OpenAI Responses Provider

Provider guide · OpenAI-compatible · 简体中文

Minimal Setup

export SIGIL_OPENAI_RESPONSES_API_KEY="sk-..."
sigil

config_version = 2

[agent]
connection = "openai-default"
model = "gpt-4.1"

[connections.openai-default]
label = "OpenAI"
provider = "openai"
protocol = "responses"
base_url = "https://api.openai.com/v1"
credential = { source = "environment", name = "SIGIL_OPENAI_RESPONSES_API_KEY" }

See openai-responses.toml for a copyable file.

Authentication

The example binds only this connection to SIGIL_OPENAI_RESPONSES_API_KEY. You can instead choose the secure credential store; sigil.toml, model cache, and session files contain no secret value. organization and project are optional connection options.

Options And Visible Limits

This connection uses the Responses route, not Chat Completions. Keep endpoint and account options on this connection so another OpenAI or compatible account cannot supply a fallback. Background requests and provider-hosted tools are not enabled.

Image attachments work only for model IDs Sigil recognizes as image-capable. Unknown names and aliases are rejected before sending. On the official endpoint and supported dated snapshot, one context-window rejection before output may trigger one compact-and-retry attempt; compatible endpoints, aliases, restored sessions, and repeated failures do not.

Verify

Run sigil doctor and confirm default=openai-default/gpt-4.1, the responses protocol, /v1 endpoint, credential source, and readiness.

Common Problems

• 404: confirm the service exposes /v1/responses, not only Chat Completions.
• Authentication: check the environment variable or repair this connection in /config; Sigil does not fall back to another connection.
• Stream ends early: confirm the endpoint emits a completed Responses event.
• Tool or image input fails: confirm the selected model supports that input.

Next: Return to Providers.