Chats

Chats are the interactive work mode in d5s. A chat keeps one session while you and the agent shape the result through multiple turns.

A persistent agent can use that same durable session as an inbox for bound external channels. Slack, linked Telegram, and email events remain audience-attributed for routing and delivery, while the coworker's bounded model context and configured capabilities are shared across its connected surfaces. Everyone who can open the agent can review activity from its currently connected Slack channels, but sending, replying, or cancelling channel work still applies Slack identity and membership checks. A manual chat does not automatically enter an external reply context. Every agent-authored Slack message carries a visible Sent by an AI agent identifier in its attribution footer.

Telegram teammate DMs are a limited, workspace-enabled beta for private text messages. A teammate must confirm their Telegram identity in the d5s browser before their messages can reach the agent. The linked DM uses the named agent's existing persistent session, but does not add group conversations, files, general interactive buttons, streaming replies, or proactive Telegram sends. When an Ask-gated connector action pauses work, the agent can show a single-use approval control in that same linked DM. See Agents for setup and limits.

External delivery evidence is intentionally conservative. An accepted send keeps its normal sent/replied activity heading; an uncertain outcome is labeled Delivery unknown in both the step and its delivery detail. Accepted by Telegram means the provider accepted the request, not that a person received or read it. Delivery unknown means d5s cannot safely tell whether the provider accepted the message, so it will not blindly repeat an ambiguous send. A temporary refusal that the provider proves happened before acceptance may receive a bounded retry after current access is checked again.

You can start a chat from the workspace home without choosing a project. d5s creates an unlisted project for it automatically, with the workspace's active connectors attached by default. Choose an existing project when the work should immediately inherit that project's instructions, files, memory, skills, and connector choices, including any connectors it removed.

Where available, a dedicated Teams identity admits mentions from activated standard channels into the same agent conversation. The agent's ordinary response remains in d5s; posting to Teams requires its explicit messaging action. Installing the app alone does not activate every channel.

Use a chat when

• the question is exploratory or changing;
• you expect to review intermediate results;
• a tool action may need approval;
• the work should remain attached to one conversation;
• you want to prototype a prompt before making it repeatable.

During a run

The conversation streams the agent's progress rather than waiting for one opaque response. Depending on the task, you may see tool activity, sub-agents, research phases, generated artifacts, connector approval cards, and a request for access to an exact external API host.

When a chat sends work to a named workspace agent, it shows a handoff card with the sender, recipient, and exact request, then continues without waiting for the agent's separate run. The request enters the agent's inbox, and the agent's final answer appears in the chat later with the agent's name and avatar. That returned answer is context for future turns but does not start a new run by itself.

The approval control beside the + menu has two modes. Ask for approval is the default and pauses when a connector action resolves to Ask. Allow all persists for this session and lets Ask actions continue without individual approval cards. It does not override an explicit Deny, change connector credentials, or expand the allowed host, method, path, sandbox, or network boundaries. Switch back to Ask whenever you want the next Ask action to pause again.

If an approval card is already waiting when you choose Allow all, d5s retries that action once. A failed retry leaves the card available so you can review or deny it manually.

An external API request is not access by itself. When the workspace policy allows member grants, the card shows the exact host, its ordinary-method and all-path HTTPS API scope, and the risk that the service may retain data, perform external actions, or return prompt injection. Allow in this chat lasts only for that chat and can be revoked. Approve only a host you trust and never send credentials or other secrets.

Messages sent while work is active are durably admitted to the session. d5s may incorporate compatible new input into the current work cycle or show it as queued for a later turn. You can cancel a queued turn or stop active work without deleting the conversation.

To shorten the model context before its automatic limit, type /, choose /compact, then submit the selected command. d5s summarizes eligible older turns without adding a user message or a normal agent reply. The timeline changes from Compacting context to Context compacted when the saved summary is ready. The command is available only for an existing, idle session; if there is not enough older context, no model call starts.

Session uploads belong to this chat. Project files and memory belong to the reusable project context. Use Save as project when the hidden project behind a standalone chat should become a named, listed workspace resource.

For a persistent agent reached through Slack, attachments appear as references first. A request inside a Slack thread can also use accessible files shared earlier in that thread. The agent downloads only files it needs; downloaded copies remain available for later work and their contents remain untrusted. Ask the agent to attach generated files to its Slack reply. Attachment failures are reported even when the text reply succeeds.

Continue, branch, or repeat

Continue in the same chat when the new request depends on its history. Branch from a completed assistant response when you want to explore a different direction without changing the original thread. Start a new chat when the work should not inherit the old transcript.

On the web, use the thumbs-up or thumbs-down icon on a completed agent response to mark it as helpful or unhelpful. Selecting an icon immediately saves your rating and keeps it selected even if you skip or close the optional details dialog. You can add reasons and a comment later, switch directions, or remove your own feedback.

In the mobile app, use Copy to copy the response text. Thumbs-up marks a completed response as helpful; tap it again to clear that rating. Thumbs-down opens What was wrong? Choose Incorrect or unhelpful, optionally add details, then submit to save a negative rating.

To report concerning content in the mobile app, open thumbs-down and choose Offensive or unsafe content. You can report a response while it is still being written. Review the selected response, add optional details, and submit. The response excerpt and your details are shared privately with the d5s team for review. Avoid including passwords or unrelated personal information. Long responses include only the first 20,000 characters; use the details field to identify the concerning part.

A report is separate from your helpful or unhelpful rating. Changing or clearing a rating does not withdraw a report. The receipt confirms that the report was received, not that review is complete. If submission fails, keep the sheet open and try again.

The conversation shows only your own selection. It does not show teammate ratings, aggregate counts, or other people's feedback details. Feedback is not currently added to the agent's next-message context.

Mention teammates

When mentions are enabled for your workspace, type @ in the composer to mention another workspace member. The picker lists members by name, with online people first, and inserts the name as plain text; the mention becomes active when you send. You can only mention members who can already open the conversation: people without access to its project appear disabled in the picker, and mentioning never grants access. Share the project first if they need to see it.

A mention renders as the person's name in blue in the message and notifies them: the conversation pins to the top of their Recents list with an @ marker until they open it, and a notification toast appears if they are elsewhere in d5s. Opening the conversation clears the mention. Mentions of you are also highlighted, so the one highlighted name in a conversation is the one addressing you.

Agents can mention workspace members the same way in their replies — for example to flag a result for someone's review. An agent mention that targets someone without access to the conversation is shown as plain text and notifies no one.

When a conversation is connected to Slack, mentions bridge both ways for people who have linked their Slack account to d5s: a Slack mention of a linked teammate becomes a d5s mention, and a d5s mention in an agent reply delivered to Slack pings them there natively. Unlinked Slack accounts keep showing as plain display names.

Correct a sent message

When Message editing is enabled in Settings → Labs, select the pencil beside one of your own messages, change the text, then select Save & send. Cancel leaves the conversation unchanged.

Cancel any queued messages before saving an edit. If messages are still queued, the edit leaves them intact and keeps your correction in the editor.

Saving replaces that message and the conversation after it. The old text and dependent responses disappear from the chat and are excluded from future conversation context. If the agent is working, d5s stops it before submitting the replacement. Attached files stay attached.

Editing does not undo actions already completed, remove files the agent created, or reverse usage charges. You cannot edit messages received from external channels or replace a conversation tail containing someone else's messages or system instructions. Some older or branched messages cannot be edited if their original tool settings are unavailable. If saving fails, the editor keeps your correction so you can retry.

Manage a chat

Open Conversation actions from a manual chat in the project Recents list, the All sessions dialog, the Chats list, the History list, or the open-chat header. From there you can:

• Move to project to keep the transcript but attach future work to another project's context;
• Rename to change the conversation title; or
• Delete to permanently remove the conversation after confirming. Deletion cannot be undone.

These conversation actions apply to manual chats. Manage automation runs, dashboard runs, and agent conversations through their parent automation, dashboard, or agent instead.

Move stable patterns onward

Once the role or trigger becomes repeatable, the same project context can support a persistent agent, scheduled automation, or refreshable dashboard. Moving to another mode reuses the project; it does not turn the chat transcript itself into standing instructions unless you explicitly carry the relevant material forward.