Proposal: configurable OpenAI-compatible service tier for llm-pi-ai

[yijigao]

Motivation

@deepseek-ai/dsh-llm-pi-ai can configure reasoning, transport, and cache behavior per provider, but there is no validated setting for the OpenAI-compatible service_tier request field. This prevents deployments from opting into priority or flex processing without patching the installed runtime.

Proposed shape

Add an optional provider-profile setting:

serviceTier: priority

Accepted values are auto, default, flex, and priority. The adapter forwards the value through pi-ai's OpenAI-compatible samplingParams as service_tier; protocols that do not consume sampling parameters ignore it. An unrestricted sampling-parameter dictionary is deliberately not exposed.

Patch and validation

A complete reference patch is available on my fork:

• Branch: https://github.com/yijigao/deepseek-harness/tree/contrib/pi-ai-service-tier
• Compare: master...yijigao:deepseek-harness:contrib/pi-ai-service-tier
• Commit: e2fc5513056088cd01afa3196e581320b1a75e06

It includes schema/type exports, adapter wiring, bilingual package docs, generated config catalogs, a bilingual Agent Note, and focused tests.

Validation completed on Windows x64:

• pi-ai adapter/config and related focused tests: passed
• pre-commit staged translation/lint/whitespace/vendor checks: passed
• full Host build and Client typecheck from a clean origin/master worktree: passed

I am sharing this as a Discussion because the repository currently does not accept external pull requests. Would this configuration surface fit the project's intended pi-ai abstraction?