Skip to content

Latest commit

 

History

History
78 lines (61 loc) · 4.65 KB

2.0.0.md

File metadata and controls

78 lines (61 loc) · 4.65 KB

Lesscode 2.0.0

Prepared 2026-11-17. The tag v2.0.0 is pushed by the owner, after v1.2.0; the release workflow then builds, signs and publishes the archives (release signing).

Proposed tag message:

Lesscode 2.0.0: model calls move to a second binary, lesscode-model (endpoint file, waits, check, measure); regenerate run needs it and loses two flags

What is new

  • A second binary, lesscode-model. It makes every model call. It reads one endpoint file (API style, base URL, model, key, an optional reasoning level), sends the request, and waits when the endpoint refuses for a rate limit: 10 seconds first, then at least twice the wait before, up to 10 minutes in all. While it waits, and while an answer takes longer than earlier ones did, it says so on the terminal. It speaks OpenAI-compatible endpoints and Anthropic's Messages API. Guide: model-endpoint.md.
  • lesscode-model check makes one small call and says whether the endpoint is usable, and if not, what the endpoint answered.
  • lesscode-model measure runs a task set several times and reports how often the first reply passed, how often a reply passed within two repairs, and calls and output tokens per accepted answer. It ends with one of three verdicts. --quick runs 5 trials.
  • Each platform has two archives, lesscode-core-<target> and lesscode-model-<target>, each with its own license inventory. The install scripts fetch the second one when LESSCODE_WITH_MODEL=1 is set.
  • One endpoint file can serve several models: LESSCODE_MODEL_NAME and LESSCODE_MODEL_REASONING_EFFORT in the environment stand before the file.

Fixed

  • The pre-push hook printed "Pre-push verification passed." before the step guard had run, so a refusal by the guard followed that line. It is printed once, after every check has agreed.
  • A failed model call says what the endpoint answered, once.

What changed for existing users

This is a major version because two things break an existing regenerate run. If you do not use regenerate run, nothing changes for you: no other command of lesscode-core calls a model.

  • regenerate run needs lesscode-model. lesscode-core no longer holds an HTTP client. Put lesscode-model beside lesscode-core or on the search path, or name it in LESSCODE_MODEL_BIN. Without it the command ends with exit code 3 and says where it looked.
  • regenerate run refuses --effort and --thinking-display. The reasoning level comes from the endpoint file. --temperature is sent only when you give it.
  • A new receipt's turn says "mode": "seam-1" and lists under sent only the settings that were sent. elapsed_s is now seconds; model_sent, model_returned, seed and seed_note are gone. Earlier receipts keep their form and still verify.
  • The stop word of an answered turn is complete for every endpoint, where it used to be the endpoint's own word (end_turn, stop).
  • Everything else is as in 1.2.0.

What it was tried on

regenerate run completed one turn on each of six cases through the two Windows binaries of a release dry run (R113): GLM 5.3 and Mistral Large on Mistral's API, gpt-6.1-sol on OpenAI's, claude-opus-5-5 on Anthropic's, and the first two again through a local endpoint. That is one turn per case; it measures no model.

Not tried: the binaries of the other four targets were built and packed, not run. The publish stage has not run with ten archives. The install scripts' path for the second binary passed a syntax check only, because no release carried that archive yet. The last four changes before this release (the two removed flags, the reason said once, and two rules of the repair loop) are covered by scripted tests, with no live call.

What is not checked

measure tells you how often a model's replies passed the checks of a task set. It does not tell you that an accepted reply is correct beyond those checks, and five trials are a first look, not a measurement.

For the owner, before tagging

  1. Tag v1.2.0 first, at db175e5 (pipeline run 43087717445 was green there), and let it publish.
  2. Tag a commit of main for which CI is green on all five platforms.
  3. Set date-released in CITATION.cff to the day of the tag if it is not 2026-11-17.
  4. git tag -a v2.0.0 -m "<the tag message above>" and git push origin v2.0.0.
  5. Approve the release environment when the workflow asks.
  6. After publication: check signatures, checksums and the attestation for all ten archives, and that Zenodo archived it. Then run install.sh and install.ps1 once with LESSCODE_WITH_MODEL=1.