ShippedCLI · Node.js

Adaptive Router for Agents

An open-source plugin set that picks a worker model tier for every coding-agent turn using deterministic local scoring instead of an LLM call. Ships as two independent plugins — one for OpenAI Codex, one for Claude Code — sharing the same routing design but adapted to each agent's own plugin, hook and CLI conventions.

Next projectEasyDPI

Adaptive Router for Agents · Semih Yıldız

Problem

Coding agents expose a single model choice for an entire session, so a one-line rename and a multi-file refactor cost the same. Routing per turn is the obvious fix, but the naive implementation — asking a model which model to use — adds latency and cost to every turn and is non-deterministic.

Approach

Routing happens in a UserPromptSubmit hook, entirely locally, with no network call:

User prompt
  → UserPromptSubmit hook
  → deterministic score + explicit override + session continuity
  → developer routing context
  → direct low-tier response OR root task calls run_routed_task exactly once
  → plugin-local agent CLI with pinned model and effort
  → worker result returned to the root task

The displayed root model never hot-swaps. For substantive routes the root task calls a local MCP worker that starts a model-pinned execution and returns its result, so the conversation stays coherent while the actual work runs on the tier it deserves.

Design decisions worth naming

Deterministic scoring, not a classifier — the same prompt always routes the same way, which makes the behaviour explainable and testable. No inference cost is added to routing itself.
Explicit overrides — slash controls (/luna, /terra, /sol, /auto, …) always beat the score, and an explicit choice is inherited conservatively across a conversation.
Turkish-aware normalization — prompts are normalized in memory for scoring without assuming ASCII input.
Sandboxed worker spawning — fixed model/effort mappings, no shell interpolation, approval: never, bounded sandbox choices. ADAPTIVE_MODEL_ROUTER_WORKER=1 bypasses both hooks inside workers to prevent recursion.
Privacy-limited telemetry — decision logs record the route, never prompt content and never prompt-derived hashes.
Graceful degradation — if a selected worker model is unavailable to the account, the coordinator continues on the current root model and reports a material fallback rather than failing the turn.

Stack

Runtime: Node.js 22+, ESM, cross-platform hooks (Windows PowerShell 5.1+, Bash on macOS/Linux)
Integration: MCP server for worker execution, agent-native hook and plugin conventions
Install: independent per-plugin installers; installing one does not affect the other

Notes

Unofficial community project, not affiliated with or endorsed by OpenAI or Anthropic.