llm-adaptive

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 the plugin log.
• 120s decision cache — keyed by user-text head plus goal fingerprint.

Requirements

• DeepSeek Harness (dsh)
• A model pool file at ~/.dsh/tools/cc-switch-sync/pool.json with:
• classifier section: url, model, key_ref (resolved against ~/.dsh/.credentials.yaml, pool api_key as fallback)
• routing.levels: low / medium / high / critical chains
• A DeepSeek API key for the classifier

The pool file is produced by the cc-switch-sync import tool (or can be authored by hand). The plugin reads it on every request, so pool edits take effect immediately.

Install

dsh plugin add llm-adaptive

or, from a local checkout:

cd ~/.dsh/profiles/web && npx pnpm@10 install   # with "llm-adaptive": "file:plugins/llm-adaptive"

Restart the dsh web service, then select adaptive（自动路由） in the /model picker.

Usage

1. Open /model and choose adaptive（自动路由）.
2. Every subsequent LLM request is classified (low/medium/high/critical) and routed to the first available provider of that level's chain.
3. Decisions are logged with level=… cause=… chain=… to ~/.dsh/hooks/plugin.log.

The explicit level models (low, medium, high, critical) are also listed in the picker for direct selection.

How it works

A custom LlmAdapter for the adaptive provider: stream() awaits classification (async generator), then forwards to the target backend via ctx.llm.prepareCall + stream (unified chunk protocol, passthrough). Request-level interception was chosen over proxy or request-layer hooks because dsh hot-swaps configuration and the prepared-call contract requires matching provider/model options.

License

MIT