subagent-isolation-guard_skill

Overview

This skill enforces physical workspace isolation for subagents and implements a semantic routing bypass to prevent routing logic from interfering with subagent sessions. It eliminates cross-agent context contamination and stops unintended model switching or session resets caused by parent-agent semantic routing.

How this skill works

The skill requires each subagent to use a dedicated agent directory so file I/O and workspaces never overlap. It also detects subagent sessions using a session_key marker (for example ":subagent:") and automatically bypasses semantic routing; when detected, the system returns a continue signal and avoids injecting declarations or applying model recommendations.

When to use it

• Deploying multiple subagents that each need their own filesystem workspace
• Preventing context bleed between independent agent instances
• Avoiding semantic-router logic from forcing model switches in child agents
• Running environments where subagents may be invoked recursively or by other agents
• Guaranteeing stable long-running subagent sessions without unexpected resets

Best practices

• Assign each subagent a unique, non-overlapping agentDir path
• Include a clear session identifier (e.g. ":subagent:") in every subagent session ID
• Maintain an explicit allowAgents list to control which subagents can be invoked
• Implement routing bypass early in the webhook or routing pipeline to avoid side effects
• Log bypass decisions and workspace assignments for auditing and debugging

Example use cases

• A parent agent dispatches specialized subagents (PM, architect) that must keep separate file workspaces to avoid write conflicts
• A webhook-based semantic router should skip policy enforcement for internal subagent sessions to prevent model changes
• Batch-processing pipelines where isolated subagents maintain persistent local caches without interference
• Testing scenarios where repeated subagent restarts must not clear their isolated context