v0.3.3 - Cached vision model discovery

Improves provider discovery performance with a 30-second cache shared across model listing and routing, including in-flight request deduplication. Documents the direct fallback's retry, middleware, and token-accounting limitations. Also synchronizes the runtime User-Agent version with package.json and strengthens release validation. All 17 tests pass on GitHub Actions.

v0.3.2 - Linux timeout fix

Fixes the vision timeout implementation so stalled registered sidecars reliably time out on Linux and GitHub Actions. Removes the unreferenced timer behavior that allowed Node.js to exit before the timeout fired. All 16 tests and package validation pass.

v0.2.0 - Composite DeepSeek Vision

English: Replaces the standalone vision model with one selectable DeepSeek V4 Flash + Vision composite model. Text goes directly to DeepSeek; image messages are analyzed by a hidden vision sidecar first. Existing vision-openai/GLM routes are reused automatically during migration. 中文：将原先独立的视觉模型改为一个可直接选择的 DeepSeek V4 Flash + Vision 组合模型。纯文字直达 DeepSeek，图片先由内部视觉模型识别；升级时会自动复用已有的 vision-openai/GLM 配置。

v0.1.0 - Initial release

Initial public release of the config-only DeepSeek Harness vision provider bundle. Includes English and Simplified Chinese documentation, OpenAI-compatible vision defaults, environment overrides, local placeholder-auth support, MIT licensing, and automated package validation.