dylan121322/llm-adaptive

README excerpt

llm-adaptive Adaptive model routing plugin for DeepSeek Harness. Adds an adaptive provider to the model picker: every LLM request is classified by a flash classifier (low / medium / high / critical) and routed to the matching backend provider through config-driven chains. Features - Per-request complexity classification — deepseek-v4-flash called directly (never through a proxy, no recursion). - Context-aware judging — injects a rolling session-goal summary plus the recent turns into the classifier prompt (continuation / wrap-up / error-loop rules). - Sticky level protection — a mid-task downgrade is held at the previous level unless the message carries explicit downgrade or wrap-up signals. - Config-driven routing chains — chains come from pool.json → routing.levels ( $active expands to the active provider, missing entries fall back to defaults); transport failures walk down the chain. - Classifier config from the pool — URL / model / key reference read from the classifier section of pool.json (no hardcoded credentials). - Fail-open — any classification failure degrades to medium ; never blocks a request. - Observable — every decision (level, cause: llm/sticky/cache) is written to…

Category