FallbackAdapter: allow a separate attempt timeout for fallback LLMs

[roshaans]

Problem

llm.FallbackAdapter applies a single attempt_timeout to every provider, so a fallback inherits the primary's latency target. For voice agents the two needs conflict:

• The primary's timeout should stay tight so a slow turn fails over quickly and the caller isn't left in silence.
• A fallback typically runs cold — no warmed prompt cache, no session affinity — so its time-to-first-token is meaningfully higher than the primary's, and it needs a longer window to actually serve.

With one shared value there is no good setting: raising it to accommodate the fallback slows every ordinary failover, while keeping it tight regularly cuts off a healthy fallback mid-attempt — turning a one-provider blip into a fully failed request and marking a working provider unavailable. We hit exactly this in production.

Proposal

An optional fallback_attempt_timeout on FallbackAdapter.__init__, applying to every LLM after the first (serving attempts and recovery checks alike):

FallbackAdapter([primary, fallback], attempt_timeout=5.0, fallback_attempt_timeout=8.0)

Default None keeps today's behavior (attempt_timeout applies to all providers), so the change is fully backward compatible.

Presumably the same reasoning applies to the STT/TTS fallback adapters, but this proposal scopes to LLM where the cold-start gap is largest.

Implementation

Draft PR with the change and tests: #6915. Happy to adjust scope/naming based on discussion here.

🤖 Generated with Claude Code