| 1 | //! Last-round coverage floor for compaction replacement history. |
| 2 | //! |
| 3 | //! Compaction may summarize older turns, but the latest user round (user |
| 4 | //! text plus following assistant/tool results) must survive verbatim, |
| 5 | //! bounded, or the pass is refused. See [`SURVIVAL_CONTRACT.md`]. |
| 6 | |
| 7 | use anyhow::Result; |
| 8 | use std::collections::HashSet; |
| 9 | |
| 10 | use codewhale_models::{ContentBlock, Message, SystemPrompt}; |
| 11 | |
| 12 | use super::{ |
| 13 | compaction_checkpoint_message, is_compaction_checkpoint_message, retained_user_messages, |
| 14 | truncate_retained_block, user_text_of, |
| 15 | }; |
| 16 | |
| 17 | const LAST_ROUND_TOOL_RESULT_MAX_CHARS: usize = 8 * 1024; |
| 18 | |
| 19 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Default)] |
| 20 | pub enum CompactionPath { |
| 21 | #[default] |
| 22 | Summary, |
| 23 | PruneOnly, |
| 24 | } |
| 25 | |
| 26 | #[derive(Debug, Clone, PartialEq, Eq, Default)] |
| 27 | pub struct CompactionCoverage { |
| 28 | pub path: CompactionPath, |
| 29 | pub last_round_messages: usize, |
| 30 | pub last_round_tool_results: usize, |
| 31 | pub last_round_assistant: bool, |
| 32 | pub dropped_messages: usize, |
| 33 | pub anchors_chars: usize, |
| 34 | /// Effective `[compaction] retained_user_message_tokens` budget this pass |
| 35 | /// spent on verbatim user messages (#5956). `0` on the prune-only path, |
| 36 | /// which never builds a replacement history. |
| 37 | pub retained_user_message_tokens: usize, |
| 38 | /// Whether `[compaction] summary_instructions` was appended to the |
| 39 | /// summarizer prompt on this pass (#5956). |
| 40 | pub operator_instructions_applied: bool, |
| 41 | } |
| 42 | |
| 43 | impl CompactionCoverage { |
| 44 | #[must_use] |
| 45 | pub fn receipt_clause(&self) -> String { |
| 46 | let path = match self.path { |
| 47 | CompactionPath::Summary => "summary", |
| 48 | CompactionPath::PruneOnly => "prune-only", |
| 49 | }; |
| 50 | let assistant = if self.last_round_assistant { |
| 51 | ", assistant" |
| 52 | } else { |
| 53 | "" |
| 54 | }; |
| 55 | let mut clause = format!( |
| 56 | "{path}; last round kept: {} messages ({} tool results{assistant})", |
| 57 | self.last_round_messages, self.last_round_tool_results |
| 58 | ); |
| 59 | if self.anchors_chars > 0 { |
| 60 | clause.push_str(&format!("; anchors {} chars", self.anchors_chars)); |
| 61 | } |
| 62 | // Name the tuning knobs so an operator who set them can tell they took |
| 63 | // effect without reading the log (#5956). The prune-only path builds no |
| 64 | // replacement history, so it reports no budget. |
| 65 | if self.retained_user_message_tokens > 0 { |
| 66 | clause.push_str(&format!( |
| 67 | "; verbatim user budget {} tokens", |
| 68 | self.retained_user_message_tokens |
| 69 | )); |
| 70 | } |
| 71 | if self.operator_instructions_applied { |
| 72 | clause.push_str("; operator instructions applied"); |
| 73 | } |
| 74 | clause |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 79 | pub struct LastCompactionSnapshot { |
| 80 | pub auto: bool, |
| 81 | pub coverage: CompactionCoverage, |
| 82 | pub messages_before: usize, |
| 83 | pub messages_after: usize, |
| 84 | } |
| 85 | |
| 86 | #[derive(Debug, Clone, PartialEq, Eq, Default)] |
| 87 | pub struct CompactionKeep { |
| 88 | pub has_checkpoint: bool, |
| 89 | pub last_round_messages: usize, |
| 90 | pub last_round_tool_results: usize, |
| 91 | pub last_round_assistant: bool, |
| 92 | } |
| 93 | |
| 94 | #[must_use] |
| 95 | pub fn inspect_compaction_keep(messages: &[Message]) -> CompactionKeep { |
| 96 | let last_round = last_round_slice(messages); |
| 97 | CompactionKeep { |
| 98 | has_checkpoint: messages.iter().any(is_compaction_checkpoint_message), |
| 99 | last_round_messages: last_round.len(), |
| 100 | last_round_tool_results: last_round.iter().flat_map(tool_result_ids).count(), |
| 101 | last_round_assistant: last_round |
| 102 | .iter() |
| 103 | .any(|message| message.role.is_assistant_like()), |
| 104 | } |
| 105 | } |
| 106 | |
| 107 | /// Workspace anchors are passive model input; only user-configured trust admits |
| 108 | /// them. Missing, linked, unreadable, or empty anchors contribute no text. |
| 109 | #[must_use] |
| 110 | pub fn pinned_anchors_text(workspace: Option<&std::path::Path>) -> Option<String> { |
| 111 | let workspace = workspace?; |
| 112 | if !crate::config::is_workspace_trusted(workspace) { |
| 113 | return None; |
| 114 | } |
| 115 | let primary = workspace.join(".codewhale").join("anchors.md"); |
| 116 | let path = if primary.symlink_metadata().is_ok() || workspace.join(".codewhale").is_symlink() { |
| 117 | primary |
| 118 | } else { |
| 119 | workspace.join(".deepseek").join("anchors.md") |
| 120 | }; |
| 121 | crate::fs_confined::read_to_string(workspace, &path) |
| 122 | .ok() |
| 123 | .map(|contents| contents.trim().to_string()) |
| 124 | .filter(|contents| !contents.is_empty()) |
| 125 | } |
| 126 | |
| 127 | fn is_plain_user_text(message: &Message) -> bool { |
| 128 | !is_compaction_checkpoint_message(message) |
| 129 | && !crate::runtime_handoff::is_runtime_owned_user_message(message) |
| 130 | && user_text_of(message).is_some() |
| 131 | } |
| 132 | |
| 133 | fn user_prompt_text_of(message: &Message) -> Option<String> { |
| 134 | is_plain_user_text(message) |
| 135 | .then(|| user_text_of(message)) |
| 136 | .flatten() |
| 137 | } |
| 138 | |
| 139 | fn last_plain_user_index(messages: &[Message], end: usize) -> Option<usize> { |
| 140 | messages[..end] |
| 141 | .iter() |
| 142 | .enumerate() |
| 143 | .rev() |
| 144 | .find_map(|(idx, message)| is_plain_user_text(message).then_some(idx)) |
| 145 | } |
| 146 | |
| 147 | fn slice_has_tool_result(messages: &[Message], start: usize) -> bool { |
| 148 | messages[start..].iter().any(|message| { |
| 149 | message |
| 150 | .content |
| 151 | .iter() |
| 152 | .any(|block| matches!(block, ContentBlock::ToolResult { .. })) |
| 153 | }) |
| 154 | } |
| 155 | |
| 156 | #[must_use] |
| 157 | pub(crate) fn last_round_start(messages: &[Message]) -> usize { |
| 158 | let Some(last_user) = last_plain_user_index(messages, messages.len()) else { |
| 159 | return 0; |
| 160 | }; |
| 161 | if slice_has_tool_result(messages, last_user) { |
| 162 | return last_user; |
| 163 | } |
| 164 | // Trailing toolless user/assistant turns still need the previous |
| 165 | // tool-bearing round; otherwise those results vanish behind the summary. |
| 166 | // If no tool round exists, keep only the latest user turn so chat-only |
| 167 | // sessions can still summarize older text. |
| 168 | let mut candidate = last_user; |
| 169 | loop { |
| 170 | let Some(prev) = last_plain_user_index(messages, candidate) else { |
| 171 | return last_user; |
| 172 | }; |
| 173 | if slice_has_tool_result(messages, prev) { |
| 174 | return prev; |
| 175 | } |
| 176 | candidate = prev; |
| 177 | } |
| 178 | } |
| 179 | |
| 180 | #[must_use] |
| 181 | pub(crate) fn last_round_range(messages: &[Message]) -> (usize, usize) { |
| 182 | let start = last_round_start(messages).min(messages.len()); |
| 183 | // A previous checkpoint can sit in the middle of an uninterrupted task. |
| 184 | // Stopping at that marker hid every tool step after the first compact |
| 185 | // from the next pass's survival checks. |
| 186 | let end = messages.len().saturating_sub(usize::from( |
| 187 | messages |
| 188 | .last() |
| 189 | .is_some_and(is_compaction_checkpoint_message), |
| 190 | )); |
| 191 | (start, end) |
| 192 | } |
| 193 | |
| 194 | /// How many messages of the open round sit in `messages` before a checkpoint. |
| 195 | #[must_use] |
| 196 | pub fn last_round_kept_count(messages: &[Message]) -> Option<usize> { |
| 197 | let checkpoint = messages |
| 198 | .iter() |
| 199 | .rposition(is_compaction_checkpoint_message)?; |
| 200 | if checkpoint == 0 { |
| 201 | return None; |
| 202 | } |
| 203 | let start = last_round_start(&messages[..checkpoint]); |
| 204 | Some(checkpoint.saturating_sub(start)) |
| 205 | } |
| 206 | |
| 207 | fn last_round_slice(messages: &[Message]) -> &[Message] { |
| 208 | let (start, end) = last_round_range(messages); |
| 209 | &messages[start..end] |
| 210 | } |
| 211 | |
| 212 | /// Retain the current user instructions and the two most recent tool |
| 213 | /// exchanges. A user round can contain thousands of steps: retaining that |
| 214 | /// entire round forever makes a continuous task impossible to compact. |
| 215 | /// Older completed exchanges are covered by the summary and durable history. |
| 216 | /// A split is legal only when all preceding tool calls have their results. |
| 217 | fn protected_last_round(messages: &[Message]) -> Vec<&Message> { |
| 218 | let round = last_round_slice(messages); |
| 219 | let mut pending = HashSet::new(); |
| 220 | let mut boundaries = Vec::new(); |
| 221 | for (idx, message) in round.iter().enumerate() { |
| 222 | let calls = tool_use_ids(message); |
| 223 | if !calls.is_empty() && pending.is_empty() { |
| 224 | boundaries.push(idx); |
| 225 | } |
| 226 | pending.extend(calls); |
| 227 | for id in tool_result_ids(message) { |
| 228 | pending.remove(&id); |
| 229 | } |
| 230 | } |
| 231 | let start = if boundaries.len() > 2 { |
| 232 | boundaries[boundaries.len() - 2] |
| 233 | } else { |
| 234 | 0 |
| 235 | }; |
| 236 | round |
| 237 | .iter() |
| 238 | .enumerate() |
| 239 | .filter_map(|(idx, message)| { |
| 240 | (!is_compaction_checkpoint_message(message) |
| 241 | && (idx >= start || is_plain_user_text(message))) |
| 242 | .then_some(message) |
| 243 | }) |
| 244 | .collect() |
| 245 | } |
| 246 | |
| 247 | pub(super) fn replacement_messages( |
| 248 | messages: &[Message], |
| 249 | retained_user_message_tokens: usize, |
| 250 | ) -> Vec<Message> { |
| 251 | let (start, _) = last_round_range(messages); |
| 252 | let mut retained = retained_user_messages(&messages[..start], retained_user_message_tokens); |
| 253 | let round = protected_last_round(messages) |
| 254 | .into_iter() |
| 255 | .cloned() |
| 256 | .collect::<Vec<_>>(); |
| 257 | retained.extend(bound_last_round(&round)); |
| 258 | // The Operate contract applies to the current tool loop as well as later |
| 259 | // turns. It must survive compaction even when old-user retention is full. |
| 260 | let current_contract = messages |
| 261 | .iter() |
| 262 | .rev() |
| 263 | .find(|message| crate::runtime_handoff::is_current_operate_contract_message(message)); |
| 264 | let contract = current_contract.or_else(|| { |
| 265 | messages |
| 266 | .iter() |
| 267 | .rev() |
| 268 | .find(|message| crate::runtime_handoff::is_operate_contract_message(message)) |
| 269 | }); |
| 270 | if let Some(contract) = contract { |
| 271 | retained.retain(|message| !crate::runtime_handoff::is_operate_contract_message(message)); |
| 272 | retained.insert(0, contract.clone()); |
| 273 | } |
| 274 | // Contributions are captured once per turn. Preserve the complete latest |
| 275 | // snapshot, including a withdrawal, rather than summarizing or truncating |
| 276 | // instructions that still apply to this tool loop. |
| 277 | if let Some(snapshot) = messages.iter().rev().find(|message| { |
| 278 | crate::runtime_handoff::extension_prompt_contributions_display(message).is_some() |
| 279 | }) { |
| 280 | retained.retain(|message| { |
| 281 | crate::runtime_handoff::extension_prompt_contributions_display(message).is_none() |
| 282 | }); |
| 283 | retained.insert(0, snapshot.clone()); |
| 284 | } |
| 285 | retained |
| 286 | } |
| 287 | |
| 288 | pub(super) fn bound_last_round(messages: &[Message]) -> Vec<Message> { |
| 289 | let mut round = messages.to_vec(); |
| 290 | for message in &mut round { |
| 291 | for block in &mut message.content { |
| 292 | if let ContentBlock::ToolResult { |
| 293 | content, |
| 294 | content_blocks, |
| 295 | .. |
| 296 | } = block |
| 297 | && truncate_retained_block("tool result", content, LAST_ROUND_TOOL_RESULT_MAX_CHARS) |
| 298 | { |
| 299 | *content_blocks = None; |
| 300 | } |
| 301 | } |
| 302 | } |
| 303 | round |
| 304 | } |
| 305 | |
| 306 | fn tool_result_ids(message: &Message) -> Vec<(codewhale_models::ToolCallKey<'_>, &str)> { |
| 307 | message |
| 308 | .content |
| 309 | .iter() |
| 310 | .filter_map(|block| match block { |
| 311 | ContentBlock::ToolResult { tool_use_id, .. } => { |
| 312 | block.tool_call_key().map(|key| (key, tool_use_id.as_str())) |
| 313 | } |
| 314 | _ => None, |
| 315 | }) |
| 316 | .collect() |
| 317 | } |
| 318 | |
| 319 | fn has_tool_result_id(message: &Message, id: &(codewhale_models::ToolCallKey<'_>, &str)) -> bool { |
| 320 | if id.0.as_str().trim().is_empty() { |
| 321 | return false; |
| 322 | } |
| 323 | message.content.iter().any(|block| { |
| 324 | matches!( |
| 325 | block, |
| 326 | ContentBlock::ToolResult { tool_use_id, .. } if block.tool_call_key() == Some(id.0) && tool_use_id == id.1 |
| 327 | ) |
| 328 | }) |
| 329 | } |
| 330 | |
| 331 | fn tool_use_ids(message: &Message) -> Vec<(codewhale_models::ToolCallKey<'_>, &str)> { |
| 332 | message |
| 333 | .content |
| 334 | .iter() |
| 335 | .filter_map(|block| match block { |
| 336 | ContentBlock::ToolUse { id, .. } => block.tool_call_key().map(|key| (key, id.as_str())), |
| 337 | _ => None, |
| 338 | }) |
| 339 | .collect() |
| 340 | } |
| 341 | |
| 342 | fn has_tool_use_id(message: &Message, id: &(codewhale_models::ToolCallKey<'_>, &str)) -> bool { |
| 343 | if id.0.as_str().trim().is_empty() { |
| 344 | return false; |
| 345 | } |
| 346 | message.content.iter().any(|block| { |
| 347 | matches!( |
| 348 | block, |
| 349 | ContentBlock::ToolUse { id: provider, .. } if block.tool_call_key() == Some(id.0) && provider == id.1 |
| 350 | ) |
| 351 | }) |
| 352 | } |
| 353 | |
| 354 | fn tool_call_identity_label(id: &(codewhale_models::ToolCallKey<'_>, &str)) -> String { |
| 355 | match id.0 { |
| 356 | codewhale_models::ToolCallKey::Execution(execution_id) => { |
| 357 | format!("execution {execution_id} (provider call {})", id.1) |
| 358 | } |
| 359 | codewhale_models::ToolCallKey::LegacyProvider(provider_id) => provider_id.to_string(), |
| 360 | } |
| 361 | } |
| 362 | |
| 363 | fn assistant_text_of(message: &Message) -> Option<String> { |
| 364 | if !message.role.is_assistant_like() { |
| 365 | return None; |
| 366 | } |
| 367 | let text = message |
| 368 | .content |
| 369 | .iter() |
| 370 | .filter_map(|block| match block { |
| 371 | ContentBlock::Text { text, .. } => Some(text.as_str()), |
| 372 | _ => None, |
| 373 | }) |
| 374 | .collect::<Vec<_>>() |
| 375 | .join("\n"); |
| 376 | let text = text.trim(); |
| 377 | (!text.is_empty()).then(|| text.to_string()) |
| 378 | } |
| 379 | |
| 380 | /// A retained copy may be truncated (`bound_last_round` caps oversized blocks), |
| 381 | /// so a prefix either way counts as survival -- but nothing weaker does. |
| 382 | fn survives(text: &str, replacement: &[Message], of: fn(&Message) -> Option<String>) -> bool { |
| 383 | replacement.iter().any(|message| { |
| 384 | of(message) |
| 385 | .is_some_and(|kept| kept == text || text.starts_with(&kept) || kept.starts_with(text)) |
| 386 | }) |
| 387 | } |
| 388 | |
| 389 | pub(crate) fn validate_last_round_coverage( |
| 390 | original: &[Message], |
| 391 | replacement: &[Message], |
| 392 | ) -> Result<()> { |
| 393 | let last_round = protected_last_round(original); |
| 394 | if last_round.is_empty() { |
| 395 | return Ok(()); |
| 396 | } |
| 397 | // Every user turn in the round, not the first one `find_map` happens to |
| 398 | // reach. `last_round_start` walks back past a toolless tail to the previous |
| 399 | // tool-bearing turn, so the round routinely spans two user messages -- and |
| 400 | // checking only the earliest let a rewrite drop the *latest* one, which is |
| 401 | // the turn this whole contract exists to keep. |
| 402 | for text in last_round.iter().copied().filter_map(user_prompt_text_of) { |
| 403 | if !survives(&text, replacement, user_prompt_text_of) { |
| 404 | anyhow::bail!( |
| 405 | "Making room stopped: a last-round user message was dropped; history was not replaced." |
| 406 | ); |
| 407 | } |
| 408 | } |
| 409 | for id in last_round.iter().copied().flat_map(tool_result_ids) { |
| 410 | if !replacement |
| 411 | .iter() |
| 412 | .any(|message| has_tool_result_id(message, &id)) |
| 413 | { |
| 414 | let label = tool_call_identity_label(&id); |
| 415 | anyhow::bail!( |
| 416 | "Making room stopped: last-round tool result {label} was dropped; history was not replaced." |
| 417 | ); |
| 418 | } |
| 419 | } |
| 420 | // The call, not just its result. Keeping a tool_result whose tool_use was |
| 421 | // summarized away leaves an orphaned result that providers reject outright. |
| 422 | for id in last_round.iter().copied().flat_map(tool_use_ids) { |
| 423 | if !replacement |
| 424 | .iter() |
| 425 | .any(|message| has_tool_use_id(message, &id)) |
| 426 | { |
| 427 | let label = tool_call_identity_label(&id); |
| 428 | anyhow::bail!( |
| 429 | "Making room stopped: last-round tool call {label} was dropped; history was not replaced." |
| 430 | ); |
| 431 | } |
| 432 | } |
| 433 | // Match the assistant's actual output. An existential "some assistant |
| 434 | // message survived" check passed on a replacement whose only assistant |
| 435 | // message was the summary the rewrite had just written. |
| 436 | for text in last_round.iter().copied().filter_map(assistant_text_of) { |
| 437 | if !survives(&text, replacement, assistant_text_of) { |
| 438 | anyhow::bail!( |
| 439 | "Making room stopped: last-round assistant output was dropped; history was not replaced." |
| 440 | ); |
| 441 | } |
| 442 | } |
| 443 | if last_round |
| 444 | .iter() |
| 445 | .any(|message| message.role.is_assistant_like()) |
| 446 | && !replacement |
| 447 | .iter() |
| 448 | .any(|message| message.role.is_assistant_like()) |
| 449 | { |
| 450 | anyhow::bail!( |
| 451 | "Making room stopped: last-round assistant output was dropped; history was not replaced." |
| 452 | ); |
| 453 | } |
| 454 | Ok(()) |
| 455 | } |
| 456 | |
| 457 | pub(crate) fn require_text_survives( |
| 458 | replacement: &[Message], |
| 459 | needle: &str, |
| 460 | label: &str, |
| 461 | ) -> Result<()> { |
| 462 | let needle = needle.trim(); |
| 463 | if needle.is_empty() { |
| 464 | return Ok(()); |
| 465 | } |
| 466 | let kept = replacement.iter().any(|message| { |
| 467 | message.content.iter().any(|block| match block { |
| 468 | ContentBlock::Text { text, .. } => text.contains(needle), |
| 469 | ContentBlock::ToolResult { content, .. } => content.contains(needle), |
| 470 | _ => false, |
| 471 | }) |
| 472 | }); |
| 473 | if !kept { |
| 474 | anyhow::bail!("Making room stopped: {label} was dropped; history was not replaced."); |
| 475 | } |
| 476 | Ok(()) |
| 477 | } |
| 478 | |
| 479 | pub(crate) fn validate_survival_contract( |
| 480 | original: &[Message], |
| 481 | replacement: &[Message], |
| 482 | anchors: Option<&str>, |
| 483 | ) -> Result<()> { |
| 484 | validate_last_round_coverage(original, replacement)?; |
| 485 | let checkpoints = replacement |
| 486 | .iter() |
| 487 | .filter(|message| is_compaction_checkpoint_message(message)) |
| 488 | .count(); |
| 489 | if checkpoints == 0 { |
| 490 | anyhow::bail!( |
| 491 | "Making room stopped: checkpoint receipt was dropped; history was not replaced." |
| 492 | ); |
| 493 | } |
| 494 | if checkpoints > 1 { |
| 495 | anyhow::bail!( |
| 496 | "Making room stopped: prior summaries were duplicated; history was not replaced." |
| 497 | ); |
| 498 | } |
| 499 | if let Some(anchors) = anchors { |
| 500 | require_text_survives(replacement, anchors, "pinned /anchor text")?; |
| 501 | } |
| 502 | Ok(()) |
| 503 | } |
| 504 | |
| 505 | pub(super) fn measure_coverage( |
| 506 | original: &[Message], |
| 507 | replacement: &[Message], |
| 508 | path: CompactionPath, |
| 509 | anchors_chars: usize, |
| 510 | ) -> CompactionCoverage { |
| 511 | let last_round = last_round_slice(replacement); |
| 512 | CompactionCoverage { |
| 513 | path, |
| 514 | last_round_messages: last_round.len(), |
| 515 | last_round_tool_results: last_round.iter().flat_map(tool_result_ids).count(), |
| 516 | last_round_assistant: last_round |
| 517 | .iter() |
| 518 | .any(|message| message.role.is_assistant_like()), |
| 519 | dropped_messages: original.len().saturating_sub(replacement.len()), |
| 520 | anchors_chars, |
| 521 | // Tuning provenance is owned by the caller that holds the |
| 522 | // `CompactionConfig`; measurement over two message lists cannot know it. |
| 523 | retained_user_message_tokens: 0, |
| 524 | operator_instructions_applied: false, |
| 525 | } |
| 526 | } |
| 527 | |
| 528 | /// Build the post-compaction history: recent plain user messages kept |
| 529 | /// verbatim within `retained_user_message_tokens`, the bounded last round, and |
| 530 | /// the checkpoint. The budget is `[compaction] retained_user_message_tokens` |
| 531 | /// (#5956); it was a hard-coded 20 000 before that key existed. |
| 532 | pub(super) fn build_replacement_history( |
| 533 | messages: &[Message], |
| 534 | checkpoint_text: &str, |
| 535 | anchors: Option<&str>, |
| 536 | retained_user_message_tokens: usize, |
| 537 | ) -> Result<Vec<Message>> { |
| 538 | let mut retained = replacement_messages(messages, retained_user_message_tokens); |
| 539 | retained.push(compaction_checkpoint_message(&SystemPrompt::Text( |
| 540 | checkpoint_text.to_string(), |
| 541 | ))); |
| 542 | validate_survival_contract(messages, &retained, anchors)?; |
| 543 | Ok(retained) |
| 544 | } |
| 545 | |
| 546 | #[cfg(test)] |
| 547 | mod tests { |
| 548 | use super::*; |
| 549 | use crate::compaction::{COMPACTION_SUMMARY_MARKER, compaction_checkpoint_message}; |
| 550 | use codewhale_models::{ContentBlock, Role}; |
| 551 | use serde_json::json; |
| 552 | |
| 553 | #[test] |
| 554 | fn coverage_requires_the_original_execution_and_provider_pair() { |
| 555 | let original: Vec<Message> = serde_json::from_value(json!([ |
| 556 | {"role":"user","content":[{"type":"text","text":"keep this exchange"}]}, |
| 557 | {"role":"assistant","content":[{"type":"tool_use","id":"wire","execution_id":"local","name":"read","input":{}}]}, |
| 558 | {"role":"user","content":[{"type":"tool_result","tool_use_id":"wire","execution_id":"local","content":"kept"}]} |
| 559 | ])).unwrap(); |
| 560 | assert!(validate_last_round_coverage(&original, &original).is_ok()); |
| 561 | for identity in [None, Some("different")] { |
| 562 | let mut replacement = original.clone(); |
| 563 | for block in replacement |
| 564 | .iter_mut() |
| 565 | .flat_map(|message| &mut message.content) |
| 566 | { |
| 567 | match block { |
| 568 | ContentBlock::ToolUse { execution_id, .. } |
| 569 | | ContentBlock::ToolResult { execution_id, .. } => { |
| 570 | *execution_id = identity.map(str::to_string) |
| 571 | } |
| 572 | _ => {} |
| 573 | } |
| 574 | } |
| 575 | assert!(validate_last_round_coverage(&original, &replacement).is_err()); |
| 576 | } |
| 577 | let mut replacement = original.clone(); |
| 578 | if let ContentBlock::ToolResult { tool_use_id, .. } = &mut replacement[2].content[0] { |
| 579 | *tool_use_id = "wrong-wire".to_string(); |
| 580 | } |
| 581 | let error = validate_last_round_coverage(&original, &replacement).unwrap_err(); |
| 582 | assert_eq!( |
| 583 | error.to_string(), |
| 584 | "Making room stopped: last-round tool result execution local (provider call wire) was dropped; history was not replaced." |
| 585 | ); |
| 586 | let mut replacement = original.clone(); |
| 587 | if let ContentBlock::ToolUse { id, .. } = &mut replacement[1].content[0] { |
| 588 | *id = "wrong-wire".to_string(); |
| 589 | } |
| 590 | let error = validate_last_round_coverage(&original, &replacement).unwrap_err(); |
| 591 | assert_eq!( |
| 592 | error.to_string(), |
| 593 | "Making room stopped: last-round tool call execution local (provider call wire) was dropped; history was not replaced." |
| 594 | ); |
| 595 | } |
| 596 | |
| 597 | #[test] |
| 598 | fn confined_pinned_anchors_require_workspace_trust() { |
| 599 | use crate::test_support::{EnvVarGuard, lock_test_env}; |
| 600 | let _lock = lock_test_env(); |
| 601 | let workspace = tempfile::tempdir().unwrap(); |
| 602 | let config = tempfile::tempdir().unwrap(); |
| 603 | let _config = EnvVarGuard::set("CODEWHALE_CONFIG_PATH", config.path().join("config.toml")); |
| 604 | let legacy = workspace.path().join(".deepseek"); |
| 605 | std::fs::create_dir(&legacy).unwrap(); |
| 606 | std::fs::write(legacy.join("anchors.md"), " legacy anchor \n").unwrap(); |
| 607 | std::fs::write(legacy.join("trusted"), "true").unwrap(); |
| 608 | assert_eq!(pinned_anchors_text(None), None); |
| 609 | assert_eq!(pinned_anchors_text(Some(workspace.path())), None); |
| 610 | |
| 611 | crate::config::save_workspace_trust(workspace.path()).unwrap(); |
| 612 | assert!(crate::config::is_workspace_trusted(workspace.path())); |
| 613 | assert_eq!( |
| 614 | pinned_anchors_text(Some(workspace.path())).as_deref(), |
| 615 | Some("legacy anchor") |
| 616 | ); |
| 617 | let primary = workspace.path().join(".codewhale"); |
| 618 | std::fs::create_dir(&primary).unwrap(); |
| 619 | std::fs::write(primary.join("anchors.md"), " primary anchor \n").unwrap(); |
| 620 | assert_eq!( |
| 621 | pinned_anchors_text(Some(workspace.path())).as_deref(), |
| 622 | Some("primary anchor") |
| 623 | ); |
| 624 | std::fs::write(primary.join("anchors.md"), " \n").unwrap(); |
| 625 | assert_eq!(pinned_anchors_text(Some(workspace.path())), None); |
| 626 | } |
| 627 | |
| 628 | #[cfg(unix)] |
| 629 | #[test] |
| 630 | fn confined_pinned_anchors_refuse_linked_files_and_directories() { |
| 631 | use crate::test_support::{EnvVarGuard, lock_test_env}; |
| 632 | use std::os::unix::fs::symlink; |
| 633 | let _lock = lock_test_env(); |
| 634 | let config = tempfile::tempdir().unwrap(); |
| 635 | let _config = EnvVarGuard::set("CODEWHALE_CONFIG_PATH", config.path().join("config.toml")); |
| 636 | for directory in [".codewhale", ".deepseek"] { |
| 637 | for linked_directory in [false, true] { |
| 638 | let workspace = tempfile::tempdir().unwrap(); |
| 639 | let outside = tempfile::tempdir().unwrap(); |
| 640 | crate::config::save_workspace_trust(workspace.path()).unwrap(); |
| 641 | assert!(crate::config::is_workspace_trusted(workspace.path())); |
| 642 | let target = outside.path().join("anchors.md"); |
| 643 | std::fs::write(&target, "separate anchor").unwrap(); |
| 644 | let parent = workspace.path().join(directory); |
| 645 | if linked_directory { |
| 646 | symlink(outside.path(), &parent).unwrap(); |
| 647 | } else { |
| 648 | std::fs::create_dir(&parent).unwrap(); |
| 649 | symlink(&target, parent.join("anchors.md")).unwrap(); |
| 650 | } |
| 651 | assert_eq!(pinned_anchors_text(Some(workspace.path())), None); |
| 652 | // A dangling preferred path must not select legacy content. |
| 653 | if directory == ".codewhale" { |
| 654 | let legacy = workspace.path().join(".deepseek"); |
| 655 | std::fs::create_dir(&legacy).unwrap(); |
| 656 | std::fs::write(legacy.join("anchors.md"), "legacy anchor").unwrap(); |
| 657 | } |
| 658 | std::fs::remove_file(&target).unwrap(); |
| 659 | assert_eq!(pinned_anchors_text(Some(workspace.path())), None); |
| 660 | } |
| 661 | } |
| 662 | } |
| 663 | |
| 664 | fn msg(role: &str, text: &str) -> Message { |
| 665 | Message { |
| 666 | role: Role::from(role), |
| 667 | content: vec![ContentBlock::Text { |
| 668 | text: text.to_string(), |
| 669 | cache_control: None, |
| 670 | }], |
| 671 | } |
| 672 | } |
| 673 | |
| 674 | fn tool_use(id: &str, name: &str, input: serde_json::Value) -> Message { |
| 675 | Message { |
| 676 | role: Role::Assistant, |
| 677 | content: vec![ContentBlock::ToolUse { |
| 678 | execution_id: None, |
| 679 | id: id.to_string(), |
| 680 | name: name.to_string(), |
| 681 | input, |
| 682 | caller: None, |
| 683 | thought_signature: None, |
| 684 | }], |
| 685 | } |
| 686 | } |
| 687 | |
| 688 | fn tool_result(id: &str, content: &str) -> Message { |
| 689 | Message { |
| 690 | role: Role::User, |
| 691 | content: vec![ContentBlock::ToolResult { |
| 692 | execution_id: None, |
| 693 | tool_use_id: id.to_string(), |
| 694 | content: content.to_string(), |
| 695 | is_error: None, |
| 696 | content_blocks: None, |
| 697 | }], |
| 698 | } |
| 699 | } |
| 700 | |
| 701 | fn checkpoint(summary: &str) -> Message { |
| 702 | compaction_checkpoint_message(&SystemPrompt::Text(format!( |
| 703 | "{COMPACTION_SUMMARY_MARKER}: {summary}" |
| 704 | ))) |
| 705 | } |
| 706 | |
| 707 | #[test] |
| 708 | fn replacement_keeps_only_latest_complete_prompt_snapshot_and_withdrawal() { |
| 709 | use crate::runtime_handoff::{ |
| 710 | extension_prompt_contributions_display, extension_prompt_contributions_runtime_message, |
| 711 | }; |
| 712 | let old = extension_prompt_contributions_runtime_message(Some("old instructions")); |
| 713 | let current_text = "current instructions ".repeat(400); |
| 714 | for current in [ |
| 715 | extension_prompt_contributions_runtime_message(Some(¤t_text)), |
| 716 | extension_prompt_contributions_runtime_message(None), |
| 717 | ] { |
| 718 | let quoted = msg( |
| 719 | "user", |
| 720 | &user_text_of(¤t).expect("runtime snapshot has text"), |
| 721 | ); |
| 722 | let original = vec![ |
| 723 | old.clone(), |
| 724 | quoted.clone(), |
| 725 | current.clone(), |
| 726 | msg("user", "Continue this task."), |
| 727 | tool_use("first", "Read", json!({"path": "first"})), |
| 728 | tool_result("first", "first output"), |
| 729 | tool_use("second", "Read", json!({"path": "second"})), |
| 730 | tool_result("second", "second output"), |
| 731 | tool_use("third", "Read", json!({"path": "third"})), |
| 732 | tool_result("third", "third output"), |
| 733 | ]; |
| 734 | let kept = replacement_messages(&original, 20_000); |
| 735 | let snapshots: Vec<_> = kept |
| 736 | .iter() |
| 737 | .filter(|message| extension_prompt_contributions_display(message).is_some()) |
| 738 | .collect(); |
| 739 | assert_eq!(snapshots, [¤t]); |
| 740 | assert!( |
| 741 | kept.contains("ed), |
| 742 | "a person's quote is ordinary user text" |
| 743 | ); |
| 744 | assert_eq!( |
| 745 | replacement_messages(&kept, 20_000) |
| 746 | .iter() |
| 747 | .filter(|message| extension_prompt_contributions_display(message).is_some()) |
| 748 | .count(), |
| 749 | 1, |
| 750 | "repeated compaction must not accumulate snapshots" |
| 751 | ); |
| 752 | } |
| 753 | } |
| 754 | |
| 755 | /// The handoff header tells the next turn what survived. It must match |
| 756 | /// what the replacement history keeps: only the last steps of a long |
| 757 | /// round, with long tool output shortened and marked. |
| 758 | #[test] |
| 759 | fn summary_header_matches_what_replacement_history_keeps() { |
| 760 | let long_output = "x".repeat(LAST_ROUND_TOOL_RESULT_MAX_CHARS * 2); |
| 761 | let original = vec![ |
| 762 | msg("user", "Fix the build."), |
| 763 | tool_use("first", "Bash", json!({"command": "cargo check"})), |
| 764 | tool_result("first", "first step output"), |
| 765 | tool_use("second", "Bash", json!({"command": "cargo build"})), |
| 766 | tool_result("second", "second step output"), |
| 767 | tool_use("third", "Bash", json!({"command": "cargo test"})), |
| 768 | tool_result("third", &long_output), |
| 769 | ]; |
| 770 | let kept = replacement_messages(&original, 20_000); |
| 771 | let kept_ids: Vec<&str> = kept |
| 772 | .iter() |
| 773 | .flat_map(tool_result_ids) |
| 774 | .map(|(_, wire)| wire) |
| 775 | .collect(); |
| 776 | assert_eq!(kept_ids, ["second", "third"], "earlier steps are dropped"); |
| 777 | assert!( |
| 778 | kept.iter() |
| 779 | .any(|m| user_text_of(m).as_deref() == Some("Fix the build.")) |
| 780 | ); |
| 781 | let shortened = kept |
| 782 | .iter() |
| 783 | .flat_map(|m| &m.content) |
| 784 | .find_map(|block| match block { |
| 785 | ContentBlock::ToolResult { |
| 786 | tool_use_id, |
| 787 | content, |
| 788 | .. |
| 789 | } if tool_use_id == "third" => Some(content.clone()), |
| 790 | _ => None, |
| 791 | }) |
| 792 | .expect("the last tool result is kept"); |
| 793 | assert!(shortened.len() < long_output.len()); |
| 794 | assert!(shortened.starts_with("[tool result retained-history truncated from")); |
| 795 | |
| 796 | let header = crate::compaction::SUMMARY_HEADER; |
| 797 | assert!( |
| 798 | header.contains("last steps of the current round"), |
| 799 | "{header}" |
| 800 | ); |
| 801 | assert!( |
| 802 | header.contains("including earlier steps of this round"), |
| 803 | "{header}" |
| 804 | ); |
| 805 | assert!( |
| 806 | header.contains("Long tool output there is shortened"), |
| 807 | "{header}" |
| 808 | ); |
| 809 | assert!( |
| 810 | header.contains("with a marker where it was cut"), |
| 811 | "{header}" |
| 812 | ); |
| 813 | assert!(!header.contains("as they were"), "{header}"); |
| 814 | } |
| 815 | |
| 816 | #[test] |
| 817 | fn coverage_floor_rejects_a_replacement_that_drops_last_round_tools() { |
| 818 | let original = vec![ |
| 819 | msg("user", "Run the failing test."), |
| 820 | msg("assistant", "Running."), |
| 821 | tool_use("live", "Bash", json!({"command": "cargo test"})), |
| 822 | tool_result("live", "test session_store::roundtrip ... FAILED"), |
| 823 | ]; |
| 824 | let gutting = vec![ |
| 825 | msg("user", "Run the failing test."), |
| 826 | checkpoint("and kept going"), |
| 827 | ]; |
| 828 | let error = validate_last_round_coverage(&original, &gutting) |
| 829 | .expect_err("dropping the last tool result must fail the coverage floor"); |
| 830 | assert!(error.to_string().contains("tool result live"), "{error}"); |
| 831 | assert!(validate_last_round_coverage(&original, &original).is_ok()); |
| 832 | } |
| 833 | |
| 834 | #[test] |
| 835 | fn coverage_floor_rejects_a_replacement_that_drops_last_round_assistant() { |
| 836 | let original = vec![ |
| 837 | msg("user", "What failed?"), |
| 838 | msg("assistant", "session_store::roundtrip panics on reload."), |
| 839 | ]; |
| 840 | let error = validate_last_round_coverage(&original, &[msg("user", "What failed?")]) |
| 841 | .expect_err("dropping last-round assistant text must fail closed"); |
| 842 | assert!(error.to_string().contains("assistant"), "{error}"); |
| 843 | } |
| 844 | |
| 845 | /// The round spans the tool-bearing turn *and* the toolless tail after it, |
| 846 | /// because `last_round_start` walks back for the tools. Checking only the |
| 847 | /// first user text it found meant a rewrite could keep the older question |
| 848 | /// and drop the one the person actually just asked. |
| 849 | #[test] |
| 850 | fn coverage_floor_rejects_a_replacement_that_drops_the_latest_user_turn() { |
| 851 | let original = vec![ |
| 852 | msg("user", "Run the suite."), |
| 853 | msg("assistant", "Running."), |
| 854 | tool_use("live", "Bash", json!({"command": "cargo test"})), |
| 855 | tool_result("live", "ok"), |
| 856 | msg("user", "Now ship it."), |
| 857 | msg("assistant", "Shipping."), |
| 858 | ]; |
| 859 | assert_eq!(last_round_start(&original), 0, "round must span both turns"); |
| 860 | |
| 861 | let drops_latest = vec![ |
| 862 | msg("user", "Run the suite."), |
| 863 | msg("assistant", "Running."), |
| 864 | tool_use("live", "Bash", json!({"command": "cargo test"})), |
| 865 | tool_result("live", "ok"), |
| 866 | checkpoint("then shipped"), |
| 867 | ]; |
| 868 | let error = validate_last_round_coverage(&original, &drops_latest) |
| 869 | .expect_err("dropping the latest user turn must fail the coverage floor"); |
| 870 | assert!(error.to_string().contains("user message"), "{error}"); |
| 871 | assert!(validate_last_round_coverage(&original, &original).is_ok()); |
| 872 | } |
| 873 | |
| 874 | /// A surviving `tool_result` whose `tool_use` was summarized away is an |
| 875 | /// orphan the provider rejects, so the floor must cover the call too. |
| 876 | #[test] |
| 877 | fn coverage_floor_rejects_a_replacement_that_drops_the_tool_call() { |
| 878 | let original = vec![ |
| 879 | msg("user", "Run the failing test."), |
| 880 | msg("assistant", "Running."), |
| 881 | tool_use("live", "Bash", json!({"command": "cargo test"})), |
| 882 | tool_result("live", "FAILED"), |
| 883 | ]; |
| 884 | let orphaned = vec![ |
| 885 | msg("user", "Run the failing test."), |
| 886 | msg("assistant", "Running."), |
| 887 | tool_result("live", "FAILED"), |
| 888 | checkpoint("and it failed"), |
| 889 | ]; |
| 890 | let error = validate_last_round_coverage(&original, &orphaned) |
| 891 | .expect_err("dropping the tool call must fail the coverage floor"); |
| 892 | assert!(error.to_string().contains("tool call live"), "{error}"); |
| 893 | } |
| 894 | |
| 895 | /// "Some assistant message survived" was satisfied by the summary the |
| 896 | /// rewrite had just written, so the round's real output could vanish. |
| 897 | #[test] |
| 898 | fn coverage_floor_rejects_assistant_output_replaced_by_a_summary() { |
| 899 | let original = vec![ |
| 900 | msg("user", "What failed?"), |
| 901 | msg("assistant", "session_store::roundtrip panics on reload."), |
| 902 | ]; |
| 903 | let summarized = vec![ |
| 904 | msg("user", "What failed?"), |
| 905 | msg("assistant", "Earlier we discussed several test failures."), |
| 906 | ]; |
| 907 | let error = validate_last_round_coverage(&original, &summarized) |
| 908 | .expect_err("substituting a summary for the round's output must fail closed"); |
| 909 | assert!(error.to_string().contains("assistant"), "{error}"); |
| 910 | } |
| 911 | |
| 912 | #[test] |
| 913 | fn survival_contract_rejects_dropped_anchors_and_receipts() { |
| 914 | let original = vec![msg("user", "Keep the pin."), msg("assistant", "Anchored.")]; |
| 915 | let without_receipt = vec![msg("user", "Keep the pin."), msg("assistant", "Anchored.")]; |
| 916 | let error = validate_survival_contract(&original, &without_receipt, Some("ship 0.9.12")) |
| 917 | .expect_err("missing checkpoint receipt must fail closed"); |
| 918 | assert!(error.to_string().contains("receipt"), "{error}"); |
| 919 | |
| 920 | let without_anchor = vec![ |
| 921 | msg("user", "Keep the pin."), |
| 922 | msg("assistant", "Anchored."), |
| 923 | checkpoint("progress without the pin"), |
| 924 | ]; |
| 925 | let error = validate_survival_contract(&original, &without_anchor, Some("ship 0.9.12")) |
| 926 | .expect_err("dropped /anchor text must fail closed"); |
| 927 | assert!(error.to_string().contains("anchor"), "{error}"); |
| 928 | } |
| 929 | |
| 930 | #[test] |
| 931 | fn last_round_starts_at_the_latest_plain_user_message() { |
| 932 | let messages = vec![ |
| 933 | msg("user", "older"), |
| 934 | msg("assistant", "working"), |
| 935 | tool_result("old", "stale"), |
| 936 | msg("user", "Run the suite now."), |
| 937 | msg("assistant", "Rerunning."), |
| 938 | tool_use("live", "Bash", json!({"command": "cargo test"})), |
| 939 | tool_result("live", "ok"), |
| 940 | ]; |
| 941 | assert_eq!(last_round_start(&messages), 3); // last user with tools |
| 942 | let (start, end) = last_round_range(&messages); |
| 943 | let kept = bound_last_round(&messages[start..end]); |
| 944 | assert!(kept.iter().any(|message| { |
| 945 | message.content.iter().any(|block| { |
| 946 | matches!( |
| 947 | block, |
| 948 | ContentBlock::ToolResult { tool_use_id, content, .. } |
| 949 | if tool_use_id == "live" && content == "ok" |
| 950 | ) |
| 951 | }) |
| 952 | })); |
| 953 | } |
| 954 | |
| 955 | #[test] |
| 956 | fn second_compaction_keeps_long_user_question_and_tool_pair_past_retention_budget() { |
| 957 | const PRODUCTION_MIN_RETAINED_TOKENS: usize = 2_000; |
| 958 | let long_question = format!( |
| 959 | "{}?", |
| 960 | "Analyze every step of this case carefully. ".repeat(400) |
| 961 | ); |
| 962 | assert!(long_question.len() > PRODUCTION_MIN_RETAINED_TOKENS * 3); |
| 963 | let original = vec![ |
| 964 | msg("user", &long_question), |
| 965 | tool_use("call_1", "Bash", json!({"command": "echo ready"})), |
| 966 | tool_result("call_1", "ready"), |
| 967 | ]; |
| 968 | let first_summary = |
| 969 | crate::compaction::build_compaction_summary_block_text("First pass complete", ""); |
| 970 | let mut first = build_replacement_history( |
| 971 | &original, |
| 972 | &first_summary, |
| 973 | None, |
| 974 | PRODUCTION_MIN_RETAINED_TOKENS, |
| 975 | ) |
| 976 | .expect("first compaction"); |
| 977 | crate::runtime_handoff::replace_agent_topology_checkpoint(&mut first, &[]); |
| 978 | assert_eq!(last_round_start(&first), 0); |
| 979 | let topology = first |
| 980 | .iter() |
| 981 | .find(|message| crate::runtime_handoff::is_agent_topology_checkpoint(message)) |
| 982 | .expect("first compaction topology checkpoint") |
| 983 | .clone(); |
| 984 | assert!( |
| 985 | crate::compaction::retained_user_messages( |
| 986 | std::slice::from_ref(&topology), |
| 987 | PRODUCTION_MIN_RETAINED_TOKENS, |
| 988 | ) |
| 989 | .is_empty(), |
| 990 | "runtime topology must not consume the older-user retention budget" |
| 991 | ); |
| 992 | |
| 993 | let second_summary = |
| 994 | crate::compaction::build_compaction_summary_block_text("Second pass complete", ""); |
| 995 | let second = build_replacement_history( |
| 996 | &first, |
| 997 | &second_summary, |
| 998 | None, |
| 999 | PRODUCTION_MIN_RETAINED_TOKENS, |
| 1000 | ) |
| 1001 | .expect("second compaction"); |
| 1002 | assert!( |
| 1003 | second.iter().any(|message| { |
| 1004 | user_text_of(message).as_deref() == Some(long_question.as_str()) |
| 1005 | }) |
| 1006 | ); |
| 1007 | assert!(second.iter().any(|message| has_tool_use_id( |
| 1008 | message, |
| 1009 | &( |
| 1010 | codewhale_models::ToolCallKey::LegacyProvider("call_1"), |
| 1011 | "call_1" |
| 1012 | ) |
| 1013 | ))); |
| 1014 | assert!(second.iter().any(|message| has_tool_result_id( |
| 1015 | message, |
| 1016 | &( |
| 1017 | codewhale_models::ToolCallKey::LegacyProvider("call_1"), |
| 1018 | "call_1" |
| 1019 | ) |
| 1020 | ))); |
| 1021 | let without_question = second |
| 1022 | .iter() |
| 1023 | .filter(|message| user_text_of(message).as_deref() != Some(long_question.as_str())) |
| 1024 | .cloned() |
| 1025 | .collect::<Vec<_>>(); |
| 1026 | assert!(validate_last_round_coverage(&first, &without_question).is_err()); |
| 1027 | } |
| 1028 | |
| 1029 | #[test] |
| 1030 | fn runtime_text_cannot_satisfy_real_user_coverage() { |
| 1031 | let runtime = crate::runtime_handoff::operate_contract_runtime_message(); |
| 1032 | let copied_text = user_text_of(&runtime).expect("runtime text"); |
| 1033 | let original = vec![msg("user", &copied_text), msg("assistant", "Acknowledged")]; |
| 1034 | let replacement = vec![runtime, msg("assistant", "Acknowledged")]; |
| 1035 | assert!( |
| 1036 | validate_last_round_coverage(&original, &replacement).is_err(), |
| 1037 | "runtime-owned text must not stand in for the user's actual prompt" |
| 1038 | ); |
| 1039 | } |
| 1040 | |
| 1041 | #[test] |
| 1042 | fn operate_contract_survives_compaction_without_spending_user_budget() { |
| 1043 | let contract = crate::runtime_handoff::operate_contract_runtime_message(); |
| 1044 | let original = vec![ |
| 1045 | contract.clone(), |
| 1046 | msg("user", "First task"), |
| 1047 | msg("assistant", "Working"), |
| 1048 | msg("user", "Continue the same task"), |
| 1049 | msg("assistant", "Continuing"), |
| 1050 | ]; |
| 1051 | let replaced = build_replacement_history( |
| 1052 | &original, |
| 1053 | &format!("{COMPACTION_SUMMARY_MARKER}: work continues"), |
| 1054 | None, |
| 1055 | 1, |
| 1056 | ) |
| 1057 | .expect("compaction must retain the active Operate contract"); |
| 1058 | assert_eq!(replaced.first(), Some(&contract)); |
| 1059 | assert_eq!( |
| 1060 | replaced |
| 1061 | .iter() |
| 1062 | .filter(|message| **message == contract) |
| 1063 | .count(), |
| 1064 | 1 |
| 1065 | ); |
| 1066 | } |
| 1067 | |
| 1068 | #[test] |
| 1069 | fn compaction_prefers_current_operate_contract_over_legacy() { |
| 1070 | let legacy = crate::runtime_handoff::legacy_operate_contract_runtime_message(); |
| 1071 | let current = crate::runtime_handoff::operate_contract_runtime_message(); |
| 1072 | let original = vec![ |
| 1073 | legacy.clone(), |
| 1074 | current.clone(), |
| 1075 | msg("user", "Continue"), |
| 1076 | msg("assistant", "Working"), |
| 1077 | ]; |
| 1078 | let replaced = build_replacement_history( |
| 1079 | &original, |
| 1080 | &format!("{COMPACTION_SUMMARY_MARKER}: work continues"), |
| 1081 | None, |
| 1082 | 1, |
| 1083 | ) |
| 1084 | .expect("current contract must survive compaction"); |
| 1085 | assert_eq!(replaced.first(), Some(¤t)); |
| 1086 | assert!(!replaced.contains(&legacy)); |
| 1087 | assert_eq!( |
| 1088 | replaced |
| 1089 | .iter() |
| 1090 | .filter(|message| crate::runtime_handoff::is_operate_contract_message(message)) |
| 1091 | .count(), |
| 1092 | 1 |
| 1093 | ); |
| 1094 | } |
| 1095 | |
| 1096 | #[test] |
| 1097 | fn last_round_walks_back_through_toolless_tails_to_the_tool_round() { |
| 1098 | let original = vec![ |
| 1099 | msg("user", "Run the failing test."), |
| 1100 | msg("assistant", "Running."), |
| 1101 | tool_use("live", "Bash", json!({"command": "cargo test"})), |
| 1102 | tool_result("live", "test session_store::roundtrip ... FAILED"), |
| 1103 | msg("user", "ok thanks"), |
| 1104 | msg("assistant", "you're welcome"), |
| 1105 | msg("user", "one more thing"), |
| 1106 | msg("assistant", "sure"), |
| 1107 | ]; |
| 1108 | assert_eq!(last_round_start(&original), 0); |
| 1109 | let next = format!("{COMPACTION_SUMMARY_MARKER}: keep the failing test result"); |
| 1110 | let replaced = build_replacement_history( |
| 1111 | &original, |
| 1112 | &next, |
| 1113 | None, |
| 1114 | crate::compaction::COMPACT_RETAINED_USER_MESSAGE_MAX_TOKENS, |
| 1115 | ) |
| 1116 | .expect("toolless tails must not drop the last tool result"); |
| 1117 | assert!(replaced.iter().any(|message| { |
| 1118 | message.content.iter().any(|block| { |
| 1119 | matches!( |
| 1120 | block, |
| 1121 | ContentBlock::ToolResult { tool_use_id, content, .. } |
| 1122 | if tool_use_id == "live" && content.contains("FAILED") |
| 1123 | ) |
| 1124 | }) |
| 1125 | })); |
| 1126 | } |
| 1127 | |
| 1128 | #[test] |
| 1129 | fn chat_only_history_keeps_the_latest_user_round() { |
| 1130 | let messages = vec![ |
| 1131 | msg("user", "hello"), |
| 1132 | msg("assistant", "hi"), |
| 1133 | msg("user", "how are you"), |
| 1134 | msg("assistant", "fine"), |
| 1135 | ]; |
| 1136 | assert_eq!(last_round_start(&messages), 2); |
| 1137 | } |
| 1138 | |
| 1139 | #[derive(serde::Deserialize)] |
| 1140 | struct FixtureMatrix { |
| 1141 | schema_version: u32, |
| 1142 | cases: Vec<FixtureCase>, |
| 1143 | } |
| 1144 | |
| 1145 | #[derive(serde::Deserialize)] |
| 1146 | struct FixtureCase { |
| 1147 | id: String, |
| 1148 | expect: String, |
| 1149 | #[serde(default)] |
| 1150 | anchors: Option<String>, |
| 1151 | original: Vec<Message>, |
| 1152 | replacement: Vec<Message>, |
| 1153 | #[serde(default)] |
| 1154 | last_round_start: Option<usize>, |
| 1155 | } |
| 1156 | |
| 1157 | #[test] |
| 1158 | fn fixture_matrix_enforces_survival_contract() { |
| 1159 | let matrix: FixtureMatrix = |
| 1160 | serde_json::from_str(include_str!("fixtures/matrix.json")).expect("matrix.json"); |
| 1161 | assert_eq!(matrix.schema_version, 2); |
| 1162 | assert!( |
| 1163 | matrix.cases.len() >= 8, |
| 1164 | "fixture matrix must cover last-round, toolless-tail, chat-only, anchor, and receipt cases" |
| 1165 | ); |
| 1166 | for case in &matrix.cases { |
| 1167 | if let Some(start) = case.last_round_start { |
| 1168 | assert_eq!( |
| 1169 | last_round_start(&case.original), |
| 1170 | start, |
| 1171 | "{} last_round_start", |
| 1172 | case.id |
| 1173 | ); |
| 1174 | } |
| 1175 | let result = validate_survival_contract( |
| 1176 | &case.original, |
| 1177 | &case.replacement, |
| 1178 | case.anchors.as_deref(), |
| 1179 | ); |
| 1180 | match case.expect.as_str() { |
| 1181 | "pass" => { |
| 1182 | result.unwrap_or_else(|error| panic!("{} should pass: {error}", case.id)); |
| 1183 | } |
| 1184 | "fail" => { |
| 1185 | result.expect_err(&format!("{} should fail closed", case.id)); |
| 1186 | } |
| 1187 | other => panic!("{}: unknown expect {other}", case.id), |
| 1188 | } |
| 1189 | } |
| 1190 | } |
| 1191 | |
| 1192 | #[test] |
| 1193 | fn second_compact_does_not_duplicate_prior_summaries() { |
| 1194 | let first = vec![ |
| 1195 | msg("user", "older"), |
| 1196 | msg("user", "Run the suite now."), |
| 1197 | msg("assistant", "Rerunning."), |
| 1198 | tool_use("live", "Bash", json!({"command": "cargo test"})), |
| 1199 | tool_result("live", "ok"), |
| 1200 | checkpoint("first handoff: suite still running"), |
| 1201 | ]; |
| 1202 | let next = format!( |
| 1203 | "{COMPACTION_SUMMARY_MARKER}: second handoff with User-pinned anchors (verbatim):\nship 0.9.12" |
| 1204 | ); |
| 1205 | let replaced = build_replacement_history( |
| 1206 | &first, |
| 1207 | &next, |
| 1208 | Some("ship 0.9.12"), |
| 1209 | crate::compaction::COMPACT_RETAINED_USER_MESSAGE_MAX_TOKENS, |
| 1210 | ) |
| 1211 | .expect("second compact must keep last round and one receipt"); |
| 1212 | let checkpoints = replaced |
| 1213 | .iter() |
| 1214 | .filter(|message| is_compaction_checkpoint_message(message)) |
| 1215 | .count(); |
| 1216 | assert_eq!(checkpoints, 1, "{replaced:?}"); |
| 1217 | assert!(replaced.iter().any(|message| { |
| 1218 | message.content.iter().any(|block| { |
| 1219 | matches!( |
| 1220 | block, |
| 1221 | ContentBlock::ToolResult { tool_use_id, .. } if tool_use_id == "live" |
| 1222 | ) |
| 1223 | }) |
| 1224 | })); |
| 1225 | require_text_survives(&replaced, "ship 0.9.12", "pinned /anchor text").unwrap(); |
| 1226 | } |
| 1227 | } |
| 1228 |