Skip to content

Releases

1.6.1

Choose a tag to compare

@github-actions github-actions released this 18 Nov 12:47

Added

  • Claude Code's header at the top of the conversation view, as its terminal draws it: the mascot, version, model and effort, your plan, the agent profile, and the folder, ending in /rc, green while Remote Control is on and gray while it's off, which opens Claude's Remote Control panel when clicked. A session that has just started shows it with a Working row while its opening is on the way, instead of saying there's no transcript yet.
  • Messages from other sessions show in the conversation view, each as a card like a sub-agent's: an envelope, "Message from" and the session's name, its first line and time, and the whole message when opened. A message that came while the session was idle shows too, and is now in the transcript note. A message this session sends stays a tool line, as in the terminal.
  • File paths in Claude's replies are links in the conversation view, as they are in the terminal: a file that exists, or a note in the vault by its path or name, opens on a click or Cmd+click, as "Open links with" is set. Paths in code blocks stay text.
  • Check and repair older transcript notes. Notes written before this release can be missing messages sent while Claude was working. "Check transcript notes for missing messages" changes nothing and writes a report of which Claude Code notes are missing them; "Repair transcript notes (put missing messages in place)" puts each back where it belongs and reports what changed. A copy of every note it changes is kept first, in a hidden .backup folder inside the transcripts folder.

Changed

  • No agent profile is now a switch of its own, off by default, on the Agent profiles page. Off, every session starts with a profile (the first one, unless you pick another default), sessions with no profile are listed under that default, and the sidebar's switch and Change agent profile… offer only your profiles. On, it works as in 1.6.0: the default can be No agent profile, and the sidebar's switch offers it while it is. If you used No agent profile in 1.6.0, turn this on to get it back.
  • The strip's ticket icon is back before the session's name: a click opens the session's ticket (or the note or folder it started from; a menu when it has several tickets), and a right-click lists the same places to open as the name's menu: the ticket, note, or folder, then its overviews and reads.
  • The strip's view switch has the terminal on the left and the conversation on the right; in 1.6.0 it was the other way around.
  • Renaming a ticket renames the sessions started from it that are open in a tab and still carry its old name: in the tab, the sidebar, and, for a running Claude session, Claude's own name for it. A name you gave a session yourself stays.
  • While Claude works, the send button is Stop only while the box is empty: type, and it's the send arrow again (the message is queued, as in Claude Code); clear the box, and it's Stop.
  • The folded tool line sits evenly between paragraphs, in a fainter gray, as in the Claude app.
  • One copy button and time per response, under Claude's last paragraph of text, as in the Claude app, not one under each part; it copies the whole response's text and shows on hover once the response has ended.
  • Images in the conversation view sit where the Claude app puts them. Images you send or paste stand above your message rather than inside it, and an image Claude reads is a small thumbnail at the end of the tool line, which stays one line.

Fixed

  • Slash and ! commands in the conversation view show as you typed them (/wrap, /model claude-sonnet-4-6, ! ls), not as Claude's tags; a built-in command's output ("Set model to …") is a quiet line, and a ! command's output folds like a tool line. Transcript notes get the same from now on.
  • A message you cancel as it sends, or rewind past, leaves the conversation view when the next one replaces it, as it leaves the terminal.
  • A file path longer than the terminal's width is a link. Claude cuts a path too long for its row in the middle, at the last column, and the terminal now reads it whole, as it already did web addresses.
  • A long message no longer shows Claude's pasted-content tags in the conversation view or its transcript note, and no longer stays grayed out as unsent.
  • A message you send while Claude is working is in its transcript note, and the conversation view shows it where Claude took it rather than leaving it grayed out at the bottom as unsent. Transcript notes also keep messages from sub-agents and other sessions, labeled with who sent them, and a message that was only an image. A background sub-agent that finishes while Claude is working ends its card. From now on, a note missing a message gets it in its place when it's next written, not at the end, with a copy of the note kept in the same .backup folder first.

1.6.0

Choose a tag to compare

@github-actions github-actions released this 18 Nov 02:32

The conversation view and pop-out mode release. The conversation view shows a Claude Code session as a chat interface rather than a terminal window: the terminal still runs underneath, but the conversation view makes it easier to interact with Claude. Pop-out mode puts Agent Manager's sidebar and tabs in a window of its own, so the main window is left for your Obsidian notes. Also a color for each agent profile, searchable settings on pages of their own, richer transcript notes, clearer notices, and one kind of scheduled run.

Added

  • Conversation view (experimental, off by default: "Conversation view (experimental)" in settings, on the Sessions page). A Claude Code session opens in a view of its conversation, and the strip's Terminal/Conversation switch goes between the two. The terminal keeps running underneath, so switching is instant and nothing is lost, and an approval can be answered in either place. With it on, the strip's buttons fold into a ⋯ menu in both views.
    • The conversation reads like a messenger: your messages in bubbles on the right, replies rendered, tool runs folded into a pill you can open. Each message shows when it was sent and a copy button on hover, and a button jumps back to the latest message. Cmd+F finds text in the view, with Search earlier turns for a long session.
    • The composer follows the Claude app's: a + for files and common commands, the text (Enter sends, Shift+Enter adds a line), and a send button that becomes Stop while Claude works. Files you add, or drop on it or anywhere on the conversation, show as chips and are copied into an attachments folder when you send, their paths added to your message: set one first (a setting, and one per agent profile; there's no default). A note dragged from Obsidian's file list is mentioned with @ instead. Paste an image, type @ to mention a vault file, or use + to mention this session's ticket or the note you last had open. Esc stops Claude and ↑ brings back messages you sent from the box.
    • Approvals, Claude's questions, and plans as cards you answer in the view. A plan can be approved into auto mode, ask before edits, or accept edits, or sent back with what to change. A Working row shows while Claude works, and a note says when Claude is asking something only the terminal can show.
    • The model and effort under the composer: pick either or both, and nothing changes until you send, when Agent Manager resumes the session once with them, for this session only (Claude's own /model and /effort change your default for every session).
    • Images, as in the Claude app. Images Claude reads, a tool's screenshots included, show as thumbnails under the tools that read them; images you send or paste show larger in your message; a reply's image of a vault file renders (anything from the web stays a link). Click any of them for a full-size view; Esc closes it.
    • Sub-agents as cards. Each sub-agent Claude starts gets a card labeled with its type in words: what it was asked, running or done and how long, its latest step while it works, and its own report when it ends. Open it for its steps; with transcript notes on, it links the sub-agent's own note. One waiting on your approval says "Waiting for you"; one that ended with its session reads "Stopped". While any are running, "N sub-agents running" under the box lists each one's task, time so far, and latest step, and goes to its card.
    • Stream replies (experimental, off by default): Claude's replies appear as it writes them, a line (for prose, a paragraph) at a time.
    • Which view a session opens in is a setting (a Claude tab, with Local server on, then reopens in the view it was last in), and "Switch the session between the terminal and the conversation view" is a command you can give a hotkey.
    • Codex and Antigravity sessions get the view to read: the conversation, tool runs folded, and the Working row, with the message box greyed out and a note to type in the terminal.
  • Pop-out mode. A button at the top right of the sidebar (or the command "Open in pop-out mode") opens an Obsidian window with a sidebar of its own and the main area's session and terminal tabs at its right, so the main window is left to your notes. Sessions keep running through the move. The main window keeps its sidebar, so you can find or start a session from either; while the window is open, new sessions, resumes, Daily Chat, scheduled runs, and terminals open there, a note opened from it opens in the main window, and the main window's button brings the window forward. The same spot in the window's sidebar (or "Dock in the main window") puts everything back, and so does closing the window, rather than closing its sessions. Dragging the sidebar into a window of your own by hand does the same, or from one window into another. The main window gives up the sessions' width when they leave and gets that back when they return, as far as the screen allows (not in full screen, where it can't change size). "Show the pop-out button" in settings hides the buttons; the commands work either way.
  • Local server (off by default; turned on and off with the conversation view, and its own switch, just above it). A small server on this computer only (127.0.0.1, a port the system picks, a secret key new each time Obsidian starts) that Claude Code's hooks report to as a session works. The conversation view needs it to be typed in and for approvals, questions, and streaming; off, the view is for reading. It never reaches the internet and costs nothing while idle. Switching it off applies to sessions started or resumed after; the listener itself closes when Obsidian restarts or the plugin reloads.
  • Sessions told about each other. With Local server on, a Claude session gets one line with the message you send when another session is in the middle of a turn in the same code repo, naming it, so it checks with that session before building or committing there. Nothing is added when no other session is working there, or for one it was told about recently.
  • "Needs you" in the sidebar. A session asking you something (an approval or a question) has an orange dot on its row, and the sidebar's header shows how many are asking; click it to go to the first.
  • Keep pane widths when resizing (off by default). Obsidian shares the width of side-by-side panes by percentage, so dragging a sidebar's edge or resizing the window grew or shrank every pane at once. With this on, only the pane next to the edge that moved changes width, in the main area and in pop-out windows, panes split inside a pane included, and in pop-out mode Agent Manager's sidebar keeps its width. Dragging the line between two panes works as it always has, and sets the width that's kept.
  • A color for each agent profile (off by default: Profile colors). Each profile gets a color of its own (the first is Obsidian's default purple), which you can change in its settings, with white or black text for a light one. A line in that color floats just above its sessions' tabs, a divider marks where one profile's tabs end and the next's begin, and its name on a session's strip and the sidebar's switch take the color.
  • Each agent profile has its own Daily Chat button and custom button. A profile's page has a "Daily chat" toggle (on by default), the button's text, and the chat's name, which replace the Sidebar page's while profiles are on. Turn on "Custom button" on a profile's page and give it a note, a base, or a link to open, and a button for it shows under the sidebar's switch while that profile is picked, so each profile can have its own dashboard.
  • Pinned can be turned off, on the Sidebar page, or per profile (on by default). Off, pinned sessions are listed in Recent and the menu has no Pin.
  • More of each session in its transcript note. Properties for when the session started and last had a turn, the model and effort most of its replies ran on, its tokens (input, cache reads, and output), the CLI's version, and the folder it ran in, kept current as turns land. Each of your messages shows its time, and a reply is marked with its model and effort wherever they change. Claude Code and Codex record all of these; Antigravity records the times and tokens. Notes already written get the properties at their next turn.
  • Keep tool commands (off by default, on the front page of settings): each shell command a Claude Code or Codex session runs is kept in its transcript note, folded under the tool's line and cut at 40 lines or 4,000 characters, with recognizable keys, tokens, and passwords masked.
  • Cmd+F in the transcript view finds text in the conversation, with Search earlier turns for a long session, whether or not the conversation view is on.
  • Recent notices. "Show recent notices", a command and an item in the menu of the sidebar's tab icon, lists the last 50 notices (copies aside) in full, with their times, so one that went while you looked away isn't lost. A row about a scheduled task opens it, and one about a session goes to its tab while it's open, for notices since the last reload. "Keep recent notices" (on by default) turns it off.
  • Scheduled tasks get their own switches. shell: false takes a run's shell away, chrome: on lets it use Claude in Chrome (with the web fully open), own-folder-only: true keeps it writing only in its own folder (it needs the web off or a list of sites, or no shell), and obsidian-cli: [unresolved, orphans] has Agent Manager answer those questions from Obsidian's link index before each run. The New Scheduled Task card has a row for each; its Web is a switch, and Reads first, the vault queries, and Refuse these tools (deny) are in Advanced Options.
  • Edit a scheduled task in a popup. Edit, on a task's page and in its approval, opens the card's rows filled in from the note, and saves only what you changed.

Changed

  • **Settings are searchable, and split...
Read more

1.5.0

Choose a tag to compare

@github-actions github-actions released this 11 Nov 02:30

The scheduled tasks release: prompts that run on a schedule, as a full Claude Code session or locked down without one, with an approval for what each can reach. Also underlined links, notifications sorted by what they mean, pictures Claude reads, a plain terminal, and selection with Shift+arrows in Claude's input.

Added

  • Scheduled tasks (off by default: "Use scheduled tasks" in settings; Claude Code only, for now). A scheduled task is a note with type: scheduled task in its properties, in a folder of its own named like it. Its properties say when it runs (schedule, in plain words such as Weekdays 7:00, or a date), how it runs (mode), and what it starts with (reads), with the prompt under ## Prompt; the template opens with a ## Human Notes section for your own notes on it, which a run is given after its prompt and which are approved with it. Its project is where it sits, as a ticket's is: with Use overviews on, the nearest folder above it with an overview (else its own folder), so one in <project>/Scheduled Tasks/<Name>/ moves and archives with its project. Names needn't be unique; while two share one, notices, notifications, commits, tabs, and lists add where each sits. One in your Scheduled Tasks folder that sits in no agent profile's folder runs as no agent profile; elsewhere a task runs as the profile whose folder holds it, or the Default agent profile when none does.
  • The Scheduled Tasks row and page. A row under the Daily Chat, above New Session and New Ticket, shows the next run (or "Running now…" while one runs), how many results you haven't opened ("3 outputs"), and how many tasks are pending. It opens a page that says what runs next and whether this computer runs the schedule, then lists every task, with the agent profile that runs it, under Active, Pending (to fix, to approve, or held), and Paused (folded). Right-clicking a task offers Open note, Run now (Stop while it runs), Start session, Approve…, Pause or Unpause, and Delete task (after asking; its folder is deleted the way Obsidian's Deleted files setting says). A task's own page shows its schedule, what's wrong with it, every folder, site, and tool a run could reach, what no run ever reads, its open follow-up ticket, and its runs, with Run now (Stop while it runs), Start session, Edit (which opens the note), and Pause or Unpause. A task's note has Run now in its header, and Stop while it runs; a Running pill at the top of the page opens the running task. An op=task link opens a task's page; no link starts a run.
  • The New Scheduled Task card, from the page's New Scheduled Task row: a name, the project, how it runs, the notes and folders it starts with (Reads first, picked from the vault as chips, written as links that follow a rename), what a run can use for that mode, the schedule (repeat, days and time, or a date), and the prompt, with Each run (a new session or continue the last), the model (Fable among them), a follow-up ticket, and Only report when something's new under a folded Advanced Options. It files the task in the chosen project's Scheduled Tasks/ and opens its page to approve.
  • Three ways to run. Full session, the default, is a real Claude Code session in a background tab, listed in the sidebar like any session, in auto mode with your own settings, hooks, skills, MCP servers, and connectors. Headless (Local Only) and Headless (Web Access) run Claude Code without a session tab, locked down to Read, Glob, Grep, and Write (plus the web tools when the web is on, and Edit, NotebookEdit, or TodoWrite when its tools names them), with no MCP servers, connectors, or hooks, writing only into its own folder. A task with no mode is headless. The card writes mode: full or mode: headless, with web off, search, or a list of sites.
  • Full session runs. The opening is typed in once the session has settled, so connectors are up, and the run is told its project's overviews (broadest first), its repo, and its reads, in its opening and pinned to its system prompt, through a compaction too. It keeps your AGENTS.md, with a paragraph in its system prompt that puts its opening ahead of the parts written for interactive sessions, telling a run with a shell to commit nothing, your work in progress included. It's watched to the end: done, finished without a result, waiting (it stopped to ask: " is waiting for you"), or stalled past the time limit, which is left open and never ended. Only the end of Claude's turn counts as finished; a tab closed or a session ended mid-turn is logged interrupted, and Stop interrupts the turn and leaves the tab open. It's never retried, a waiting or stalled run doesn't take a runs-at-once slot, and a stalled run lets the Mac sleep. Its tab stays open afterward, or closes after a clean run with tab: close; a continued task closes its last run's tab first when that tab is idle and untouched, and otherwise waits.
  • What a full session run can't do. It refuses connectors' sending, replying, forwarding, sharing, scheduling, and deleting (drafts, labels, and reads stay, apart from Slack's draft tool: a run's Slack drafts go in its result), Claude in Chrome, and the Keychain's password commands. Connectors are off unless a task sets connectors: on, the card starts a full session task with web: off, and a task can add deny: (Bash, WebFetch, WebSearch, Write, Edit, NotebookEdit, Agent, Skill). web is off, search, or a list of sites, held by a hook on fetches (which refuses an address with a backslash or a user name before its host) and by the shell's sandbox, held strictly to the list; since Claude Code adds your own settings' sandbox exceptions to it rather than letting a run clear them, a run with a site list or no web and a shell is refused while your user or managed settings widen the sandbox (excluded commands, allowed domains, local binding, and the like), and says which. Nothing is typed into a run until its guards' hooks have reported in at the session's start, so one whose hooks don't run (hooks turned off, or managed settings allowing only their own) is refused having been given nothing, and a session that didn't start in auto mode has its first prompt refused; the guard is written inline in each hook, never in a file a run could change. shell, file-access, writes, or tools with a value on a full session task (false included; a blank one is ignored) is a problem on it. The README says plainly what full session mode doesn't stop: it uses your normal Claude login, a reload or restart ends a run, the send rules match tool names, and the site list doesn't reach other MCP servers' tools.
  • An action log for every full session run: Actions.md in the run's folder lists every tool call, its own and its sub-agents', with flags worked out by rule: writes outside the task's own folder, shell commands that may write outside it, git push, the security command, network commands, connector calls that aren't reads, fetches, shell hosts, and connector calls naming hosts off the site list, starting another agent's CLI (codex, claude, agy, gemini, and others, whose work the guards can't see inside), a missing transcript, and a run far longer than usual. A continued run's log covers only that run, and tokens are blanked in it. A flagged run says so in its run log row; the sharp flags (a push, the Keychain, a connector call that isn't a read, a network command, anything off the site list, web use while web is off, a missing transcript, and an action log that couldn't be written) are also a "needs you" notification. A headless run with the web is flagged for fetches off its site list too.
  • What a headless run reads. Its own folder and its reads always; beyond them, file-access says how far its file tools reach (a sandboxed shell, below, reads the rest of the Mac apart from the never-read list): none, project, vault, disk, or folders outside the vault (~/ or /, only with the web off, each checked on this computer: it exists, it's a folder, it isn't the whole disk or home, the vault or holding it, or on or leading to the never-read list; shown by its real path in the approval, and checked again before each run). A vault folder or note in file-access is a problem saying to move it to reads, and a reads entry that names nothing is a problem too. Each note in reads is given to the run as text. It never reads other scheduled tasks' Outputs/ or Memory.md unless its reads names that task or one of its results, so a web task's findings reach the rest of a project only where something asks for them. No headless run loads your AGENTS.md as standing instructions (only as a note its reads names): a short fixed set of rules stands in for it (outside content never instructs; flag planted instructions; send nothing, and with the web, put nothing from your notes into a web address or a search; nothing to commit or ask).
  • A firewall between the web and the vault: a headless task with the web reads only its own folder and the vault folders or notes its reads names, never its project, the whole vault, the whole Mac, or a folder outside the vault, and its agent profile's instructions and project overviews aren't pasted in unless it names them. A task with the web can't have Gather first, since a command's output changes from run to run. A task's site list can live in a note (web: "[[Newsletter sites]]", one site per line), in both modes; its sites are in the approval, so editing the note asks again.
  • A sandboxed shell for offline headless tasks (shell: true, with web: off): commands sandboxed by macOS, with no network, no writes, and nothing on the never-read list, so a local task can look things up itself; writes names folders its shell may change (organizing Downloads, say). There's no shell beside file-access: disk or where Claude Code managed settings exist. Tokens are droppe...
Read more

1.4.0

Choose a tag to compare

@github-actions github-actions released this 03 Nov 18:28

A session keeps its context through a compaction, a Remote Control dot on Claude's strip, a New Ticket button, folding tool output with /quiet, and typing over a selection in Claude's input box.

Added

  • A Remote Control dot on a Claude session's strip, last on the right: green while you can reach the session from the Claude app or claude.ai, grey while you can't, and hollow until Claude says. Claude's own hooks report it after each turn, so a session disconnected from the phone or claude.ai shows it at its next turn; one you disconnect from Claude's panel in the tab shows it at once. Clicking it types /remote-control: on a grey dot that connects the session, and on a green one it opens Claude's Remote Control panel, with the session's link, a QR code, and Disconnect (on a hollow one, whichever applies). Right-click it for Copy claude.ai link, while it's green. "Show Remote Control on the strip" in settings (on by default) turns it off. It only shows what Claude's own Remote Control setting did.
  • A session keeps its context through a compaction: a session started from a ticket, note, or folder now has where it started, its project's overviews, and its repo written into its system prompt as paths, beside its agent, so it still knows them after the conversation is compacted. Claude gets it on every launch, resume, and side chat (--append-system-prompt) and takes it up at its next compaction; Codex gets it at launch, with the agent, and keeps it for the conversation's life; Antigravity still gets it in the opening only. Each part is included only while its part of the opening is on, so emptying a template still means the session isn't told it. What a session was launched from is recorded by its id, follows it to a new id and through a rename, and never takes a ticket attached later, which sent the session nothing; a session started before this falls back to its transcript note's source.
  • The note icon shows a session's context: hovering the strip's ticket, note, or folder icon lists where the session was started from, its overviews, its reads, and its repo, and says whether the session keeps them through a compaction (Claude and Codex, for a session started in 1.4) or has them in the first message only (Antigravity, a session started before 1.4, a re-hosted process). Right-click it to open any of them. A session with a ticket attached but started from nothing says no context was sent.
  • reads on a ticket: a list of notes or folders (wikilinks or vault paths) the session should read too, added after the overviews in the opening and the system prompt. A folder is named as a folder, not opened up; an entry that names nothing is left out and said in a notice at launch.
  • A link that opens today's Daily Chat: obsidian://agent-manager?vault=<vault>&op=daily, with &agent=<profile> to pick an agent profile's chat whatever the sidebar's switch shows. It does what the Daily Chat button does (goes to the chat's tab, resumes it, or starts it, never a second one), so an agent can open it for you from another session. It doesn't ask first, since it can only reach the day's one chat and carries no text of its own.
  • The New Session button can be turned off, and renamed: "Show the new session button" and "New session button name" in settings. Turned off, sessions still start from the command palette, a note's or folder's menu, and Daily Chat.
  • A New Ticket button, a row under New Session while tickets are on. It opens a card with the ticket's name, its project (picked from a list, never guessed; while agent profiles are on, the switched profile's projects; never one inside a folder named Archive), notes, and the CLI it starts on (your default to start, written to its backend property), beside the buttons. Create makes it in the project's Tickets folder and opens it; Launch ("Show the launch button", on unless turned off) also starts its main session on the CLI picked, named for the ticket. A ticket always needs a name. "Show the new ticket button" and "New ticket button name" in settings, and "Where the card's notes go": first, right under the properties, unless you name a heading in your ticket template for them to go under.
  • Typing, Cmd+V, and Cmd+X over a selection in Claude's input box, as in any text field. Typing over a selection replaces it, Cmd+V pastes over it, and Cmd+X cuts it (a copy, then a delete). It works on a selection made with a drag, a double click (a word), a triple click (a line), or Cmd+A. Keys you type while the selection is being deleted wait their turn, so fast typing lands in order. A selection in a reply is left alone, and with nothing selected, Cmd+X only copies.
  • Fold tool output, with /quiet: "Fold tool output" under Advanced & experimental in settings (on by default) adds /quiet to Claude sessions, and a Quiet button to the strip above them. It folds each tool's output into one line under its header (a file edit to +3 −1 or "New file, 12 lines", a command to "48 lines of output", a file read to "Read 120 lines"), and back again, redrawing the whole session at once, mid-turn too; /quiet on and /quiet off set it. Only the tools that produce output fold (file edits, commands, file reads, web fetches and searches, MCP tools); what Claude writes to you stays: its text, its questions, plans, task lists, links, and subagents' reports, and any errors. The transcript is untouched, and Ctrl+O still shows everything. Claude Code remembers the choice between sessions. The Quiet button's icon shows which way it is. It works through a second small Claude Code mod that ships with Agent Manager, loaded only into sessions Agent Manager starts, and needs Claude Code's fullscreen rendering to fold output already shown.
  • Make a ticket from this session, in a session tab's menu while tickets are on: it opens the New Ticket card with the session's project filled in (when it started from a note or folder in one; a Daily Chat's is left to pick), and Create attaches the session to the new ticket as Linked, so its lines stay current there.

Changed

  • The Quiet button sits just left of the transcript button on the strip, so the last three are Quiet, the transcript, and the Remote Control dot.

  • Typing while Agent Manager types into a session stops its Enter: a command from the strip, Wrap, or the opening is never sent with text you typed while it went in. Your text is put back in the prompt, and the opening tries again. And two lines Agent Manager types at once (a click on the Remote Control dot, then Quiet) never run together: the second waits, or says to try again.

  • Claude's Remote Control panel counts as a question the session is asking, so nothing from the strip is typed into it while it's open.

  • Right-click the transcript button to open the transcript note. The button above a session still shows the rendered transcript on a click; its right-click menu now has Open transcript note, which was only in the rendered transcript's own right-click menu.

  • A pinned session shows a faint pin at the right of its row, as Obsidian marks a pinned note.

  • New Session and New Ticket share one line while both are shown, each card opening full width below them.

  • The sidebar's header is roomier: the title is a size larger, and the custom button is larger, with no border until you hover it.

  • New Session is a row under Daily Chat, and its card opens under it and picks an agent profile. The card grows down from under the row as it opens: the name, then, while there's more than one profile, a dropdown for the one the session starts as (set to the switch's choice until you pick another), then the CLI beside Start. The row takes the card's color while the card is open.

  • The Tickets button is now a custom button, at the top right of the sidebar: its icon, text, and what it opens are each a setting ("Custom button icon", "Custom button text", "Custom button opens", a note, a base, or a link), so it can open anything. While tickets are on it shows a ticket and opens the tickets note, which the Open tickets command opens too ("Tickets note", which was "Tickets button opens"). "Show the tickets button" and "Tickets button name" are gone rather than carried over: if you'd hidden or renamed the button, do it again with "Show the custom button" and "Custom button text" (empty, it says Tickets).

  • An agent profile's Daily Chat is named Daily Chat - {{profile}} - {{date}} by default (was {{profile}} Daily Chat - {{date}}), so every profile's daily chats sort together and are quick to find in a list, the Claude app's on your phone included. A profile with its own Daily Chat name keeps it. A profile on the default whose chat for today already ran under the old name starts a new one today; earlier chats keep their names.

  • Pictures Claude sends you have a setting of their own, "Show pictures Claude sends you", under Advanced & experimental in settings with the other Claude Code mod, which says what it does: a picture Claude sends is drawn in the session, framed, with Open, Save to Downloads, and Copy path, through a small Claude Code mod that comes with Agent Manager. It was tied to the terminal's picture setting, and still needs it.

  • The Daily Chat button can be turned off, with "Show the daily chat button" in settings, as the Tickets button already could.

  • Set up this vault creates an agent profile's files before you add the profile: it drafts each profile's instructions note and home folder, makes them on your yes, and only then asks you to add the profile in settings, whose pickers need them to exist.

  • The rendered transcript loads earlier turns as you scroll up. A long session's transcript still opens with its latest turns (the last 512 KB), and scrolling near the top now brings in the earlier ones, back to the start, without moving what you're reading. It used to stop there and point you to the trans...

Read more

1.3.1

Choose a tag to compare

@github-actions github-actions released this 23 Oct 21:07

Fixes from Obsidian's automated review, two picture fixes, and the README on more than one computer and your phone.

Fixed

  • A long caption on a picture Claude sends is cut short before its size, which no longer wraps onto a second line.
  • Copy path on a picture shows one notice, the picture's own (as Save to Downloads does), not a second one from Agent Manager.

Changed

  • The README says to turn Agent Manager on in one computer only, how to keep it off on another under Obsidian Sync, and how to reach sessions on the main computer through Remote Control or SSH. Your phone and Known limits say what can and can't be done from a phone or another computer: a running Claude session can be continued there, but no session can be started in Agent Manager from elsewhere yet.
  • The terminal's pictures layer, link tooltip, and color lookup are made with Obsidian's own element helpers, and a leftover package.json in src/ is gone, from Obsidian's automated review.

1.3.0

Choose a tag to compare

@github-actions github-actions released this 23 Oct 19:55

Onboarding: Set up this vault, with starter skills.

Added

  • Set up this vault (experimental). A session, named "Vault setup", that helps you set the vault up for agents: it asks what the vault is for and what's off limits, looks around, and offers an AGENTS.md, a rule for where files go, and optional extras (projects, skills, a wrap skill, tickets, agent profiles), each written only when you say yes. It reads its instructions from a guide the plugin keeps in its own folder (guides/vault-setup.md), so the opening is short. In the sidebar while there's no AGENTS.md, under Advanced & experimental in settings, and in the command palette; when a setup session ran before, it offers to resume it or start fresh. Rather than only describe projects and tickets, it offers an example project (an overview, a decisions note, a sample ticket its overview lists, and a ticket for the setup itself, with the setup session as its main session), ends its report with what to open and try first and an offer to answer questions (from the README, which ships beside the guide), keeps what it says short for people new to all this, and recommends which of your own folders could be projects. It opens with a short welcome, offers git first (explained for people who've never used it), and asks whether to go step by step or set things up on its own and report back; either way it asks before changing anything already in the vault. It stays experimental: an AI does the work, so results vary with the CLI and model.
  • Starter skills, wrap and second-opinion, in skills/: written into the plugin's folder, offered by Set up this vault, and copied into your skills folder only when you say yes. Once copied they're yours; the plugin never changes a copy. Both run only when asked. The second opinion's commands keep each reviewer as close to read-only as its CLI allows (tested on Claude Code, Codex, and Antigravity).
  • Tips for each CLI in the README.
  • Attach to a ticket as its main session, in a session tab's menu: the ticket's Resume opens this session, and a main session it had becomes Previous main. Only when asked; a plain attach still never takes main.

Changed

  • The settings' Advanced section is now Advanced & experimental.
  • Overviews are now Project overviews, in settings and the folder menu (Add a project overview to this folder), so it's clear what they are.
  • Turning tickets on turns project overviews on too, since a ticket's session reads its project's overview. Project overviews can still be turned off on their own.
  • A link that starts a session asks first, since a link can come from a web page or a note someone else wrote. Starting from a ticket (the tickets board's Launch) doesn't ask. A resume link asks only when the session ran outside this vault.
  • Cmd+A in Claude Code selects what you've typed in its input box. In its full-screen mode it's Claude Code's own selection, made as a drag would make it, so Backspace deletes it and Cmd+C copies it as they do after a drag; otherwise the terminal selects it, and Backspace or Delete clears the input without interrupting Claude Code. Elsewhere Cmd+A no longer selects the whole terminal; that's Select all in the terminal's right-click menu.
  • Every copy in the terminal shows a notice with the start of what was copied: yours (Cmd+C, Ctrl+C, Copy in the menu) and a program's, so text you didn't copy yourself is never a surprise on a paste.
  • The README and the plugin's description say desktop and macOS only, for now.
  • Set up and Add (an agent profile) in settings show a clear hover, as the terminal's confirm buttons do; Obsidian's own barely shows on a plain button.

Removed

  • Create AGENTS.md, the sidebar's starter file, replaced by Set up this vault.

Fixed

  • Cmd+clicking a folder inside the vault shows it in Obsidian's file explorer, not Finder.
  • A path with an apostrophe in a name (Projects/Sam's Project/Overview.md) can be Cmd+clicked; a path after a possessive (README.md's) links without the 's.
  • Hovering a long line in the terminal no longer checks the same file paths again for every row: a line's paths are found once and kept for a moment.
  • The terminal reads macOS's default PATH once per load, not on every launch.
  • An agent stopped with Ctrl+C right after it starts is no longer started again without the terminal's additions.
  • A closed terminal tab no longer stays hooked into its tab's close.
  • The session list stays fast with thousands of Codex sessions: it no longer reads every session file again once there are more than 4,096.
  • A session still waiting for its transcript looks for it less often as time goes on, rather than searching every CLI's sessions several times a second.
  • Renaming or moving a transcript note no longer starts a second one: the session keeps writing to it wherever it is.
  • Renaming the transcripts folder, or the folder for sent files, in the file explorer updates the setting, so new transcripts don't go to a new copy of the old folder.
  • A ticket that appears without you making it (an agent writing it, sync, git) no longer opens a "which project?" picker; only one you're working in does.
  • Resuming a Claude session with an agent profile reads only the end of a long transcript.
  • A double quote in a session's name no longer cuts the name short for Codex and Antigravity.
  • A line break in a value an opening fills in (a multi-line repo: property) no longer sends the opening early.
  • Added folders saves when you leave the field, not on every keystroke.
  • A transcript note written to by the session and by its transcript view at once no longer loses one of the writes: both take the same lock.
  • A folder setting (the ticket template, the sent-files folder, the tickets note) that points outside the vault is ignored rather than followed.
  • A session id that isn't a well-formed id is never passed to the CLI on a resume or a fork.

1.2.0

Choose a tag to compare

@github-actions github-actions released this 23 Oct 19:55

Agent profiles.

Added

  • Agent profiles. Turn on Use agent profiles and add one (a name, an instructions note, an optional home folder, and its own Daily Chat name), and sessions start with it: Claude Code gets it as an agent (--agents and --agent), Codex as developer instructions, and every CLI's opening names it and its note. There are none until you add one.
    • A session started from a note or folder in a profile's home folder starts with that profile (the deepest home wins); one started outside every home starts with the Default agent profile, No agent profile unless set. A session started from the sidebar, Daily Chat, or the New session command starts with the profile picked on the sidebar's switch.
    • The switch, at the top of the sidebar: two choices are a pill whose purple thumb slides to the one picked, more a dropdown. Either is the height of the System Dashboard and Daily Chat buttons below it, in their titles' type. It picks who new sessions and Daily Chat start as, and Pinned and Recent list only that choice's sessions. A session with no agent profile (or one the plugin didn't start) is listed under the Default agent profile; while that's No agent profile, No agent profile is a choice of its own on the switch, with its own list and Daily Chat.
    • One Daily Chat per profile per day, named by the profile's own Daily Chat name ({{profile}} Daily Chat - {{date}} by default). The sidebar's button says whose it is (Alpha Daily Chat), changing with the switch; with No agent profile picked it's Daily Chat.
    • Transcripts per profile, in a folder of the profile's name in the transcripts folder, with an agent property. Renaming a profile renames its folder.
    • Side chats and Continue Session keep the profile of the session they came from.
    • Move to agent profile…, in a session's menu on the sidebar and on its strip, files it under another profile: its list, its transcript folder, and what sessions continued from it start with. The conversation itself keeps the profile it started with, since neither Claude Code nor Codex changes it on a resume.
    • A new opening part, Agent profile, sent first: "You are the {{profile}} agent…".
    • Add an agent profile opens a popup for its name, instructions note, home folder, and Daily Chat name; it sits at the top of the section, above Default agent profile.
    • Each profile can be turned off without losing its settings: it leaves the switch, homes, and the Default agent profile, and its sessions are listed as if they had no agent profile until it's back on.
    • Words: what you add in settings is an agent profile everywhere you see it, since every session is an agent; a profile is only who it starts as.

Changed

  • Settings: each CLI's first switch is Enable (was "Offer for new sessions"); Default CLI, now Default backend (CLI), lists only the enabled CLIs and says it's what a ticket with no backend starts on; Openings moved down, just above Advanced.
  • Daily Chat sits above the Tickets button.

1.1.0

Choose a tag to compare

@github-actions github-actions released this 23 Oct 19:55

Tickets, with overviews to go with them.

Added

  • Tickets. Turn on Use tickets and a note with type: ticket is a ticket: a session started from it gets the ticket opening, which says to read it for context and wait. The ticket template (Templates/Ticket.md by default, a setting) is never one.
    • Turning tickets on creates Tickets.base in the AGENTS.md folder and the ticket template, each only where it's missing. The base lists every ticket in three views (Open, grouped by project; Done; This folder), with Done, Chat (a Launch or Resume pill), Ticket, Project (worked out from its folder), CLI, Created, and Updated.
    • A ticket has one main session: launching it again while that session is open, or still starting, goes to its tab. Its backend property picks the CLI.
    • Each session started from a ticket, or found in one on a resume, keeps its lines in the ticket: a resume link under ## Session Links, labeled with its name, its CLI, and its kind (Main, Previous main, Side session, Linked, Re-hosted, or Continued), and a link to its transcript note under ## Transcripts. Each line is matched by the session's full id, so a rename updates it in place.
    • Attach to a ticket, in a session tab's menu: pick any ticket, and the session's lines go on it as Linked, kept current while it's open. It never takes the ticket's main session, and a session can be attached to several.
    • A ticket made outside any Tickets/ folder (the base's New button puts one in the default folder for new notes) gets one question: which project it's for, from the folders with an overview, or leave it where it is. Picking one moves it into that project's Tickets/ folder, keeping links to it.
    • A Tickets button above Daily Chat opens the tickets base, or a note set in settings, under a name set there. It can be turned off. A note it opens keeps it when renamed.
    • The strip shows a ticket icon for a session started from a ticket or attached to one; it opens the ticket.
    • New ticket (a command, and "New ticket here" in a folder's menu) makes one from the template in the folder's Tickets/ folder, with its name ready to type.
  • Overviews. Turn on Use overviews and a folder with an Overview.md (the name is a setting) is a project: a session started from a note or a folder below it is first told to read every overview from the AGENTS.md folder down, broadest first, and the repo the nearest one names in its repo property. Both parts of that opening are settings.
  • Add an overview to this folder, in a folder's menu while overviews are on, makes one from a template (a setting, with a short one built in).
  • Wrap, on every session's strip: types the Wrap message setting (a request for a summary by default; empty hides the button). A message like /wrap runs a skill, and goes to Codex as $wrap , the way Codex invokes one.
  • Continue Session replaces the strip's Side chat button. On a session started from a note it offers a side session or a new main session, on any CLI; on any other session, a new session. Then two choices: Keep full context (a copy of the conversation, on the same CLI) or Read the transcript (on another), and the light one: Read the ticket, Read the note, or Read the last message only. A new main session takes the note's main session over once it starts.
  • Files Claude sends you are saved in the vault. Each file a Claude session sends (its SendUserFile tool) is copied where Obsidian puts a new attachment for the note the session was started from, or into a folder set in settings, and shown in the session's transcript note. A copy never overwrites, and one that fails is tried again on the next turn. It can be turned off.

Changed

  • Every setting that takes a folder or a note suggests them as you type, on Obsidian's own suggestion list, alphabetical.
  • Settings show a section's own settings only while its switch is on: overviews, tickets, the Tickets button, and each CLI's while it's offered, and turning one on or off keeps your place on the page.
  • The plugin no longer reads Obsidian's deprecated workspace.activeLeaf: the focused tab comes from the active view (getActiveViewOfType), in the main area or a sidebar alike.
  • A session started from a folder records the folder in its transcript note's source (as its path with a trailing slash), so a resume knows where it started.
  • The starter AGENTS.md tells agents where files go: in a Files/ folder beside the note they belong to, since agents don't read Obsidian's attachment setting.

1.0.1

Choose a tag to compare

@github-actions github-actions released this 23 Oct 19:32

Added

  • The pictures Claude sends you, in the terminal. A Claude Code mod ships inside the plugin and draws each PNG a Claude session sends (SendUserFile) as a framed picture, with Open, Save to Downloads, and Copy path, capped in height so scrolling stays smooth. The plugin writes it into its own folder and passes it to every Claude start and resume with --plugin-dir while "Show pictures" is on; nothing is downloaded or written to ~/.claude. Its look and some of its speed-ups come from show-your-work (MIT, credited). The mod is also a marketplace for installing it by hand (claude-mod/); a copy installed that way and the bundled one don't draw a picture twice.

Fixed

  • A tab whose saved backend name is one every JavaScript object inherits ("constructor") now falls back to Claude, as an unknown name always did, instead of breaking the tab (found by Codex reviewing the TypeScript conversion).
  • From Obsidian's plugin checker: the manifest's description no longer says "Obsidian"; the plugin's own CSS has no !important (two remain in xterm's, which the build appends) (a hidden row uses Obsidian's own toggle, and the sidebar's tab-title rule wins by being specific); the starting-folder picker asks Obsidian for folders (getAllFolders) instead of listing every file in the vault.
  • A picture a program draws (kitty graphics) no longer shows a pointing hand, since clicking it does nothing, nor a thin colored line beside it.
  • Large pictures scroll smoothly: each row of a picture is drawn at once, not each cell (about 2 ms a paint for a full-width screenshot).

Changed

  • The release workflow uses current actions on Node 24 (actions/checkout@v7, actions/setup-node@v7, actions/attest@v4) and builds with Node 24, ahead of GitHub retiring Node 20.
  • The notice shown when the login shell's environment can't be read explains VAULT_TERMINAL_ENV_PROBE itself, rather than pointing at the README.
  • The plugin is TypeScript now: the plugin class and its parts (src/plugin/), the session tab (src/sessions/, with its transcript overlay and its note write-back), the sidebar (src/browser/), the transcript writer and the import, the settings tab, the dialogs, Live Preview's link pills, the constants, and the shared data types, checked by the typecheck and by lint rules that refuse untyped values there. No change in behavior, except that opening a transcript note whose path is a folder now says the note isn't written yet instead of trying to open the folder.

1.0.0

Choose a tag to compare

@github-actions github-actions released this 21 Oct 21:48

Added

  • A terminal inside Obsidian for Claude Code, Codex, and Antigravity, with sessions in a sidebar: a New Session button that opens the new-session card (a name and the CLI; the session is told by default to greet you and ask what you'd like, reading nothing beyond AGENTS.md until you say), Daily Chat, then Pinned (in an order you drag) and Recent (open sessions first, then every other session in the vault, newest first, 20 at a time, with a search over their names).
  • A dot on every session: green while it's open in a tab, an empty ring when it isn't.
  • Daily Chat: one row, showing the date (it moves to the new day on its own), with a sun that's grey until the day's chat has run, yellow once it has, and green while it's open, that goes to the day's chat, resumes it, or starts it, named by a setting (Daily Chat - {{date}}), with its own opening (by default, that it's for quick questions that may have nothing to do with each other).
  • The AGENTS.md folder: every session starts in one folder, the vault's root by default. A session started in a folder below the root isn't given the rest of the vault. The first-run check offers to create a short AGENTS.md there when there's none. Sessions started in any folder the plugin has used stay listed after the setting changes.
  • A first-run check for Python 3 and the three CLIs. Only the installed CLIs are offered; when none is, or Python is missing, the sidebar says what to install, with each CLI's install command.
  • Start a session from any note or folder: the "Start session from this note" command, "Start agent session" in the file menu, or a link (obsidian://agent-manager?vault=<vault>&op=new&note=<path>). It's named for the note or folder, sent an opening that points at it (a setting per kind), and a note gets one property, main-session, the session's resume link. Its name is fixed, so a base written for one vault works in any. The latest session started from a note holds it. Claude's link is written at launch; Codex's and Antigravity's as soon as their transcript appears, with no need for a second turn. The session's transcript note links back to the note in its source property.
  • Settings laid out for this plugin: general settings first (default CLI, the AGENTS.md folder, transcripts folder, Daily Chat name, the terminal split, added folders), then the openings, each CLI (offered or not, whether it's installed and how to install it, its command and location), the agents' environment, the terminal, and Advanced.
  • Resume a session from the sidebar, from a link (obsidian://agent-manager?vault=<vault>&op=resume&session=<id>), or by pressing Enter after it ends. A reloaded window brings its session tabs back.
  • Every session saved as a note in one folder (default Agent Transcripts), append-only, and a rendered transcript beside the terminal.
  • Side chat, on every session's strip: a copy of the conversation so far in a new tab, made by the CLI's own fork (Claude's --fork-session, Codex's fork, and Antigravity's /fork, typed into a new tab on the conversation). If Antigravity's fork doesn't answer, the side chat is a new session that reads the transcript instead, and says so.
  • A session that Remote Control or a compaction moves to a new id is followed: its tab, pin, name, and transcript stay with it.
  • Sessions started outside the plugin are listed too, named by their first message.
  • The licenses of xterm.js and its three addons at the top of main.js and styles.css, since an install gets only those files, and in THIRD_PARTY_NOTICES.
  • Agent Manager links in a note, to resume a session or start one, show as pills, in Live Preview and reading view.
  • A README for someone starting from an empty vault: install, the first session, starting from a note, a Bases board, links, and everything the plugin reads, runs, and writes outside the vault.