mirror of
https://github.com/openai/codex.git
synced 2026-05-04 05:11:37 +03:00
126bd6e7a8
## Summary This PR aims to improve integration between the realtime model and the codex agent by sharing more context with each other. In particular, we now share full realtime conversation transcript deltas in addition to the delegation message. realtime_conversation.rs now turns a handoff into: ``` <realtime_delegation> <input>...</input> <transcript_delta>...</transcript_delta> </realtime_delegation> ``` ## Implementation notes The transcript is accumulated in the realtime websocket layer as parsed realtime events arrive. When a background-agent handoff is requested, the current transcript snapshot is copied onto the handoff event and then serialized by `realtime_conversation.rs` into the hidden realtime delegation envelope that Codex receives as user-turn context. For Realtime V2, the session now explicitly enables input audio transcription, and the parser handles the relevant input/output transcript completion events so the snapshot includes both user speech and realtime model responses. The delegation `<input>` remains the actual handoff request, while `<transcript_delta>` carries the surrounding conversation history for context. Reviewers should note that the transcript payload is intended for Codex context sharing, not UI rendering. The realtime delegation envelope should stay hidden from the user-facing transcript surface, while still being included in the background-agent turn so Codex can answer with the same conversational context the realtime model had.
codex-api
Typed clients for Codex/OpenAI APIs built on top of the generic transport in codex-client.
- Hosts the request/response models and request builders for Responses and Compact APIs.
- Owns provider configuration (base URLs, headers, query params), auth header injection, retry tuning, and stream idle settings.
- Parses SSE streams into
ResponseEvent/ResponseStream, including rate-limit snapshots and API-specific error mapping. - Serves as the wire-level layer consumed by
codex-core; higher layers handle auth refresh and business logic.
Core interface
The public interface of this crate is intentionally small and uniform:
-
Responses endpoint
- Input:
ResponsesApiRequestfor the request body (model,instructions,input,tools,parallel_tool_calls, reasoning/text controls).ResponsesOptionsfor transport/header concerns (conversation_id,session_source,extra_headers,compression,turn_state).
- Output: a
ResponseStreamofResponseEvent(both re-exported fromcommon).
- Input:
-
Compaction endpoint
- Input:
CompactionInput<'a>(re-exported ascodex_api::CompactionInput):model: &str.input: &[ResponseItem]– history to compact.instructions: &str– fully-resolved compaction instructions.
- Output:
Vec<ResponseItem>. CompactClient::compact_input(&CompactionInput, extra_headers)wraps the JSON encoding and retry/telemetry wiring.
- Input:
-
Memory summarize endpoint
- Input:
MemorySummarizeInput(re-exported ascodex_api::MemorySummarizeInput):model: String.raw_memories: Vec<RawMemory>(serialized astracesfor wire compatibility).RawMemoryincludesid,metadata.source_path, and normalizeditems.
reasoning: Option<Reasoning>.
- Output:
Vec<MemorySummarizeOutput>. MemoriesClient::summarize_input(&MemorySummarizeInput, extra_headers)wraps JSON encoding and retry/telemetry wiring.
- Input:
All HTTP details (URLs, headers, retry/backoff policies, SSE framing) are encapsulated in codex-api and codex-client. Callers construct prompts/inputs using protocol types and work with typed streams of ResponseEvent or compacted ResponseItem values.