dsh-llm-ai

In the DeepSeek Harness ecosystem, the LLM inference layer has been tightly coupled to pi-ai, constraining provider coverage, protocol extensibility, and failure diagnostics. dsh-llm-ai replaces that dependency entirely: it sources every provider and model fact from the community-maintained models.dev registry, executes requests through a native fetch + SSE streaming pipeline with zero pi-ai code paths, and remains fully compatible with the existing ctx.llm seam.

Three capabilities stand out. First, dynamic routing: the providers dictionary merges with user settings per provider, so adding, reshaping, or retiring a gateway takes effect on the very next request with no process restart. All 203 models.dev providers are declared in the configuration directory, honestly marked where a protocol family is not yet serviceable. Second, reasoning and compatibility control: tri- or quad-level reasoning declarations map each selectable level to its exact wire spelling; the three compat switches (maxTokensField, supportsDeveloperRole, thinkingFormat) resolve through a model → route → protocol-default fallback chain, and unknown keys are refused loudly rather than silently dropped. Third, robust runtime semantics: 14 stable error codes partition retryable from non-retryable faults, a stream idle watchdog guards against hung connections, image offload and raw-string tool-call arguments work out of the box, and usage plus cache-hit accounting are always on. The registry is cached to disk, so an offline boot still serves the last good snapshot.

This plugin fits the following situations: you are running DSH Desktop or the Harness 0.1.1-rc line and need multi-provider openai-completions inference without pi-ai protocol handcuffs; you operate in offline or low-bandwidth environments and still want a working model catalog; you want clearly classified error codes and credentials that never appear as plaintext in configuration files. If you are already on llm-pi-ai and it is stable for you, this plugin offers a non-breaking replacement path — credential references, route naming, and settings-layer semantics all remain backward-compatible.