Files
llmx/codex-rs/core/src/client_common.rs
2025-10-05 16:41:55 -07:00

551 lines
17 KiB
Rust

use crate::client_common::tools::ToolSpec;
use crate::error::Result;
use crate::model_family::ModelFamily;
use crate::protocol::RateLimitSnapshot;
use crate::protocol::TokenUsage;
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::ResponseItem;
use futures::Stream;
use serde::Deserialize;
use serde::Serialize;
use serde_json::Value;
use std::borrow::Cow;
use std::collections::HashSet;
use std::ops::Deref;
use std::pin::Pin;
use std::task::Context;
use std::task::Poll;
use tokio::sync::mpsc;
/// Review thread system prompt. Edit `core/src/review_prompt.md` to customize.
pub const REVIEW_PROMPT: &str = include_str!("../review_prompt.md");
/// API request payload for a single model turn
#[derive(Default, Debug, Clone)]
pub struct Prompt {
/// Conversation context input items.
pub input: Vec<ResponseItem>,
/// Tools available to the model, including additional tools sourced from
/// external MCP servers.
pub(crate) tools: Vec<ToolSpec>,
/// Whether parallel tool calls are permitted for this prompt.
pub(crate) parallel_tool_calls: bool,
/// Optional override for the built-in BASE_INSTRUCTIONS.
pub base_instructions_override: Option<String>,
/// Optional the output schema for the model's response.
pub output_schema: Option<Value>,
}
impl Prompt {
pub(crate) fn get_full_instructions<'a>(&'a self, model: &'a ModelFamily) -> Cow<'a, str> {
let base = self
.base_instructions_override
.as_deref()
.unwrap_or(model.base_instructions.deref());
// When there are no custom instructions, add apply_patch_tool_instructions if:
// - the model needs special instructions (4.1)
// AND
// - there is no apply_patch tool present
let is_apply_patch_tool_present = self.tools.iter().any(|tool| match tool {
ToolSpec::Function(f) => f.name == "apply_patch",
ToolSpec::Freeform(f) => f.name == "apply_patch",
_ => false,
});
if self.base_instructions_override.is_none()
&& model.needs_special_apply_patch_instructions
&& !is_apply_patch_tool_present
{
Cow::Owned(format!("{base}\n{APPLY_PATCH_TOOL_INSTRUCTIONS}"))
} else {
Cow::Borrowed(base)
}
}
pub(crate) fn get_formatted_input(&self) -> Vec<ResponseItem> {
let mut input = self.input.clone();
// when using the *Freeform* apply_patch tool specifically, tool outputs
// should be structured text, not json. Do NOT reserialize when using
// the Function tool - note that this differs from the check above for
// instructions. We declare the result as a named variable for clarity.
let is_freeform_apply_patch_tool_present = self.tools.iter().any(|tool| match tool {
ToolSpec::Freeform(f) => f.name == "apply_patch",
_ => false,
});
if is_freeform_apply_patch_tool_present {
reserialize_shell_outputs(&mut input);
}
input
}
}
fn reserialize_shell_outputs(items: &mut [ResponseItem]) {
let mut shell_call_ids: HashSet<String> = HashSet::new();
items.iter_mut().for_each(|item| match item {
ResponseItem::LocalShellCall { call_id, id, .. } => {
if let Some(identifier) = call_id.clone().or_else(|| id.clone()) {
shell_call_ids.insert(identifier);
}
}
ResponseItem::CustomToolCall {
id: _,
status: _,
call_id,
name,
input: _,
} => {
if name == "apply_patch" {
shell_call_ids.insert(call_id.clone());
}
}
ResponseItem::CustomToolCallOutput { call_id, output } => {
if shell_call_ids.remove(call_id)
&& let Some(structured) = parse_structured_shell_output(output)
{
*output = structured
}
}
ResponseItem::FunctionCall { name, call_id, .. }
if is_shell_tool_name(name) || name == "apply_patch" =>
{
shell_call_ids.insert(call_id.clone());
}
ResponseItem::FunctionCallOutput { call_id, output } => {
if shell_call_ids.remove(call_id)
&& let Some(structured) = parse_structured_shell_output(&output.content)
{
output.content = structured
}
}
_ => {}
})
}
fn is_shell_tool_name(name: &str) -> bool {
matches!(name, "shell" | "container.exec")
}
#[derive(Deserialize)]
struct ExecOutputJson {
output: String,
metadata: ExecOutputMetadataJson,
}
#[derive(Deserialize)]
struct ExecOutputMetadataJson {
exit_code: i32,
duration_seconds: f32,
}
fn parse_structured_shell_output(raw: &str) -> Option<String> {
let parsed: ExecOutputJson = serde_json::from_str(raw).ok()?;
Some(build_structured_output(&parsed))
}
fn build_structured_output(parsed: &ExecOutputJson) -> String {
let mut sections = Vec::new();
sections.push(format!("Exit code: {}", parsed.metadata.exit_code));
sections.push(format!(
"Wall time: {} seconds",
parsed.metadata.duration_seconds
));
let mut output = parsed.output.clone();
if let Some(total_lines) = extract_total_output_lines(&parsed.output) {
sections.push(format!("Total output lines: {total_lines}"));
if let Some(stripped) = strip_total_output_header(&output) {
output = stripped.to_string();
}
}
sections.push("Output:".to_string());
sections.push(output);
sections.join("\n")
}
fn extract_total_output_lines(output: &str) -> Option<u32> {
let marker_start = output.find("[... omitted ")?;
let marker = &output[marker_start..];
let (_, after_of) = marker.split_once(" of ")?;
let (total_segment, _) = after_of.split_once(' ')?;
total_segment.parse::<u32>().ok()
}
fn strip_total_output_header(output: &str) -> Option<&str> {
let after_prefix = output.strip_prefix("Total output lines: ")?;
let (_, remainder) = after_prefix.split_once('\n')?;
let remainder = remainder.strip_prefix('\n').unwrap_or(remainder);
Some(remainder)
}
#[derive(Debug)]
pub enum ResponseEvent {
Created,
OutputItemDone(ResponseItem),
Completed {
response_id: String,
token_usage: Option<TokenUsage>,
},
OutputTextDelta(String),
ReasoningSummaryDelta(String),
ReasoningContentDelta(String),
ReasoningSummaryPartAdded,
WebSearchCallBegin {
call_id: String,
},
RateLimits(RateLimitSnapshot),
}
#[derive(Debug, Serialize)]
pub(crate) struct Reasoning {
#[serde(skip_serializing_if = "Option::is_none")]
pub(crate) effort: Option<ReasoningEffortConfig>,
#[serde(skip_serializing_if = "Option::is_none")]
pub(crate) summary: Option<ReasoningSummaryConfig>,
}
#[derive(Debug, Serialize, Default, Clone)]
#[serde(rename_all = "snake_case")]
pub(crate) enum TextFormatType {
#[default]
JsonSchema,
}
#[derive(Debug, Serialize, Default, Clone)]
pub(crate) struct TextFormat {
pub(crate) r#type: TextFormatType,
pub(crate) strict: bool,
pub(crate) schema: Value,
pub(crate) name: String,
}
/// Controls under the `text` field in the Responses API for GPT-5.
#[derive(Debug, Serialize, Default, Clone)]
pub(crate) struct TextControls {
#[serde(skip_serializing_if = "Option::is_none")]
pub(crate) verbosity: Option<OpenAiVerbosity>,
#[serde(skip_serializing_if = "Option::is_none")]
pub(crate) format: Option<TextFormat>,
}
#[derive(Debug, Serialize, Default, Clone)]
#[serde(rename_all = "lowercase")]
pub(crate) enum OpenAiVerbosity {
Low,
#[default]
Medium,
High,
}
impl From<VerbosityConfig> for OpenAiVerbosity {
fn from(v: VerbosityConfig) -> Self {
match v {
VerbosityConfig::Low => OpenAiVerbosity::Low,
VerbosityConfig::Medium => OpenAiVerbosity::Medium,
VerbosityConfig::High => OpenAiVerbosity::High,
}
}
}
/// Request object that is serialized as JSON and POST'ed when using the
/// Responses API.
#[derive(Debug, Serialize)]
pub(crate) struct ResponsesApiRequest<'a> {
pub(crate) model: &'a str,
pub(crate) instructions: &'a str,
// TODO(mbolin): ResponseItem::Other should not be serialized. Currently,
// we code defensively to avoid this case, but perhaps we should use a
// separate enum for serialization.
pub(crate) input: &'a Vec<ResponseItem>,
pub(crate) tools: &'a [serde_json::Value],
pub(crate) tool_choice: &'static str,
pub(crate) parallel_tool_calls: bool,
pub(crate) reasoning: Option<Reasoning>,
pub(crate) store: bool,
pub(crate) stream: bool,
pub(crate) include: Vec<String>,
#[serde(skip_serializing_if = "Option::is_none")]
pub(crate) prompt_cache_key: Option<String>,
#[serde(skip_serializing_if = "Option::is_none")]
pub(crate) text: Option<TextControls>,
}
pub(crate) mod tools {
use crate::openai_tools::JsonSchema;
use serde::Deserialize;
use serde::Serialize;
/// When serialized as JSON, this produces a valid "Tool" in the OpenAI
/// Responses API.
#[derive(Debug, Clone, Serialize, PartialEq)]
#[serde(tag = "type")]
pub(crate) enum ToolSpec {
#[serde(rename = "function")]
Function(ResponsesApiTool),
#[serde(rename = "local_shell")]
LocalShell {},
// TODO: Understand why we get an error on web_search although the API docs say it's supported.
// https://platform.openai.com/docs/guides/tools-web-search?api-mode=responses#:~:text=%7B%20type%3A%20%22web_search%22%20%7D%2C
#[serde(rename = "web_search")]
WebSearch {},
#[serde(rename = "custom")]
Freeform(FreeformTool),
}
impl ToolSpec {
pub(crate) fn name(&self) -> &str {
match self {
ToolSpec::Function(tool) => tool.name.as_str(),
ToolSpec::LocalShell {} => "local_shell",
ToolSpec::WebSearch {} => "web_search",
ToolSpec::Freeform(tool) => tool.name.as_str(),
}
}
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
pub struct FreeformTool {
pub(crate) name: String,
pub(crate) description: String,
pub(crate) format: FreeformToolFormat,
}
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
pub struct FreeformToolFormat {
pub(crate) r#type: String,
pub(crate) syntax: String,
pub(crate) definition: String,
}
#[derive(Debug, Clone, Serialize, PartialEq)]
pub struct ResponsesApiTool {
pub(crate) name: String,
pub(crate) description: String,
/// TODO: Validation. When strict is set to true, the JSON schema,
/// `required` and `additional_properties` must be present. All fields in
/// `properties` must be present in `required`.
pub(crate) strict: bool,
pub(crate) parameters: JsonSchema,
}
}
pub(crate) fn create_reasoning_param_for_request(
model_family: &ModelFamily,
effort: Option<ReasoningEffortConfig>,
summary: ReasoningSummaryConfig,
) -> Option<Reasoning> {
if !model_family.supports_reasoning_summaries {
return None;
}
Some(Reasoning {
effort,
summary: Some(summary),
})
}
pub(crate) fn create_text_param_for_request(
verbosity: Option<VerbosityConfig>,
output_schema: &Option<Value>,
) -> Option<TextControls> {
if verbosity.is_none() && output_schema.is_none() {
return None;
}
Some(TextControls {
verbosity: verbosity.map(std::convert::Into::into),
format: output_schema.as_ref().map(|schema| TextFormat {
r#type: TextFormatType::JsonSchema,
strict: true,
schema: schema.clone(),
name: "codex_output_schema".to_string(),
}),
})
}
pub struct ResponseStream {
pub(crate) rx_event: mpsc::Receiver<Result<ResponseEvent>>,
}
impl Stream for ResponseStream {
type Item = Result<ResponseEvent>;
fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>> {
self.rx_event.poll_recv(cx)
}
}
#[cfg(test)]
mod tests {
use crate::model_family::find_family_for_model;
use pretty_assertions::assert_eq;
use super::*;
struct InstructionsTestCase {
pub slug: &'static str,
pub expects_apply_patch_instructions: bool,
}
#[test]
fn get_full_instructions_no_user_content() {
let prompt = Prompt {
..Default::default()
};
let test_cases = vec![
InstructionsTestCase {
slug: "gpt-3.5",
expects_apply_patch_instructions: true,
},
InstructionsTestCase {
slug: "gpt-4.1",
expects_apply_patch_instructions: true,
},
InstructionsTestCase {
slug: "gpt-4o",
expects_apply_patch_instructions: true,
},
InstructionsTestCase {
slug: "gpt-5",
expects_apply_patch_instructions: true,
},
InstructionsTestCase {
slug: "codex-mini-latest",
expects_apply_patch_instructions: true,
},
InstructionsTestCase {
slug: "gpt-oss:120b",
expects_apply_patch_instructions: false,
},
InstructionsTestCase {
slug: "gpt-5-codex",
expects_apply_patch_instructions: false,
},
];
for test_case in test_cases {
let model_family = find_family_for_model(test_case.slug).expect("known model slug");
let expected = if test_case.expects_apply_patch_instructions {
format!(
"{}\n{}",
model_family.clone().base_instructions,
APPLY_PATCH_TOOL_INSTRUCTIONS
)
} else {
model_family.clone().base_instructions
};
let full = prompt.get_full_instructions(&model_family);
assert_eq!(full, expected);
}
}
#[test]
fn serializes_text_verbosity_when_set() {
let input: Vec<ResponseItem> = vec![];
let tools: Vec<serde_json::Value> = vec![];
let req = ResponsesApiRequest {
model: "gpt-5",
instructions: "i",
input: &input,
tools: &tools,
tool_choice: "auto",
parallel_tool_calls: true,
reasoning: None,
store: false,
stream: true,
include: vec![],
prompt_cache_key: None,
text: Some(TextControls {
verbosity: Some(OpenAiVerbosity::Low),
format: None,
}),
};
let v = serde_json::to_value(&req).expect("json");
assert_eq!(
v.get("text")
.and_then(|t| t.get("verbosity"))
.and_then(|s| s.as_str()),
Some("low")
);
}
#[test]
fn serializes_text_schema_with_strict_format() {
let input: Vec<ResponseItem> = vec![];
let tools: Vec<serde_json::Value> = vec![];
let schema = serde_json::json!({
"type": "object",
"properties": {
"answer": {"type": "string"}
},
"required": ["answer"],
});
let text_controls =
create_text_param_for_request(None, &Some(schema.clone())).expect("text controls");
let req = ResponsesApiRequest {
model: "gpt-5",
instructions: "i",
input: &input,
tools: &tools,
tool_choice: "auto",
parallel_tool_calls: true,
reasoning: None,
store: false,
stream: true,
include: vec![],
prompt_cache_key: None,
text: Some(text_controls),
};
let v = serde_json::to_value(&req).expect("json");
let text = v.get("text").expect("text field");
assert!(text.get("verbosity").is_none());
let format = text.get("format").expect("format field");
assert_eq!(
format.get("name"),
Some(&serde_json::Value::String("codex_output_schema".into()))
);
assert_eq!(
format.get("type"),
Some(&serde_json::Value::String("json_schema".into()))
);
assert_eq!(format.get("strict"), Some(&serde_json::Value::Bool(true)));
assert_eq!(format.get("schema"), Some(&schema));
}
#[test]
fn omits_text_when_not_set() {
let input: Vec<ResponseItem> = vec![];
let tools: Vec<serde_json::Value> = vec![];
let req = ResponsesApiRequest {
model: "gpt-5",
instructions: "i",
input: &input,
tools: &tools,
tool_choice: "auto",
parallel_tool_calls: true,
reasoning: None,
store: false,
stream: true,
include: vec![],
prompt_cache_key: None,
text: None,
};
let v = serde_json::to_value(&req).expect("json");
assert!(v.get("text").is_none());
}
}