v0.3.1 — session-scoped analysis capsule

The paste-analysis progress capsule now follows its owning session: switching sessions hides it, switching back re-shows it while the analysis still runs. Previously it floated globally over every session.

v0.3.0 — intent-aware vision paste + original-image fetch

Highlights

• Intent-aware paste analysis: your composer draft steers the vision route's focus (error text for debugging, outfit details for styling) instead of a generic description
• Clean chat bubbles: bridged user messages render only your words plus thumbnails; the vision analysis lives in the thumbnail lightbox
• New vision_image_fetch tool: materialize a pasted image's durable vision-subagent:// reference into the workspace .dsh-vision/ directory — original bytes for editing or full-fidelity inspection
• maxDepth fix (from 5347596): vision subagent spawn no longer rejected by the absolute depth cap

Full details in CHANGELOG.md.

v0.2.0 — Codex-style paste UX

• Web paste bridge: paste/drop images into the composer; analyzed by the vision route on an isolated context, only text enters the conversation
• Progress capsule, thumbnail gallery, analysis lightbox for bridged user rows
• Host endpoint POST /vision-subagent/v1/web-paste with session authorization, bounded uploads, client-disconnect handling
• New config field: maxTokens

v0.1.0 — vision_agent tool + web paste bridge

• vision_agent tool: delegate image reading to a one-shot subagent on a configurable vision route (MiniMax / Kimi / any OpenAI-compatible provider)
• Image admission: extension allow-list, workspace containment, symlink rejection, byte caps
• Durable attachments via the harness attachment seam; keys stay in route credential references