Allow native OpenAI Responses API to use custom base URL

[chyax98]

Problem

The native openai-api provider uses the OpenAI Responses API, but its base URL is currently hardcoded to https://api.openai.com/v1.

That means users with a local/proxied Responses API endpoint cannot configure openai-api to use it. The available openai-compatible configuration path is not equivalent because it targets the OpenAI-compatible chat/completions protocol rather than the native Responses API path.

Reproduction

1. Run a local Responses API endpoint at http://127.0.0.1:8317/v1.
2. Configure OPENAI_API_KEY and try to set OPENAI_BASE_URL, OPENAI_API_BASE, or JCODE_OPENAI_API_BASE.
3. Run:

jcode -p openai-api -m gpt-5.5 run 'hello'

Actual behavior

openai-api still sends requests to https://api.openai.com/v1/responses.

Expected behavior

openai-api should continue using /responses, but allow a custom API base URL for API-key based usage.

Notes

This should remain distinct from openai-compatible, which is a separate provider/protocol path.

[1jehuang]

Triage: confirmed reproducible. The native openai-api provider (OpenAI Responses API path) hardcodes its base in crates/jcode-base/src/provider/openai.rs:

const OPENAI_API_BASE: &str = "https://api.openai.com/v1";

and responses_url() always uses it for non-ChatGPT (API-key) mode, so there is no env/config hook for a custom Responses endpoint. There is no OPENAI_BASE_URL / JCODE_OPENAI_API_BASE lookup anywhere on this path (confirmed: zero matches). The openai-compatible provider is a different protocol (chat/completions), so it is not an equivalent substitute as you noted.

Fix path is clear and low-risk: in API-key mode (not ChatGPT/Codex OAuth mode), resolve the base from an env var (e.g. OPENAI_BASE_URL / OPENAI_API_BASE / JCODE_OPENAI_API_BASE) before falling back to the hardcoded default, keeping the /responses suffix intact. ChatGPT-mode (CHATGPT_API_BASE) should stay fixed.

Labeling enhancement / priority: medium / triage: reproducible.

[1jehuang]

Fixed. The native openai-api (Responses API) base URL is now overridable for API-key usage.

Change

responses_url() resolves the base from, in order: JCODE_OPENAI_API_BASE > OPENAI_BASE_URL > OPENAI_API_BASE, falling back to https://api.openai.com/v1. jcode still appends /responses, derives the WebSocket and /compact endpoints from the same base, and points the /models catalog probe at it. ChatGPT/Codex OAuth mode stays pinned to the fixed Codex backend; a malformed override is logged and ignored (never fatal); a trailing /responses is trimmed.

Example:

export OPENAI_API_KEY=...                 # API-key mode
export JCODE_OPENAI_API_BASE=http://127.0.0.1:8317/v1
jcode -p openai-api -m gpt-5.5 run hello  # -> POST http://127.0.0.1:8317/v1/responses

Verification (live, against a mock Responses server)

• With the override: jcode -p openai-api run hello POSTed to http://127.0.0.1:8317/v1/responses and rendered the mock reply. Mock log: HIT POST /v1/responses.
• Without it (control): requests went to api.openai.com (OpenAI returned "Incorrect API key"), and the local mock received zero hits.
• 3 unit tests added: override precedence/validation, WS+/compact derivation, and that ChatGPT/OAuth mode ignores the override. Full openai suite (80 tests) passes.
• Documented in OAUTH.md.

Committed in d8a19a3a; lands in the next release.