Built on top of @codragraph/cli MCP tools (graph-aware code intelligence) and works with any inference provider (Claude, Codex, OpenCode, OpenAI, Anthropic, Gemini, ...).

Status

Developer preview. The package ships with single-proposer search, multi-role swarm search (Explorer + Exploiter + Critic), versioned recipe memory keyed on graph snapshots, and CLI / MCP entry points.

See RFC.md for the full design.

Concept

A harness is the code around a fixed base model that decides what to store, retrieve, and present at each step. Different harnesses produce different (accuracy, token-cost, latency) tradeoffs for the same task family.

codragraph-harness search runs an outer optimization loop:

Start with seed harnesses (zero-shot, few-shot, graph-aware).
Score each on a search-set of tasks → 3-vector (accuracy, tokens, latencyMs).
An agentic proposer (Claude Code by default) reads the filesystem of all prior candidates' source + traces + scores and writes new harness variants.
Each new harness is validated, scored, added to the Pareto frontier.
Loop for N iterations.
Return the non-dominated frontier.

Reference: Meta-Harness paper, arXiv 2603.28052.

Usage (planned)

codragraph-harness search \
  --task ./tasks/codebase-qa/ \
  --seeds zero-shot,few-shot,graph-aware \
  --iterations 20 \
  --proposer claude-code \
  --output ./runs/2026-04-29/

import { search } from "@codragraph/harness";

const frontier = await search({
  taskSet: "./tasks/codebase-qa/",
  iterations: 20,
  proposer: "claude-code",
});