Feature: per-task-class model lanes with ordered fallbacks for swarm workers

[ianalitis]

Problem

Swarm workers get one default (agents.swarm_model) plus a per-spawn model override, and quota fallback is a single global list (provider.quota_fallback). A coordinator that routes by task class (cheap readers, routine implementation, hard iteration, long-context review, frontier review) has to repeat the model id in every spawn and cannot declare an ordered fallback per class. With model prices and availability changing weekly, those ids end up scattered across prompts and plans.

Proposal (small, opt-in)

[agents.lanes]
reader = ["opencode-go:mimo-v2.5", "opencode-go:deepseek-v4.1-flash"]
builder = ["opencode-go:deepseek-v4.1-flash", "opencode-go:glm-5.3-flash"]
reviewer = ["openai-oauth:gpt-5.6-terra"]

• The swarm tool accepts lane = "reader" as an alternative to model; it resolves to the first available route in the list.
• Moving to the next entry happens only before a spawn starts, is recorded in the spawn's metadata and is visible in the UI. There is no silent mid-turn substitution, which keeps the existing no-hidden-failover behavior.
• No [agents.lanes] table means the behavior is unchanged.

Why upstream

It's generic (any provider mix), small (config parsing plus one resolution step at spawn), and removes a class of prompt-level bookkeeping. I'm happy to implement it if the shape fits. Filing it as an issue first per CONTRIBUTING.

[1jehuang]

Thanks! The lane config shape is a design decision, so this is flagged for maintainer review before implementation. See also #1521, which asks for quota-aware routing on top of a similar fallback chain.

---

Jcode agent (automated triage), on behalf of @1jehuang