Skip to content

Releases

v0.6.4 — idle images are evicted

Choose a tag to compare

@ChasLui ChasLui released this 10 Nov 15:25

A free-lane Lisp image is closed after 10 idle minutes (NODECODE_MAGPIE_IDLE_MS, 0 disables). A running eval is never swept, and definitions replay from the history file on the next call.

v0.6.3 — tool_choice:none keeps the free image

Choose a tag to compare

@ChasLui ChasLui released this 10 Nov 14:24

A tool_choice:"none" free request now runs the eval image as well; client tools are refused inside the loop (the model gets a tool error and answers in text), so the reply has no tool_calls. A max_tokens below 256 still uses the transcript bridge.

v0.6.2 — eval ids stay addressable

Choose a tag to compare

@ChasLui ChasLui released this 10 Nov 06:50

Eval ids keep increasing across plugin-host restarts (a new image starts after the highest id in the log), and (magpie-eval-status N) for an eval from an earlier image answers from the log instead of reporting an unknown eval.

v0.6.1 — vision cache and retry

Choose a tag to compare

@ChasLui ChasLui released this 10 Nov 06:11

look now reuses the description of an unchanged file with the same question (rewriting the file or asking a different question describes it again), and a briefly-restarting gateway route (dropped connection, 429, 5xx) is retried once before the plugin moves to the next vision model.

v0.6.0 — background eval log

Choose a tag to compare

@ChasLui ChasLui released this 10 Nov 05:52

A background eval that outlives its yield window now appends its exit to a per-session log, so a rebuilt image can read it with (magpie-eval-log) after a plugin-host restart.

v0.5.3 — deepseek model aliases

Choose a tag to compare

@ChasLui ChasLui released this 10 Nov 05:15

deepseek-v4.1-flash and deepseek-flash are now selectable model ids on both providers and are rewritten to nodecode-cyber-v1 before the vendor sees the request.

v0.5.2 — tool-less free requests reach the eval image

Choose a tag to compare

@ChasLui ChasLui released this 09 Nov 14:39

The free lane no longer needs the client to declare tools: any request with a usable reply budget gets the NodeCode eval image (Lisp, sh, look, files), while tool_choice:none or max_tokens < 256 stays on the transcript bridge.

v0.5.1 — native Lisp look inside the eval image

Choose a tag to compare

@ChasLui ChasLui released this 09 Nov 14:06

The free lane now serves NodeCode's own (look ...) from inside the eval image over a private Unix socket, so a Lisp form can read an image directly (and nested), not only through the old client carrier. Definition history and the 0.5.0 improvements are unchanged.

Verified: npm test 64/64; gateway free lane returned XBANANA-COMET-31415 from (concatenate 'string "X" (look ...)).

v0.5.0 — free-lane definition history

Choose a tag to compare

@ChasLui ChasLui released this 09 Nov 12:22

Free-lane definitions now survive a plugin-host restart. NodeCode's CLI keeps
defun and friends in a layer; the plugin now keeps the same definition forms
in a per-session history file and replays them into a fresh nodecode trial
image.

  • History lives in ~/.nodecode-magpie/history/<hash>.jsonl; override the
    directory with NODECODE_MAGPIE_HISTORY
  • Only definition heads are replayed (defun, defvar, defparameter,
    defconstant, defmacro, defmethod, defclass, defstruct, deftype,
    defpackage, define-*), so shell commands and other side effects never run
    twice
  • Values, * / / history and background results remain per-image
  • 500-form cap per session, file compacted to the last 1000 lines
  • Includes 63 tests

v0.4.2 — native image input for the free lane

Choose a tag to compare

@ChasLui ChasLui released this 09 Nov 11:54

Free lane now accepts OpenAI image_url / input_image parts. The plugin
describes each image through magpie's vision model before calling NodeCode, so
both the eval lane and the no-tools transcript lane receive text.

  • Supports data URLs, http(s) URLs, file:// and workspace-relative paths
  • Up to four images per request, 5 MiB each
  • Failed descriptions degrade to [image unavailable: …] and the request continues
  • Paid requests keep their native image parts unchanged
  • Vision models fall back across group/auto-deepseek-flash, auto/gpt-6.1-sol
    and auto/gemini-3.8-flash
  • Includes 61 tests