Deepseek invokes Copilot AI agents as subagents instead of deepseek models

[MehmeDBDuran]

When in the plan mode, if the the deepseek invokes subagents, the copilot AI agents are used instead of the deepseek models. It consumes the copilot quota instead of Deepseek quota.

[Vizards]

I traced this through the VS Code source (commit 7d7e9513af). Three subagent models are currently hard-wired to Copilot's backend and bypass LanguageModelChatProvider:

• chat.tools.riskAssessment.model → hardcoded vendor: copilot
• chat.searchSubagent.model → uses ProxyAgenticEndpoint (Copilot agentic proxy, vscode-agentic-search-router-a)
• chat.executionSubagent.model → same pattern (exec-subagent-router-a)

These resolve endpoints directly rather than going through the provider API this extension implements, so subagents always hit Copilot infra even when the main agent is DeepSeek.

The VS Code team is actively working on BYOK — I expect these to become BYOK-routable once that lands.

[jaoyina]

What happens if I don't have copilot subscription? the sub agent task will fail? how the main agent handles this?

[Vizards]

What happens if I don't have copilot subscription? the sub agent task will fail? how the main agent handles this?

Based on the VS Code team's current BYOK progress, I expect we'll be able to override all three of those subagent models with BYOK models before June 1st.

If you want something that works right now, there's a workaround using a custom agent file with model: "DeepSeek V4 Flash" — the main agent can delegate simple tasks to it via the run_subagent tool. See #87 (comment) for an example. However, I haven't tested this in plan mode yet, so YMMV.