v4.100.255

Agents

• No more back and forth over check_file at the end of a run. A model
  that had finished was asked to check files it had already checked, and
  files it had deleted.
  • A command that fails already runs the checker on the file it names, and
    that now counts as checking it.
  • Files that are no longer on disk are left out of the reminder, and
    check_file now says a missing file is missing instead of calling it
    clean.
  • The reminder is marked as coming from Cerebriline. Models answered it as
    if you had asked ("The user is asking me to run check_file").
• A finished report is no longer quoted back as a promise. "This lets the
  player…" was read as "let's", and the reply was nudged as if the model had
  announced work and stopped. The nudge also no longer says "for the second
  turn running" when it is the first.

Compaction

• Summaries cite the right tool calls in small windows. The numbered
  list of calls the summary writer cites against was cut to the ledger's
  budget, which at a 32k window dropped the oldest calls of the stretch
  being summarised and renumbered the rest. Summaries then cited by
  counting, and the calls placed under their sentences were the wrong ones.
  The list now holds every call of the stretch; long inputs are shortened
  to keep it small. What is carried to the next summary is unchanged.

History

• Adding a tag no longer crashes the extension. Saving a tag or a
  favourite reloaded the whole history by opening every conversation's
  messages, and on a large history the extension ran out of memory. Tags
  added while it was stuck were lost. It now reloads the list from the
  saved summaries only.
• The home list shows a conversation's error count, as the full
  history already did.