2.3.0

A reworked chat composer, especially on mobile, agents that keep their memory and skills up to date, and release notes inside the plugin.

New

• Chat: attachments as preview tiles. Images show a thumbnail and PDFs their first page; other files show a card. Click a tile to open the file, or use its corner × to remove it (#526).
• Chat: Take photo on mobile. The attach menu has a Take photo row that goes straight to the camera (#520).
• Chat: one jump-to-bottom button replaces the message-navigation arrows. It appears once you scroll away from the end, including while a reply streams below the fold (#522, #523, #527, #528).
• What's new, in the plugin. After an update, a notice opens a What's new tab with the notes for every release since the one you last saw. The Show release notes command and a row in Settings → Troubleshooting open it any time (#533).
• Update notices. Once a day the plugin asks GitHub whether a newer version exists and says so, with a link to Community plugins. It sends nothing about you or your vault, stays quiet while Obsidian's own automatic update check is on, and can be turned off under Settings → General → Privacy (#534).

Improvements

• Chat: the composer grows as you type. It expands up to two thirds of the chat pane before it scrolls, so the separate fullscreen editor is gone. Enter on a list line continues the list; Enter on an empty item ends it (#521).
• Chat: the mobile composer keeps up with the keyboard, even in very large chats. It now slides with the keyboard as it opens and closes, and Obsidian's formatting toolbar is hidden while you type in the chat (#524, #525).
• Agents: memory in view. The system prompt now carries an index of your memory notes (each note's description, or the whole note when it's marked always: true), so the agent knows what it remembers without searching. Guidance now routes user facts to memory and task lessons into the skill that was used (#514).
• Agents: skills are revised in place. manage_skills gains a patch operation that changes one passage instead of rewriting the whole skill, and it only edits a skill the agent has loaded in that chat (#515).
• Agents: skill writing is on by default. manage_skills now ships enabled, so the agent can put what it learned into the skill it used; turn it off per agent in the Tools modal (#517).
• Agents: skills are deleted only when you ask. The agent removes a skill only when you ask for that skill to be deleted, whoever wrote it, and never to tidy up on its own. Core skills can't be deleted (#536).
• Agents: fewer narrating sentences. The agent explains a tool call only when that tells you something new, and skips it for a single obvious lookup (#518, #519).
• Agents: lighter vault listings. list_directory at the root returns a folder overview with file counts instead of every file, and every listing stays within a size budget. The explore-vault and Dataview skills now lean on tags, properties and queries first (#513).
• Settings: the prompt comparison dialog only compares. It shows your version against the default, with Open note and Use default; you edit the real note (#529).

Fixes

• Chat: chats whose first turn failed get a proper title. They are named New Chat (failed) and titled on their next successful turn, so new chats stop piling up as New Chat (2), (3), … (#530).
• Chat: stream start-up failures fall back gracefully, and very large tool outputs are capped in their preview instead of stalling the chat (#510, by @AdamsGH).
• Settings: the prompt "Review" notice clears as soon as you choose Use default, instead of staying until a reload (#529).

Thanks to @AdamsGH for the first correct diagnosis of the cross-realm AbortSignal failure and for #510.