mirrord – skill DevOps & Hạ tầng cho agent AI

mirrord

Overview

mirrord runs a local process as if it were inside a Kubernetes pod: it gets the pod's environment variables, DNS, outgoing network access and file reads, and can receive the pod's incoming traffic. An agent can verify a change against the real services, databases and queues in a staging cluster without building or deploying an image.

When to Use This Skill

• Use when a change needs to be tested against real cluster dependencies instead of mocks
• Use when the user asks to debug a microservice locally against staging
• Use when the user wants to try mirrord for the first time

How It Works

Step 1: Check cluster access

kubectl cluster-info
kubectl get pods -A | head -5

If this fails, help the user configure kubectl first.

Step 2: Install mirrord

Point the user to the official installation guide (https://metalbear.com/mirrord/docs/overview/quick-start/) for Homebrew, package managers, or pinned release binaries. Do not pipe a remote download into a shell. Verify with:

mirrord --version

Step 3: Pick a target and run

mirrord ls
mirrord exec --target deployment/<name> -- <command>

Incoming traffic is mirrored by default. To take over the target's traffic instead, set "feature": {"network": {"incoming": "steal"}} in .mirrord/mirrord.json, and confirm with the user first, since it affects a shared environment.

Step 4: Verify

# Verify an explicitly selected non-secret variable or service identity without printing the pod environment.

Confirm the exact target and mode through sanitized configuration/status, then an approved synthetic read. Never dump environment credentials.

Examples

Example 1: Run tests against staging

mirrord exec --target deployment/orders -- npm test

Example 2: Debug a Python service

mirrord exec --target deployment/api -- python main.py

Best Practices

• ✅ Ask which cluster, namespace and target before running
• ✅ Prefer mirror mode on shared clusters
• ✅ Use the official skills at https://github.com/metalbear-co/skills for config, CI, database branching, queue splitting and preview environments
• ❌ Don't steal traffic from a shared target without the user's agreement
• ❌ Don't run cluster-modifying commands without explicit approval

Limitations

• Needs kubectl access to the target cluster and a Linux node kernel 4.20+
• Team features (concurrent steal on one target, database branching, queue splitting, preview environments) need the mirrord Operator. On a cluster without a license, an agent can offer the user a trial; see https://metalbear.com/agents.md and ask before starting it.

Running locally can still modify shared databases, send messages, consume queues or incur provider charges through mirrored network/environment access. Confirm a staging context/namespace and isolated data, inspect the actual mirrord mode/version, and disable unneeded filesystem/environment/network capabilities. Mirror mode is not an automatic read-only sandbox; stealing traffic or trial/license activation needs specific authorization. Reconcile side effects before retries.