support different model for subagent

[paul2xf]

What feature would you like to see?

k3 cost too much，please allow user to set the subagent model to k2.7

Additional information

No response

[Alyosha28]

+1 with a concrete use case and evidence from the current build (0.26.0).

Use case — cost-routed orchestration. Main agent on kimi-code/k3 for planning and hard reasoning; swarm subtasks like web fetches, page summarization, data extraction, and translation are simple, self-contained legs that a cheap external model (e.g. DeepSeek via an openai-type provider alias) handles fine. Today that is impossible inside one session:

• [subagent] in config.toml only exposes timeout_ms — no model field (docs)
• Agent / AgentSwarm workers always inherit the session model
• Model switching exists only at session level (-m, /model, default_model)

What would unblock it (any of):

1. [subagent] model = "<alias>" as a default, or per-type bindings like [subagent.explore] model = "..." (as proposed in per-workspace model bindings for subagent types/子agent指定模型与思考强度功能 #1927)
2. A model parameter on the Agent / AgentSwarm tool so the orchestrator can choose per dispatch
3. A stable, documented mid-session model-switch API — ACP already has unstable_setSessionModel (session/set_model); a hook/script-callable equivalent would let the community build routers

Prior art: Claude Code supports per-subagent model: frontmatter in agent definitions, plus a smallFastModel setting for auxiliary tasks.

Workaround in use today: dispatching simple legs to headless kimi -m deepseek/v4-pro -p "<task>" processes from the main agent. It works, but loses session-context sharing, permission inheritance, wire records, and the unified task UI.

中文摘要：希望能给 subagent / swarm worker 单独指定模型 alias——主 Agent 用 k3 编排，简单的抓取/摘要/抽取子任务跑便宜的第三方模型。目前 [subagent] 只有 timeout_ms，模型只能在会话级别切换。