Skip to main content
AI SRE is available to accounts on an On-call Pro or higher subscription, with no application needed. Starting September 16, 2026, 08:00 Beijing time, it must be activated in the console before use and is billed on actual usage; see billing terms.

​
Overview

Incident response usually happens right inside your IM group — alerts come in, the team discusses in the channel, and someone spins up a war room. AI SRE puts the agent directly into that collaboration chain: you never have to switch to the console; you can summon it from IM to investigate, and your teammates can follow every step of its analysis in real time. AI SRE’s IM integration has two trigger modes:

@Mention Summon

@mention AI SRE in a group chat or DM and describe the problem to start or continue an investigation. It replies in-thread, and the conversation context is tied to that IM session.

War Room Auto-Diagnosis

When an IM war room is opened for an incident, AI SRE runs a preliminary diagnosis and posts the findings back to the war room if the integration keeps Automatically start AI incident analysis enabled by default — no one needs to manually summon it.

​
Supported IM Platforms

AI SRE’s IM interaction covers four major platforms. Each platform supports inbound @mentions (webhook), historical message reads, and outbound replies: A “processing receipt” is the immediate signal the bot gives once it has your message; how it looks on each platform is described under Processing Receipt.
IM interaction requires that you have already connected the corresponding platform’s bot in Flashduty (the same IM bot used for alert notifications and collaboration). Complete the bot configuration in Flashduty’s IM integration first — only then can AI SRE send and receive messages on that platform.

​
AI SRE Settings in IM Integrations

Open the corresponding IM integration under On-call → Integration Center → Integration List → Instant Messaging and configure its AI SRE behavior under Enhanced features. The following switches are on by default:
  • Automatically start AI incident analysis: available after you enable War Room. When the integration creates a war room, AI SRE starts one preliminary diagnosis and posts the result back to the room.
  • Allow AI SRE conversations in IM: available without enabling War Room. It controls both entry points — group @mentions and direct chats with the bot; in Slack it also controls the /fd command. Turning it off makes group @mentions and direct chats with the bot (and, in Slack, /fd) silent. In WeCom’s Third-party App mode, only direct chats with the bot are available; group @mentions require Custom App Integration mode.
  • Use thread replies in normal group chats: available only for Feishu/Lark and Slack. When enabled, each normal-group thread has an independent session; War Rooms still receive replies directly in the group.
If AI SRE is not enabled for the account, these switches do not take effect. After AI SRE is enabled, each IM integration follows its own settings.

​
@Mention Summon

In a connected IM group where Allow AI SRE conversations in IM is enabled, @mention the bot and type your question (for example, “@AI SRE check why the payment service’s 5xx rate is spiking”) — the message is forwarded to AI SRE for processing via the platform webhook:
1

Detect the mention

The platform distinguishes between group @mentions and direct messages. Once mentioned, AI SRE deduplicates the event, puts a “processing” receipt on that message (see Processing Receipt), and determines whether the message contains an imperative instruction.
2

Bind to a session

AI SRE locates a session using “account + platform + chat” as the key: multiple @mentions within the same IM chat are routed to the same agent session, preserving context. For a new chat, the platform supplies the opening message of that IM thread and any associated incident as initial context.
3

Read context and reply

AI SRE reads the thread / chat history to build context, investigates autonomously, and posts its findings in-thread. Other people @mentioned in the original message are carried over into the reply, making it easy for multiple parties to collaborate in the group.
The reply mode is configurable (off / first / all), controlling whether AI SRE @mentions the person who asked, and whether it replies in-thread or in the main channel. In a busy large group, in-thread replies keep the investigation discussion focused without flooding the channel.

​
Processing Receipt

Once the bot has your message, it first puts an emoji reaction on the message you sent, as an immediate “received, working on it” signal. The emoji is picked at random from that platform’s own pool rather than being fixed: On Slack this sits alongside a native “is working on it” status, which the bot also shows. The reaction is removed once the answer is delivered. A single failed removal (a platform hiccup or rate limit) does not leave the emoji on the message forever: the next successful delivery retries the removal with the exact emoji that was added, so stuck receipts are eventually cleared. WeCom has no emoji API, so its “processing” receipt is a placeholder message (a random canned acknowledgement such as “On it…”), which the task checklist card and the final answer then replace in place.

​
How a Reply Reaches the Chat

AI SRE’s replies in IM are delivered only through the reply tool — ordinary assistant text is private working text and is never posted automatically. Once you hand it an investigation, every message that appears in the group is one it deliberately submitted, and it is one of two kinds:
  • Result: an answer to your concrete question, or an actionable interim finding.
  • Progress: a one-or-two-sentence update while work is still going on — confirmed progress, what it is currently waiting on, a real blocker. A progress message is not the answer: it does not finish the turn, and the result is delivered separately as usual.
If a turn ends naturally with the conclusion written as ordinary text instead of submitted through reply, the system adds one automatic delivery-only correction: that correction exposes reply alone and runs no investigative or mutating operations; if nothing has been submitted even then, the turn ends with an error — that draft never appears in the group. Non-natural endings — a cancellation, an interruption by a new message, a dispatched subagent, a wait on a subagent’s reply, a truncated output — do not trigger this correction. This delivery contract applies only to human-initiated IM requests: console sessions still answer with streaming text, and war-room auto-diagnosis is an unattended background turn that also keeps native streaming output.
A single reply also has a soft reading-length budget: when it is too long, reply refuses it once and suggests trimming the repetition, or publishing the detailed evidence and comparisons as an artifact and sending only the key conclusions plus the artifact link into the chat. That is why a long analysis in IM usually arrives as “conclusions + artifact link” rather than one very long message — see Artifacts. If you explicitly ask to have the full report pasted directly into the chat, this constraint does not apply.

​
War Room Auto-Diagnosis

When you open a war room for an incident in IM and the integration keeps Automatically start AI incident analysis enabled, AI SRE intervenes automatically — no one needs to @mention it:
1

Create the war room

Create a war room (Lark / DingTalk / WeCom / Slack group) as part of the incident collaboration workflow.
2

Trigger diagnosis in the background

Once the war room is created, the platform triggers a preliminary diagnosis as a non-blocking background task, so AI SRE enters the investigation with the full context of that incident.
3

Post the findings

When the diagnosis is complete, AI SRE sends the analysis results as a message back to the war room. Before anyone has started investigating manually, the first-cut analysis is already in the group.
War room auto-diagnosis is bound to the incident context: when the investigation begins, the corresponding incident_id is attached to the run so AI SRE can read the incident details, timeline, and recent changes. For details on how incidents / war rooms interact with A2A, see Agent · Incident & War Room Integration.

​
Standing Tasks and the Monitoring Card

When the agent’s current turn ends but a standing process-type task it started is still running (e.g. a monitor watcher or a background command), AI SRE posts a monitoring card to the chat:
Each standing task gets one line: tasks with a deadline show the remaining minutes (under one minute rounds up to one minute); tasks without a deadline show “indefinitely”. The card is re-posted whenever the standing-task list changes.
While any standing task is alive, the IM session’s root message stays open (the incident view is not closed): later turns triggered by task notifications keep posting into the same chat — no re-@mention needed.
Notification rounds follow a silent “no message = no news” semantics: if a notification round has nothing new to deliver, AI SRE closes that round silently — no placeholder receipt is posted to the chat, and the monitoring card is not re-posted either. A new message appears in the chat only when there is a real new finding.

​
Progress Messages During a Long Investigation

When an investigation runs long, the chat can stay quiet for a while — interim prose is not delivered, and the monitoring card only appears once a standing task exists. So when an IM request has been silent for about 60 seconds with no public reply submitted, the system may add a transient reminder telling the model how long you have already been waiting, and lets the model decide whether to post a progress message:
  • Reminder cadence: two reminders for the same request are at least 3 minutes apart, and one uninterrupted silence stretch gets at most two reminders. Any successfully submitted reply (result or progress) opens a new silence stretch and resets the count, but does not shorten that 3-minute minimum gap.
  • Where the waiting time starts: at the moment your message entered the platform’s pending queue, so queueing time counts — not when the agent actually started working on it.
  • A reminder sends nothing by itself: it only tells the model how long you have waited; it never writes anything to the group on the model’s behalf. When the wait was already explained and nothing has changed, or the model is waiting on information from you, sending nothing is the correct choice.
  • A progress message is not the answer: it says “still investigating, what it is stuck on, what it checks next”, does not finish the turn, and the result follows separately.
The “silence” here is not the same thing as the “no message = no news” rule above: that rule is about a notification round with nothing new, which posts no placeholder receipt; this one is about a person waiting while the agent has posted nothing public yet.
Waiting reminders appear only on attended IM requests; the console, API, and automation channels have none.

​
Connections and Authorization

When a credential is missing, the console renders inline cards (“Authorize [resource name] to continue”, “Connect [vendor] to continue” — see Console). IM and API channels get no card — the agent hands the matter to you according to this session’s channel, as follows: On the automation channel nobody is watching and no reply will ever arrive: the agent does not wait and does not phrase anything as a question — it records the missing connection or authorization as a blocker in its final report and delivers whatever the evidence already supports.

​
In-session Switch Commands

IM sessions have no picker UI like the console, but they support two slash commands that let you switch bindings mid-conversation — neither command is available for console sessions, whose environment and team bindings are set at creation and cannot be changed afterwards.

​
/env — Switch Runner Environment

Send /env <target> to the bot in a group or DM to rebind the current IM session to a different execution environment. The conversation history and context are preserved in full. After the switch, the previous environment’s working directory is no longer accessible: files created or modified during the conversation in the old environment are gone, and skill and knowledge files are re-staged in the new environment on demand. <target> takes three forms: Send /env with no argument and the bot lists every target currently available: shared environments, your teams’ environments, and the cloud sandbox (the default sandbox plus any named templates). If a named sandbox template is not found, the bot replies “No cloud sandbox template named “x”. Use /env sandbox for the default sandbox.” along with the list of available templates.
The switch takes effect after the current turn completes and before the next message is processed — it does not interrupt a tool call that is already running. If the target runner is offline or not found, the command returns an error immediately and the binding remains unchanged.
On-premises deployments have no cloud sandbox. On-premises installs run no cloud sandbox infrastructure, so /env shows no sandbox option, no sandbox entries in its list, and no sandbox wording — you can only switch between BYOC runners and Auto.

​
/scope — Switch Team Scope

Send /scope <team-name> (or /scope personal to return to personal scope) to rebind the current IM session to a different team without ending the conversation. Team binding is a usage operation: any account member can bind a session to any team in the account — this matches the war-room pattern where responders are often not members of the incident’s team. After the switch, AI SRE:
  1. Refreshes the memory snapshot for the new team scope (memories from the previous team scope no longer apply to this session).
  2. Queues the new team’s knowledge bundle to be mounted and injected into context when the next message is processed.
  3. Re-checks whether the session’s pinned execution environment is still available: if the pin is a BYOC runner or cloud sandbox template that belongs exclusively to another team, the execution environment is reset to Auto and the bot replies ""x” does not belong to the current team; the runtime environment has been reset to Auto.” Shared and account-level environments are unaffected, and an inconclusive check (for example, the template list cannot be fetched) leaves the pin untouched.
Knowledge bundles already mounted into this session are not removed — mounts are conversation-scoped, so switching back to a previously-mounted team will not duplicate the mount reminder.

​
/feedback — Rate an AI SRE Reply

Once an AI SRE reply lands in a group chat, anyone in the group can rate it — not only whoever started the investigation. Send /feedback to the bot to rate or comment on the reply you are looking at:
  • Entry and resolution: /feedback is backed by the AI SRE side (POST /safari/session/latest-reply → POST /safari/feedback/create); it resolves “the message you are looking at” into the corresponding reply event even if you never took part in that conversation.
  • Eligibility: not limited to session members — any account member who can use AI SRE may rate (read-only group-chat readers included); an AI SRE answer posted into a group chat is public, and the reader who spotted the error is exactly whose opinion we want.
  • How to rate: choose like / dislike / none; a comment-only submission (no rating) is allowed and does not wipe an earlier rating you gave, while an explicit none clears the existing rating.
  • Stored per rater: each reply is keyed by (reply event, rater) — when several people rate the same reply, each gets its own row and they do not overwrite one another.
  • Console side: the console conversation header also offers thumbs-up feedback on replies (see Feedback), sharing the same rating system as IM /feedback.

​
Relationship to Console Sessions

Whether you start a session from IM or the console, it’s the same AI SRE: messages are handled one at a time, with streaming output, automatic context compaction for long conversations, and optional team and environment bindings. The only difference is the entry point —
  • Console: ask questions one at a time in the Chat workspace and view the full tool-call and sub-session panels; environment and team bindings are fixed at creation and cannot be changed.
  • IM: @mention it in the group where your team already works; findings are posted back to the thread. Use /env and /scope to switch bindings mid-conversation. Great for grabbing a quick analysis at the incident scene, then diving deeper in the console.
For more about sessions, see Conversations.

Overview

Learn about AI SRE’s overall capabilities, typical use cases, and console navigation.

Console

Learn about sessions, streaming output, cancellation, and context compaction.

Agent

Inbound Agent Card and incident / war room integration.

Usage Insights

Use /insight to review the past 30 days of sessions and surface operational friction.