Dividing UserMsgs into categories to send it back to the tui (#3127)

This PR does the following:

- divides user msgs into 3 categories: plain, user instructions, and
environment context
- Centralizes adding user instructions and environment context to a
degree
- Improve the integration testing

Building on top of #3123

Specifically this
[comment](https://github.com/openai/codex/pull/3123#discussion_r2319885089).
We need to send the user message while ignoring the User Instructions
and Environment Context we attach.
This commit is contained in:
Ahmed Ibrahim
2025-09-03 22:34:50 -07:00
committed by GitHub
parent f2036572b6
commit 2b96f9f569
12 changed files with 175 additions and 34 deletions

View File

@@ -6,7 +6,6 @@ use codex_apply_patch::APPLY_PATCH_TOOL_INSTRUCTIONS;
use codex_protocol::config_types::ReasoningEffort as ReasoningEffortConfig;
use codex_protocol::config_types::ReasoningSummary as ReasoningSummaryConfig;
use codex_protocol::config_types::Verbosity as VerbosityConfig;
use codex_protocol::models::ContentItem;
use codex_protocol::models::ResponseItem;
use futures::Stream;
use serde::Serialize;
@@ -20,10 +19,6 @@ use tokio::sync::mpsc;
/// with this content.
const BASE_INSTRUCTIONS: &str = include_str!("../prompt.md");
/// wraps user instructions message in a tag for the model to parse more easily.
const USER_INSTRUCTIONS_START: &str = "<user_instructions>\n\n";
const USER_INSTRUCTIONS_END: &str = "\n\n</user_instructions>";
/// API request payload for a single model turn
#[derive(Default, Debug, Clone)]
pub struct Prompt {
@@ -68,17 +63,6 @@ impl Prompt {
pub(crate) fn get_formatted_input(&self) -> Vec<ResponseItem> {
self.input.clone()
}
/// Creates a formatted user instructions message from a string
pub(crate) fn format_user_instructions_message(ui: &str) -> ResponseItem {
ResponseItem::Message {
id: None,
role: "user".to_string(),
content: vec![ContentItem::InputText {
text: format!("{USER_INSTRUCTIONS_START}{ui}{USER_INSTRUCTIONS_END}"),
}],
}
}
}
#[derive(Debug)]

View File

@@ -107,6 +107,7 @@ use crate::safety::assess_command_safety;
use crate::safety::assess_safety_for_untrusted_command;
use crate::shell;
use crate::turn_diff_tracker::TurnDiffTracker;
use crate::user_instructions::UserInstructions;
use crate::user_notification::UserNotification;
use crate::util::backoff;
use codex_protocol::config_types::ReasoningEffort as ReasoningEffortConfig;
@@ -482,6 +483,7 @@ impl Session {
InitialHistory::New => None,
InitialHistory::Resumed(items) => Some(sess.build_initial_messages(items)),
};
let events = std::iter::once(Event {
id: INITIAL_SUBMIT_ID.to_owned(),
msg: EventMsg::SessionConfigured(SessionConfiguredEvent {
@@ -540,7 +542,7 @@ impl Session {
// TODO: Those items shouldn't be "user messages" IMO. Maybe developer messages.
let mut conversation_items = Vec::<ResponseItem>::with_capacity(2);
if let Some(user_instructions) = turn_context.user_instructions.as_deref() {
conversation_items.push(Prompt::format_user_instructions_message(user_instructions));
conversation_items.push(UserInstructions::new(user_instructions.to_string()).into());
}
conversation_items.push(ResponseItem::from(EnvironmentContext::new(
Some(turn_context.cwd.clone()),

View File

@@ -8,12 +8,10 @@ use crate::shell::Shell;
use codex_protocol::config_types::SandboxMode;
use codex_protocol::models::ContentItem;
use codex_protocol::models::ResponseItem;
use codex_protocol::protocol::ENVIRONMENT_CONTEXT_CLOSE_TAG;
use codex_protocol::protocol::ENVIRONMENT_CONTEXT_OPEN_TAG;
use std::path::PathBuf;
/// wraps environment context message in a tag for the model to parse more easily.
pub(crate) const ENVIRONMENT_CONTEXT_START: &str = "<environment_context>";
pub(crate) const ENVIRONMENT_CONTEXT_END: &str = "</environment_context>";
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, DeriveDisplay)]
#[serde(rename_all = "kebab-case")]
#[strum(serialize_all = "kebab-case")]
@@ -79,7 +77,7 @@ impl EnvironmentContext {
/// </environment_context>
/// ```
pub fn serialize_to_xml(self) -> String {
let mut lines = vec![ENVIRONMENT_CONTEXT_START.to_string()];
let mut lines = vec![ENVIRONMENT_CONTEXT_OPEN_TAG.to_string()];
if let Some(cwd) = self.cwd {
lines.push(format!(" <cwd>{}</cwd>", cwd.to_string_lossy()));
}
@@ -101,7 +99,7 @@ impl EnvironmentContext {
{
lines.push(format!(" <shell>{shell_name}</shell>"));
}
lines.push(ENVIRONMENT_CONTEXT_END.to_string());
lines.push(ENVIRONMENT_CONTEXT_CLOSE_TAG.to_string());
lines.join("\n")
}
}

View File

@@ -2,6 +2,8 @@ use crate::protocol::AgentMessageEvent;
use crate::protocol::AgentReasoningEvent;
use crate::protocol::AgentReasoningRawContentEvent;
use crate::protocol::EventMsg;
use crate::protocol::InputMessageKind;
use crate::protocol::UserMessageEvent;
use crate::protocol::WebSearchEndEvent;
use codex_protocol::models::ContentItem;
use codex_protocol::models::ReasoningItemContent;
@@ -17,15 +19,37 @@ pub(crate) fn map_response_item_to_event_messages(
show_raw_agent_reasoning: bool,
) -> Vec<EventMsg> {
match item {
ResponseItem::Message { content, .. } => {
let mut events = Vec::new();
for content_item in content {
if let ContentItem::OutputText { text } = content_item {
events.push(EventMsg::AgentMessage(AgentMessageEvent {
message: text.clone(),
}));
}
ResponseItem::Message { role, content, .. } => {
// Do not surface system messages as user events.
if role == "system" {
return Vec::new();
}
let events: Vec<EventMsg> = content
.iter()
.filter_map(|content_item| match content_item {
ContentItem::OutputText { text } => {
Some(EventMsg::AgentMessage(AgentMessageEvent {
message: text.clone(),
}))
}
ContentItem::InputText { text } => {
let trimmed = text.trim_start();
let kind = if trimmed.starts_with("<environment_context>") {
Some(InputMessageKind::EnvironmentContext)
} else if trimmed.starts_with("<user_instructions>") {
Some(InputMessageKind::UserInstructions)
} else {
Some(InputMessageKind::Plain)
};
Some(EventMsg::UserMessage(UserMessageEvent {
message: text.clone(),
kind,
}))
}
_ => None,
})
.collect();
events
}

View File

@@ -34,6 +34,7 @@ mod mcp_tool_call;
mod message_history;
mod model_provider_info;
pub mod parse_command;
mod user_instructions;
pub use model_provider_info::BUILT_IN_OSS_MODEL_PROVIDER_ID;
pub use model_provider_info::ModelProviderInfo;
pub use model_provider_info::WireApi;

View File

@@ -0,0 +1,42 @@
use serde::Deserialize;
use serde::Serialize;
use codex_protocol::models::ContentItem;
use codex_protocol::models::ResponseItem;
use codex_protocol::protocol::USER_INSTRUCTIONS_CLOSE_TAG;
use codex_protocol::protocol::USER_INSTRUCTIONS_OPEN_TAG;
/// Wraps user instructions in a tag so the model can classify them easily.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
#[serde(rename = "user_instructions", rename_all = "snake_case")]
pub(crate) struct UserInstructions {
text: String,
}
impl UserInstructions {
pub fn new<T: Into<String>>(text: T) -> Self {
Self { text: text.into() }
}
/// Serializes the user instructions to an XML-like tagged block that starts
/// with <user_instructions> so clients can classify it.
pub fn serialize_to_xml(self) -> String {
format!(
"{USER_INSTRUCTIONS_OPEN_TAG}\n\n{}\n\n{USER_INSTRUCTIONS_CLOSE_TAG}",
self.text
)
}
}
impl From<UserInstructions> for ResponseItem {
fn from(ui: UserInstructions) -> Self {
ResponseItem::Message {
id: None,
role: "user".to_string(),
content: vec![ContentItem::InputText {
text: ui.serialize_to_xml(),
}],
}
}
}