| 1 | //! Deterministic integrity repair for persisted model-visible tool history. |
| 2 | //! |
| 3 | //! Session JSON predates a durable per-call journal, so process exit can leave |
| 4 | //! a `tool_use` without its terminal `tool_result`. Provider APIs reject that |
| 5 | //! shape. This module repairs the existing message format without changing its |
| 6 | //! schema and returns a bounded diagnostic receipt for every mutation. |
| 7 | |
| 8 | use std::collections::{HashMap, HashSet}; |
| 9 | |
| 10 | use codewhale_models::Role; |
| 11 | use codewhale_models::{ContentBlock, Message}; |
| 12 | |
| 13 | const CRASH_REPAIR_CONTENT: &str = |
| 14 | "Tool call interrupted by process exit; terminal status: crashed_and_repaired."; |
| 15 | |
| 16 | #[derive(Debug, Clone, Default, PartialEq, Eq)] |
| 17 | pub(crate) struct ToolRepairReceipt { |
| 18 | pub(crate) repaired_call_ids: Vec<String>, |
| 19 | pub(crate) duplicate_result_ids: Vec<String>, |
| 20 | pub(crate) orphan_result_ids: Vec<String>, |
| 21 | } |
| 22 | |
| 23 | impl ToolRepairReceipt { |
| 24 | #[must_use] |
| 25 | pub(crate) fn is_empty(&self) -> bool { |
| 26 | self.repaired_call_ids.is_empty() |
| 27 | && self.duplicate_result_ids.is_empty() |
| 28 | && self.orphan_result_ids.is_empty() |
| 29 | } |
| 30 | |
| 31 | fn visible_message(&self) -> String { |
| 32 | format!( |
| 33 | "[tool_history_repair] Repaired {} crashed tool call(s); quarantined {} duplicate and {} orphan terminal result(s).", |
| 34 | self.repaired_call_ids.len(), |
| 35 | self.duplicate_result_ids.len(), |
| 36 | self.orphan_result_ids.len(), |
| 37 | ) |
| 38 | } |
| 39 | } |
| 40 | |
| 41 | /// Repair tool-use/result integrity in place. |
| 42 | /// |
| 43 | /// The first terminal result after a known call and before the next assistant |
| 44 | /// turn is retained. Results that precede their call, arrive after a later |
| 45 | /// assistant turn, reference no call, or repeat a retained result are |
| 46 | /// quarantined by removing them from model-visible history. Every dangling |
| 47 | /// call receives a synthetic error result directly after its assistant call |
| 48 | /// message. A visible system receipt makes the repair apparent after resume. |
| 49 | pub(crate) fn repair_tool_call_pairs(messages: &mut Vec<Message>) -> ToolRepairReceipt { |
| 50 | repair_tool_call_pairs_inner(messages, true) |
| 51 | } |
| 52 | |
| 53 | /// Repair an ephemeral provider request without appending a trailing receipt. |
| 54 | /// |
| 55 | /// Anthropic-style APIs interpret a final assistant message as a completion |
| 56 | /// prefill. Pair repair must therefore leave the synthetic user tool result as |
| 57 | /// the request tail; the durable session-facing path owns the visible receipt. |
| 58 | pub(crate) fn repair_tool_call_pairs_for_provider( |
| 59 | messages: &mut Vec<Message>, |
| 60 | ) -> ToolRepairReceipt { |
| 61 | repair_tool_call_pairs_inner(messages, false) |
| 62 | } |
| 63 | |
| 64 | fn repair_tool_call_pairs_inner( |
| 65 | messages: &mut Vec<Message>, |
| 66 | append_visible_receipt: bool, |
| 67 | ) -> ToolRepairReceipt { |
| 68 | let mut pending_call_message = None; |
| 69 | let mut pending_call_ids = Vec::new(); |
| 70 | let mut retained_for_pending = HashSet::new(); |
| 71 | let mut missing_by_message: HashMap<usize, Vec<String>> = HashMap::new(); |
| 72 | let mut repaired_call_ids = Vec::new(); |
| 73 | let mut duplicate_result_ids = Vec::new(); |
| 74 | let mut orphan_result_ids = Vec::new(); |
| 75 | let mut keep_results = HashSet::new(); |
| 76 | let mut result_ordinal = 0usize; |
| 77 | |
| 78 | for (message_index, message) in messages.iter().enumerate() { |
| 79 | if message.role == "assistant" |
| 80 | || message.role == codewhale_models::INTERRUPTED_ASSISTANT_ROLE |
| 81 | { |
| 82 | record_missing_results( |
| 83 | pending_call_message, |
| 84 | &pending_call_ids, |
| 85 | &retained_for_pending, |
| 86 | &mut missing_by_message, |
| 87 | &mut repaired_call_ids, |
| 88 | ); |
| 89 | pending_call_ids = message |
| 90 | .content |
| 91 | .iter() |
| 92 | .filter_map(|block| match block { |
| 93 | ContentBlock::ToolUse { id, .. } => Some(id.clone()), |
| 94 | _ => None, |
| 95 | }) |
| 96 | .collect(); |
| 97 | pending_call_message = (!pending_call_ids.is_empty()).then_some(message_index); |
| 98 | retained_for_pending.clear(); |
| 99 | } |
| 100 | for block in &message.content { |
| 101 | let ContentBlock::ToolResult { tool_use_id, .. } = block else { |
| 102 | continue; |
| 103 | }; |
| 104 | let ordinal = result_ordinal; |
| 105 | result_ordinal = result_ordinal.saturating_add(1); |
| 106 | |
| 107 | let follows_known_call = pending_call_message |
| 108 | .is_some_and(|call_index| call_index < message_index) |
| 109 | && pending_call_ids.iter().any(|id| id == tool_use_id); |
| 110 | if !follows_known_call { |
| 111 | orphan_result_ids.push(tool_use_id.clone()); |
| 112 | } else if !retained_for_pending.insert(tool_use_id.clone()) { |
| 113 | duplicate_result_ids.push(tool_use_id.clone()); |
| 114 | } else { |
| 115 | keep_results.insert(ordinal); |
| 116 | } |
| 117 | } |
| 118 | } |
| 119 | record_missing_results( |
| 120 | pending_call_message, |
| 121 | &pending_call_ids, |
| 122 | &retained_for_pending, |
| 123 | &mut missing_by_message, |
| 124 | &mut repaired_call_ids, |
| 125 | ); |
| 126 | |
| 127 | let receipt = ToolRepairReceipt { |
| 128 | repaired_call_ids, |
| 129 | duplicate_result_ids, |
| 130 | orphan_result_ids, |
| 131 | }; |
| 132 | if receipt.is_empty() { |
| 133 | return receipt; |
| 134 | } |
| 135 | |
| 136 | let original = std::mem::take(messages); |
| 137 | let mut rebuilt = Vec::with_capacity( |
| 138 | original |
| 139 | .len() |
| 140 | .saturating_add(receipt.repaired_call_ids.len()), |
| 141 | ); |
| 142 | let mut seen_result_ordinal = 0usize; |
| 143 | |
| 144 | for (message_index, message) in original.into_iter().enumerate() { |
| 145 | let missing_after_message = missing_by_message |
| 146 | .remove(&message_index) |
| 147 | .unwrap_or_default(); |
| 148 | let mut filtered = message; |
| 149 | filtered.content.retain(|block| { |
| 150 | if matches!(block, ContentBlock::ToolResult { .. }) { |
| 151 | let keep = keep_results.contains(&seen_result_ordinal); |
| 152 | seen_result_ordinal = seen_result_ordinal.saturating_add(1); |
| 153 | keep |
| 154 | } else { |
| 155 | true |
| 156 | } |
| 157 | }); |
| 158 | if !filtered.content.is_empty() { |
| 159 | rebuilt.push(filtered); |
| 160 | } |
| 161 | |
| 162 | if !missing_after_message.is_empty() { |
| 163 | rebuilt.push(Message { |
| 164 | role: Role::User, |
| 165 | content: missing_after_message |
| 166 | .into_iter() |
| 167 | .map(|tool_use_id| ContentBlock::ToolResult { |
| 168 | tool_use_id, |
| 169 | content: CRASH_REPAIR_CONTENT.to_string(), |
| 170 | is_error: Some(true), |
| 171 | content_blocks: None, |
| 172 | }) |
| 173 | .collect(), |
| 174 | }); |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | if append_visible_receipt { |
| 179 | rebuilt.push(Message { |
| 180 | role: Role::Assistant, |
| 181 | content: vec![ContentBlock::Text { |
| 182 | text: receipt.visible_message(), |
| 183 | cache_control: None, |
| 184 | }], |
| 185 | }); |
| 186 | } |
| 187 | *messages = rebuilt; |
| 188 | receipt |
| 189 | } |
| 190 | |
| 191 | fn record_missing_results( |
| 192 | call_message: Option<usize>, |
| 193 | call_ids: &[String], |
| 194 | retained_results: &HashSet<String>, |
| 195 | missing_by_message: &mut HashMap<usize, Vec<String>>, |
| 196 | repaired_call_ids: &mut Vec<String>, |
| 197 | ) { |
| 198 | let Some(message_index) = call_message else { |
| 199 | return; |
| 200 | }; |
| 201 | let missing = call_ids |
| 202 | .iter() |
| 203 | .filter(|id| !retained_results.contains(*id)) |
| 204 | .cloned() |
| 205 | .collect::<Vec<_>>(); |
| 206 | if missing.is_empty() { |
| 207 | return; |
| 208 | } |
| 209 | repaired_call_ids.extend(missing.iter().cloned()); |
| 210 | missing_by_message.insert(message_index, missing); |
| 211 | } |
| 212 | |
| 213 | #[cfg(test)] |
| 214 | mod tests { |
| 215 | use serde_json::json; |
| 216 | |
| 217 | use super::*; |
| 218 | |
| 219 | fn tool_call(id: &str) -> Message { |
| 220 | Message { |
| 221 | role: Role::Assistant, |
| 222 | content: vec![ContentBlock::ToolUse { |
| 223 | id: id.to_string(), |
| 224 | name: "read_file".to_string(), |
| 225 | input: json!({"path": "README.md"}), |
| 226 | caller: None, |
| 227 | thought_signature: None, |
| 228 | }], |
| 229 | } |
| 230 | } |
| 231 | |
| 232 | fn tool_result(id: &str, content: &str) -> Message { |
| 233 | Message { |
| 234 | role: Role::User, |
| 235 | content: vec![ContentBlock::ToolResult { |
| 236 | tool_use_id: id.to_string(), |
| 237 | content: content.to_string(), |
| 238 | is_error: None, |
| 239 | content_blocks: None, |
| 240 | }], |
| 241 | } |
| 242 | } |
| 243 | |
| 244 | fn text(role: &str, content: &str) -> Message { |
| 245 | Message { |
| 246 | role: Role::from(role), |
| 247 | content: vec![ContentBlock::Text { |
| 248 | text: content.to_string(), |
| 249 | cache_control: None, |
| 250 | }], |
| 251 | } |
| 252 | } |
| 253 | |
| 254 | #[test] |
| 255 | fn well_formed_history_is_unchanged() { |
| 256 | let mut messages = vec![tool_call("call-1"), tool_result("call-1", "ok")]; |
| 257 | let before = messages.clone(); |
| 258 | |
| 259 | let receipt = repair_tool_call_pairs(&mut messages); |
| 260 | |
| 261 | assert!(receipt.is_empty()); |
| 262 | assert_eq!(messages, before); |
| 263 | } |
| 264 | |
| 265 | #[test] |
| 266 | fn repeated_provider_call_id_is_scoped_to_each_assistant_turn() { |
| 267 | let mut messages = vec![ |
| 268 | tool_call("call-reused"), |
| 269 | tool_result("call-reused", "hydrated"), |
| 270 | tool_call("call-reused"), |
| 271 | tool_result("call-reused", "executed"), |
| 272 | ]; |
| 273 | let before = messages.clone(); |
| 274 | |
| 275 | let receipt = repair_tool_call_pairs_for_provider(&mut messages); |
| 276 | |
| 277 | assert!(receipt.is_empty()); |
| 278 | assert_eq!(messages, before); |
| 279 | } |
| 280 | |
| 281 | #[test] |
| 282 | fn repairs_dangling_calls_beside_their_assistant_message() { |
| 283 | let mut messages = vec![ |
| 284 | tool_call("call-1"), |
| 285 | text("assistant", "later assistant text"), |
| 286 | ]; |
| 287 | |
| 288 | let receipt = repair_tool_call_pairs(&mut messages); |
| 289 | |
| 290 | assert_eq!(receipt.repaired_call_ids, vec!["call-1"]); |
| 291 | assert_eq!(messages[1].role, "user"); |
| 292 | assert!(matches!( |
| 293 | &messages[1].content[0], |
| 294 | ContentBlock::ToolResult { |
| 295 | tool_use_id, |
| 296 | content, |
| 297 | is_error: Some(true), |
| 298 | .. |
| 299 | } if tool_use_id == "call-1" && content.contains("crashed_and_repaired") |
| 300 | )); |
| 301 | assert_eq!(messages.last().expect("receipt").role, "assistant"); |
| 302 | } |
| 303 | |
| 304 | #[test] |
| 305 | fn provider_repair_never_appends_an_assistant_prefill_receipt() { |
| 306 | let mut messages = vec![tool_call("call-1")]; |
| 307 | |
| 308 | let receipt = repair_tool_call_pairs_for_provider(&mut messages); |
| 309 | |
| 310 | assert_eq!(receipt.repaired_call_ids, vec!["call-1"]); |
| 311 | assert_eq!(messages.last().expect("synthetic result").role, "user"); |
| 312 | assert!(!messages.iter().any(|message| { |
| 313 | message.content.iter().any(|block| { |
| 314 | matches!( |
| 315 | block, |
| 316 | ContentBlock::Text { text, .. } |
| 317 | if text.contains("[tool_history_repair]") |
| 318 | ) |
| 319 | }) |
| 320 | })); |
| 321 | } |
| 322 | |
| 323 | #[test] |
| 324 | fn quarantines_orphan_and_duplicate_results_without_losing_other_blocks() { |
| 325 | let mut mixed_result = tool_result("call-1", "duplicate"); |
| 326 | mixed_result.content.push(ContentBlock::Text { |
| 327 | text: "keep me".to_string(), |
| 328 | cache_control: None, |
| 329 | }); |
| 330 | let mut messages = vec![ |
| 331 | tool_result("orphan", "bad"), |
| 332 | tool_call("call-1"), |
| 333 | tool_result("call-1", "first"), |
| 334 | mixed_result, |
| 335 | ]; |
| 336 | |
| 337 | let receipt = repair_tool_call_pairs(&mut messages); |
| 338 | |
| 339 | assert_eq!(receipt.orphan_result_ids, vec!["orphan"]); |
| 340 | assert_eq!(receipt.duplicate_result_ids, vec!["call-1"]); |
| 341 | let result_contents: Vec<_> = messages |
| 342 | .iter() |
| 343 | .flat_map(|message| &message.content) |
| 344 | .filter_map(|block| match block { |
| 345 | ContentBlock::ToolResult { content, .. } => Some(content.as_str()), |
| 346 | _ => None, |
| 347 | }) |
| 348 | .collect(); |
| 349 | assert_eq!(result_contents, vec!["first"]); |
| 350 | assert!(messages.iter().any(|message| { |
| 351 | message |
| 352 | .content |
| 353 | .iter() |
| 354 | .any(|block| matches!(block, ContentBlock::Text { text, .. } if text == "keep me")) |
| 355 | })); |
| 356 | } |
| 357 | |
| 358 | #[test] |
| 359 | fn repair_is_idempotent() { |
| 360 | let mut messages = vec![tool_call("call-1"), tool_result("orphan", "bad")]; |
| 361 | |
| 362 | let first = repair_tool_call_pairs(&mut messages); |
| 363 | let after_first = messages.clone(); |
| 364 | let second = repair_tool_call_pairs(&mut messages); |
| 365 | |
| 366 | assert!(!first.is_empty()); |
| 367 | assert!(second.is_empty()); |
| 368 | assert_eq!(messages, after_first); |
| 369 | } |
| 370 | |
| 371 | #[test] |
| 372 | fn result_preceding_its_call_is_orphaned_and_call_is_repaired() { |
| 373 | let mut messages = vec![tool_result("call-1", "too early"), tool_call("call-1")]; |
| 374 | |
| 375 | let receipt = repair_tool_call_pairs(&mut messages); |
| 376 | |
| 377 | assert_eq!(receipt.orphan_result_ids, vec!["call-1"]); |
| 378 | assert_eq!(receipt.repaired_call_ids, vec!["call-1"]); |
| 379 | assert!(messages.iter().any(|message| { |
| 380 | message.content.iter().any(|block| { |
| 381 | matches!( |
| 382 | block, |
| 383 | ContentBlock::ToolResult { content, .. } |
| 384 | if content.contains("crashed_and_repaired") |
| 385 | ) |
| 386 | }) |
| 387 | })); |
| 388 | } |
| 389 | |
| 390 | #[test] |
| 391 | fn result_after_a_later_assistant_turn_is_quarantined_as_too_late() { |
| 392 | let mut messages = vec![ |
| 393 | tool_call("call-1"), |
| 394 | text("assistant", "a later model turn"), |
| 395 | tool_result("call-1", "too late"), |
| 396 | ]; |
| 397 | |
| 398 | let receipt = repair_tool_call_pairs(&mut messages); |
| 399 | |
| 400 | assert_eq!(receipt.orphan_result_ids, vec!["call-1"]); |
| 401 | assert_eq!(receipt.repaired_call_ids, vec!["call-1"]); |
| 402 | assert!(!messages.iter().any(|message| { |
| 403 | message.content.iter().any(|block| { |
| 404 | matches!(block, ContentBlock::ToolResult { content, .. } if content == "too late") |
| 405 | }) |
| 406 | })); |
| 407 | } |
| 408 | } |
| 409 |