Custom connection reliability (LM Studio and other local/custom
endpoints)
Several fixes for the Custom connection mode, surfaced while testing
against LM Studio's OpenAI-compatible and Anthropic-compatible
endpoints:
Fixed an HTTP 400 (parameters.properties required) that broke every
custom OpenAI-compatible connection with tool calling enabled —
no-argument tool schemas now always include an explicit properties: {}
instead of omitting it.
The base URL is now used exactly as entered for both OpenAI-compatible
and Anthropic-compatible custom connections — removed the automatic
/v1 guessing/insertion that produced inconsistent behavior depending
on whether you included /v1 yourself.
Fixed a crash (AttributeError: 'AnthropicHttpClient' object has no
attribute 'config') that broke every Anthropic-wire-format connection,
built-in or custom.
Fixed a hang on Anthropic-compatible custom endpoints (e.g. LM Studio)
caused by unconditionally sending a thinking field the server didn't
support — it's now only sent when explicitly configured.
Smarter max output tokens
Replaced the flat 16384-token output cap with per-model limits: known
providers (OpenAI, Anthropic, Gemini, DeepSeek, etc.) use their real
documented caps, and unrecognized/custom models (local connections)
scale with the guessed context window instead.
yes
ultramenid
2026-07-08T23:57:45.189192+00:00
3.22.0
4.99.0
None
no
Plugin Tags