dcdpr / jp / 25105079435

66%

push

github

feat(llm, openai): Add GPT-5.5 and GPT-5.5 Pro model support (#585)

GPT-5.5 and GPT-5.5 Pro are now recognised by the OpenAI provider.
GPT-5.5 follows the same streaming path as its predecessors, while
GPT-5.5 Pro requires the non-streaming Responses API — a constraint
captured by the new `STREAMING_UNSUPPORTED` feature flag. When that flag
is present the provider fetches the full response in one shot and
synthesises the same streaming event sequence (`OutputItemAdded`,
`OutputTextDelta`, `ReasoningSummaryTextDelta`, `OutputItemDone`, …)
that the rest of the pipeline already understands, so no downstream
changes were needed.

The `classify_stream_error` helper was also tightened: it now inspects
both the `type` and `code` fields of in-stream errors, and extracts a
`retry_after` duration from the message body when present. This fixes a
case where OpenAI's per-minute token/request rate-limits (delivered as
`type=tokens, code=rate_limit_exceeded`) were surfaced as a
non-retryable `Other` error instead of the correct `RateLimit` kind. An
unknown error type that carries a retry hint is now treated as transient
so the retry layer can engage.

Project aliases are updated: `openai`, `chatgpt`, `gpt`, and `gpt5` now
point to `gpt-5.5`, with new `gpt5-pro` and `gpt-pro` aliases added for
`gpt-5.5-pro`.

Signed-off-by: Jean Mertz <git@jeanmertz.com>

106 of 143 new or added lines in 1 file covered. (74.13%)

23559 of 36615 relevant lines covered (64.34%)

190.32 hits per line