feat(agent): add VS Code Copilot multi-agent support and phase-based model routing

[Snakeblack]

Pre-flight Checklist

• I have searched existing issues and this is not a duplicate
• I understand that PRs will be rejected if the linked issue does not have status:approved

🔍 Affected Area

Other (Agent / VS Code Copilot Integration)

💡 Problem Statement

Currently, Gentle AI does not support multi-agent orchestration or phase-based model routing when running inside VS Code Copilot. In internal/agents/vscode/adapter.go, the capability is explicitly disabled:

func (a *Adapter) SupportsSubAgents() bool {
	return false
}

While other environments (like Claude Code) allow configuring different models per phase in SDD workflows by rewriting placeholders in localized agent files (~/.claude/agents/*.md), the VS Code Copilot integration lacks any multi-agent capability.

Historically, a workaround suggested for this limitation was to use Copilot subscriptions inside alternative clients like OpenCode or Pi Agent (which already support these workflows). However, in many corporate environments, security and compliance policies strictly prohibit the use of unofficial clients like OpenCode or Pi Agent. Dev teams in these enterprises are mandated to standardize exclusively on the official VS Code Copilot extension.

Without native multi-agent support in the official VS Code Copilot extension, dev teams in compliance-restricted companies cannot adopt or use Gentle AI's multi-phase workflows (like SDD) effectively. This limits Gentle AI's adoption in the enterprise space.

We need a way to:

1. Optimize Costs: Running simple mechanical tasks like code generation/applying (sdd-apply) or testing/verification (sdd-verify) on cheaper models instead of wasting premium credits.
2. Improve Accuracy: Restricting instructions to a single, monolithic agent context instead of routing specialized phases to subagents configured with focused instructions and agent skills.

📦 Proposed Solution

We propose implementing multi-agent support in the VS Code Copilot adapter by enabling SupportsSubAgents() = true and managing dynamic model routing at the orchestrator layer.

Since VS Code Copilot extensions are compiled and distributed as static plugins, we cannot dynamically rewrite local agent files like in Claude Code. However, VS Code Copilot's runtime execution engine supports dynamic overrides: when calling subagents via the underlying runSubagent tool/API, the model can be overridden by passing a dynamic model parameter.

To support this:

1. Enable SupportsSubAgents() = true in the VS Code adapter.
2. Add support to resolve model mappings per phase (e.g., from openspec/config.yaml or workspace configurations).
3. Have the orchestrator agent inject the target model parameter dynamically when delegating tasks to subagents via the tool call.

📎 Additional Context

For details on how VS Code Copilot supports agent plugins, custom agents, subagents, and skills, see the official documentation (ordered by relevance):

• VS Code Copilot - Subagents (Architectural reference for orchestration and subagent delegation)
• VS Code Copilot Customization - Custom Agents (Defining custom agents inside the VS Code ecosystem)
• VS Code Copilot Customization - Agent Plugins (Packaging and bundling custom agents as plugins)
• VS Code Copilot Customization - Agent Skills (Specifying targeted capabilities per agent/subagent)
• VS Code Copilot Customization - Custom Instructions (Managing workspace-level rules)

[Alan-TheGentleman]

Closing as duplicate of #504, which already has status:approved and an open PR (#505) implementing exactly this scope: enable SupportsSubAgents() = true in the VS Code adapter, ship 10 embedded .agent.md templates, and wire phase-based model routing via the named-profile mapping (provider/model → "Display Name (copilot)").

Recommend tracking the work on #504 / #505. If you've spotted a piece of scope that #505 misses, drop a comment there with the specific gap — easier than fork-tracking two issues. Thanks!