Raw chain-of-thought was written to both `content` (`reasoning_text`) and
`summary` (`summary_text`), and the stream emitter sent the same delta on
`response.reasoning_text.delta` *and* `response.reasoning_summary_text.delta`.
Clients that render both channels therefore printed every thinking chunk
twice — most visibly the Codex CLI, whose thinking panel repeated itself.
OpenAI keeps the two channels distinct: `content` carries the raw CoT while
`summary` is the summarised view. Emit the thinking on `content` only:
- `openai_responses_reasoning_text_fields` becomes
`openai_responses_reasoning_text_parts`, returning just the `content` array;
reasoning items keep `summary: []` (or a provider-supplied summary).
- The Responses stream emitter emits `response.reasoning_text.delta` /
`.done` and no longer mirrors them onto the summary events. The reasoning
`output_item.added` no longer announces a `reasoning_summary_part`.
- The provider-state reasoning reader accepts `content` (`reasoning_text`)
first and falls back to `summary`, so it also understands items produced by
older Aether versions; its state field is renamed accordingly.
- Non-streaming builders (Chat -> Responses, manual Responses response, Grok
gateway) place the thinking on `content` and leave `summary` empty.
Tests cover the raw thinking appearing exactly once in the emitted stream.
Gemini runs `googleSearch` inside Google. The search leaves no
client-visible tool call, and the evidence arrives only as
`candidates[].groundingMetadata`. Every cross-format target dropped it
wholesale, so a grounded answer reached OpenAI- and Claude-shaped clients
as prose that names its sources with nothing structured behind it: no
`annotations`, no `citations`, no `url_citation`. Callers that verify
grounding — the common "did this model actually search?" check — saw a
200 with no evidence and had to treat the answer as ungrounded.
Adapters now normalise `groundingMetadata` into neutral citations and
each target renders its own family's standard shape: `url_citation`
annotations for `openai:chat` and `openai:responses`, and
`web_search_result_location` citations on the text block for
`claude:messages`. Gemini reports segment bounds as UTF-8 byte offsets
while both targets count characters, so the bounds are converted rather
than copied.
Streaming is covered too, since that is what grounded traffic actually
uses. A new `CanonicalStreamEvent::Citations` carries the neutral list
once the answer text is whole — the offsets index into the finished
answer, so it rides just ahead of `Finish` rather than as a delta per
chunk — and each client emitter renders it: `delta.annotations` chunks,
`response.output_text.annotation.added` events (also kept on the finished
message item so clients that only read `response.completed` see them),
and `citations_delta` content block deltas.
For reference, CLIProxyAPI projects grounding only in its
antigravity→Claude translator, and only when the client declared a typed
`web_search_*` tool; its OpenAI and plain Gemini translators have no
grounding handling at all. The citation shape here matches theirs, but
the coverage is deliberately wider: all three targets, streaming and
non-streaming, with no dependency on a declared tool.
Co-Authored-By: Claude Opus 5 <[email protected]>
OpenAI Responses treats reasoning.content as the raw chain-of-thought
and summary as a skim view. Aether was dumping thinking into summary
and leaving content null, which hid the thinking panel in desktop UIs.
Put reasoning_content / equivalent text into reasoning_text content
parts, and copy the same text into summary_text so CLI clients still
work. Stream emitters now send both reasoning_text and summary events.
Consolidate subscription usage policy enforcement, privacy-safe persistence, and gateway security hardening into one reviewable change.
Includes bounded HTTP and execution envelopes, header and protocol guards, DNS and relay validation, authentication and secret projection hardening, secure backup/install paths, and regression coverage.
Kayphoon