Auxiliary model system
DoloresCaritasAngelus/DSH-AUX · skills
DeepSeek Harness' auxiliary model system: unified aux-LLM routing (per-task model, timeouts, concurrency, failure cooldown, main model fallback) + visionanalyze/webextract/compresstext tools, settings page and ses.
★ 5GitHub stars
0Forks
2026-08-16Last updated
JavaScriptLanguage
MITLicense
Key features
- Unified auxiliary LLM routingeach type of task can be configured with independent model/timeout/concurrency; failure automatically downgrades the main model; continuous failure enters cooling; each call writes session events, which can be audited
- vision_analyzeImage analysis (supports multiple pictures in parallel): "What is in this picture?" "Read the chart values" "Compare the two pictures"
- Zero third-party dependencies at runtime: peerDependencies are all DSH official packages (included in the environment), no dependencies, and no need to install any additional third-party runtime libraries.
- Route parsingExplicitly configure the task default session main model; if the auxiliary model fails, the main model will be automatically downgraded.
- Three out-of-the-box toolsvision_analyze (image analysis), web_extract (web page extraction + summary), compress_text (long text compression)
Requirements
- Zero third-party dependencies during runtime: peerDependencies are all DSH official packages (included in the environment), without dependencies.
Install command
dsh plugin --profile web add github:DoloresCaritasAngelus/DSH-AUX