codex/codex-rs/utils/output-truncation/src/lib.rs

//! Helpers for truncating tool and exec output using [`TruncationPolicy`](codex_protocol::protocol::TruncationPolicy).

use codex_protocol::models::FunctionCallOutputContentItem;
pub use codex_utils_string::approx_bytes_for_tokens;
pub use codex_utils_string::approx_token_count;
pub use codex_utils_string::approx_tokens_from_byte_count;
use codex_utils_string::truncate_middle_chars;
use codex_utils_string::truncate_middle_with_token_budget;

pub use codex_protocol::protocol::TruncationPolicy;

pub fn formatted_truncate_text(content: &str, policy: TruncationPolicy) -> String {
    if content.len() <= policy.byte_budget() {
        return content.to_string();
    }

    let total_lines = content.lines().count();
    let result = truncate_text(content, policy);
    format!("Total output lines: {total_lines}\n\n{result}")
}

pub fn truncate_text(content: &str, policy: TruncationPolicy) -> String {
    match policy {
        TruncationPolicy::Bytes(bytes) => truncate_middle_chars(content, bytes),
        TruncationPolicy::Tokens(tokens) => truncate_middle_with_token_budget(content, tokens).0,
    }
}

pub fn formatted_truncate_text_content_items_with_policy(
    items: &[FunctionCallOutputContentItem],
    policy: TruncationPolicy,
) -> (Vec<FunctionCallOutputContentItem>, Option<usize>) {
    let text_segments = items
        .iter()
        .filter_map(|item| match item {
            FunctionCallOutputContentItem::InputText { text } => Some(text.as_str()),
            FunctionCallOutputContentItem::InputImage { .. }
            | FunctionCallOutputContentItem::EncryptedContent { .. } => None,
        })
        .collect::<Vec<_>>();

    if text_segments.is_empty() {
        return (items.to_vec(), None);
    }

    let mut combined = String::new();
    for text in &text_segments {
        if !combined.is_empty() {
            combined.push('\n');
        }
        combined.push_str(text);
    }

    if combined.len() <= policy.byte_budget() {
        return (items.to_vec(), None);
    }

    let mut out = vec![FunctionCallOutputContentItem::InputText {
        text: formatted_truncate_text(&combined, policy),
    }];
    out.extend(items.iter().filter_map(|item| match item {
        FunctionCallOutputContentItem::InputImage { image_url, detail } => {
            Some(FunctionCallOutputContentItem::InputImage {
                image_url: image_url.clone(),
                detail: *detail,
            })
        }
        FunctionCallOutputContentItem::EncryptedContent { encrypted_content } => {
            Some(FunctionCallOutputContentItem::EncryptedContent {
                encrypted_content: encrypted_content.clone(),
            })
        }
        FunctionCallOutputContentItem::InputText { .. } => None,
    }));

    (out, Some(approx_token_count(&combined)))
}

pub fn truncate_function_output_items_with_policy(
    items: &[FunctionCallOutputContentItem],
    policy: TruncationPolicy,
) -> Vec<FunctionCallOutputContentItem> {
    let mut out: Vec<FunctionCallOutputContentItem> = Vec::with_capacity(items.len());
    let mut remaining_budget = match policy {
        TruncationPolicy::Bytes(_) => policy.byte_budget(),
        TruncationPolicy::Tokens(_) => policy.token_budget(),
    };
    let mut omitted_text_items = 0usize;

    for item in items {
        match item {
            FunctionCallOutputContentItem::InputText { text } => {
                if remaining_budget == 0 {
                    omitted_text_items += 1;
                    continue;
                }

                let cost = match policy {
                    TruncationPolicy::Bytes(_) => text.len(),
                    TruncationPolicy::Tokens(_) => approx_token_count(text),
                };

                if cost <= remaining_budget {
                    out.push(FunctionCallOutputContentItem::InputText { text: text.clone() });
                    remaining_budget = remaining_budget.saturating_sub(cost);
                } else {
                    let snippet_policy = match policy {
                        TruncationPolicy::Bytes(_) => TruncationPolicy::Bytes(remaining_budget),
                        TruncationPolicy::Tokens(_) => TruncationPolicy::Tokens(remaining_budget),
                    };
                    let snippet = truncate_text(text, snippet_policy);
                    if snippet.is_empty() {
                        omitted_text_items += 1;
                    } else {
                        out.push(FunctionCallOutputContentItem::InputText { text: snippet });
                    }
                    remaining_budget = 0;
                }
            }
            FunctionCallOutputContentItem::InputImage { image_url, detail } => {
                out.push(FunctionCallOutputContentItem::InputImage {
                    image_url: image_url.clone(),
                    detail: *detail,
                });
            }
            FunctionCallOutputContentItem::EncryptedContent { encrypted_content } => {
                out.push(FunctionCallOutputContentItem::EncryptedContent {
                    encrypted_content: encrypted_content.clone(),
                });
            }
        }
    }

    if omitted_text_items > 0 {
        out.push(FunctionCallOutputContentItem::InputText {
            text: format!("[omitted {omitted_text_items} text items ...]"),
        });
    }

    out
}

pub fn approx_tokens_from_byte_count_i64(bytes: i64) -> i64 {
    if bytes <= 0 {
        return 0;
    }

    let bytes = usize::try_from(bytes).unwrap_or(usize::MAX);
    i64::try_from(approx_tokens_from_byte_count(bytes)).unwrap_or(i64::MAX)
}

#[cfg(test)]
mod truncate_tests;