@tailored-ai/provider-openai

OpenAI model provider for Tailored AI agents — GPT and o-series models with the request shaping current OpenAI models require.

This plugin provides the openai provider id (#236 moved it out of core). Existing providers.openai config keeps working, and gains:

• Reasoning-model handling — o-series and gpt-5 family models reject temperature and the deprecated max_tokens; the plugin omits temperature for them and always uses max_completion_tokens. The built-in sends both, which errors on these models.
• Org/project headers — OpenAI-Organization and OpenAI-Project for multi-org accounts and usage attribution.
• params.extra passthrough — e.g. reasoning_effort.
• Responses API — reasoning and tool calls in the same turn, for the models chat completions refuses (below).
• Streaming (chatStream) and model discovery (listModels).

Which endpoint

/v1/chat/completions refuses function tools alongside any reasoning effort on gpt-5.4, gpt-5.5 and the gpt-5.6 family, and does not serve gpt-5.3-codex at all. TAI always sends tools, so on those models reasoning and tool use are mutually exclusive there. /v1/responses has no such restriction. Measured 2026-08-05, tools present:

| model | /v1/chat/completions | /v1/responses | |---|---|---| | gpt-5.4, gpt-5.5, gpt-5.6-* | reasoning impossible | every effort accepted | | gpt-5.3-codex | not served | accepted | | gpt-5-mini, o4-mini | rejects none | rejects none | | gpt-5.3-chat-latest | only medium | only medium |

api: auto (the default) sends the first two rows to /v1/responses and everything else to chat completions, so nothing moves that was already working. It only auto-routes against api.openai.com — a baseUrl pointing at Azure, a proxy or a local gateway stays on chat completions unless you set api: responses explicitly.

Per-model reasoning quirks are learned from the API's own refusals rather than hardcoded: an effort a model rejects is replaced with the nearest one it accepts (so off on gpt-5-mini becomes minimal, not high), remembered for the process, and reported once.

Install

tai plugin install @tailored-ai/provider-openai

Configure

plugins:
  - "@tailored-ai/provider-openai"

providers:
  openai:
    apiKey: "${OPENAI_API_KEY}"        # https://platform.openai.com/api-keys
    defaultModel: "gpt-5-mini"

agent:
  defaultProvider: openai

| Field | Required | Notes | |---|---|---| | apiKey | yes | Use ${OPENAI_API_KEY} to read from the environment. | | defaultModel | yes | Any chat-completions model id. tai edit lists them via listModels. | | organization | no | OpenAI-Organization header. | | project | no | OpenAI-Project header. | | reasoningModels | no | Extra model-id prefixes to treat as reasoning models (built-in heuristic covers o* and gpt-5*). | | baseUrl | no | Default https://api.openai.com/v1. Override for proxies. | | api | no | auto (default), chat, or responses. See which endpoint. | | responsesModels | no | Extra model-id prefixes to route to /v1/responses under api: auto. | | reasoningSummary | no | auto (default), concise, detailed, off. The readable reasoning trace TAI captures. Orgs not verified for summaries are detected and downgraded automatically. | | store | no | Let OpenAI retain responses server-side. Default false, matching chat completions. |

Development

pnpm --filter @tailored-ai/provider-openai run build
pnpm --filter @tailored-ai/provider-openai run test

License

MIT