Releases · EmotionG/dsh-llm-config

Multi-provider LLM configuration for the DeepSeek Harness LLM seam, now on DSH 0.2.

Requires DSH 0.2.0-rc.2..0.2.x (peerDependencies are >=0.2.0-rc.2 <0.3.0; 0.1.x runtimes are not supported).

Highlights since 0.1.0

DSH 0.2.0-rc.2 support — the plugin moved to the 0.2 message and settings models: the profile entry is the settings namespace, providers is declared .volatile() and re-synced on loader/volatile-update, and the hand-written page is kept with settings.configure({ auto: false }).
OpenAI Responses is a real wire format — openai-responses now sends input items, instructions, max_output_tokens and flattened tool declarations to /responses, and translates the response.* event family by output_index. It no longer reuses the Chat Completions body and event vocabulary.
Tool results as first-class role: 'tool' messages, and a history that may lead with a system message, are handled per protocol (Anthropic merges consecutive results into one user turn; Gemini resolves the function name a result answers). developer tool-change messages are refused with UNSUPPORTED_CONTENT.
Image input resolves through the framework attachment service into a full ImageRequestTarget; offloaded images are projected to placeholder text before the image gate.
Per-provider retry policy (retry: <n> or { mode: always }) resolved through the framework's own resolveRetryPolicy.
Model requests go through the global fetch, so interceptors and proxies see them.

Install

dsh plugin --profile web add github:EmotionG/dsh-llm-config

Or use the prebuilt tarball attached below — it skips the build-approval step.

Verification

npm test covers smoke / client / adapter / transport / mount / sync / channel / rpc / enrich / runtime / compat / responses.

Navigation Menu