Pair

Pair Programming Skill

Launch a pair programming session where a second AI agent watches your work and provides real-time feedback on bugs, security issues, and missed edge cases.

Commands

• /pair or /pair start - Start with default backend (opus)
• /pair opus - Start with Claude Opus as pair (Task tool)
• /pair codex - Start with Codex as pair (background CLI)
• /pair stop - End the current session
• /pair status - Check if pairing is active

Starting a Session

When the user invokes /pair:

Step 1: Generate Session ID

SESSION_ID=$(openssl rand -hex 4)
echo "Session ID: $SESSION_ID"

Step 2: Start the Bridge Server

SESSION_ID="$SESSION_ID" pair-bridge start

Verify it started:

SESSION_ID="$SESSION_ID" pair-bridge status

Step 3: Spawn the Pair Agent

Determine the backend from the command (default is opus):

• /pair or /pair opus → Use Task tool
• /pair codex → Use Bash background process

For Opus (default): Use Task Tool

Spawn the pair agent using the Task tool with these parameters:

• subagent_type: "general-purpose"
• model: "opus"
• run_in_background: true
• prompt: The pair programmer prompt below

Pair Programmer Prompt for Task Tool:

You are a pair programmer watching another Claude agent work. Your job is to catch bugs, security issues, and missed edge cases.

## Connection
Session ID: {SESSION_ID}
Socket: /tmp/claude-pair-{SESSION_ID}.sock

## Your Loop

Run this continuously:

1. Wait for activity:
   ```bash
   SESSION_ID="{SESSION_ID}" pair-bridge wait $LAST_SEEN

This blocks until the main agent does something. Start with LAST_SEEN=0.

2. Parse the JSON response - it contains tool calls with:

   • tool: Which tool was used (Edit, Bash, Read, etc.)
   • input: The tool's parameters
   • output_summary: First 2KB of output
   • sequence: Incrementing ID (use as next LAST_SEEN)

3. Analyze for issues:

   • Bugs or logic errors
   • Security vulnerabilities
   • Missed edge cases
   • Breaking API changes
   • Missing test updates

4. If you spot an issue, emit feedback:

   SESSION_ID="{SESSION_ID}" pair-bridge emit feedback '{"severity":"high","message":"Description","context":{"file":"path.ts","line":42}}'
   

5. Update LAST_SEEN and loop back to step 1.

Severity Levels

• high: Stop and fix now (security, data loss, breaking changes)
• medium: Fix soon (missing error handling, test gaps)
• low: Consider later (minor style issues)

Guidelines

• Stay quiet on routine operations
• Quality over quantity - one good catch beats ten nitpicks
• Ask questions if intent is unclear
• Focus on what the main agent might miss

Begin watching now.

#### For Codex: Use Bash Background Process

Spawn Codex as a background process:

```bash
codex --quiet "You are a pair programmer. Session: {SESSION_ID}. Run: while true; do ACTIVITY=\$(SESSION_ID={SESSION_ID} pair-bridge wait \$LAST_SEEN); # analyze and emit feedback; done" &
CODEX_PID=$!
echo "Codex pair agent started with PID: $CODEX_PID"

Store the PID for later cleanup.

Step 4: Confirm to User

Tell the user:

• Session ID
• Which backend is running
• That their pair will provide feedback as they work

Stopping a Session

When the user says /pair stop:

1. Stop the bridge server:

   SESSION_ID="$SESSION_ID" pair-bridge stop
   

2. If using Codex, kill the background process:

   kill $CODEX_PID 2>/dev/null
   

3. Summarize the session (activity count, feedback given)

Checking Status

When the user says /pair status:

SESSION_ID="$SESSION_ID" pair-bridge status

Report:

• Whether pairing is active
• Session uptime
• Activity count
• Pending feedback count

How Feedback Appears

The hooks handle feedback injection automatically:

• pair-emit-activity.sh (PostToolUse): Sends your tool calls to the bridge
• pair-check-feedback.sh (PreToolUse): Shows pair feedback before your next action

When the pair spots an issue, you'll see:

<pair-feedback>
PAIR PROGRAMMER FEEDBACK:
This SQL query concatenates user input - use parameterized queries
Location: src/db/queries.ts:45
SEVERITY: HIGH - Address before continuing
</pair-feedback>

Example Session

User: /pair

Agent: Starting pair programming session...
  Session ID: a1b2c3d4
  Bridge: /tmp/claude-pair-a1b2c3d4.sock
  Backend: opus (Task tool)

[Spawns background Task with pair programmer prompt]

Pair programming active! I'll work on your task while my pair watches for issues.

User: Refactor the auth module to use JWT

Agent: [Reads files, makes edits...]

<pair-feedback>
PAIR PROGRAMMER FEEDBACK:
You removed the rate limiting check - this protected against brute force attacks.
Location: src/auth/login.ts:45
SEVERITY: HIGH - Address before continuing
</pair-feedback>

Agent: Good catch from my pair! Let me restore that rate limiting...

User: /pair stop

Agent: Ending pair programming session...
  Duration: 12 minutes
  Tool calls observed: 34
  Feedback provided: 2 (1 high, 1 medium)

Session ended.