You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Optional structured vision pre-scan 1+x (experimental, #136): add structuredVisionBootstrap (off by default). Each image task first gets one task-independent structured visual baseline, then the chat model must make at least one task-directed evidence/deepening vision-tool call before it may answer (x >= 1). The baseline exposes stable fields such as visual_kind, regions, visible_text, entities, relationships, uncertainties, and recommended_followups. When the structured flow actually needs OCR, auto mode prefers vision-model OCR to reduce noisy local OCR on Chinese/UI screenshots. The normal vision tools and default tools-first flow are unchanged.
Settings are reorganized around user intent (#139): Quick Start and the three-step first-run guide remain, while implementation vocabulary is replaced with outcome-oriented labels — the “vision backend chain” becomes “Vision model”; the first screen focuses on Auto Vision groups, vision tools, structured 1+x pre-scan, whole-turn vision routing, and the vision-model chain. Advanced controls are grouped into Performance / Compatibility / Network / Developer settings, connection testing moves next to vision-model selection, and version/diagnostics move lower. Whole-turn routing remains a normal toggle; existing config keys, defaults, provider behavior, and runtime routing semantics are unchanged.
Lighter unsaved-settings controls (#136 / #139): keep the always-reachable Unsaved / Discard / Save behavior and draft retention on failed writes, but replace the full-width sticky banner with a compact top-right sticky action group, removing the heavy blur, full-row divider, and large shadow so it interferes less with the settings content.
Windows “Open log folder” compatibility (#112): numeric explorer.exe relay exit codes are treated as a successful shell hand-off; true spawn-level failures fall back to cmd /c start, and failures from both launch paths carry a machine-readable error code. macOS/Linux keep strict native non-zero handling, with regression coverage across Windows, macOS, and Ubuntu.
dsh-web-ui / dsh-web-ui-all coexistence note: both READMEs now explain that the upstream dsh-tool-describe-image send hook may rewrite an image into a describe-image reference before Vision Router receives the original image block. Current dsh-web-ui can disable “Rewrite images to describe-image references on send” (interceptImageSend: false), read dynamically on every send with no hook reinstall or DSH restart.
Expanded design-lineage attribution: both READMEs explicitly credit Anionex/agent-vision-toolkit and dsh-vision-toolkit for influence on the deep-vision tool layer, UI-restoration/pixel-diff loop, progressive tool exposure, and parts of the tool decomposition/naming, while distinguishing the independently implemented turn-level/tools-first routing, DSH admission/wrapping, multi-backend fallback, built-in free chain, attachment/image memory, caching, and runtime resilience.
270 automated tests pass: the Node 22 / Node 24 test matrix and Ubuntu / macOS / Windows packed-host + shared-sharp regression jobs all pass, covering the settings redesign, 1+x flow, and log-folder compatibility changes.
Commits
feat: integrate #98 local vision features on current main (#141)
