@veyorhq/openai

@veyorhq/openai provides OpenAI-backed implementations for Veyor actors.

It exposes two separate adapters:

• @veyorhq/openai/openai creates an Effect AI model layer for @veyorhq/ai-effect.
• @veyorhq/openai/codex runs a Veyor actor through the local Codex CLI.

[!IMPORTANT] This package is under active development. It is private, remains at 0.0.0, and is not published to a package registry.

OpenAI API

Use openai(model) with agent(...) when the actor should call the OpenAI API through Effect AI.

import { agent } from "@veyorhq/ai-effect";
import { openai } from "@veyorhq/openai/openai";
import { Writer } from "./actors.js";

const WriterImpl = agent(Writer, openai("gpt-4o-mini"), {
  prompt: "Return a concise implementation plan.",
});

The adapter reads OPENAI_API_KEY through Effect Config. Missing or invalid configuration fails when the model layer is used.

Codex CLI

Use codex(actor, model, options) when a Codex task should implement the actor.

import { codex } from "@veyorhq/openai/codex";
import { Planner } from "./actors.js";

const PlannerImpl = codex(Planner, "gpt-5.6-sol", {
  effort: "xhigh",
  cwd: process.cwd(),
  sandbox: "workspace-write",
  instructions: "Plan the assignment. Do not edit files.",
});

The adapter:

1. converts the actor output contract to JSON Schema;
2. writes that schema to a scoped temporary directory;
3. runs codex exec with structured output enabled;
4. parses and validates the final JSON value;
5. passes the Veyor task's abort signal into the Effect runtime.

Options

| Option | Purpose | | -------------- | ------------------------------------------------------------------- | | effort | Sets Codex model reasoning effort. | | verbosity | Sets model verbosity. | | instructions | Adds package- or actor-specific instructions before the assignment. | | cwd | Sets the Codex working directory. | | sandbox | Selects read-only, workspace-write, or danger-full-access. |

The Codex executable must be installed, authenticated, and available on PATH. The adapter defaults to the workspace-write sandbox; choose a narrower sandbox when the actor does not need writes.

When several actors share the same model and effort, capture that once with codex.preset(...) instead of repeating it per actor:

import { codex } from "@veyorhq/openai/codex";
import { Planner } from "./actors.js";

const planWorker = codex.preset({
  model: "gpt-5.6-sol",
  effort: "xhigh",
  sandbox: "workspace-write",
});

const PlannerImpl = planWorker(Planner, {
  instructions: "Plan the assignment. Do not edit files.",
});

codex.preset is sugar over codex — per-call options (including model) shallow-merge over the preset, with the per-call value winning.

Current limits

• Both adapters require output contracts that can produce JSON Schema.
• Codex aggregation currently returns an empty placeholder.
• The OpenAI model adapter inherits the unfinished aggregation support from @veyorhq/ai-effect.
• Retry, fallback, and budget policy belong to the workflow and are not added by this package.

See @veyorhq/ai-effect, @veyorhq/core, and the software factory example.