mirror of
https://github.com/openai/codex.git
synced 2026-09-06 15:29:32 +00:00
## What changed - Add a positive `output_token_limit` setting to each entry under an MCP server's `tools` configuration. - Apply the most restrictive limit when plugin and user policies overlap, while keeping approval policy independent. - Carry the effective MCP output budget in conversation history so tool output, post-tool hook responses, and resumed sessions use the same truncation limit. ## Testing - Cover configuration parsing, serialization, schema validation, and plugin policy merging. - Cover MCP output below and above the configured limit, post-tool hook responses, and session resume. GitOrigin-RevId: d0beb4fca9ba6055d9e1d31c137373b465d50d61
574 lines
19 KiB
Rust
574 lines
19 KiB
Rust
use crate::original_image_detail::sanitize_original_image_detail;
|
|
use crate::session::session::Session;
|
|
use crate::session::step_context::StepContext;
|
|
use crate::session::turn_context::TurnContext;
|
|
use crate::turn_diff_tracker::TurnDiffTracker;
|
|
use crate::unified_exec::format_output_omission_marker;
|
|
use crate::unified_exec::resolve_max_tokens;
|
|
use codex_protocol::ResponseItemId;
|
|
use codex_protocol::mcp::CallToolResult;
|
|
use codex_protocol::models::FunctionCallOutputBody;
|
|
use codex_protocol::models::FunctionCallOutputContentItem;
|
|
use codex_protocol::models::FunctionCallOutputPayload;
|
|
use codex_protocol::models::ResponseInputItem;
|
|
use codex_protocol::models::ResponseItem;
|
|
use codex_protocol::models::function_call_output_content_items_to_text;
|
|
use codex_tools::LoadableToolSpec;
|
|
use codex_tools::ToolName;
|
|
use codex_utils_audio::estimate_audio_token_count;
|
|
use codex_utils_output_truncation::TruncationPolicy;
|
|
use codex_utils_output_truncation::approx_token_count;
|
|
use codex_utils_output_truncation::formatted_truncate_text;
|
|
use codex_utils_output_truncation::truncate_function_output_payload;
|
|
use codex_utils_output_truncation::truncate_text;
|
|
use serde::Serialize;
|
|
use serde_json::Value as JsonValue;
|
|
use std::num::NonZeroUsize;
|
|
use std::sync::Arc;
|
|
use std::time::Duration;
|
|
use tokio::sync::Mutex;
|
|
use tokio_util::sync::CancellationToken;
|
|
|
|
pub use codex_tools::ToolOutput;
|
|
pub use codex_tools::ToolPayload;
|
|
|
|
pub(crate) fn boxed_tool_output<T>(output: T) -> Box<dyn ToolOutput>
|
|
where
|
|
T: ToolOutput + 'static,
|
|
{
|
|
Box::new(output)
|
|
}
|
|
|
|
pub type SharedTurnDiffTracker = Arc<Mutex<TurnDiffTracker>>;
|
|
|
|
#[derive(Clone, Debug, Eq, PartialEq)]
|
|
pub enum ToolCallSource {
|
|
Direct,
|
|
DirectPlaintextMessage,
|
|
CodeMode {
|
|
/// Runtime cell that issued the nested tool request.
|
|
cell_id: String,
|
|
/// Code-mode's per-cell tool invocation id. This is useful for
|
|
/// debugging the JS/runtime bridge, but it is not the Codex tool call id
|
|
/// because the runtime id only needs to be unique within one cell.
|
|
runtime_tool_call_id: String,
|
|
},
|
|
}
|
|
|
|
#[derive(Clone)]
|
|
pub struct ToolInvocation {
|
|
pub session: Arc<Session>,
|
|
// TODO(sayan): Remove this compatibility field once handlers use `step_context.turn`.
|
|
pub turn: Arc<TurnContext>,
|
|
pub(crate) step_context: Arc<StepContext>,
|
|
pub cancellation_token: CancellationToken,
|
|
pub tracker: SharedTurnDiffTracker,
|
|
pub call_id: String,
|
|
pub tool_name: ToolName,
|
|
pub source: ToolCallSource,
|
|
pub payload: ToolPayload,
|
|
}
|
|
|
|
impl ToolInvocation {
|
|
/// Returns the Responses item that requested this call or started its code-mode cell.
|
|
pub(crate) async fn originating_item_id(&self) -> Option<ResponseItemId> {
|
|
if let ToolCallSource::CodeMode { cell_id, .. } = &self.source {
|
|
return self
|
|
.session
|
|
.services
|
|
.code_mode_service
|
|
.cell_originating_item_id(&codex_code_mode::CellId::new(cell_id.clone()));
|
|
}
|
|
|
|
self.session
|
|
.clone_history()
|
|
.await
|
|
.raw_items()
|
|
.rev()
|
|
.find_map(|item| match item {
|
|
ResponseItem::FunctionCall { id, call_id, .. }
|
|
| ResponseItem::CustomToolCall { id, call_id, .. }
|
|
if call_id == &self.call_id =>
|
|
{
|
|
id.clone()
|
|
}
|
|
_ => None,
|
|
})
|
|
}
|
|
}
|
|
|
|
#[derive(Clone, Debug)]
|
|
pub struct McpToolOutput {
|
|
pub result: CallToolResult,
|
|
pub tool_input: JsonValue,
|
|
pub wall_time: Duration,
|
|
pub original_image_detail_supported: bool,
|
|
pub truncation_policy: TruncationPolicy,
|
|
}
|
|
|
|
impl ToolOutput for McpToolOutput {
|
|
fn log_output(&self) -> String {
|
|
// Logging has its own budget; do not first apply the model-context budget.
|
|
let output = self.result.log_output();
|
|
let wall_time_seconds = self.wall_time.as_secs_f64();
|
|
let header = format!("Wall time: {wall_time_seconds:.4} seconds\nOutput:");
|
|
if output.is_empty() {
|
|
header
|
|
} else {
|
|
format!("{header}\n{output}")
|
|
}
|
|
}
|
|
|
|
fn success_for_logging(&self) -> bool {
|
|
self.result.success()
|
|
}
|
|
|
|
fn fallback_token_limit_override(&self) -> Option<usize> {
|
|
Some((self.truncation_policy * 1.2).token_budget())
|
|
}
|
|
|
|
fn to_response_item(&self, call_id: &str, _payload: &ToolPayload) -> ResponseInputItem {
|
|
ResponseInputItem::FunctionCallOutput {
|
|
call_id: call_id.to_string(),
|
|
output: self.response_payload(),
|
|
}
|
|
}
|
|
|
|
fn code_mode_result(&self, payload: &ToolPayload) -> JsonValue {
|
|
self.result.code_mode_result(payload)
|
|
}
|
|
|
|
fn post_tool_use_input(&self, _payload: &ToolPayload) -> Option<JsonValue> {
|
|
Some(self.tool_input.clone())
|
|
}
|
|
|
|
fn post_tool_use_response(&self, _call_id: &str, _payload: &ToolPayload) -> Option<JsonValue> {
|
|
serde_json::to_value(&self.result).ok()
|
|
}
|
|
}
|
|
|
|
impl McpToolOutput {
|
|
fn response_payload(&self) -> FunctionCallOutputPayload {
|
|
let mut payload = self.result.as_function_call_output_payload();
|
|
if let Some(items) = payload.content_items_mut() {
|
|
sanitize_original_image_detail(self.original_image_detail_supported, items);
|
|
}
|
|
|
|
let wall_time_seconds = self.wall_time.as_secs_f64();
|
|
let header = format!("Wall time: {wall_time_seconds:.4} seconds\nOutput:");
|
|
|
|
match &mut payload.body {
|
|
FunctionCallOutputBody::Text(text) => {
|
|
if text.is_empty() {
|
|
*text = header;
|
|
} else {
|
|
*text = format!("{header}\n{text}");
|
|
}
|
|
}
|
|
FunctionCallOutputBody::ContentItems(items) => {
|
|
items.insert(0, FunctionCallOutputContentItem::InputText { text: header });
|
|
}
|
|
}
|
|
|
|
// History receives this budget in tokens. Code Mode keeps the raw result.
|
|
truncate_function_output_payload(
|
|
&mut payload,
|
|
self.truncation_policy * 1.2,
|
|
estimate_audio_token_count,
|
|
);
|
|
payload
|
|
}
|
|
}
|
|
|
|
#[derive(Clone)]
|
|
pub struct ToolSearchOutput {
|
|
pub tools: Vec<LoadableToolSpec>,
|
|
}
|
|
|
|
impl ToolOutput for ToolSearchOutput {
|
|
fn log_output(&self) -> String {
|
|
let tools = self
|
|
.tools
|
|
.iter()
|
|
.map(|tool| {
|
|
serde_json::to_value(tool).unwrap_or_else(|err| {
|
|
JsonValue::String(format!("failed to serialize tool_search output: {err}"))
|
|
})
|
|
})
|
|
.collect();
|
|
JsonValue::Array(tools).to_string()
|
|
}
|
|
|
|
fn success_for_logging(&self) -> bool {
|
|
true
|
|
}
|
|
|
|
fn to_response_item(&self, call_id: &str, _payload: &ToolPayload) -> ResponseInputItem {
|
|
ResponseInputItem::ToolSearchOutput {
|
|
call_id: call_id.to_string(),
|
|
status: "completed".to_string(),
|
|
execution: "client".to_string(),
|
|
tools: self
|
|
.tools
|
|
.iter()
|
|
.map(|tool| {
|
|
serde_json::to_value(tool).unwrap_or_else(|err| {
|
|
JsonValue::String(format!("failed to serialize tool_search output: {err}"))
|
|
})
|
|
})
|
|
.collect(),
|
|
}
|
|
}
|
|
}
|
|
|
|
pub struct FunctionToolOutput {
|
|
pub body: Vec<FunctionCallOutputContentItem>,
|
|
pub success: Option<bool>,
|
|
pub post_tool_use_response: Option<JsonValue>,
|
|
}
|
|
|
|
impl FunctionToolOutput {
|
|
pub fn from_text(text: String, success: Option<bool>) -> Self {
|
|
Self {
|
|
body: vec![FunctionCallOutputContentItem::InputText { text }],
|
|
success,
|
|
post_tool_use_response: None,
|
|
}
|
|
}
|
|
|
|
pub fn from_content(
|
|
content: Vec<FunctionCallOutputContentItem>,
|
|
success: Option<bool>,
|
|
) -> Self {
|
|
Self {
|
|
body: content,
|
|
success,
|
|
post_tool_use_response: None,
|
|
}
|
|
}
|
|
|
|
pub fn into_text(self) -> String {
|
|
function_call_output_content_items_to_text(&self.body).unwrap_or_default()
|
|
}
|
|
}
|
|
|
|
impl ToolOutput for FunctionToolOutput {
|
|
fn log_output(&self) -> String {
|
|
function_call_output_content_items_to_text(&self.body).unwrap_or_default()
|
|
}
|
|
|
|
fn success_for_logging(&self) -> bool {
|
|
self.success.unwrap_or(true)
|
|
}
|
|
|
|
fn to_response_item(&self, call_id: &str, payload: &ToolPayload) -> ResponseInputItem {
|
|
function_tool_response(call_id, payload, self.body.clone(), self.success)
|
|
}
|
|
|
|
fn post_tool_use_response(&self, _call_id: &str, _payload: &ToolPayload) -> Option<JsonValue> {
|
|
self.post_tool_use_response.clone()
|
|
}
|
|
}
|
|
|
|
pub struct ApplyPatchToolOutput {
|
|
pub text: String,
|
|
}
|
|
|
|
impl ApplyPatchToolOutput {
|
|
pub fn from_text(text: String) -> Self {
|
|
Self { text }
|
|
}
|
|
}
|
|
|
|
impl ToolOutput for ApplyPatchToolOutput {
|
|
fn log_output(&self) -> String {
|
|
self.text.clone()
|
|
}
|
|
|
|
fn success_for_logging(&self) -> bool {
|
|
true
|
|
}
|
|
|
|
fn to_response_item(&self, call_id: &str, payload: &ToolPayload) -> ResponseInputItem {
|
|
function_tool_response(
|
|
call_id,
|
|
payload,
|
|
vec![FunctionCallOutputContentItem::InputText {
|
|
text: self.text.clone(),
|
|
}],
|
|
Some(true),
|
|
)
|
|
}
|
|
|
|
fn post_tool_use_response(&self, _call_id: &str, _payload: &ToolPayload) -> Option<JsonValue> {
|
|
Some(JsonValue::String(self.text.clone()))
|
|
}
|
|
|
|
fn code_mode_result(&self, _payload: &ToolPayload) -> JsonValue {
|
|
JsonValue::Object(serde_json::Map::new())
|
|
}
|
|
}
|
|
|
|
pub struct AbortedToolOutput {
|
|
pub message: String,
|
|
}
|
|
|
|
impl ToolOutput for AbortedToolOutput {
|
|
fn log_output(&self) -> String {
|
|
self.message.clone()
|
|
}
|
|
|
|
fn success_for_logging(&self) -> bool {
|
|
false
|
|
}
|
|
|
|
fn to_response_item(&self, call_id: &str, payload: &ToolPayload) -> ResponseInputItem {
|
|
match payload {
|
|
ToolPayload::ToolSearch { .. } => ResponseInputItem::ToolSearchOutput {
|
|
call_id: call_id.to_string(),
|
|
status: "completed".to_string(),
|
|
execution: "client".to_string(),
|
|
tools: Vec::new(),
|
|
},
|
|
_ => function_tool_response(
|
|
call_id,
|
|
payload,
|
|
vec![FunctionCallOutputContentItem::InputText {
|
|
text: self.message.clone(),
|
|
}],
|
|
/*success*/ None,
|
|
),
|
|
}
|
|
}
|
|
}
|
|
|
|
#[derive(Debug, Clone, PartialEq)]
|
|
pub struct ExecCommandToolOutput {
|
|
pub event_call_id: String,
|
|
pub chunk_id: String,
|
|
pub wall_time: Duration,
|
|
/// Raw bytes returned for this unified exec call before any truncation.
|
|
pub raw_output: Vec<u8>,
|
|
pub truncation_policy: TruncationPolicy,
|
|
pub max_output_tokens: Option<usize>,
|
|
pub process_id: Option<i32>,
|
|
pub exit_code: Option<i32>,
|
|
pub original_token_count: Option<usize>,
|
|
/// Bytes omitted by the output collection cap before model-facing truncation.
|
|
pub output_omitted_bytes: Option<NonZeroUsize>,
|
|
pub hook_command: Option<String>,
|
|
}
|
|
|
|
impl ToolOutput for ExecCommandToolOutput {
|
|
fn log_output(&self) -> String {
|
|
// The telemetry budget must not inherit the model's output-token limit.
|
|
let mut output = String::from_utf8_lossy(&self.raw_output).into_owned();
|
|
if let Some(omitted_bytes) = self.output_omitted_bytes {
|
|
let marker = format_output_omission_marker(omitted_bytes.get());
|
|
if !output.contains(&marker) {
|
|
output = format!("{marker}\n{output}");
|
|
}
|
|
}
|
|
format!("{}\n{output}", self.response_header())
|
|
}
|
|
|
|
fn success_for_logging(&self) -> bool {
|
|
true
|
|
}
|
|
|
|
fn to_response_item(&self, call_id: &str, payload: &ToolPayload) -> ResponseInputItem {
|
|
function_tool_response(
|
|
call_id,
|
|
payload,
|
|
vec![FunctionCallOutputContentItem::InputText {
|
|
text: self.response_text(),
|
|
}],
|
|
Some(true),
|
|
)
|
|
}
|
|
|
|
fn post_tool_use_id(&self, call_id: &str) -> String {
|
|
if self.event_call_id.is_empty() {
|
|
call_id.to_string()
|
|
} else {
|
|
self.event_call_id.clone()
|
|
}
|
|
}
|
|
|
|
fn post_tool_use_input(&self, _payload: &ToolPayload) -> Option<JsonValue> {
|
|
self.hook_command
|
|
.as_ref()
|
|
.map(|command| serde_json::json!({ "command": command }))
|
|
}
|
|
|
|
fn post_tool_use_response(&self, _call_id: &str, _payload: &ToolPayload) -> Option<JsonValue> {
|
|
if self.process_id.is_some() || self.hook_command.is_none() {
|
|
return None;
|
|
}
|
|
|
|
Some(JsonValue::String(
|
|
self.truncated_output_with_policy(self.model_output_policy()),
|
|
))
|
|
}
|
|
|
|
fn code_mode_result(&self, _payload: &ToolPayload) -> JsonValue {
|
|
#[derive(Serialize)]
|
|
struct UnifiedExecCodeModeResult {
|
|
#[serde(skip_serializing_if = "Option::is_none")]
|
|
chunk_id: Option<String>,
|
|
wall_time_seconds: f64,
|
|
#[serde(skip_serializing_if = "Option::is_none")]
|
|
exit_code: Option<i32>,
|
|
#[serde(skip_serializing_if = "Option::is_none")]
|
|
session_id: Option<i32>,
|
|
#[serde(skip_serializing_if = "Option::is_none")]
|
|
original_token_count: Option<usize>,
|
|
output: String,
|
|
}
|
|
|
|
let result = UnifiedExecCodeModeResult {
|
|
chunk_id: (!self.chunk_id.is_empty()).then(|| self.chunk_id.clone()),
|
|
wall_time_seconds: self.wall_time.as_secs_f64(),
|
|
exit_code: self.exit_code,
|
|
session_id: self.process_id,
|
|
original_token_count: self.original_token_count,
|
|
output: match self.max_output_tokens {
|
|
Some(max_tokens) => self.truncated_output(max_tokens),
|
|
None => String::from_utf8_lossy(&self.raw_output).to_string(),
|
|
},
|
|
};
|
|
|
|
serde_json::to_value(result).unwrap_or_else(|err| {
|
|
JsonValue::String(format!("failed to serialize exec result: {err}"))
|
|
})
|
|
}
|
|
}
|
|
|
|
impl ExecCommandToolOutput {
|
|
fn model_output_policy(&self) -> TruncationPolicy {
|
|
let requested_policy = TruncationPolicy::Tokens(resolve_max_tokens(self.max_output_tokens));
|
|
if requested_policy.byte_budget() < self.truncation_policy.byte_budget() {
|
|
requested_policy
|
|
} else {
|
|
self.truncation_policy
|
|
}
|
|
}
|
|
|
|
pub(crate) fn truncated_output(&self, max_tokens: usize) -> String {
|
|
self.truncated_output_with_policy(TruncationPolicy::Tokens(max_tokens))
|
|
}
|
|
|
|
fn truncated_output_with_policy(&self, policy: TruncationPolicy) -> String {
|
|
let text = String::from_utf8_lossy(&self.raw_output).to_string();
|
|
let Some(omitted_bytes) = self.output_omitted_bytes else {
|
|
return formatted_truncate_text(&text, policy);
|
|
};
|
|
|
|
let marker = format_output_omission_marker(omitted_bytes.get());
|
|
if text.len() <= policy.byte_budget() {
|
|
return if text.contains(&marker) {
|
|
text
|
|
} else {
|
|
format!("{marker}\n{text}")
|
|
};
|
|
}
|
|
|
|
let original_token_count = self
|
|
.original_token_count
|
|
.unwrap_or_else(|| approx_token_count(&text));
|
|
let truncated = truncate_text(&text, policy);
|
|
let omission_notice = if truncated.contains(&marker) {
|
|
String::new()
|
|
} else {
|
|
format!("{marker}\n")
|
|
};
|
|
format!(
|
|
"Warning: truncated output (original token count: {original_token_count})\n{omission_notice}\n{truncated}"
|
|
)
|
|
}
|
|
|
|
fn response_header(&self) -> String {
|
|
let mut sections = Vec::new();
|
|
|
|
if !self.chunk_id.is_empty() {
|
|
sections.push(format!("Chunk ID: {}", self.chunk_id));
|
|
}
|
|
|
|
let wall_time_seconds = self.wall_time.as_secs_f64();
|
|
sections.push(format!("Wall time: {wall_time_seconds:.4} seconds"));
|
|
|
|
if let Some(exit_code) = self.exit_code {
|
|
sections.push(format!("Process exited with code {exit_code}"));
|
|
}
|
|
|
|
if let Some(process_id) = &self.process_id {
|
|
sections.push(format!("Process running with session ID {process_id}"));
|
|
}
|
|
|
|
if let Some(original_token_count) = self.original_token_count {
|
|
sections.push(format!("Original token count: {original_token_count}"));
|
|
}
|
|
|
|
sections.push("Output:".to_string());
|
|
sections.join("\n")
|
|
}
|
|
|
|
fn response_text(&self) -> String {
|
|
let header = self.response_header();
|
|
let output_budget = (self.truncation_policy * 1.2)
|
|
.byte_budget()
|
|
.saturating_sub(header.len().saturating_add(/*rhs*/ 1));
|
|
let mut policy = self.model_output_policy();
|
|
let mut output = self.truncated_output_with_policy(policy);
|
|
|
|
// History applies this same serialization budget to the complete response.
|
|
// Reserve room for metadata, warning headers, and the truncation marker so
|
|
// it does not truncate an already-truncated output a second time.
|
|
while output.len() > output_budget && policy.byte_budget() > 0 {
|
|
let excess_bytes = output.len() - output_budget;
|
|
policy = match policy {
|
|
TruncationPolicy::Bytes(bytes) => {
|
|
TruncationPolicy::Bytes(bytes.saturating_sub(excess_bytes))
|
|
}
|
|
TruncationPolicy::Tokens(tokens) => TruncationPolicy::Tokens(
|
|
tokens.saturating_sub(TruncationPolicy::Bytes(excess_bytes).token_budget()),
|
|
),
|
|
};
|
|
output = self.truncated_output_with_policy(policy);
|
|
}
|
|
|
|
format!("{header}\n{output}")
|
|
}
|
|
}
|
|
|
|
fn function_tool_response(
|
|
call_id: &str,
|
|
payload: &ToolPayload,
|
|
body: Vec<FunctionCallOutputContentItem>,
|
|
success: Option<bool>,
|
|
) -> ResponseInputItem {
|
|
let body = match body.as_slice() {
|
|
[FunctionCallOutputContentItem::InputText { text }] => {
|
|
FunctionCallOutputBody::Text(text.clone())
|
|
}
|
|
_ => FunctionCallOutputBody::ContentItems(body),
|
|
};
|
|
|
|
if matches!(payload, ToolPayload::Custom { .. }) {
|
|
return ResponseInputItem::CustomToolCallOutput {
|
|
call_id: call_id.to_string(),
|
|
name: None,
|
|
output: FunctionCallOutputPayload { body, success },
|
|
};
|
|
}
|
|
|
|
ResponseInputItem::FunctionCallOutput {
|
|
call_id: call_id.to_string(),
|
|
output: FunctionCallOutputPayload { body, success },
|
|
}
|
|
}
|
|
|
|
#[cfg(test)]
|
|
#[path = "context_tests.rs"]
|
|
mod tests;
|