Overview
A session is one complete conversation between you and AI SRE. It holds every message you send, the agent’s streaming replies, tool calls made along the way, and any artifacts the agent produces (such as code, reports, charts, or Skill archives). Each session is isolated, with its own context, bound team, and environment. You switch between sessions in the left sidebar, and exchange messages, read replies, and view artifacts in the central chat area. Switching sessions does not interrupt a running turn; when you come back, the streaming output is still there.Creating and managing sessions
Click New Chat in the left sidebar to start a session. The list is sorted by recent activity; click Show more for older sessions. A new session shows four suggestion cards above the input: Investigate an incident, Query data with natural language, Reduce alert noise and Schedule an automated inspection. Clicking a card fills in a prompt you can edit before sending.Search and filter
The search box filters by session name. Click the Filter icon at the top right of the list to combine these filters (a dot on the icon means filters are active):Session visibility and permissions
Sessions are never accessible across accounts. Within an account:
Pinning affects only your own list. Account Owners and admins can manage team sessions but cannot read other members’ personal sessions.
Sharing a session
If you can manage a session, click Copy share link in the chat header.- Only members signed in to the same account can open the link; anonymous access is not supported. The link stays the same while sharing is on.
- People with the link see the whole session read-only (including subagent details) and cannot continue it. They can click Fork into a new session to continue in their own session.
- Click Remove share to invalidate the link immediately; sharing again creates a new link.
- Personal sessions are visible only to their creator by default, so a share link is how other members of the account can view one.
Incognito sessions cannot be shared. Sharing is enabled on the root session only; subagent (Subagent / A2A) sessions cannot be shared on their own.
Per-session actions
A new session gets an automatic title after its first turn; you can rename it at any time.
List status indicators
Session credit spend
The credits badge next to the input shows what this session has spent; hover it to see the account balance and this period’s pay-as-you-go usage:- Only members who can continue the session see these figures: read-only collaborators and anyone arriving through a share link do not see what the session cost.
- Billing rules, the trial grant, packages and the pay-as-you-go switch are in AI SRE overview · Activation and billing.
Sending messages and streaming responses
Type a message in the input and press Enter. The input supports Markdown; type/ for the Skill and command menu. When a new console version is released, a notice above the input asks you to refresh.
Predicted completion
When a turn ends and you leave the input empty, the input shows grey text predicting what you might send next. That text is not your input yet: it fills the input only when you accept it, and you can ignore it and keep typing. A prediction is requested once, at the moment the turn ends, so opening an existing session shows none, and none is generated once the session’s last reply is more than 1 hour old; a session with no agent reply yet gets none either. The grey text disappears as soon as what you type no longer matches it and comes back when you clear the input. Each prediction is an extra model request that consumes tokens.- Accept: with the caret at the end of the input, press Tab or →. The whole prediction fills the input with the caret at the end.
- Dismiss: press Esc. While you are editing an earlier message for a rewind, Esc cancels the rewind instead.
- Send: sending exactly the predicted text counts as accepting; sending anything else counts as ignoring.
Attachments and context references
Uploading attachments
Uploading attachments
Click the plus button at the bottom left of the input, or paste a screenshot.
- Supported types: images, PDF, text / Markdown / CSV / HTML, Office documents, archives (
.zip/.rar), and any UTF-8 text file (such as.go,.py,.yaml). Empty files are rejected. - Size and count: up to 50 MB per file; up to 9 attachments per message, 50 MB in total.
- Images: an image whose longest side is above 8192 px, whose shortest side is below 11 px, or whose aspect ratio exceeds 200:1 is scaled down, padded with white, and re-encoded, so its stored format and file extension can change (for example a GIF is stored as PNG and keeps only its first frame). Such an image above 40 megapixels is rejected; send a smaller one. A corrupt image or an unrecognized format variant fails to upload. Images inside a folder are stored as sent.
- Folders: choose Upload folder from the plus menu or drag a folder onto the input. A folder can hold up to 50 files (
node_modulesand dot-files are not counted). A larger folder is rejected as a whole, and individual files that do not qualify are skipped with the reason listed. A folder counts as 1 attachment and keeps its directory structure. - How the agent reads them: attachments are provided as text. For text over 64 KB only the beginning is provided directly, and the agent reads the full file from the sandbox when needed. Archives and folders are not expanded up front; the agent unpacks and reads them in the sandbox as needed.
Context references
Context references
You can open AI SRE from the AI entry on an incident, alert, alert rule, monitored object, data source, or RUM issue detail page (for example AI analysis on an incident, AI fix on a RUM issue). AI SRE then adds that object to the input as a reference capsule and sends it to the agent with your message. Click a single-object capsule to open that object, or remove it before sending.These entry points usually pre-fill a question as well (for example, asking the agent to separate confirmed facts from inferences); edit it as you like before sending. No question is pre-filled when several references are brought in at once.To reference several incidents at once, select them in the incident list and click AI analysis in the action bar. The selection becomes one Incident list (N) capsule in the input, with a pre-filled batch question: how the incidents relate to each other and how to group them, then the likely root cause, impact, and remediation priority per group. After you send the message, clicking that capsule opens the list: each row shows a severity dot, the ID number, the title, and the progress (Triggered / Processing / Closed), and clicking a row opens that incident in a new tab. The Status when sent label at the top of the list means these statuses are a snapshot taken when the message was sent; later changes to an incident do not update the list.Entering from the Let AI tune noise reduction button in the header of Channel details → Noise reduction → Alert grouping / Flapping detection also adds that channel to the input as a reference capsule and pre-fills a question that starts with
/channel-noise. This is the only entry that can start /channel-noise: typing it in an IM, automation, or API session is rejected. See Skills.In any session, type @ to search incidents and insert a reference. Typing an email address does not trigger the search.Real-time streaming output
The agent’s reply streams in real time, including text, tool calls and reasoning. The run state is kept if you switch pages and come back.While a turn runs, the Send button becomes a Stop button. Stopping interrupts the turn immediately and marks it as interrupted.
Queueing messages while running
While the current turn is executing, you can keep sending messages; they queue and run in order after the current turn. Queued messages appear above the input, where you can edit, remove or Clear all of them. Once the agent has started on a message it can no longer be withdrawn; use Stop to interrupt. Once the agent has handed the work to a subagent and its own turn has ended, the session still counts as running (the list still shows the spinning circle, and the button beside the input is still Stop), but the main agent is idle: a message you send then does not queue, it starts a new turn immediately. The same holds after a page refresh, so the session is never mistaken for “executing” and your messages are not held back.Environment initialization
The first time a session runs, an environment initialization card shows the sandbox being prepared and then collapses to one line:
Sessions on a self-hosted BYOC Runner show only the step of connecting to the runner.
Tool calls and artifacts
Tools the agent calls (reading and writing files, querying monitoring, running commands, MCP tools, and so on) are collapsed in the conversation by default; click to see input and output.Todo list
For tasks with three or more steps, the agent shows a progress badge (such as “Step X / N”). Click it to see the task list with each step’s status and priority. If the turn ends with a step unfinished, that step shows as paused and continues only after you send a new message. Todo lists are available in console and IM sessions only, not in automation, API or subagent sessions.Agent questions
When the agent needs clarification, it adds a question card — single choice, multiple choice or free text. You can use the keyboard (↑↓ to choose, Enter to confirm, Esc to skip); optional questions can be skipped with ✕. If you send a new message instead of answering, the question is treated as skipped and your message runs immediately. Answering a question that was already skipped or answered elsewhere sends your answer as a regular message.Operating console pages
Some console pages declare that AI can operate them: each turn sends the agent the actions available on the page you are viewing, so on these pages the agent can read and fill in the form directly instead of only giving written advice. The console pages that declare page actions:
Other pages have no page actions, so the agent can only give written advice. Page actions exist only in console sessions, not in IM, API, or automation sessions.
Actions on the Alert grouping and Flapping detection pages work like this:
- Reading the form (fields, current values, options, whether writable) needs no authorization.
- Filling a field only writes the value into the form; it does not submit. Review it and click Save yourself.
- Saving, and other actions that take effect immediately, each pop up their own confirmation card in the conversation (above the input), listing “old value → new value” with Cancel and the action’s own verb (such as Save).
When authorization is required
When a tool or MCP call needs a credential or OAuth authorization, an “Authorize [resource name] to continue” card appears:- Secret-based: click the button, paste the API key or token and save; the task continues automatically.
- OAuth-based: click Authorize, complete the third-party authorization, then click Continue task. If the authorization link expires, start the task again.
Only the console shows the authorization card. In IM and API sessions the agent posts the OAuth link; finish authorizing in a browser and reply that you have authorized. Secrets must be entered in this same session in the web console — never paste a secret into the chat. Automation sessions do not wait and list the missing authorization as a blocker. See IM Platforms.
When a connection is required
You might ask the agent to use a vendor that is in the plugin marketplace but not yet connected, for example “connect Datadog for me.” A “Connect [vendor] to continue” card then appears, so you don’t have to find the vendor in the marketplace first.- Click Connect and fill in credentials and run the connection test in the dialog.
- Once connected, the card shows “[vendor] connected”; click Continue task and the agent resumes the original task.
Only the console shows the connection card. In IM and API sessions the agent posts a link; connect and reply “continue”. Automation sessions do not wait and list the missing connection as a blocker. See IM Platforms.
Subagents
When the agent delegates work, a subagent line appears in the conversation. It shows the name, execution environment, status, and duration. After the run, it also shows tool calls, token usage, and any failure reason. The line carries no type badge; the type (Agent / A2A) appears in the title of the side panel. Click it to open the subagent session side by side; it can be expanded to full screen. A running subagent can also report progress to the main agent. A report appears in the main conversation as a Subagent report line: collapsed it is just a robot icon and the title, and clicking it shows the full text, starting with the subagent’s name (generic names such asagent and general are not shown) and followed by the body as an Output block. While the subagent is still running, the report first appears as the same line under that dispatch; once the permanent record settles, this line disappears, so a report never shows twice. After a page refresh only the permanent record remains.
A subagent may also need you:
- Awaiting authorization: click Authorize, then Continue task; the subagent resumes the original task.
- Awaiting input: read and answer the subagent’s question in the task panel.
Task panel and background tasks
Besides subagents, the agent can run long commands in the background (builds, bulk data pulls) or start monitors that watch a condition until it is met or a deadline passes. Click the Tasks button in the chat header to see all tasks in the session, grouped as Needs you / Running / Finished. The number on the button counts tasks that need you or are running, and turns red when something needs you. You can also click the line in the conversation that started a task to open its details.- Subagent details show the subagent’s transcript.
- Background command and monitor details show status, run time, deadline, exit code, the full command and an event timeline.
- Stop task or Stop Subagent in the detail header stops only that task, not the main session.
Artifacts preview
Click a file the agent produced to preview it on the right:
Previews support Copy and Download. All artifacts can also be managed on the Artifacts page in the left navigation — see Artifacts.
Status page update drafts
The agent can draft an announcement for a public status page: a new incident or maintenance event, or an update to an existing one. Drafts are never published automatically — you review and publish them in the console, because publishing notifies subscribers immediately and cannot be undone. The draft card shows exactly what would be published: title, status or schedule, the Markdown message, affected components and their status, and the target status page and event type. If the message uses formatting the status page does not support (such as headings, code blocks, quotes or images), the card lists it, because it would be lost on publishing. Click Review → to open the status page publishing form in a new tab with the draft filled in. After publishing, the button becomes Published and links to the published event; if the draft has expired or is missing, ask the agent to draft it again.IM and API sessions show no card; the agent puts a “Review and publish” link at the start of the reply and summarizes the draft. If saving the draft failed, the agent summarizes it in text and asks you to check it in the console.
Message actions
Hover over a message to:Editing an earlier message rewinds the session: everything after it is replaced, and related background tasks and subagents are stopped. You cannot edit while that turn is still running or while the session is busy (for example, compacting); try again shortly.
Forking a session
After a turn completes, click Fork next to the reply. Choose the new session’s scope (personal or a team you can access) and runtime environment, then confirm to create and open it.- The new session contains the conversation and tool calls up to that turn; from then on the two sessions are independent.
- If the source team or environment is not available to you, the new session uses personal scope or Auto instead.
- Completed or failed subagents are copied with their results; subagents still running at fork time show as interrupted and can be started again.
- You can fork only from completed turns of a main session, not from subagent sessions.
Session feedback
Use Helpful or Not helpful in the chat header to rate the whole session. With Not helpful you can pick reasons (wrong diagnosis / root cause, did not find the problem, unusable remediation, off-topic) and add a comment.Context compaction
As a conversation grows, AI SRE automatically summarizes older history while keeping recent content and key conclusions, so the conversation carries on. Content folded out of context, such as Skills or MCP tool definitions, is reloaded by the agent when needed.
Progress is shown while compacting; afterwards the Context percentage in the chat header drops (hover for token usage).
Compaction only replaces what is sent to the model; the original messages and tool results stay stored. When an investigation depends on an exact value (an ID, a host, a metric value, a threshold, a log line, the text of an error, a command), the agent looks the original text up instead of repeating the summary’s wording; if the value really is not there, it says so plainly rather than guessing. Images and attachments leave only a placeholder with no original text, and a truncated large output keeps only the file path of its full result.
Choosing a runtime environment
When creating a session, choose the runtime environment in the input area to decide where the agent’s tools, Skills and MCP run: Auto (default), the cloud sandbox, or an online self-hosted BYOC Runner. For the selection order, offline handling, and why the environment is fixed once the session is created, see Environments · Selecting an environment in a session.Binding a team
When creating a session you can bind a team; the session then loads that team’s knowledge, Skills and MCP. Your choice is remembered. In a personal session without a team, the agent can load knowledge from your teams when needed; a team session uses only shared content and that team’s content. For scope and permission rules, see Context.Session source
Each session records where it came from, which affects some behavior:Reply language
The reply language is set when the session is created and used for the whole session.Integrating sessions through the API
This section is for developers integrating AI SRE sessions through the Open API.Session run state
BothPOST /safari/session/get and POST /safari/session/list return two run-state fields, and they mean different things:
After the agent hands work to a subagent, its own turn ends first: during that window
is_running stays true while turn_in_flight is false. Read turn_in_flight to know whether a message sent now will queue, and is_running to know whether the session is still working. Both are computed in the response, so they describe the moment you query.
Withdrawing a queued message
Thepending_messages field in the POST /safari/session/get response lists queued messages the agent has not read yet, in execution order. Before the agent reads one, its author can withdraw it with POST /safari/session/message/cancel, passing session_id and invocation_id (the invocation ID returned when the message was sent). The result field means:
Only regular human messages can be withdrawn, not question answers or messages produced internally. The caller needs permission to chat in the session.
Error markers in stream frames
Frames in the session stream may carryis_error and error_code, marking a runtime error the platform surfaces to the user. A frame without them may still contain diagnostics; they also differ from the error line that marks a broken export stream below.
Session data export
POST /safari/session/export streams all events from a session as NDJSON (application/x-ndjson), one JSON object per line, for archiving, offline analysis, or feeding session data into external systems. See Export session transcript for the endpoint.
- The first line is always
session_meta(session metadata), followed by the session events. Each inlined child session also starts with its ownsession_metaline, which carriesparent_session_id. - Each line’s
typenames the event type; all values:session_meta,user_message,system_reminder(system-injected context, not user input),tool_call,subagent_dispatch,final_answer,llm_call,agent_text,session_fork,model_error(a model turn that failed with an error code: the code is incontent, the message inerror, andstatusiserror), anderror, which only ever appears as the last line of a stream (see the Warning below).seqstarts at 1 and restarts for each child session. - A tool call’s request and response are separate lines. Pair them by
call_id, not by line order. A remote A2A agent’s results are exported astool_calllines too, and such a line may have no matching request line. - When a model writes prose and calls tools in the same response, the prose comes out as its own
agent_textline first, followed by that response’s tool-call lines — the order you saw in the session. usageholds the event’s token counts and appears exactly once per event: on its own line for a text-only event, and on the firsttool_call(orsubagent_dispatch) line for a response that carried tool calls.tsis the event’s write time (RFC 3339). The export doesn’t include the serving model identifier.
Related pages
Overview
Learn about AI SRE’s positioning, capabilities, and use cases.
Usage Insights
Generate team incident-handling and operational insights from session data.
Manage Knowledge
Provide domain knowledge to sessions, loaded by team; personal sessions can mount their creator’s teams’ knowledge on demand.
Skills
Reusable skills invoked through slash commands.
MCP (External Tools)
Connect external systems through MCP to extend the agent’s tool capabilities.
IM Platform
Mention the agent in Slack / Feishu / DingTalk / WeCom to troubleshoot, with automatic war-room diagnosis.