OKLMdocumentation v1.0.3
On this page

OKLM 1.0.2 release notes

What changed between 1.0.1 (2026-11-17) and 1.0.2: the chats on a server that is not OKLM's, the packages built and published by CI, and two defects that kept 1.0.1 from building on Debian 13. What 1.0.0 holds, the platforms and the promise of compatibility are in release-notes-1.0.0.md; they stand.

Status: released on 2026-11-17, tag v1.0.2.

The number says fixes; most of this is new (options of oklm chat, fields of the chat window, functions of the library). All of it is added: nothing of the command line, of the /v1/ routes or of the settings files was removed or changed its meaning. One command that was a usage error now runs: oklm chat with neither -m nor --url takes the settings' baseUrl when there is one, as oklm agent does.

Building on Debian 13

  • Fixed: the graphical tools of 1.0.1 did not build with GCC 14 (Debian 13's): a model's name was copied into a shorter field of the chat window without its length, which that compiler refuses. make docker-build failed there for the same reason.
  • Fixed: the Windows build of 1.0.1 did not build with MinGW-w64 12 (Debian 13's), whose header already names stat _stat64.

The packages of 1.0.1 built on the release machine were not affected.

A server that is not OKLM's

The agent talked to any OpenAI-compatible server; the chats now do (roadmap C11). Every proof ran on llama.cpp's server on the machine itself and on stand-ins: the checks reach no hosted endpoint on purpose, a key and a conversation would leave the machine. A hosted server (Mistral's, another) is the user's to try.

  • oklm chat --url URL [--api-key KEY] [--remote-model NAME]: the terminal chat, line by line and full screen, on such a server. The key comes from the option, else OKLM_API_KEY, else OPENAI_API_KEY. An oklm serve tells its context; another server's is -c, then the settings' contextWindow, then 32768. What needs the model in the process says so in a line (/rag, /decide, /tool, a picture, a sound or a video), and the options that need it are refused at start (inference.md).
  • The chat window on such a server finds out once, on linking, which of OKLM's own routes the server has, and says in a line that it is not an OKLM server and what that leaves out: no figures, no memory line, no model loaded in place of its own. The context is the window's setting, which the line that warns of a cut follows; a refusal that tells the server's context is taken for it. A server that does not answer when the window links is taken for neither kind until it answers (gui.md).
  • The models a server lists (GET /v1/models) are chosen from a list: in the terminal chat at start and with /model, in the chat window's side panel and in its Connect window (List models). /model completes among them. With several models and input that is not typed, oklm chat --url refuses and names --remote-model: it does not take the first of a list.
  • Servers kept with their key: the Connect window saves a server with a name, a model and a context; it is listed in the chooser of a window started with nothing named. The key is in the tools' settings file, which is its owner's alone (mode 0600; on Windows an access list with the owner's entry alone), shown masked and written in no output. Forgetting a server removes its key.
  • A member of a request a server refuses by name (400 or 422: stream_options, chat_template_kwargs, top_k, seed...) is left out and the request sent again at once, in the chat window, oklm chat --url and oklm agent --url; a line says once which option has no effect on that server. A refusal that names none stays an error.
  • A full context there: a reply that ends without the server's count is counted by estimate, so a compaction is made by itself on every server; a cut reply is gone on with as an open message, and, where a server refuses that, closed with a request to go on.
  • A sound goes to such a server as a WAV or an MP3 (another format as the WAV of its samples); it went with its file's extension as the format's name.

Packages

  • CI builds and publishes them for a version tag: the job packages builds the .deb and the AppImage in a Debian 13 container (oklm_1.0.2-trixie_amd64.deb, oklm-1.0.2-trixie-x86_64.AppImage) and the Windows installer (oklm-1.0.2-setup.exe), with SHA256SUMS-1.0.2, and attaches them to the tag's release (build.md). This tag is the job's first run on the server; it was run in a container on the release machine before.
  • Built on the release machine as before, by make deb, make appimage and make installer-windows into dist/: oklm_1.0.2_amd64.deb, oklm-1.0.2-x86_64.AppImage, oklm-1.0.2-setup.exe.

Checks before the tag

The whole suite, the graphical tools' checks on the CPU and on the device, the checks of the chats on a server, the static analysis and the documentation's lint ran on the tree of the tag less its last commit (this status); the test programs changed since 1.0.1 ran again under ThreadSanitizer, AddressSanitizer and Wine, and the three packages were built in a Debian 13 container (roadmap.md).

Known limits

Those of 1.0.1 stand. Of these changes:

  • No published audio model was served by a server that is not OKLM's: the sound's part of a request was read on a stand-in.
  • The owner-only access list of the settings file on Windows is set by calls that succeed under Wine, which keeps no access list: Windows itself was not run.
  • On a server, oklm chat does not ask for tool calls, /decide or /rag, and sends no picture; a reply cut inside its reasoning is asked again, not gone on with.
  • The size of a conversation on a server is an estimate between the server's counts.
  • No https server and no hosted endpoint were run by the checks.