Embedding metaharness In A Host Runtime

The optimization kernel is already optimize_harness(...) in metaharness.api. Host agent runtimes should call that kernel. They should not grow a proprietary proposer backend inside this repository.

Three Host Adapters, One Kernel

1. Python import. An in-process host (Omnigent worker, AgentSky runner, or a custom loop) imports optimize_harness, points it at a baseline workspace, and reads OptimizeResult.
2. Omnigent skill. An Omnigent agent wraps metaharness run as a skill. See examples/embed_runtime/omnigent_skill/SKILL.md.
3. UHP harness metaharness.optimize. HarnessRouter-style hosts should expose a UHP method that forwards the same request shape into optimize_harness or metaharness run.

AgentSky should call the same Python or UHP contract. This repository does not wrap AgentSky's cloud API as a proposer, and it does not add a first-class HarnessRouter proposer backend.

A tiny host that already has agent files and calls the kernel is examples/embed_runtime/optimize_host.py.

Request / Response Sketch

This is a documentation sketch, not a new wire protocol. Hosts can map it onto Python kwargs, a UHP method, or metaharness run plus a project directory.

{
  "request": {
    "baseline_files": {
      "routing.json": "{\"default\": {\"harness\": \"pi\"}, \"rules\": []}",
      "route.py": "# stable router"
    },
    "tasks": [
      {
        "id": "route-simple",
        "type": "command",
        "command": "python route.py --prompt \"say hello and list files\""
      }
    ],
    "budget": 1,
    "write_scope": {
      "allowed_write_paths": [{"path": "routing.json", "class": "middleware"}],
      "write_scope_mode": "single-class"
    },
    "leakage_gate": true
  },
  "response": {
    "best_candidate_id": "c0001",
    "best_objective": 1.0,
    "diffs": ["routing.json"],
    "ledger_path": "runs/host-run/indexes/leaderboard.json"
  }
}

The coding-tool project layout (metaharness.json, tasks.json, baseline/) is the usual way to supply those fields on disk. OptimizeResult returns the winning candidate id, objective, workspace, and run directory; diffs and the ledger live under that run directory.

What This Repo Does Not Do

• No AgentSky cloud proposer.
• No HarnessRouter / UHP proposer backend in-tree.
• Omnigent remains an experimental proposer (OmnigentCliBackend) and can be a target (routing tables, agent configs). Those are different roles.