• LLeMbas 2.0.0-beta.4 Pre-Release

    Homer released this 2026-11-07 22:37:29 +00:00 | -57 commits to main since this release

    LLeMbas 2.0.0-beta.4

    A beta: install by tag, by hand. No update feed offers it.

    The harness spec is 3.1.0, pinned to LLeMbas CLI v2.0.0-beta.2.

    Changed

    • A chat's system prompt is made once and sent unchanged (harness spec 3.1.0, "the prefix
      stays still"). It was composed again for every reply, so anything in it changing -- the date at
      midnight, a memory, the plan, Plan mode, a personality -- made the model's server read the whole
      chat again: on a long chat, minutes on a local model and the full input price on a cloud one.
      It is now kept with the chat and made again only after a compaction, for another model, when the
      offered tools change, or when you edit the chat's own system prompt. What changes during a chat
      reaches the model as a short note after your next message instead -- the date (on the first
      message and when it changes), a memory you edited, Plan mode switched, a plan you chose, a
      personality or custom instructions changed. The note is kept with that message and sent the same
      way from then on; it is not shown in the transcript. An instance's, a folder's or a model's
      system prompt edited during a chat applies from its next compaction or a new chat.
    • The tool list stays the same for the whole chat. python is offered from the start wherever
      Python in the browser is switched on, and says it is unavailable while no page of the chat is
      open; Plan mode no longer takes the tools that change things out of the list but refuses them
      when called, and plan_submit is offered in every mode and says so outside Plan mode; reading
      skills stays offered with none. The last round after a limit is reached, or after a plan is put
      aside, keeps the list too: it is told the tools are withdrawn, and a call in it is not run.
    • Compaction asks for its summary as the next turn of the same conversation -- the same system
      prompt, tools and messages, then the summary request -- instead of a separate request with the
      conversation pasted into it, which was a full re-read and, capped to fit, lost the middle of a
      long chat. The summary request's tokens, cached ones included, are counted in usage like a
      reply's.
    • schedule_list starts with the time it is now, since the system prompt no longer carries a
      clock.

    Fixed

    • Saving a file that is not there yet makes it, instead of answering 403 "No such file or
      directory".
    Downloads