As a nabledge-6 user, I want skills to run in separate contexts so that main conversation remains focused and efficient

[kiyotis]

Situation

nabledge-6 skills currently execute in the main agent context, consuming significant context window space:

• Large-scale file searches (Grep across hundreds of files)
• Multiple file reads for code analysis
• Complex handler chain analysis with dependency tracking
• All intermediate results and verbose outputs remain in main context

This context bloat degrades response quality, latency, and cost efficiency.

Pain

Users:

• Slower responses as context fills with search results
• Degraded answer quality due to context pollution
• Higher costs from excessive token consumption
• Difficulty maintaining conversation focus

Developers:

• Hard to debug context-related issues
• Complex context management in skill design
• Inability to run resource-intensive analysis without side effects

Benefit

Users:

• Fast, focused responses with clean context
• High-quality answers from unpolluted reasoning space
• Lower costs through efficient token usage
• Clear conversation flow without intermediate noise

Developers:

• Clean separation of analysis vs. presentation
• Easier debugging with isolated execution contexts
• Maintainable skills with predictable context behavior
• Ability to add resource-intensive features without hesitation

Success Criteria

Context Isolation

• nabledge-6 skill workflows execute in separate context on both Claude Code and GitHub Copilot
• Main conversation context contains only summary results, not intermediate search/analysis output
• Context token usage for typical handler search reduced by 80%+ compared to current implementation

Cross-Platform Support

• Single maintenance point for workflow logic (no duplication between platforms)
• Both Claude Code and GitHub Copilot users can invoke separate context execution
• Output format and quality consistent across platforms

Maintenance Efficiency

• Changes to workflow logic require updating single source file
• No platform-specific workflow duplication
• Testing process verifies both platforms work correctly

User Experience

• Claude Code: Automatic delegation to separate context (no user action required)
• GitHub Copilot: Clear invocation method documented and easy to use
• Response quality measurably improved through context pollution reduction

---

Research Findings (Reference)

Claude Code

• Task tool: Can delegate to custom agents in separate contexts
• Agent location: .claude/agents/name.md
• Invocation: Automatic via Task tool with subagent_type parameter
• Built-in agents: Explore, Plan, general-purpose

GitHub Copilot

• /agent command: Manual invocation of specialized agents
• Agent location: .github/agents/name.agent.md
• Invocation: User types /agent name or copilot --agent=name
• Built-in agents: Explore, Task, general-purpose, Code-review
• Context isolation: Agents run in separate context (documented)

Platform Comparison

Aspect	Claude Code	GitHub Copilot
Separate context	✅ Automatic	✅ When /agent used
Custom agents	✅ .claude/agents/	✅ .github/agents/
Invocation method	Automatic delegation	Manual /agent command
Context preservation	✅ Yes	✅ Yes

Key Constraint

GitHub Copilot limitation: Skills cannot programmatically invoke agents. Users must manually use /agent command.

This creates tension between:

• Ease of use: Claude Code auto-delegates, GHC requires manual step
• Maintenance: Need consistent workflow logic across platforms
• User experience: Different invocation patterns between platforms