Sub-agents inherit full skill context, degrading performance

When Claude spawns sub-agents (via the Agent tool), each sub-agent receives the entire gstack context in its system prompt — all skill descriptions, instructions, and MCP tool listings. This
significantly bloats the sub-agent's context window, leaving less room for the actual task and degrading output quality.

Expected behavior:
Sub-agents should receive a minimal context — only what's relevant to their specific task. Skills and MCP tool listings that aren't needed by the sub-agent should be excluded or lazily loaded.

Current behavior:
Every sub-agent gets the full list of 50+ skill descriptions, MCP server instructions, and all deferred tool names, regardless of whether they're relevant. A simple "Explore" agent searching for a
file pattern gets the same gstack payload as the main conversation.

Impact:
- Sub-agents produce worse results due to context dilution
- Slower responses from processing unnecessary tokens
- Makes complex multi-agent workflows (parallel research, test runners) less reliable

Possible approaches:
- Filter system-reminder content before injecting into sub-agent prompts
- Only pass skills/tools that match the sub-agent type
- Lazy-load skill descriptions (keep only names, fetch full description on demand)