Context

Once the single-file markdown format for sub-agents lands (#647) and multi-model provider architecture exists (#648), the natural next step is letting each sub-agent declare its preferred model inside its frontmatter. Example:

---
name: summarizer
description: Summarize documents and content concisely
tools: [file_read]
model: local-qwen-small    # or "anthropic/claude-haiku"
timeoutSeconds: 60
---

You are a summarization specialist...

Today the only knob is modelRole: Compaction | Main | Fallback, which routes through the three hard-coded roles in NetclawChatClientProvider. That's too coarse: "Compaction" doesn't mean anything semantic about which model — it means "whatever we happened to wire to the Compaction role". Once there are multiple models in play, agents should be able to name their preferred provider/model directly.

Blocked on

• feat(providers): concurrent multi-model / multi-provider architecture for per-role and per-agent routing #648 — concurrent multi-model / multi-provider architecture. Without that, this issue can't be implemented meaningfully; there's only one model in the daemon.

Also gated on

• feat(subagents): single-file markdown format + contextual prompt + project-scoped discovery #647 — single-file markdown format for sub-agents. The model: field lives in frontmatter, which this issue assumes as the authoring surface.

Proposal

Add an optional model: field to sub-agent frontmatter. Fall back to modelRole when absent (backward compatible). When both are set, model wins.

Schema:

• model: string — a name that resolves to a concrete chat client via the registry from feat(providers): concurrent multi-model / multi-provider architecture for per-role and per-agent routing #648
• If unresolved, fail loudly on startup (not silently fall back)

Related

• feat(subagents): single-file markdown format + contextual prompt + project-scoped discovery #647 — single-file markdown format + contextual prompt (prerequisite for the authoring surface)
• feat(providers): concurrent multi-model / multi-provider architecture for per-role and per-agent routing #648 — concurrent multi-model architecture (prerequisite for routing)