Sessions

A session is one Claude Code conversation running in a topic folder. Sessions are persistent: close the browser, restart the server, come back next month – the conversation continues where it stopped.

Layout

• Left pane: the project / topic tree. Badges show running sessions and unread answers.
• Main area: one or more columns, one session each. Open several sessions side by side, resize and reorder the columns, or detach one into its own window.
• Grid: with two sessions or more, the grid button at the top right stacks them two per column (four sessions = 2x2). You can also drop a column on the lower part of another, or pick Put under the column on the left in its … menu. Only the session you clicked shows its full message box; the others show a slim Write… bar - click it to write there. Drag the line between two stacked sessions to share the height (double-click: half and half). The same button, or Give it its own column, puts them back side by side.
• Ready-made layouts: hold the mouse button down on a session's grip (the dots left of its title): small layouts appear, click the zone where it should go (the others fill the rest). Or drag the session by its grip: a layouts bar slides down at the top of the sessions, drop it on a zone. Rest the pointer on the grid button to arrange every open session at once. With the keyboard: … menu > Layouts….
• Fold the sidebar with the button at its bottom; fold the session tree with the button at its top left (a slim rail stays to unfold it).
• On a phone or a tablet (the touch layout), one session fills the screen and the app opens on the last one. ☰ at the top left opens all your sessions: a search field, New session (in the topic of the session on screen), the All / Running / Unread chips, then your projects and topics. Press and hold a session, a topic or a project for its menu (rename, favorite, archive, delete...). The session's … menu also holds Plan mode and Telegram notifications. ⊞ at the top right opens the rest of the app (see Settings).

Move a session to another topic

Under the session's title, its project / topic is a button: click it, type to search, pick another topic. Or drag the session in the session list and drop it on another topic (or on any session of that topic) - it is the same move. A session that is working cannot move until it finishes; inside its own topic the order follows the list's sort, so dropping there changes nothing. The session moves there (the list updates) and Claude continues the same conversation, now working in the new topic's folder with that topic's context. Possible once Claude has finished (not during a turn or while messages wait). Its Telegram thread is renamed. Files it created stay where they are.

The sessions list

At the top of the list:

• Projects / Sessions (the button with two icons, the lit one is the current view): click to switch.
• Projects: the sessions grouped by project and topic (the default).
• Sessions: every session in one list, whatever its project, most used first: the sessions you talk to the most lately come first (a long session used this week before a one-off from this morning; an old one slowly goes down). One line each: the title and its # topic on the right; hover it for project › topic.
• Search (the magnifier): opens the filter; type to narrow the list, Esc or × closes it.
• New project (the folder with a +).

The bar at the bottom of the list:

1. In Projects, the layout - click to switch:

In Sessions, the sort instead: most used, recent activity, A → Z, creation date (click to change). 2. List options (the sliders icon): tick the Favorites and Recent blocks to show them at the top of the list, and how many recent sessions to show (3, 5, 10 or 20). The Favorites block only shows once you have a favorite.

• Project emoji: the … menu of a project -> Set an emoji… (a list of common ones, or type / paste any emoji). It shows before the project in the list and labels, and its Telegram threads (🛒 acme/blog · Launch plan). It is saved with the project folder (kept by renames, the trash and the backup).
• Favorites: add a session with its … menu (or a right-click) -> Add to favorites; same menu to remove it. A star marks favorite sessions. Favorites are saved on the server: the same on your computer, your phone and the installed app, and a continued session (cont.) stays a favorite.
• Recent: the most recently active sessions (a favorite is not repeated there while Favorites is shown).
• The view, the layout, the sort and the blocks are remembered by this browser. Searching shows every match, whatever the view.

Starting a session

Pick a topic, click New session, write your message. The session runs in the topic folder, so Claude automatically loads the workspace, project and topic memory (see Memory).

Its title is short: the start of the message shows at once, then a shortened title replaces it (greetings and "can you…" removed, cut on a word): "Tu vas me créer une session d'automatisation de E-ink" becomes Créer une session d'automatisation. With Titles written by Claude on (Settings -> Claude, off by default: one quick call to the smallest model per new session, counted in your Claude plan), the title names the subject instead - Automatisation E-ink - a few seconds later (if the call fails, the message is shortened). The same goes for a session started from Telegram (its thread is renamed too) and for an empty session on its first message. A title you typed, a Telegram thread you named, or a rename made meanwhile, is never replaced. Turn it off in Settings -> Claude (Short session titles): the title is then the start of the first message.

The composer

• Enter sends, Shift+Enter adds a new line.
• Attachments: drag and drop files, paste an image, or use the paperclip. Files are saved in <topic>/attachments/YYYY-MM/ and their paths are given to Claude. With many files, they sit in a small strip of two rows that scrolls (the newest in view), so the conversation keeps its room.
• Queue: you can send while Claude is still working. Messages wait in a card above the composer, with their full text, and are sent one after the other; its × clears them. The card shows every waiting message, also after a reload or on another device; one that came from Telegram, a scheduled job or Harness itself says so in front of its text.
• Stop interrupts the running turn.

Steer while Claude works. A message sent during a turn waits in the queue above the composer. Click Steer (in the waiting-messages card) to use it immediately: the current turn is interrupted, Claude keeps everything it has done so far (same conversation) and continues the task taking your new message into account. Several queued messages are sent together. Edit puts the waiting messages back into the composer at once (attachments included) so you can change them before sending again; if one of them started meanwhile, only the others stay in the composer. The × of the card empties the queue at once too (it comes back if the server refuses). Stop still cancels the turn and the queue.

Ask without interrupting. Question (in the same card) asks the waiting messages on the side: Claude answers about the session - where it is at, why it did something, what is left - without stopping its task. The answer opens in a larger panel above the composer; close it with its ×. It is not added to the conversation and Claude's work is not affected (it answers from a copy of the conversation, read-only).

Quote part of an answer. Select text in one of Claude's replies (or in a side-question answer) and click Quote above the selection (below it on a phone). The text goes into a Quote card above the composer; quote several passages if you like, and remove one with its × (or all with the card's ×). When you send, the quotes go first, as Markdown blockquotes (> …), followed by what you typed - you can also send the quotes alone. Quotes are kept with the draft and shown as quote blocks in your message.

What you see

• Answers stream live (markdown, code, images).
• Tool calls (reading files, running commands, editing) appear as compact cards; expand them to see input and output. Edits are shown as diffs.
• Diagrams: to explain how something works, Claude can draw a small diagram (a flowchart, a sequence, the states of something...) right in the answer. Click it to open it in the file viewer like a picture (zoom, fullscreen, ← / → with the other files of the conversation, Source for its code); hover it for Show the code and Copy. On Telegram it arrives as its code.
• Git graphs: when Claude talks about a repository, it shows the real graph of its branches and commits (read live); click a commit for its files. See Git & GitHub.
• Pictures and videos named in an answer show as thumbnails under it (click to open), and audio files as a small player you can play right there (its name opens the file viewer); audio you send plays the same way. Thumbnails follow the files: when Claude rewrites a picture under the same name, the new version shows as soon as the tool has run (also in the viewer and the HUD).
• Right click a picture, a video, an attachment or a file path (long press on Android): Open, Download, Quote (adds the file to your next message as one line with its path - Claude reads it only if needed, the picture does not fill the context) and Copy path.
• File paths are clickable and open the file viewer: ← / → go through every file of the conversation (see Files).
• The header shows the model that actually answered, the cost and a context gauge (how full Claude's context window is, measured against the real window of the model in use - it follows when you switch model).
• A line under each answer shows the model, the time it took, its cost, the tokens (input → output) and the share of input served from cache. Hover (or tap) any of them to see what it means: the token count splits uncached input, cache reads, cache writes and output.
• About the cost: it is what the answer would cost at Anthropic API prices, computed by Claude Code. Signed in with a Claude subscription (Pro / Max) it is not billed - it only shows how much you used; signed in with an API key it is billed to your Anthropic account. The tooltip says which case applies to your server.
• Choose what the line shows: click the ⚙ at its end (it appears on hover; always there on a touch screen), or Settings → General → Line under each answer. The choice applies to every device; hiding the cost also hides the session total in the header.

Plan mode

With plan mode on, Claude changes nothing before you agree. It reads and explores what it needs, then answers with a plan: goal, steps, files concerned, risks and open questions. It cannot write files or run commands that modify anything in this mode.

• Turn it on with the plan button (checklist icon) in the session header, or tick Plan mode when you start a new session. The composer shows a Plan mode chip (click it to leave).
• Once the turn is over, a slim Plan ready line appears on top of the composer (waiting messages take its place while there are some):
• Adjust (or just write): say what should change - Claude sends the updated plan. Repeat as often as you like.
• Validate: posts Plan validated - go ahead., plan mode goes off and Claude carries out the latest version of the plan.
• ×: leave plan mode without running the plan.
• The HUD shows the plan as the answer and has a Validate the plan button.
• Plan mode applies to every message of the session, Telegram included; validate the plan from the web app or the HUD. Compact and hand off still work normally.

HUD mode

A small floating bar to drive one session while you work in other apps. Turn it on in Settings → Features → HUD mode, then click the HUD button in a session header (or Open the HUD in its menu).

• While Claude works, it shows your prompt and Claude's last three steps (tools, thinking). When it is done, it shows the answer with its images - one at a time, ‹ › to go through them; a click opens the file in the app. The arrow next to the send button folds it into a slim bar (steps only) and back.
• Enter sends; during a turn the message waits in the queue and Steer sends it right away. The square button stops the turn.
• Share my screen (the screen icon): pick a screen or a window once; from then on a capture is attached to every message you send from the HUD, so Claude sees what you see. When it helps, Claude answers with a copy of the capture annotated (arrows, boxes) to show you where to look. The capture usually goes straight to Claude with your message, so it sees your screen from its very first step. Captures are saved with the session's attachments (<topic>/attachments/). Click the icon again, or use the browser's Stop sharing, to stop.
• At the top: Back to the app brings the main window forward, × closes the HUD; hover the answer to copy it.
• In Chrome and Edge the bar stays on top of every app and can be moved anywhere, a second screen included, and it closes with the app tab. Other browsers open it in a small window of its own.
• The HUD only talks to Claude: it does not control your computer. It is not available on a phone.

Models

Each session can use its own model (menu in the session header). Empty = the default model from Settings -> Claude. The label always shows the model that really answered the last turn; if it differs from your choice, a marker tells you. Each model has its own prompt cache: on a session past 50k tokens of context, switching asks first, because the new model reads the whole conversation once without cache (Compact first to make the switch cheap).

Compact vs hand off

When a conversation gets long, the context fills up and answers get slower and more expensive.

Automatic compaction (Settings -> Claude, 400k tokens by default): when a turn ends with more context than the threshold, Harness compacts the session by itself before your next message - a note in the conversation says so. Long sessions then stop re-sending their whole history on every call, which roughly halves the tokens they use. Pick a higher threshold (or Off) for a session that needs every detail of a very long exchange. A single long turn (many tool calls in a row) is compacted along the way too, near the same threshold, instead of growing up to the model's limit. maxxing sessions token-report shows, on your own sessions, what each threshold would have saved.

Compact when you come back (Settings -> Claude, 200k tokens by default): Claude keeps the start of a conversation in a cache for a few minutes (up to an hour). When you write to a session that has been quiet for longer and holds at least that much context, the cache is gone and the whole history would be re-sent at full price - so Harness summarizes the session first (a note says "Back after 2 h: ...") and your message runs in a fresh thread seeded with the summary. You wait for one summary turn before the reply; every turn after it is much smaller. Messages from scheduled jobs never trigger it, nor does a session whose context is mostly what every new thread loads anyway (CLAUDE.md, memory): summarizing it would save nothing. Pick Off to keep the full thread whatever the pause. Stop during the summary cancels only the summary, your message stays: it runs right after, in the full thread.

Edit or delete a message that got no answer

A message you stopped before Claude answered (it shows Stopped. under it) can be fixed or removed. Right-click it (long press on Android), or use the buttons next to its time (always shown on a phone): - Edit (pencil): its text, its quotes and its attachments come back in the message box (before what you had already typed), and the message leaves the conversation. Fix it and send it again. - Delete this message (bin): it leaves the conversation with its "Stopped." line; Undo in the notice brings it back.

Only the conversation shown here changes: what Claude already received stays in its own memory of the session.

Start a new session from a message

Right-click any message (long press on a phone) - yours, even answered, or a reply of Claude - and pick New session from this message: a new session of the same topic opens with the message already in its box (a reply comes as a quote). Read it, change it if you like, and send it. Copy is in the same menu. A selected piece of text keeps the browser's own menu (and the Quote button), a file keeps its own menu (Open, Download, Quote, Copy path).

Notify, archive, delete

• Notify (bell): when a turn ends, a short message is posted in the topic's Telegram thread.
• In the browser: Settings -> Notifications -> Notify me when a session finishes shows a system notification when the tab is in the background (a small banner with Open otherwise), with an optional chime. A failed turn is notified too; scheduled jobs are not. Test checks that it works in this browser. Unread answers show a green round counter in the sidebar, on the ☰ button and the tree; it can pulse until you read them (Pulsing badge). Click a counter - running or finished, in the sidebar or on a project / topic of the tree - and the session tree shows only those sessions, unfolded. Open all opens them next to the columns already open; × brings the whole tree back (typing a search does too). The Sessions link itself brings back your last layout.
• Archive hides a session from the list without deleting it (see the Archived filter).
• Delete stops a running turn and removes the session files.

Search

Ctrl/⌘+K (or the search bar at the top of the sidebar) is the one search of the app. It finds:

• sessions, by their title and by what you and Claude wrote;
• memory (memory.md, CLAUDE.md), tasks (title, notes, labels) and the content of the text files of the workspace (markdown, code, config…), plus file and folder names. Every memory.md / CLAUDE.md is always found; in a very big workspace each topic gets its fair share of the other files, and data files (.json, .csv, .xml, .svg) are left out. New files show up within 10 minutes, memory at once;
• pages, settings, topics and actions.

Type a few words in any order: each word may be the start of a word ("fact" finds "facture"), accents do not matter, and the best matches come first with the matched words highlighted. A message is searchable a second after it is written; edited files within half a minute.

Sessions from Telegram and crons

Telegram conversations and Claude cron jobs create normal sessions (marked with their source). A Telegram session can be continued from the web UI and the other way round.

Keyboard shortcuts

Good to know

• Each turn runs in its own systemd unit (maxxing-<id>): it survives a server restart.
• Claude Code keeps transcripts for ~10 years (cleanupPeriodDays is set to 3650), so old sessions can always be resumed. If a transcript is missing, a fresh Claude context is started automatically and a note is shown.

Close all / Open all (Vertical layout): the quick button at the bottom of the session list folds every project and topic as soon as something is open, and unfolds everything when all is folded.