Files
codex/codex-rs/codex-api/src/endpoint/realtime_websocket/methods_v2.rs
T
guinness-oai 126bd6e7a8 Update realtime handoff transcript handling (#18597)
## 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.
2026-04-20 14:04:09 -07:00

157 lines
7.5 KiB
Rust

use crate::endpoint::realtime_websocket::methods_common::REALTIME_AUDIO_SAMPLE_RATE;
use crate::endpoint::realtime_websocket::protocol::AudioFormatType;
use crate::endpoint::realtime_websocket::protocol::ConversationContentType;
use crate::endpoint::realtime_websocket::protocol::ConversationFunctionCallOutputItem;
use crate::endpoint::realtime_websocket::protocol::ConversationItemContent;
use crate::endpoint::realtime_websocket::protocol::ConversationItemPayload;
use crate::endpoint::realtime_websocket::protocol::ConversationItemType;
use crate::endpoint::realtime_websocket::protocol::ConversationMessageItem;
use crate::endpoint::realtime_websocket::protocol::ConversationRole;
use crate::endpoint::realtime_websocket::protocol::NoiseReductionType;
use crate::endpoint::realtime_websocket::protocol::RealtimeOutboundMessage;
use crate::endpoint::realtime_websocket::protocol::RealtimeOutputModality;
use crate::endpoint::realtime_websocket::protocol::RealtimeSessionMode;
use crate::endpoint::realtime_websocket::protocol::RealtimeVoice;
use crate::endpoint::realtime_websocket::protocol::SessionAudio;
use crate::endpoint::realtime_websocket::protocol::SessionAudioFormat;
use crate::endpoint::realtime_websocket::protocol::SessionAudioInput;
use crate::endpoint::realtime_websocket::protocol::SessionAudioOutput;
use crate::endpoint::realtime_websocket::protocol::SessionAudioOutputFormat;
use crate::endpoint::realtime_websocket::protocol::SessionFunctionTool;
use crate::endpoint::realtime_websocket::protocol::SessionInputAudioTranscription;
use crate::endpoint::realtime_websocket::protocol::SessionNoiseReduction;
use crate::endpoint::realtime_websocket::protocol::SessionToolType;
use crate::endpoint::realtime_websocket::protocol::SessionTurnDetection;
use crate::endpoint::realtime_websocket::protocol::SessionType;
use crate::endpoint::realtime_websocket::protocol::SessionUpdateSession;
use crate::endpoint::realtime_websocket::protocol::TurnDetectionType;
use serde_json::json;
const REALTIME_V2_OUTPUT_MODALITY_AUDIO: &str = "audio";
const REALTIME_V2_OUTPUT_MODALITY_TEXT: &str = "text";
const REALTIME_V2_TOOL_CHOICE: &str = "auto";
const REALTIME_V2_BACKGROUND_AGENT_TOOL_NAME: &str = "background_agent";
const REALTIME_V2_BACKGROUND_AGENT_TOOL_DESCRIPTION: &str = "Send a user request to the background agent. Use this as the default action. Do not rephrase the user's ask or rewrite it in your own words; pass along the user's own words. If the background agent is idle, this starts a new task and returns the final result to the user. If the background agent is already working on a task, this sends the request as guidance to steer that previous task. If the user asks to do something next, later, after this, or once current work finishes, call this tool so the work is actually queued instead of merely promising to do it later.";
const REALTIME_V2_INPUT_TRANSCRIPTION_MODEL: &str = "gpt-4o-mini-transcribe";
pub(super) fn conversation_item_create_message(text: String) -> RealtimeOutboundMessage {
RealtimeOutboundMessage::ConversationItemCreate {
item: ConversationItemPayload::Message(ConversationMessageItem {
r#type: ConversationItemType::Message,
role: ConversationRole::User,
content: vec![ConversationItemContent {
r#type: ConversationContentType::InputText,
text,
}],
}),
}
}
pub(super) fn conversation_handoff_append_message(
handoff_id: String,
output_text: String,
) -> RealtimeOutboundMessage {
RealtimeOutboundMessage::ConversationItemCreate {
item: ConversationItemPayload::FunctionCallOutput(ConversationFunctionCallOutputItem {
r#type: ConversationItemType::FunctionCallOutput,
call_id: handoff_id,
output: output_text,
}),
}
}
pub(super) fn session_update_session(
instructions: String,
session_mode: RealtimeSessionMode,
output_modality: RealtimeOutputModality,
voice: RealtimeVoice,
) -> SessionUpdateSession {
match session_mode {
RealtimeSessionMode::Conversational => SessionUpdateSession {
id: None,
r#type: SessionType::Realtime,
model: None,
instructions: Some(instructions),
output_modalities: Some(vec![output_modality_value(output_modality).to_string()]),
audio: SessionAudio {
input: SessionAudioInput {
format: SessionAudioFormat {
r#type: AudioFormatType::AudioPcm,
rate: REALTIME_AUDIO_SAMPLE_RATE,
},
noise_reduction: Some(SessionNoiseReduction {
r#type: NoiseReductionType::NearField,
}),
transcription: Some(SessionInputAudioTranscription {
model: REALTIME_V2_INPUT_TRANSCRIPTION_MODEL.to_string(),
}),
turn_detection: Some(SessionTurnDetection {
r#type: TurnDetectionType::ServerVad,
interrupt_response: true,
create_response: true,
silence_duration_ms: 500,
}),
},
output: Some(SessionAudioOutput {
format: Some(SessionAudioOutputFormat {
r#type: AudioFormatType::AudioPcm,
rate: REALTIME_AUDIO_SAMPLE_RATE,
}),
voice,
}),
},
tools: Some(vec![SessionFunctionTool {
r#type: SessionToolType::Function,
name: REALTIME_V2_BACKGROUND_AGENT_TOOL_NAME.to_string(),
description: REALTIME_V2_BACKGROUND_AGENT_TOOL_DESCRIPTION.to_string(),
parameters: json!({
"type": "object",
"properties": {
"prompt": {
"type": "string",
"description": "The user request to delegate to the background agent."
}
},
"required": ["prompt"],
"additionalProperties": false
}),
}]),
tool_choice: Some(REALTIME_V2_TOOL_CHOICE.to_string()),
},
RealtimeSessionMode::Transcription => SessionUpdateSession {
id: None,
r#type: SessionType::Transcription,
model: None,
instructions: None,
output_modalities: None,
audio: SessionAudio {
input: SessionAudioInput {
format: SessionAudioFormat {
r#type: AudioFormatType::AudioPcm,
rate: REALTIME_AUDIO_SAMPLE_RATE,
},
noise_reduction: None,
transcription: Some(SessionInputAudioTranscription {
model: REALTIME_V2_INPUT_TRANSCRIPTION_MODEL.to_string(),
}),
turn_detection: None,
},
output: None,
},
tools: None,
tool_choice: None,
},
}
}
fn output_modality_value(output_modality: RealtimeOutputModality) -> &'static str {
match output_modality {
RealtimeOutputModality::Text => REALTIME_V2_OUTPUT_MODALITY_TEXT,
RealtimeOutputModality::Audio => REALTIME_V2_OUTPUT_MODALITY_AUDIO,
}
}
pub(super) fn websocket_intent() -> Option<&'static str> {
None
}