Chat

The Chat page is the main agent conversation view and where most users spend their time. Regular users also get their own Tasks, Cronjobs, Memory, Email, Account and personal Settings tabs; instance-wide pages are admin-only — see Multi-User.

It's a streaming, persistent conversation with the configured agent — messages flow over a WebSocket, attachments upload through a separate REST endpoint, and the page reconnects automatically if the socket drops.

Toolbar

The top of the page has a slim toolbar (PageHeader). On mobile its action buttons collapse into the global header bar so the second row disappears.

Left side — connection dot:

Dot color	State
Green	Connected. Messages flow live.
Amber	Connecting / reconnecting. The WebSocket is mid-handshake.
Gray	Disconnected. Send is disabled until the socket comes back.

Next to it, a context meter shows how full the agent's context window is, e.g. 33k / 262k with a ring. Click it for details: model, used tokens with a bar, and where automatic compaction starts (the marker on the bar). The ring turns amber at 80% of the compaction threshold (of the window when automatic compaction is off) and red while automatic compaction is paused. A ~ marks an estimate: right after a new session or a compaction there is no model response to measure yet. The meter updates after every answer and compaction, and appears once the agent has run since the server started. On mobile it shows as a ring next to the action buttons.

Right side — actions:

• Stop — aborts whatever the agent is currently doing (only enabled while a response is streaming). Equivalent to typing /stop in chat or hitting the Kill button on the parent task.
• New Session — closes the current chat session and starts a fresh one. The divider appears immediately; a Session Summary card may fill in shortly after once background generation completes (see Memory System → session summaries for how that works). The next message you send opens a new session.
• Settings (gear icon) — display filters, see below.

Display filters

The gear icon opens a popover with four toggles. Settings are persisted to localStorage per browser, so each operator can have their own view.

Filter	Default	What it shows
Tool Calls	On	Collapsible cards for each tool the agent invoked.
Injections	Off	Background-task results and periodic status heartbeats.
Session Summaries	Off	The summary card above each session divider.
Thinking blocks	On	The collapsed Thinking card before each assistant turn (if any).

Hiding a category never deletes anything — toggling the filter back on brings the messages right back.

Message types

The chat renders many distinct shapes besides plain text. Knowing them helps when scanning a long conversation.

Plain messages

User messages are right-aligned with a colored bubble; assistant messages are left-aligned with a neutral bubble. Both render Markdown (assistant fully, user verbatim) and support code blocks, lists, and inline formatting.

A small timestamp below each message shows the local time. For assistant messages, a speaker icon appears alongside if Text-to-Speech is enabled — click it to read the message aloud (or stop playback).

Telegram-tagged messages

Messages that came in via Telegram show a small Telegram badge on the avatar and a "via Telegram" label above the bubble. If a Telegram message was a reply to an earlier one, a quoted line appears at the top of the bubble (Replying to: "..."). Assistant messages that were also delivered to Telegram are tinted blue and labeled the same way.

Thinking cards

When the agent uses a reasoning model, its thinking appears as a collapsed card titled Thinking before the actual reply. Click to expand. While the model is still thinking, a pulsing dot row indicates streaming. Hide all thinking cards via the Thinking blocks filter.

Tool call cards

Every tool the agent invokes shows up as a collapsible card with the tool name and an icon. Expanding the card shows:

• Input — the arguments passed to the tool.
• Output — the tool's result, or an error message in red if the call failed.

Two special cases get richer rendering:

• Memory edits — when the agent edits SOUL.md, MEMORY.md, AGENTS.md, a daily file, or a wiki page, the card renders an inline diff instead of raw input/output.
• Skill loads — when the agent loads a skill (e.g. tasks-and-cronjobs), the card title shows Load Skill: <name> with a puzzle icon.
• Codemode — when the experimental codemode tool is on, a codemode call renders as a card with the script, the nested tool calls (live-updated while the script runs) and the output; collapsed, it shows the nested-call count and tool names.

Background task cards

Two flavors, both shown only when Injections is enabled:

• Periodic update — a slim single-line row showing a running task's progress: runtime, tool-call count, token usage. Non-collapsible.
• Task result — a collapsible card with the task name, duration, and final status (Completed, Failed, Question). Expanding it shows the task summary the agent injected into the chat.

See Tasks & Cronjobs concept for what these messages actually mean.

Session dividers

When you start a new session — or the agent does, after a long inactivity gap — a thin horizontal line appears with a small "New Session" label. Above it, a collapsible Session Summary card may be present (only when Session Summaries is enabled). The summary is what the next session sees as historical context.

Composer

The composer at the bottom has, from left to right:

• Brain icon — quick switch for the thinking level; it lists only the levels the active model supports. The icon color encodes the level the model actually runs with (gray → red → purple for max); if the stored level isn't supported, a note below the list shows which level is used instead. Changes apply immediately to the next turn.

  For regular users this changes only your thinking level. For admins it changes the global default that every user without a personal override inherits. For a per-task override, use Cronjobs or create_task instead.

• Text area — multi-line input. Enter sends, Shift+Enter inserts a newline. The box auto-resizes up to ~150px and then scrolls.

• Paperclip — opens a file picker. You can also drag and drop files onto the chat area; an overlay appears while dragging.

• Microphone (if Speech-to-Text is enabled) — push-to-talk: hold to record, release to transcribe. The button color encodes recording state, transcribing state, and any error (permission denied, recording too short, generic error).

• Send — disabled when the textarea is empty or the connection is down.

On mobile the mic and send buttons swap based on whether you've typed anything: empty input shows the mic, any text shows send. This mirrors Telegram and WhatsApp.

Attachments

Pending files appear as chips above the composer. Click × on a chip to remove it before sending. Once sent, attachments render below the message body — images inline, other files as download chips.

What the agent receives:

• Images (PNG, JPEG, GIF, WebP, TIFF, AVIF) are attached so a vision-capable model can see them. Before sending, Axiom applies the EXIF rotation, scales the image down to at most 2000×2000 px and keeps it under about 4.5 MB. When the image is resized, the agent gets a note with the original dimensions. The file on disk is not changed.
• HEIC photos (iPhone default) cannot be decoded by the bundled image library. They are referenced by path only; convert them to JPEG first if the agent should see them.
• Older images are dropped from a request once the conversation's images exceed about 16 MB or 50 images in total. The agent gets a note instead and can re-open the file with read_file.
• Every upload comes with its path under /data/uploads/, so the agent can pass it to other tools (for example generate_image for edits, or send_file_to_user).
• Images the model cannot accept (SVG, corrupt files, formats the server cannot decode) are not attached; the agent only gets the path. This keeps one bad upload from breaking the rest of the conversation.
• With a text-only model, attached images are replaced by a short "image omitted" placeholder.

Where attachments end up on the backend, how long they're kept, and the upload limit live in Settings → Agent → Upload retention.

Slash commands

Messages that start with / are intercepted by the chat client and answered locally instead of being forwarded to the LLM. The same vocabulary is shared with the Telegram bot, so muscle memory carries between surfaces.

Command	What it does
/help	List all available slash commands.
/new	Summarize the current session, persist it, and start a fresh conversation.
/stop	Abort the current agent turn and clear queued work. (alias /kill)
/tasks	Show running and recent background tasks (read-only).
/cronjobs	Show configured cronjobs and their next run time. (alias /cron)
/model	Show or switch the provider and model — yours for regular users, the global one for admins. (alias /provider)
/thinking	Show or set the thinking level — yours for regular users, the global one for admins.
/compact	Summarize older messages to free up context. (/compact [instructions])
/context	Show how full the context window is, the model and where compaction starts.
/skill	Load a skill into the current conversation. (/skill:<name> [prompt])

On the web chat, /model (and its alias /provider) responds as an interactive button group: provider picker → model picker → confirmation. Selecting a model writes providers.json and switches the active model for the next turn.

/compact replaces the older part of the conversation with a summary and keeps the most recent messages — see Context Compaction. Optional instructions steer the summary (/compact keep the PR numbers). It waits for a running answer, shows a Context compacted (368k → ~41k tokens) divider you can expand to read the summary, and can be stopped with /stop. The ~ marks an estimate; with the next answer the divider switches to the size the provider reported. Automatic compactions show the same divider.

/thinking accepts off, minimal, low, medium, high, xhigh, or max. Levels the active model doesn't support are rounded to the nearest supported one. Like the brain-icon selector, it changes the global main-agent setting and applies to the next turn.

/skill:<name> injects a skill's SKILL.md into the conversation so the agent follows it from now on — the same instructions it would read when it picks the skill itself. Both self-created agent skills (by name) and enabled installed skills (by owner/name) can be loaded. Typing /skill: opens an autocomplete dropdown listing every loadable skill; pick one with ↑/↓ and Tab/Enter. Append a prompt on the same line (/skill:deploy ship v2 to staging) to load the skill and run the request in one go, or send just /skill:<name> and the agent confirms the skill is loaded and waits for your next message. A bare /skill shows a button picker. Unknown names reply with the list of available skills.

Unknown commands respond with a hint pointing at /help and are not sent to the agent. To send a literal message that starts with a slash to the agent anyway, prefix it with a second slash — inputs starting with // bypass the slash-command interception and are delivered to the agent verbatim.

Reconnecting

If the WebSocket drops (network blip, server restart), the connection dot turns amber and the client reconnects automatically with backoff. Messages you sent before the drop arrive when the socket is back. The Send button stays disabled while disconnected.

Leaving while the agent works

Closing the tab or locking your phone does not stop a reply: the backend finishes the turn, stores it, and shows it when you come back. With push notifications enabled, you also get a notification once the reply is ready — but only while no chat page of yours is visible.

Keyboard shortcuts

Shortcut	Action
Enter	Send message.
Shift+Enter	Newline in the composer.
Cmd/Ctrl+C	Copy selected message as Markdown.

See also

• Memory System — how session summaries, sessions, and reply context are persisted.
• Tasks & Cronjobs — the architecture behind the task-result and heartbeat cards.
• Settings → Agent — provider, model, language, reasoning level.
• Settings → Speech-to-Text and Text-to-Speech — voice input and output.
• Settings → Telegram — how Telegram-tagged messages get into the chat.
• App & Push Notifications — install Axiom on your phone and get notified when a reply is ready.