Skill: slaude — Slack-Mirrored Claude Code Sessions

Skill: slaude — Slack-Mirrored Claude Code Sessions

When to Use

Use this skill when asked to start, stop, join, or manage Claude Code sessions mirrored to Slack via slaude. Typical triggers:

• "Start a Claude session in #channel"
• "Mirror this to Slack"
• "Join that Slack thread with a new agent"
• "Resume the session"
• "Stop/quit the session"
• "Kill this session" / "End the session" / "Stop yourself"
• "List running sessions" / "What sessions are active"

Prerequisites

• slaude binary must be installed and in PATH (or built locally from the slagent repo)
• Slack credentials configured via slaude auth
• Target Slack channel must be accessible to the authenticated user

Verify setup:

slaude status
slaude version           # check installed version

The version is also injected into the agent's system prompt.

Starting a Session

Start a new Claude Code session mirrored to a Slack thread:

slaude start #channel "your prompt here"
slaude start https://team.slack.com/archives/C123 "your prompt here"
slaude start @user "your prompt here"

Key flags:

Flag	Description
[target]	Positional: Slack URL, #channel, @user, or channel ID
--locked	Lock to owner only (default for start)
--observe	Observe mode: read all messages, only respond to authorized users
--open	Open for all thread participants
--debug	Write raw JSON events for troubleshooting
--no-bye	Skip goodbye message on exit
--dangerous-auto-approve green	Auto-approve read-only operations (overrides classifier.yaml)
--dangerous-auto-approve yellow	Auto-approve local writes too (overrides classifier.yaml)
--dangerous-auto-approve-network known	Auto-approve network to known hosts (overrides classifier.yaml)
--dangerous-auto-approve-network any	Auto-approve all network access (overrides classifier.yaml)

Access mode flags (--locked, --observe, --open) are mutually exclusive. When no flag is given — interactive terminal prompts Closed, oBserve, or open? [cBo]; non-interactive: start defaults to --locked, join/resume default to --observe.

Everything after -- passes directly to Claude Code. Use this for Claude-specific flags:

slaude start #channel -- --permission-mode plan "design the API"
slaude start #channel -- --model sonnet "quick fix"

Process management: slaude start runs as a foreground process. It blocks until the session ends. To manage it as a sub-agent:

• The process PID can be captured: slaude start ... & echo $!
• Send SIGINT to gracefully stop: kill -INT $PID
• The process exits with code 0 on normal completion

Session Management

List running sessions

slaude ps

Prints a table of all active slaude sessions on this machine:

PID     EMOJI                 CHANNEL           UPTIME
------  --------------------  ----------------  ------
12345   :fox_face: (fox)      #engineering      5m30s
67890   :dog:                 #backend          1m12s

Kill a session

slaude kill fox           # by emoji shortcode
slaude kill :fox_face:    # by emoji with colons
slaude kill 12345         # by PID

Sends SIGINT to the target session. First SIGINT interrupts the current Claude turn; the session then waits for new input or exits gracefully.

Self-termination: A running slaude session can kill itself by running slaude kill with its own emoji or PID. This is useful when Claude is asked to "stop", "quit", or "end the session" and wants to act autonomously:

# Claude finds its own session and kills it
MY_EMOJI=$(slaude ps | awk 'NR>2 && $2 ~ /fox/ {print $2}')
slaude kill "$MY_EMOJI"

Or more directly, if you know the emoji:

slaude kill fox

Stopping a Session

There are four ways to stop a running session:

1. From Slack (recommended)

Type in the Slack thread:

• quit — ends the session (owner only)
• stop — interrupts the current turn but keeps the session alive (anyone)
• :emoji: quit — quit a specific instance in multi-agent threads

2. Via slaude kill (for self-termination or remote stop)

slaude kill fox       # stop session with emoji "fox"
slaude kill 12345     # stop session by PID

Use slaude ps first to find the target emoji or PID.

3. Via signal

Send SIGINT directly to the slaude process:

kill -INT $PID

First SIGINT interrupts the current Claude turn. The session then waits for new input or exits.

4. Close stdin

If running interactively, Ctrl+C or closing stdin will end the session.

Joining an Existing Thread

Add a new agent instance to an existing Slack thread:

slaude join https://team.slack.com/archives/C123/p1234567890 "help with tests"

• Each instance gets a unique emoji identity (e.g. fox, dog, koala)
• Defaults to --observe (non-interactive): reads all messages, responds only to authorized users
• Use --locked to start locked to owner only; --open to open for everyone

Resuming a Session

Resume a previously exited session:

slaude resume https://team.slack.com/archives/C123/p1234567890#fox@1700000005.000000 -- --resume SESSION_ID

The resume URL (including #instanceID@lastTS fragment) and session ID are printed when a session exits.

Multi-Instance Threads

Multiple slaude instances can share one Slack thread. Each gets a unique emoji.

Addressing instances from Slack:

• :fox_face: do something — addressed to the fox instance only
• :dog: /compact — send a command to the dog instance only
• Messages without prefix — broadcast to all instances

Thread Access Control

Command	Effect
:emoji: /open	Open thread for everyone
:emoji: /open @user1 @user2	Allow specific users (additive)
:emoji: /lock	Lock to owner only (resets all, disables observe)
:emoji: /lock @user	Ban a specific user
:emoji: /observe	Toggle observe mode on/off

Observe mode — an orthogonal flag on top of locked/open/selective access:

• Off (default): non-authorized messages filtered out, agent ignores them
• On: all messages delivered for passive learning; agent still only responds to authorized users; non-authorized users get an ephemeral "not authorized" if they try to interact directly

Thread title encodes access state:

• 🔒🧵 Topic — locked (owner only)
• 👀🧵 Topic — observe (locked + reading all)
• 🧵 Topic — open for all
• 🧵 @user1 @user2 Topic — selective (specific users)
• 👀🧵 @user1 @user2 Topic — selective + observe

/lock and /open (without args) disable observe. Each instance manages access independently — /open and /lock on a joined/resumed instance only affect that instance's in-memory state.

Sub-Agent Management (for OpenClaw)

When managing slaude sessions as sub-agents from OpenClaw:

Lifecycle Pattern

# 1. Start session as background process, capture PID
slaude start "$CHANNEL" --open \
  --dangerous-auto-approve green \
  --dangerous-auto-approve-network known \
  -- "$PROMPT" &
SLAUDE_PID=$!

# 2. Monitor — process exits when session completes
wait $SLAUDE_PID
EXIT_CODE=$?

# 3. Or stop early via signal
kill -INT $SLAUDE_PID

Key Integration Points

1. No external API — slaude is process-local. Control is via:

   • Slack messages in the thread (quit, stop)
   • Unix signals to the process (SIGINT)
   • slaude kill <emoji|PID> — from the orchestrator or from within the session itself

2. Session output — On exit, slaude prints resume info to stderr:

   To resume: slaude resume URL#id@ts -- --resume SESSION_ID
   

   Capture stderr to extract session ID and resume URL for later resumption.

3. Thread URL — Printed to stderr on startup. Capture to track which Slack thread maps to which sub-agent.

4. Multiple agents per thread — Use slaude join to add more agents to the same thread. Each gets an independent emoji identity.

5. Auto-approve for unattended operation — For sub-agents running without human oversight, consider:

   --dangerous-auto-approve yellow \
   --dangerous-auto-approve-network known
   

   This auto-approves local reads/writes and known-host network access. Unknown hosts and destructive operations still require Slack approval.

Orchestration Example

# Start agent A on backend work
slaude start #engineering --open -- "implement the API endpoint for /users" &
PID_A=$!

# Start agent B on frontend work in the same thread
THREAD_URL="https://team.slack.com/archives/C123/p1234567890"
slaude join "$THREAD_URL" "build the React component for user list" &
PID_B=$!

# Wait for both to finish
wait $PID_A $PID_B

Health Checking

# Check if session is still running (via slaude ps)
slaude ps

# Or check by PID directly
kill -0 $SLAUDE_PID 2>/dev/null && echo "running" || echo "exited"

Troubleshooting

Issue	Fix
slaude: command not found	Install via brew install sttts/slagent/slaude or go install github.com/sttts/slagent/cmd/slaude@latest
No channels listed	Run slaude auth to set up credentials
Permission requests stuck	Check Slack thread for ✅/❌ reaction prompts
Wrong emoji syntax	Use :shortcode: (Slack shortcode syntax), not Unicode emoji
Session won't quit	Only the session owner can quit. Others can only stop (interrupt).