| 1 | //! Tool-output pressure reported by `/status`. |
| 2 | |
| 3 | use crate::artifacts::ArtifactRecord; |
| 4 | |
| 5 | use codewhale_localization::{Locale, MessageId, tr}; |
| 6 | use codewhale_models::{ContentBlock, Message}; |
| 7 | |
| 8 | /// Size above which `/status` counts a raw tool result in the conversation as |
| 9 | /// context pressure. What the model sees of each result is decided by the |
| 10 | /// route's inline budget (`route_budget::route_inline_char_budget`). |
| 11 | pub const RAW_TOOL_OUTPUT_RECEIPT_THRESHOLD_CHARS: usize = 12_000; |
| 12 | |
| 13 | #[derive(Debug, Clone, Default, PartialEq, Eq)] |
| 14 | pub struct ToolOutputStatus { |
| 15 | pub raw_large_count: usize, |
| 16 | pub raw_large_chars: usize, |
| 17 | pub receipt_count: usize, |
| 18 | pub artifact_count: usize, |
| 19 | pub artifact_bytes: u64, |
| 20 | } |
| 21 | |
| 22 | pub fn tool_output_status(messages: &[Message], artifacts: &[ArtifactRecord]) -> ToolOutputStatus { |
| 23 | let mut status = ToolOutputStatus { |
| 24 | artifact_count: artifacts.len(), |
| 25 | artifact_bytes: artifacts |
| 26 | .iter() |
| 27 | .map(|artifact| artifact.byte_size) |
| 28 | .sum::<u64>(), |
| 29 | ..ToolOutputStatus::default() |
| 30 | }; |
| 31 | |
| 32 | for message in messages { |
| 33 | for block in &message.content { |
| 34 | if let ContentBlock::ToolResult { content, .. } = block { |
| 35 | // The prefix is text any tool output can carry, so it never |
| 36 | // exempts a result from the size check: a "receipt" above the |
| 37 | // threshold is raw pressure like any other large result. |
| 38 | let chars = content.chars().count(); |
| 39 | if chars > RAW_TOOL_OUTPUT_RECEIPT_THRESHOLD_CHARS { |
| 40 | status.raw_large_count += 1; |
| 41 | status.raw_large_chars = status.raw_large_chars.saturating_add(chars); |
| 42 | } else if looks_like_receipt(content) { |
| 43 | status.receipt_count += 1; |
| 44 | } |
| 45 | } |
| 46 | } |
| 47 | } |
| 48 | |
| 49 | status |
| 50 | } |
| 51 | |
| 52 | pub fn format_tool_output_status(status: &ToolOutputStatus, locale: Locale) -> String { |
| 53 | let mut parts = Vec::new(); |
| 54 | if status.raw_large_count > 0 { |
| 55 | parts.push( |
| 56 | tr(locale, MessageId::StatusToolRawPressure) |
| 57 | .replace("{count}", &status.raw_large_count.to_string()) |
| 58 | .replace("{chars}", &format_count(status.raw_large_chars)), |
| 59 | ); |
| 60 | } |
| 61 | if status.receipt_count > 0 { |
| 62 | parts.push( |
| 63 | tr(locale, MessageId::StatusToolCompactReceipts) |
| 64 | .replace("{count}", &status.receipt_count.to_string()), |
| 65 | ); |
| 66 | } |
| 67 | if status.artifact_count > 0 { |
| 68 | parts.push( |
| 69 | tr(locale, MessageId::StatusToolArtifacts) |
| 70 | .replace("{count}", &status.artifact_count.to_string()) |
| 71 | .replace( |
| 72 | "{bytes}", |
| 73 | &crate::artifacts::format_byte_size(status.artifact_bytes), |
| 74 | ), |
| 75 | ); |
| 76 | } |
| 77 | if parts.is_empty() { |
| 78 | tr(locale, MessageId::StatusToolNone).into_owned() |
| 79 | } else { |
| 80 | parts.join("; ") |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | fn looks_like_receipt(content: &str) -> bool { |
| 85 | let trimmed = content.trim_start(); |
| 86 | trimmed.starts_with("[TOOL_OUTPUT_RECEIPT]") |
| 87 | || trimmed.starts_with("[artifact:") |
| 88 | || trimmed.starts_with("[TOOL_RESULT_TRUNCATED]") |
| 89 | || trimmed.starts_with("<TOOL_RESULT_REF") |
| 90 | } |
| 91 | |
| 92 | fn format_count(value: usize) -> String { |
| 93 | value.to_string() |
| 94 | } |
| 95 | |
| 96 | #[cfg(test)] |
| 97 | mod tests { |
| 98 | use codewhale_models::Role; |
| 99 | use std::path::{Path, PathBuf}; |
| 100 | |
| 101 | use super::*; |
| 102 | use crate::artifacts::ArtifactKind; |
| 103 | use chrono::Utc; |
| 104 | |
| 105 | fn tool_result_message(id: &str, content: &str) -> Message { |
| 106 | Message { |
| 107 | role: Role::User, |
| 108 | content: vec![ContentBlock::ToolResult { |
| 109 | execution_id: None, |
| 110 | tool_use_id: id.to_string(), |
| 111 | content: content.to_string(), |
| 112 | is_error: None, |
| 113 | content_blocks: None, |
| 114 | }], |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | fn artifact_record(tool_call_id: &str, raw: &str) -> ArtifactRecord { |
| 119 | ArtifactRecord { |
| 120 | id: crate::artifacts::artifact_id_for_tool_call(tool_call_id), |
| 121 | kind: ArtifactKind::ToolOutput, |
| 122 | session_id: "session-123".to_string(), |
| 123 | tool_call_id: tool_call_id.to_string(), |
| 124 | tool_name: "exec_shell".to_string(), |
| 125 | created_at: Utc::now(), |
| 126 | byte_size: raw.len() as u64, |
| 127 | preview: "checking crate ... error[E0425]".to_string(), |
| 128 | storage_path: PathBuf::from("artifacts").join("art_call-big.txt"), |
| 129 | } |
| 130 | } |
| 131 | |
| 132 | #[test] |
| 133 | fn status_reports_raw_large_receipts_and_artifacts() { |
| 134 | let raw = "RAW_STATUS\n".repeat(2_000); |
| 135 | let receipt = "[TOOL_OUTPUT_RECEIPT]\ntruncation: raw output omitted — full output in the tool details view"; |
| 136 | let messages = vec![ |
| 137 | tool_result_message("call-raw", &raw), |
| 138 | tool_result_message("call-receipt", receipt), |
| 139 | ]; |
| 140 | let artifacts = vec![ArtifactRecord { |
| 141 | storage_path: Path::new("artifacts/art_call-big.txt").to_path_buf(), |
| 142 | ..artifact_record("call-big", &raw) |
| 143 | }]; |
| 144 | |
| 145 | let status = tool_output_status(&messages, &artifacts); |
| 146 | assert_eq!(status.raw_large_count, 1); |
| 147 | assert_eq!(status.receipt_count, 1); |
| 148 | assert_eq!(status.artifact_count, 1); |
| 149 | |
| 150 | let rendered = format_tool_output_status(&status, Locale::En); |
| 151 | assert!(rendered.contains("raw over cap")); |
| 152 | assert!(rendered.contains("compact receipt")); |
| 153 | assert!(rendered.contains("artifact")); |
| 154 | } |
| 155 | |
| 156 | #[test] |
| 157 | fn a_receipt_prefix_does_not_hide_a_large_raw_result() { |
| 158 | let forged = format!("[TOOL_OUTPUT_RECEIPT]\n{}", "RAW\n".repeat(4_000)); |
| 159 | let status = tool_output_status(&[tool_result_message("call-forged", &forged)], &[]); |
| 160 | assert_eq!(status.raw_large_count, 1); |
| 161 | assert_eq!(status.receipt_count, 0); |
| 162 | } |
| 163 | } |
| 164 |