v1.5.0

• Added a promptless /api/generate request that warms and loads the model, returning done_reason:"load" instead of a 400.
• Added real quantization in /api/show model_info (general.file_type) derived from the model; the key is omitted when the quant is unknown instead of reporting a placeholder.
• Fixed a non-2xx upstream reply masquerading as a fabricated success on streaming requests; the real status and message now surface.
• Fixed non-streaming error responses with a non-JSON body collapsing to 500, losing the real status such as a 429.
• Fixed total_duration excluding the model load time on the native chat path after a cold load.
• Fixed --api-key rejecting the Anthropic SDK's x-api-key header; it is now accepted alongside Authorization: Bearer.
• Fixed /api/version inheriting the --api-key gate; the endpoint now serves without auth.
• Fixed Ollama's max thinking level erroring on LM Studio; it is now clamped to high.
• Fixed negative num_predict (-1/-2) and negative max_tokens being forwarded instead of treated as no-limit sentinels.
• Fixed strict:true being sent for a JSON schema with optional fields; it is now emitted only when every property is required.
• Fixed OpenAI response_format:{"type":"json_object"} 400ing on /v1/chat/completions; it is now translated to a json_schema envelope.
• Fixed streaming tool calls being emitted more than once; merged calls go out in one chunk and the final chunk carries no tool calls.
• Fixed same-name parallel tool call results correlating to the wrong call; the nth result now maps to the nth same-named call.
• Fixed image data: URLs being double-wrapped on the native and OpenAI-compatible chat paths.
• Fixed a failed model unload being logged as success, stacking a fresh instance on one never released during a num_ctx reload.
• Fixed /api/pull passing through LM Studio's raw downloading status; it now emits Ollama's pulling prefix.
• Fixed model resolution for multi-quant variants and KV-cache offload readback.
• Fixed /api/show hardcoding general.quantization_version to 2; the key is now omitted.
• Fixed /api/tags fabricating modified_at for orphan aliases; every entry now has the same shape.
• Fixed /api/web_fetch running HTML-to-markdown conversion on JSON or plain-text bodies; only text/html is converted now.
• Fixed /api/web_search forwarding a non-Ollama-shaped provider response; it now returns 502.
• Fixed /v1 passthrough not resolving /api/copy virtual aliases, so a copied alias 400/404d on /v1 routes.
• Changed /api/create to warn when stored inputs (template, adapters, seeded messages) can't reach LM Studio inference.
• Changed the license to dual MIT OR Apache-2.0.

What's Changed

• docs(api): refresh upstream API doc mirrors by @github-actions[bot] in #6

Full Changelog: v1.4.0...v1.5.0