feat: `sub` hot swapping

[gkgoat1]

Problem

sub mode currently just reroutes traffic, which can interfere with people like me with actual, valid, Claude Code subscriptions

Proposed solution

I want to have sub mode fall back to another provider (OpenRouter, Codex, etc.) when Claude hits limits or otherwise fails, and potentially also support making any provider be substitutable with another.

Token / quality impact

It trims thinking blocks--it has to--so tokens have to be reduced. The idea I had is mostly around having a bunch of subscriptions that hot swap when limits hit, to reduce agent transfer churn.

Are you willing to contribute this?

• I'd like to implement this myself and open a PR
• I'm reporting the idea; happy for a maintainer or someone else to build it

[fkiene]

Most of this landed in v0.10.0.

Fallback instead of always-on reroute. llmtrim sub mode fallback keeps Anthropic as the backend that serves your turns; the reroute only fires when Anthropic can't serve one — quota, payment, throttling, overload, or a transient 5xx. Anthropic is retried once first. (The old spelling on-error still works and now prints a notice.)

Hot swapping across several subscriptions. That's the new part: llmtrim sub chain codex,kimi (also LLMTRIM_SUB_CHAIN=codex,kimi or [sub] chain = ["codex", "kimi"]). On a failed turn each provider in the chain is tried in order, with bounded backoff that honors Retry-After, until one answers. A failure reported inside an HTTP 200 stream counts as a failure too, and is caught before any of it reaches Claude Code — so the swap happens mid-turn rather than surfacing a broken stream to you. With no chain set, the active sub provider is the chain, so nothing changes for existing setups.

A few reroute bugs that made the fallback path unusable went out in the same release: Codex thinking blocks no longer 502 (signatures now map both ways), the Codex prompt cache no longer goes cold every turn, and the status line stopped claiming a reroute that never happened — in fallback mode it now shows the backend that actually served the turn, read from the ledger, not the one in your config.

What isn't there yet: OpenRouter, and arbitrary provider-to-provider substitution. The chain is limited to the providers sub knows (codex, kimi). Leaving this open for those two.