Feature Description

Add an explicit auxiliary.compression.context_length config option.

Example:

auxiliary:
  compression:
    provider: custom
    model: gpt-5.4
    base_url: https://example-proxy/v1
    context_length: 500000

Motivation

Today, compression runtime context length is inferred indirectly:

• from model.context_length
• or from custom_providers[].models[*].context_length
• or from provider metadata probing
• or from fallback defaults

That works, but it is awkward when the compression model is treated as a first-class runtime of its own.

This is especially useful for custom providers where /models may not expose a context_length field, or may expose model IDs that do not exactly match the configured runtime.

The recent bug fix in #8786 makes override handling more consistent, but an explicit field would still improve clarity and reduce ambiguity.

Proposed Solution

Support a direct override for the compression auxiliary runtime:

auxiliary:
  compression:
    context_length: 500000

Suggested precedence:

1. auxiliary.compression.context_length
2. explicit model/runtime override (model.context_length or matching custom_providers[].models[*].context_length)
3. provider metadata probing
4. default fallback

Benefits

• makes compression runtime configuration explicit
• avoids relying on provider metadata for custom endpoints
• improves discoverability for users configuring large-context compression models
• decouples compression runtime context from the main runtime when they differ

Related

• Bug: Bug: compression runtime context_length overrides are not applied consistently #8785
• Fix PR: fix: unify context length overrides for compression runtimes #8786