hermes - 💡(How to fix) Fix feat: Enable DashScope explicit context caching for alibaba provider [1 participants]

Problem

The alibaba provider plugin sends system/user content as plain strings, which prevents DashScope from using its explicit context caching feature. This feature reduces input token costs by 90% (hit cost: 10% of standard price).

What is Needed

DashScope explicit cache requires the cache_control marker on content:

The alibaba provider currently sends content as a plain string. It needs a prepare_messages() method that:

1. Converts string content to array-of-dicts format
2. Adds cache_control to the last part of the system message

Reference

The qwen-oauth provider already implements this pattern (see plugins/model-providers/qwen-oauth/init.py). The same logic should be applied to the alibaba provider.

Test Results

Verified manually with qwen3.6-plus on International (Singapore) region:

Request 1: cached_tokens=0, cache_creation=1204 (cache created) Request 2: cached_tokens=1204, cache_creation=0 (90% discount) Request 3: cached_tokens=1204, cache_creation=0 (cache hit)

Pricing Impact

For a typical session with 10K input tokens of stable prefix:

• Without cache: 10,000 x $0.4/M = $0.004 per request
• With explicit cache: 10,000 x $0.04/M = $0.0004 per request
• Savings: ~90% on input tokens

Note

Implicit caching is NOT supported for qwen3.6-plus on International region, so explicit caching is the only way to get cache discounts.