cut-off tool call) were sent as user
  messages. Models took them as instructions from you, and on Gemma 4 each
  one also wiped the model's earlier reasoning from the prompt, so it kept
  starting over with "The user wants to …". They now arrive with the tool
  result they follow.

• The editor's "retyped N characters" hint points at the right line. It
  appeared after any long edit that left brackets unbalanced, usually naming
  a fix for some other line, and its "the call above" read as the model's
  own edit. It now appears only when the edit touched the line the bracket
  check has a one-character fix for, and names that line and the call.