v0.1.38

Fixed

• Persisted the final model-visible image evidence across DSH Profile restarts, so historical images no longer consume vision quota again or change the main model’s cached conversation prefix.
• Preserved both successful descriptions and [vision unavailable: ...] results byte-for-byte; only requests that fail before producing any model-visible result are retried.
• Bound persisted evidence to the Session lifecycle, attachment, focus prompt, credential, and output-affecting runtime settings to prevent replay across incompatible configurations.

v0.1.37

Added

• Added a bilingual, screenshot-based AIHubMix guide covering signup through the Inferera entry, API key creation, free Gemini 3.7 Flash model selection, exact Vision Toolkit settings, and troubleshooting.

Changed

• Replaced the Groq tutorial link in Vision Settings with the AIHubMix guide and select the matching English or Chinese page from the configured vision-output language.
• Updated the English and Chinese READMEs to use the Inferera signup entry and feature the AIHubMix guide.

v0.1.36

Added

• The bundled standalone Python now downloads from the domestic mirror (Tencent Cloud COS) first and falls back to the GitHub release when the mirror is unreachable, so users in China no longer need GitHub connectivity for the first-run Python bootstrap.
• The pinned assets/python-bootstrap.json gained an optional mirrorBaseUrl, and all eight platform archives are hosted on the mirror.
• The locked runtime dependencies (Pillow, NumPy, vtracer) are installed from the Tencent Cloud PyPI mirror first and fall back to the official PyPI index.

v0.1.35

Changed

• Made the fast restore mode trigger more sensitive: a floating "快速还原为 HTML" / "快速生成" / "quick restore" control visible in the reference image now counts as a speed signal.

v0.1.34

Changed

• Transparent variant routing is now on by default: imageInputVariants.hidden defaults to true, so image-input variant routes keep the original provider and model display names and the model selector shows one entry per model out of the box. Users who prefer the explicit (Vision Toolkit) entries can disable the “透明变体路由” setting (advanced settings → image input) to restore the previous behavior.

v0.1.33

Added

• Transparent variant routing (imageInputVariants.hidden, off by default): image-input variant routes keep the original provider/model display names, and the browser hides the upstream text-only twins so the model selector shows one entry per model. Pasted images, image history, and the built-in read_image tool keep working on text-only models; opening the selector no longer flashes a duplicate group because hiding is synchronous DOM reconciliation. Disabling the setting restores the explicit (Vision Toolkit) entries.
• Settings UI: “透明变体路由” checkbox under advanced settings → image input, with bilingual copy.

Changed

• Lowered the built-in free vision service daily quota to 100 requests.

Fixed

• Toggling transparent routing is display-only: it no longer rebuilds or re-verifies the vision runtime.
• The browser display-config cache is invalidated on Settings saves, and an in-flight response can no longer repopulate it with a stale flag.
• Restoring upstream model entries after transparent routing is disabled, and guarding the selector integrator against duplicate installs.

v0.1.32

Fixed

• Fixed the compressed-image cache silently missing on Windows when cache file paths exceeded the 260-character MAX_PATH limit; cache keys now use shorter 64-bit digests and are versioned as v2, so old oversized entries are pruned automatically.
• Made the portable package verification and the test suite Windows-compatible, including npm.cmd invocation, path-separator handling, Python bootstrap fixture layout, a profile E2E prompt that avoids newline-carrying argv, and restart-helper test skips where automatic restart is intentionally unavailable.
• Routed Windows pnpm batch shims through cmd.exe so plugin updates work when the harness resolves pnpm to a pnpm.CMD path.
• Added a Windows CI job that runs the portable-package build, tests, and verification on windows-latest.
• Fixed an intermittent NO_ADAPTER failure on image-input variant routes (vision-toolkit-<provider>) after adapter re-registration, model switches, or hot reload: wrappers now survive transient registry gaps, are re-registered when the live registry drops them, and self-heal on a periodic sweep.

v0.1.31

Changed

• Renamed the bundled Skill from vision-tools to vision-skills, so the model-facing name describes the capability instead of the underlying tools. Sessions created before the rename still restore activation from legacy vision-tools history; new sessions invoke /vision-skills.

v0.1.30

Changed

• Raised the default vision operation timeout from 15 seconds to 30 seconds for both semaphore queueing and tool execution.
• Removed the practical global ceiling on the built-in free vision service (raised from 5,000 to 1,000,000,000 requests per UTC day) while keeping the per-client daily and burst quotas.

v0.1.29

Added

• Install and use with zero Python setup. When no system Python 3.11+ is available, the plugin downloads a pinned, sha256-verified standalone Python 3.13 build (about 35 MB) on first use and prepares its isolated runtime with it, so new users no longer need to install Python first. A system Python or an explicit runtime.python override still takes precedence, and a committed manifest plus scripts/python-bootstrap.mjs keeps the pinned build auditable and updatable.

Fixed

• Keep the vision_toolkit_activate bootstrap callable until the end of the model step when the Skill and the bootstrap are invoked in parallel, preventing a race that surfaced as unknown tool "vision_toolkit_activate" while the Skill call was already activating the visual tools.
• Reword the vision-tools Skill description so screenshot-to-UI restoration reliably triggers visual-tool activation.