LM Studio strict chat templates

The LM Studio (local) preset enables singleLeadingSystemMessage: true for the openai-chat adapter. This combines text-only system/developer instructions into one leading system message with blank lines between messages. It supports custom model IDs, including quantization and checkpoint suffixes, without relying on the model-name compatibility selector.

For an existing renamed/custom LM Studio provider, add this boolean explicitly in the provider editor. The stored value is independent of the provider name, hostname, and port. A same-destination save preserves an omitted setting; a POST that replaces the destination must resubmit it. PATCH accepts a boolean or null to clear the override. Clearing it on the canonical LM Studio provider restores the preset default.

Set singleLeadingSystemMessage: false to disable merging for a template that accepts chronological instruction messages. Other presets keep their defaults. foldDeveloperRoleToSystem continues to control only the developer role; enabling single-message compatibility takes precedence over that role choice and the model-name-based late-instruction workaround.

Hoisting later instructions changes their chronological scope and can invalidate prompt-cache prefixes. Translated requests preserve IR instruction order: base system prompts first, followed by timeline developer instructions. The Responses parser already collects input system items into the base prompt, so their original positions relative to developer items are not recovered. Direct Chat passthrough preserves the original relative order of system/developer messages. Both paths normalize the outbound copy and keep user messages and tool pairs in order.

Non-text system/developer parts in direct Chat fail explicitly rather than being silently discarded. Translated developer documents keep their file parts in a user-compatible message at the existing timeline slot while their text joins the leading system message. Developer image/audio messages keep their existing user-compatible mapping. No LM Studio runtime is required to configure this flag; the served model’s template still determines which other wire features it accepts.