Variant limit overrides ignored for maxOutputTokens and compaction #48323

Description

When a model variant defines its own limit (e.g. a longer output cap), it is never used.
The variant is only merged into request options (session/llm/request.ts:80-91); a
nested variant.limit ends up as an inert key in provider options.

Why is this needed? Most models become more expensive after 272k token, i want to have variants to limit this without redefining it as a different model.

Both values derived from the output limit still use only the base model:

• maxOutputTokens sent to the provider (request.ts:129, via transform.ts:1468:
  Math.min(model.limit.output, outputTokenMax) || outputTokenMax)
• compaction reserved/usable (processor.ts step-finish -> overflow.ts:
  cfg.compaction?.reserved ?? min(20_000, maxOutputTokens(model)))

Suggested fix: resolve the effective output limit as variant -> model -> default before
clamping. maxOutputTokens already has everything it needs at request.ts:80 (the variant is
already resolved there); overflow.ts would additionally need the active variant passed in.

Sketch:

export function maxOutputTokens(model, outputTokenMax = OUTPUT_TOKEN_MAX, variant?) {
  const limit = variant?.limit?.output ?? model.limit.output
  return Math.min(limit, outputTokenMax) || outputTokenMax
}

Plugins

(none)

OpenCode version

1.18.30

Steps to reproduce

1. Define a model with a variant that has its own output limit:

   {
     "provider": {
       "myprovider": {
         "models": {
           "mymodel": {
             "variants": { "long": { "limit": { "output": 32000 } } }
           }
         }
       }
     }
   }

2. Use the model with the long variant.

3. Request is still capped at default output tokens; compaction reserved is also computed from default, not 32000.

Screenshot and/or share link

(n/a - request/token behavior, no visible symptom)

Operating System

Windows 11

Terminal

VS Code integrated terminal (opencode via VS Code CLI extension)