v0.5.1

• Image examples and answer guidance place illustrations, captions, and source links in separate Markdown paragraphs.
• Shorter project READMEs, with detailed measurements in dated evaluation reports.

v0.5.0

• Complete Python backend source, dependency lock, deployment configuration, and tests maintained alongside the Harness plugin.
• Direct backend installation from the repository, with CI covering plugin packaging and backend behavior.
• Harness comparison with shared prompts, full answers, and timing after model warm-up.
• Updated English and Chinese documentation and a smaller plugin installation archive.

v0.4.0

• On-demand vLLM startup and GPU-memory release after an idle period, supervised independently of retrieval workers.
• Redis leases protect active requests and expire after worker loss; lifecycle operations are serialized across processes.
• GPU contention, insufficient memory and model startup failures return unavailability without publishing new retrieval tasks or exposing partial evidence.
• Shared-GPU preemption stops only the configured CiteWeave model container; unrelated processes are never terminated.
• Deployment branches, lifecycle configuration, service supervision and cold-start behavior documented. Failed plugin calls are not retried or used as answer evidence.

v0.3.0

• Durable Celery retrieval jobs with a Redis broker and PostgreSQL task results, compatible with the existing synchronous evidence API.
• Bounded queue admission, single-worker processing, task lookup and queued cancellation, persistent service volumes, and result retention cleanup.
• Pinned infrastructure images and hash-locked queue dependencies.
• Empty search results and unreadable sources return evidence diagnostics instead of a generic service error; translated discovery can fall back to the original language.
• Harness declares retrieval calls non-concurrent and shows backend error details, request IDs and queue-admission guidance without automatically resubmitting failed requests.

v0.2.0

• Optional Qwen3.5-9B community AWQ 4-bit backend for text assistance and visual descriptions on RTX 3060 12 GB.
• Reproducible model download with an immutable revision and source.json, plus a digest-pinned vLLM Compose service.
• OpenAI-compatible text and vision adapter, CPU embeddings and existing CrossEncoder, model-aware caching, and API readiness checks.
• Model-server credentials remain separate from CiteWeave workspace authentication and Harness provider configuration.
• vLLM input-token budgeting for the 8,192-token context; the original DeepSeek API behavior is retained.
• Deployment instructions, adapter regression checks and representative live evidence results. The adapter applies to a separate backend clone; the Harness tool contract is unchanged.

v0.1.2

• Illustrated examples covering geology, wildlife identification, and bridge engineering.
• Source-image attribution and licenses distributed with the documentation.
• Backend deployment guide for RTX 3060 12 GB using API text generation and local vision models.
• Documentation and example images included in the installation archive.

v0.1.1

• Plugin configuration page for the shared Harness Desktop/Web client.
• Backend address, masked access-token entry, timeout, and image-limit controls.
• Immediate application of saved configuration to new tool calls.
• Write-only token handling: blank input preserves the saved token; a separate action clears it.
• GUI installation and configuration instructions, with command-line and environment options retained.

v0.1.0

• Native DeepSeek Harness plugin bundle for 0.2.1-rc.2.
• Web evidence and sourced image retrieval through citeweave_search.
• Native image attachment storage, structured attachment references, and inline Markdown output.
• Backend authentication, configurable image limits and timeout, and retrieval diagnostics.
• English and Chinese documentation, localized plugin metadata, automated tests, and versioned release archives.