Chat & Conversations
Starting a conversation
When you log in, Codumentor opens your most recent conversation. To start a fresh one, click New conversation at the top of the thread list (left sidebar).
Each conversation is independent — it has its own context and history. You can have as many as you like.
Sending messages
Type your message in the input area at the bottom and press Enter to send. For multi-line messages, press Shift+Enter to add a new line.
The assistant streams its response in real time. You'll see text appear as it's generated, along with any tool calls it makes.
Interrupting and redirecting
While the assistant is responding, the Send button changes to Stop. Click it (or press Enter with an empty input) to cancel the current response.
If you type plain text while the assistant is responding, Enqueue (or Enter) adds instructions to the current task without stopping its model call or tools. The assistant reads queued instructions, in order, before its next model call. A confirmation appears after the server accepts them. If the turn has already ended, the text stays in the input so you can send it normally.
While a turn is running and you have a draft, the send button has a dropdown arrow on its right. Choose Enqueue or Stop & Send there; choosing an option changes the button without sending. Click the main button (or press Enter) to perform the displayed action. Stop & Send cancels the current task and sends the draft as a new turn. Each new plain-text draft defaults to Enqueue; after the turn finishes, the button returns to Send. Messages with images or @-references use Stop & Send, and the dropdown explains why Enqueue is unavailable. Typed messages in voice mode keep their existing held-message behavior.
The steering queue is temporary, text-only, and limited to 32 pending instructions (16,000 characters each). Queued instructions are not separate conversation turns; stopping the task, an error, a server restart, or reaching the step limit can discard instructions not yet read.
Thread list
The left sidebar shows all conversations, including pinned conversations, ordered by your selected attention preference. Pinned conversations are also shown in a separate section at the bottom. Each thread shows:
- A title (auto-generated from the first message, or renamed by you)
- A time indicator (e.g., "5m ago", "Yesterday")
On a user-home instance (community: install once, run from any folder) the sidebar defaults to This folder — conversations tagged with the current workspace. All shows every conversation, including untagged ones from before this filter existed. Workspace-home servers have a single workspace, so the toggle is hidden.
Thread actions
Right-click a thread (or use the context menu) to:
- Pin / Unpin — pinned conversations appear in a separate section at the bottom of the sidebar
- Rename — give the conversation a custom title
- Delete — soft-delete the conversation (you can undo immediately)
On mobile, tap the hamburger menu icon in the header to open the thread list.
Navigating within a conversation
Scroll up to see earlier messages. A scroll-to-bottom button appears when you've scrolled away from the latest messages.
Deep links
Every conversation has a shareable URL like /c/abc123. You can bookmark these or share them (see Sharing).
You can also link directly to a specific message: /c/abc123/t/msg456 — this scrolls to and highlights that message when opened.
Tool calls in the transcript
When the assistant uses tools (searching code, running commands, writing files), you'll see collapsible cards in the transcript showing:
- The tool name and what it's doing
- The result or output
- Status (running, completed, or failed)
Click on a tool card to expand or collapse the details. When there are many tool calls in sequence, they're grouped together — click to expand the group.
Mermaid diagrams in assistant messages
Codumentor can render Mermaid code fences directly in assistant responses. For example:
flowchart TD
A[Start] --> B[Next Step]
Behavior notes:
- Mermaid rendering is available in the Web UI.
- During streaming, Mermaid blocks are shown as raw source first, then rendered after the response completes.
- If diagram rendering fails, the original Mermaid source remains visible so the content is not lost.
Max turns
The assistant has a configurable limit on how many tool-calling steps it can take per message. If it reaches this limit, you'll see a banner with a Continue button. Click it to let the assistant keep working.
Voice conversation
When your administrator has enabled Voice Mode and configured speech, a Start voice conversation button appears on the chat input. It opens a full-screen overlay: just start talking. The assistant waits until you have finished a request, then works and speaks the answer back. Captions stay on screen.
While it is working you can keep talking to refine the same task, change the goal, or say stop. Tool approvals still appear in the overlay as Approve / Deny (spoken yes/no is off unless your administrator turned it on).
This is different from push-to-talk (hold the mic button to dictate one message into the input) and from read aloud (play a finished message). Click End to leave the overlay; the conversation keeps the turns that were sent.
The browser will ask for microphone permission. Use a secure origin (HTTPS, or localhost). If the button is missing, Voice Mode is not available on this instance — ask your administrator.
Screen-free use and browser limits
Voice mode is a browser feature, not a native mobile background-call service. A hidden tab, another app, a locked screen, power saving, or a lost connection can suspend microphone capture, timers, or playback. Keep the page visible for reliable use. The screen wake lock is best-effort and does not guarantee locked-screen operation. A spoken warning is attempted when the page becomes hidden, but audio itself may already be suspended.
- Ask “what is the task status?” or “repeat that” for a conversational status or repetition. Say “stop the task” to cancel work; say “end the voice call” to close voice only, without cancelling work. Recognition and intent interpretation depend on the voice models.
- Approvals include the command/file and supplied prompt or Auto explanation. With
voice_approvals: false(the default), the task still waits for the on-screen Approve / Deny buttons. Voice mode never enables approvals automatically. Do not approve a command you do not understand; ask for clarification first. - Classic attempts spoken feedback when an utterance is lost, so you know to repeat it. Terminal failures use best-effort browser speech synthesis because the remote voice connection may have failed; browser voice availability and playback policies vary. Visible errors and Retry remain available.
- An idle hidden call checks in after five minutes, then closes after another minute without speech or a return to the page. Active agent work, ongoing speech, and pending approvals prevent idle closure. This ends only voice, not agent work. Browser timer suspension may delay the check. Existing server/provider session limits still apply and are not extended.
- Pause remains a screen control: in Live it mutes microphone and speaker, so a spoken resume cannot be heard. This is not a microphone-free or guaranteed locked-screen mode.