v4.100.253

Settings

• A status dot beside the API Provider label. Grey until checked, green
  when the provider's endpoint answers with the URL and key you set, red when
  it does not answer, refuses the key, or no longer has the selected model.
  Hover it to see what was asked and what came back; click it to check again.
  It also checks when the settings open, when you switch provider or load a
  profile, and a moment after you change the provider's settings or model.
  • The check reads the provider's model list and never sends a prompt, so it
    costs no tokens and does not load a model on Ollama, xOllama or opencoti.
  • The model is looked up on Ollama, xOllama, OpenAI Compatible servers,
    opencoti, LM Studio and LiteLLM, which list exactly the names they serve;
    when it is missing, the tooltip names what the server has. A llama.cpp
    server with one model loaded answers to any name, so a different name
    there is not marked red. Cloud APIs (Anthropic, OpenRouter, Gemini and
    others) accept names their lists leave out, so their model is not checked.
  • Providers that have no such list to ask (Bedrock, Vertex, Claude Code, the
    ChatGPT subscription, the Cline account) stay grey, and the tooltip says
    so. They are checked by their first request, as before.
  • The Vision and agent tabs show no dot: they keep settings of their own,
    and the check reads the main ones.
  • It is not checked while a conversation on that provider is running: the
    conversation's own requests already show whether it answers. The dot
    keeps its last answer, and the tooltip says it was not checked.
• Save model requests. A new switch in Settings > General. When it is on,
  every request sent to the model is saved exactly as sent, one file each,
  under logs/requests/<conversation>/ in Cerebriline's data folder
  (~/.cerebriline/data unless you moved it). It is for replaying a turn that went
  wrong, such as garbled text at the start of a reply, to tell the model's
  own output from a server fault. Off by default.
  • Only the request body is saved. Headers are not, so no API key or token
    is written. The body does hold the whole conversation, including the
    contents of files the model read.
  • Files are never deleted. Each one holds the whole prompt, so a long
    conversation can fill hundreds of MB. Turn it off once you have the turn.
  • On Ollama and xOllama it applies from the next request; on other
    providers, from the next conversation.

xOllama

• The context size you set is the window you get, or the most the server
  has. With "Book a context window" off, a 256,000 setting ran on about
  113,000: nothing was asked, and the server sized each request from the
  prompt and the output cap. Now a size you set is always asked for. The
  server grants it in full, or the largest window it has free, as long as the
  conversation still fits. It is no longer all or nothing.
  • "Never go below" still sets the smallest window you accept. It is now
    shown on xOllama whether or not booking is on, because it applies to any
    size asked.
  • With no size set, "Book a context window" books the model's own window,
    also in full or the most the server has. Off, the server sizes each
    request as before.
  • With booking on, the window is held for the conversation and freed when
    you close it. With booking off, each request gets its own window, freed
    as soon as the request ends, on xOllama v0.40.4-xollama.2 or later (the
    opencoti c13 r3 pin); an older xOllama holds it until the conversation is
    closed.
• A task runs on the context size the panel shows. A size typed into the
  panel reached the panel and the compaction marker, but a task read the
  size saved in the profile, and kept using it until the profile was saved
  again: a panel reading 128,000 ran a task at 256,000. The panel's setting
  now wins, and the profile fills in only what the panel leaves unset.
  • A conversation that already holds a window asks again when the size
    set has changed since. A window the server granted smaller than asked
    is still kept.
  • Done and Update now wait for the context size to be saved before they
    close or save the profile.
• A background request no longer costs the conversation its cache. A
  note that condenses capped reasoning, or a summary, used to ask for a
  window of its own beside the conversation. When the conversation held its
  whole window, the conversation was closed to make room, and its next turn
  re-read the whole prompt (57,865 tokens in one run). Now such a request
  runs inside the conversation's own window while the conversation is idle,
  and nothing is closed.

opencoti

• A tool call the server can't read no longer ends the turn with an
  error. When a model wrote a tool call with arguments the server couldn't
  parse (for example a Windows path written with single backslashes), the
  stream ended with an error and the call's text was lost. With an opencoti
  server that supports it, the server now sends the call's text instead, and
  Cerebriline tells the model its call could not be read, so it can write it
  again. The half-written call is dropped, so it isn't counted twice.

Chat

• An edit to a line range shows the line it changes. When the model
  edited part of a line, such as removing one } at a given column, the
  card read "Cerebriline wants to edit this file (LINE 90):" with nothing
  under it. An edit that added text showed only the added text. The card
  now shows the line as it stands and as the edit leaves it, so you can see
  what is removed and what is around it. The header names the column, for
  example "LINE 90, COLUMN 382". Files over 4 MB keep the old card. This
  now also covers the cards that still came up empty: a column edit that
  also names the text it removes, a deletion of matched text, and an insert
  at a column, which showed only the inserted text.
• Reasoning that looped and was removed now shows a line in the chat.
  When a turn's thinking repeated one passage over and over, it was taken
  out of the conversation and replaced by a short note, but the only sign
  was that note at the end of the thinking block. The chat now gets a line
  like the other guards, for example "the reasoning repeated one passage 181
  times — removed from the conversation". When the model is also asked to
  try a different approach (once per stretch of looping), the line says so.

Reading files

• A new task no longer inherits the last task's reads. What the model had
  read was remembered for as long as the window was open, across tasks. A
  file put back between two runs was then reported on the new task's first
  read as "changed since you last looked at it… something outside this
  session wrote to it", and edits could be accepted for lines the new task
  had never read. Each task now starts having read nothing.
• Asking for a file's original before any edit shows the file. With no
  change in progress, revision: "original" on a file nothing had written
  answered "This transaction created that file", about a file the model had
  just read, and the model went round on it. It now shows the file as it
  stands, and says that nothing has written it yet.
• Putting a file back no longer reads as someone else's change. After
  the model restored a file with restore_file, its next read said the file
  had "changed since you last looked at it… something outside this session
  wrote to it". It now reads as the model's own change, as an edit does.
• A patch the model applied no longer reads as someone else's change.
  After apply_patch, the next read said the file had "changed since you
  last looked at it… something outside this session wrote to it", and the
  next edit was refused in the same words. A patch is now recorded as the
  model's own write, the same way an edit is. When the patch moved lines,
  an edit aimed at the old line numbers is still refused, now with the real
  reason: the file's length changed.
• The bracket check's fix stays on the line it names. On a long line,
  the check could suggest moving a bracket "to column 393" of a 388-column
  line, a place that does not exist. It also never tried the end of the
  line, which on that line was the actual fix. Both are fixed: a suggested
  column is always on the line, and the end of the line is tried.
• Columns on line 1 of a file with a byte-order mark match the editor.
  Files saved "UTF-8 with BOM", common on Windows, start with an invisible
  mark. VS Code does not count it; the tools did, so every column on line 1
  was off by one from what the editor and the error list showed, and an
  edit at the column a diagnostic named landed one character early. A
  patch's context on line 1 also failed to match. The tools now leave the
  mark out, as VS Code does, and keep it in the file when they write.
• A task opened from History starts having read nothing. Only "New
  Task" cleared what the last task had read, so opening a task from History
  kept the previous task's reads, with the same false "changed since you last
  looked" on its first read.

Editing files

• Another program's change to a file is no longer slipped past the
  model. When a file changed outside the session after the model read it,
  grep and apply_patch recorded the file as freshly seen, and the next
  edit went through on lines the model had never looked at. grep now says
  the file changed and asks for a fresh read; apply_patch refuses the
  whole patch, before writing any file of it.
• A file created or moved by a patch can be edited straight away. After
  apply_patch added a file, the next edit was refused with "has not been
  read in this session", about a file the model had just written. A moved
  file now counts as read as far as its source was. A move onto a file the
  model had read no longer lets an edit land on the moved-in lines it never
  saw.
• A file the editor creates is protected like one it read. Another
  program's change to a file the model had just created went unnoticed, and
  the next edit landed on it.
• A discarded attempt no longer reads as someone else's change. When the
  change protocol put files back after a failed attempt, the next read said
  "something outside this session wrote to it".
• $ matches at the end of a line in Windows files. grep, sed and
  awk kept the carriage return at the end of each line of a CRLF file, so
  a pattern ending in $ found nothing, awk counted one character too
  many, and sed wrote lines it added with Unix endings. The bracket check
  also suggested a column one past the end of the line on these files, which
  the editor then refused.
• A column edit never cuts an emoji in two. An emoji and some other
  characters take two columns. An edit between the two halves corrupted the
  character; it is now refused, naming the columns on either side. Text
  inserted at a column takes the file's line endings.
• The editor will not overwrite what a read cut off. read_files shows
  a line only up to 2,000 characters, but an edit could reach past that
  point, and replacing the whole line erased the part never shown. Those
  edits are now refused, with the way to make them: a column inside the part
  shown, or an edit by text. An edit inside the part shown moves that limit
  with the text, so a deletion cannot bring unseen characters into reach,
  and a line shortened to what the model has seen is no longer limited.
  The same goes for grep, which prints 400 characters around a match on a
  long line: an edit by column outside those characters is refused. A line
  that read_files showed in full stays editable after a grep of it, and
  reading the line with read_files lifts the limit.
• Trying a proposed check no longer blocks the next edit. The change
  protocol runs a proposed check against the files as they were, then puts
  the edits back. The files ended as they were, but the next edit was
  refused as "changed since you last read it". Another program's change made
  before the check is still caught.
• An empty file that was read can be edited. Reading an empty file
  recorded nothing, so every edit of it was refused as unread, however often
  the model read it.

Capped thinking

• The note that replaces a capped think is longer, and written as the
  thinking it replaces. When a turn's thinking ran to its budget, the next
  request carried a short note in its place, about 450 tokens for a think of
  45,000. On some models that note sent the next turn round in a loop. The
  note is now asked for in the first person and the present continuous ("I'm
  reading game.js around line 274…"), the same voice as the replay
  compaction. It is given 1,000 to 2,000 tokens, about a tenth of the think,
  so it can carry every call the turn made with what came back, what the
  turn settled and what it ruled out. It is shorter only when the window has
  no room for that much.
  • The same applies to the note written for a reply cut off at the output
    limit.
  • If you saved your own Capped Thinking Prompt (Settings > Features), it is
    kept as you wrote it. The length is still asked for alongside it. "Load
    default" fills in the new prompt.

Plan mode

• Switching to act mode when already there no longer ends the run. If
  the mode had already been switched to act when the model called "switch to
  act mode", the call answered as if it had worked, and the run stopped with
  nothing switched. The call now reports that it is already in act mode, and
  the model carries on. The extension and the CLI now share one version of
  the tool.

Generation notices

• Notices work with opencoti's new notice lines. opencoti and xOllama
  can write a short line into the model's output as its thinking or answer
  room runs out, so the model wraps up instead of being cut off. The line is
  now a plain sentence in the model's own voice, such as "I'm writing the
  conclusion now.", and the server says exactly which line it wrote.
  Cerebriline hides that line from the chat and keeps it in the
  conversation, as before. Lines from servers that still use the
  <engine_status> form are hidden too; they used to show in the chat.
  • When you reopen a conversation, a line that matches one of the server's
    built-in notice lines, or one of yours, is hidden. Only a whole line is
    matched, never part of a sentence.
• Choose what the notices say. Settings > Features > Generation Notice
  Wording has one box for the thinking notices and one for the answer
  notices. Write one line per step, in step order. Line 1 is the first step
  (60% by default), and a blank line keeps the server's own line.
  {words} is replaced with the words left, {left} with the tokens left
  and {total} with the phase's budget. "Load default" fills in the
  server's lines to start from. It applies from the next request, on servers
  with the new notice lines; older servers use their own.
• A copy bar for each phase. Some models now and then write a notice
  line out again, word for word, sometimes many times over. With the Copy
  bar switch on for thinking or for the answer, the server stops the model
  repeating a notice line in that phase, and the model writes a short line
  of its own instead. The thinking switch is on by default, as that is
  where the repeats happen; the answer switch is off. They take effect
  from the next request, and only on servers that offer the bar (opencoti r5
  and later).