# Changelog

## 2.12.0 — 2026-10-18

- Add independent automatic update checks (on by default, every 6 hours) and automatic installation (off by default), manual checks/install, SHA256 verification, package backups and rollback. Active uploads/queues/generation defer installation; never restart DSH automatically.
- Protect favorite/archive metadata: validate records, stop on corruption or unexpected loss, preserve a pre-write backup, and expose explicit recovery while retaining damaged evidence and image originals.
- Replace repeated full gallery reads with an incremental index, active-job refresh and periodic reconciliation. Revision/delta snapshots send only changed records during progress polling.
- Publish precompiled immutable market snapshots, a native format-1 index, stable-name tarball and checksums. Expand agent onboarding, deployment, presets, update and recovery instructions; replace personal deployment naming with generic examples and enforce privacy scans.
- Existing settings, presets, model paths, images and remote deployments are retained; remote server scripts and model dependencies are not silently upgraded.

## 2.11.0 — 2026-10-17

- Enable native Qwen 2.1 reference editing on Apple Silicon through MFLUX QwenImage21Edit. Reuse the configured complete vision-capable checkpoint, image resources and LoRA channel; preserve existing variation/preset semantics.
- Inspect local API and vision tensors without loading the model or downloading weights; share backend-specific reference-edit availability between creation settings, agent resources and loaded presets. Explain missing components before startup without backend fallback.
- Keep full denoising, prefix KV reuse and disabled step skipping/Turbo. Reload the model when switching native editing, standard generation or controlled inpainting. Upgrade only known idle workers to protocol v8.
- Refresh the repository landing page with a concrete native-edit example, bilingual quick start, supported-backend matrix and discoverable feature terms.

## 2.10.0 — 2026-10-17

- Add explicit native reference editing to the updated Qwen remote gateway, creation settings, agent tools and shared presets. Use Qwen 2.1 image+text conditioning and a full sampling schedule for instruction/style edits; keep all legacy variation preset semantics unchanged.
- Enforce native-edit capability checks and incompatible-setting errors. Current local MLX and older gateways fail clearly instead of silently degrading or switching backends.
- Include the 2.9 image-import improvements below. Update skill routing to choose native editing for requested style/semantic changes and avoid repeated near-identical retries.
- Verify native DSH agent selection, real 40-step remote editing, synchronized chat/creation progress and completed images, and resource-mode controls while preserving existing configuration, presets and original gallery files.

## 2.9.0 — 2026-10-17

- Add `qwen_image_import` for public image URLs, host-local image files and verified conversation image attachments. Return managed IDs for img2img, ControlNet, masks and shared presets without manual gallery imports, model startup or SSH.
- Reuse streamed size limits, image normalization and atomic resource storage. Bound downloads, check every redirect, pin validated DNS, reject private-network destinations and clean aborted partial imports. Handle Fake-IP DNS with per-request HTTPS resolution without changing proxy/DNS settings.
- Document the search → direct image URL → import → plugin generation workflow in tool schemas and the bundled skill. Keep live progress, conversation image attachments and gallery synchronization on the existing durable queue.
- Distinguish legacy partial-noise variations from native Qwen 2.1 reference editing. The updated remote gateway conditions on both image and text and samples the full schedule; expose explicit edit mode in advanced settings, agent tools and presets. Preserve old preset semantics, reject unsupported local/old-gateway requests instead of silently degrading.
- Preserve connection configuration, presets, gallery works and panel drafts. Imported reference inputs remain separate from generated artwork.

## 2.8.0 — 2026-10-17

- Add plugin-owned on-demand SSH forwarding: passive gallery/resource/preset/config reads do not connect. Keep leases through upload, queueing, sampling and saving; confirm remote model shutdown before closing only the owned SSH process.
- Show the shared idle timeout near the top of settings in both local and remote modes, persist it to the Qwen gateway for independent crash/sleep cleanup, and retain the original configuration field for compatibility.
- Add `qwen_image_connection` for revision-checked live local/remote switching and idle-time edits through DSH's settings service; prevent switching during active jobs without touching credentials, presets or images.
- Document content-addressed LoRA transfer and backend-specific seed reproducibility. No change to model weights, quantization, sampler or quality parameters.

## 2.7.0 — 2026-10-17

- Add a private ComfyUI Qwen 2.1 gateway with native Union control/inpainting and checked LoRA loading, content-addressed streaming assets and an idempotent task protocol.
- Enable remote advanced creation and unchanged preset references in both panel and agent tools after capability negotiation; retain ordinary OpenAI Images compatibility.
- Synchronize real sampling steps, transfer/loading phases, cancellation, saved PNGs and reference comparisons with chat and the gallery. Add remote memory release without touching other GPU services.
- Preserve existing configuration, preset IDs/defaults and gallery data; extend skill instructions with remote capability discovery, automatic resource sync and supported schedulers.

## 2.6.0 — 2026-10-16

- Show resolved connection values and their inheritance source; save edited settings atomically without pinning defaults. Keep secrets masked and report running worker paths separately.
- Add a reusable resource catalog covering imported inputs, default/custom LoRA folders and gallery images, without starting inference.
- Add full creation presets with rename, default, delete/undo and portable JSON import/export. Match references by streamed SHA-256; report missing resources instead of silently dropping effects.
- Add negative prompt, CFG and scheduler controls with backend-specific capability validation and full parameter reuse.
- Simplify advanced settings into aligned sections with contextual help, explicit preset reload and refreshed imported-resource lists.
- Expose resource discovery and shared preset management as agent tools; document configuration-only intent, full settings replacement, revision conflicts and the generate/load/save workflow.
- Use worker v7 for aligned LoRA roots and runtime configuration reporting. Preserve existing settings, image files and job records.

## 2.5.0 — 2026-10-16

- Bring LoRA, ControlNet and masked inpainting into the creation page advanced settings, with native file pickers, previews, per-LoRA toggles and weights, and persistent draft selections.
- Import selected bytes into private local assets: normalize PNG/JPEG/WebP without loading MLX, stream LoRA uploads with content deduplication, and keep inputs out of the artwork gallery. No arbitrary local-path reading endpoint.
- Snapshot imported image inputs before task admission; preserve advanced parameters when reusing results. Validate incomplete combinations and remote-mode limits before submission.
- Use worker protocol v6 for imported inputs; gracefully replace known idle older workers without interrupting another session’s active generation.

## 2.4.0 — 2026-10-16

- Unify the DSH skill catalog and instructions from one bundled source. Image requests use the plugin queue; prompt-only requests do not submit jobs, and failures do not silently fall back to independent scripts.
- Add read-only `qwen_image_status` to recover an accepted task and its visible conversation attachment without starting inference or creating another gallery item.
- Read exact task records for conversation polling, tool waits and image lookup instead of scanning every gallery image.
- Follow jobs that finish while the creation page is unmounted; respect cleared previews. Preserve HTTP status when old hosts return non-JSON errors.
- Complete the earlier local 2.3.0 tool-only experiment: validate and dispatch LoRA, ControlNet and mask parameters through the persistent worker, with input snapshots, allowed LoRA directory, model cache isolation and exact preservation of black mask pixels at output size.
- Reject incompatible parameter combinations and unsupported remote advanced requests instead of silently generating a plain image. Advance the local worker protocol to v5 so an old worker cannot acknowledge unsupported parameters.


## 2.2.0 — 2026-10-16

First independently maintained public release as `dsh-qwen-studio` (display name: 图像).

- Direct creation workspace in the left navigation, without a chat-model intermediary.
- In-place progress and results, reference-based variations, parameter recall and removable preview history.
- Persistent local jobs with idempotent receipts, serialized inference, cancellation and interrupted-job reporting.
- On-demand MFLUX worker with configurable idle shutdown and manual unload.
- Privacy gallery enabled by default: unstarred works stay private; starred works appear in the normal gallery.
- Batch favorites/archive/trash/restore and confirmed permanent deletion.
- DSH theme/font integration, compact single-row navigation, independently scrolling editor and visible submit footer.
- Step/seed help and a separate advanced settings page; portal tooltips avoid scroll-container clipping.
- Creation preview follows new active conversation jobs; conversation progress cards track exact invocation receipts and retire on result.
- Model-invocable `qwen_generate_image`, automatically registered `qwen-image` skill and conversation image cards.
- Local negative prompt, true CFG and scheduler tool parameters recovered from the previously applied runtime patch.
- Self-hosted OpenAI-compatible Images adapter; unsupported local-only parameters fail explicitly in remote mode.
- Independent package/repository metadata, reproducible dependency lock, CI, release tarball and package-content checks.

Earlier local builds used `2.1.0-local.3`; the package ID, configuration keys and data format remain compatible.