dsh-provider-passport

A generic OpenAI-compatible connectivity check can light up every green, yet the very first real task in DeepSeek Harness still fails because the endpoint rejects a subtle request-dialect choice: developer versus system, max_completion_tokens versus max_tokens, or flags like store, reasoning_effort, and stream_options.include_usage. dsh-provider-passport closes that last-mile gap. Before your first production call breaks, it runs a bounded, review-first preflight—no more than 10 fixed requests, 1 output token each, 12-second timeout—to discover the minimum compatibility profile the endpoint actually accepts, rather than leaving you to trial-and-error.

The workflow is deliberately conservative. The plugin proposes only the compatibility fields that Harness currently exposes (maxTokensField, supportsDeveloperRole, supportsStore, supportsReasoningEffort, supportsUsageInStreaming), lets you preview each change in the settings panel, and writes them to a single model of your choice. Immediately after, it verifies the result through the real Harness LLM runtime. If verification fails, you cancel, or Harness itself rejects a proposed field during atomic settings validation, the previous configuration is restored automatically and no dirty state is left behind. Credentials and custom headers never leave memory, and the redacted report strips out endpoint URLs, model IDs, request and response bodies before anything is copied.

This plugin is aimed at developers who run self-hosted, enterprise, or third-party OpenAI Chat Completions endpoints behind DeepSeek Harness for inference orchestration. It is not a general API conformance suite, a model-capability prober, or a routing proxy; it owns exactly one job—making sure the request shape Harness sends is one your endpoint will accept, and giving you a safe, auditable, roll-back-able path to get there.