| 1 | //! Privacy-first session failure diagnostics (#2022). |
| 2 | //! |
| 3 | //! This module intentionally consumes loose JSONL event shapes instead of one |
| 4 | //! exact persisted-session schema. Runtime logs, tool audits, and future bug |
| 5 | //! exports can all emit slightly different records; the classifier only needs |
| 6 | //! redacted handles, aggregate counts, and broad failure classes. |
| 7 | |
| 8 | use std::collections::{BTreeMap, BTreeSet}; |
| 9 | use std::fmt; |
| 10 | |
| 11 | use serde::{Deserialize, Serialize}; |
| 12 | use serde_json::Value; |
| 13 | |
| 14 | use crate::error_taxonomy::{ErrorCategory, classify_error_message}; |
| 15 | |
| 16 | #[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Serialize, Deserialize)] |
| 17 | #[serde(rename_all = "snake_case")] |
| 18 | pub(crate) enum SessionFailureClass { |
| 19 | CommandExit, |
| 20 | Network, |
| 21 | SandboxApproval, |
| 22 | MissingDependency, |
| 23 | Timeout, |
| 24 | BackgroundJob, |
| 25 | ToolSchema, |
| 26 | Model, |
| 27 | Unknown, |
| 28 | UnclosedTurn, |
| 29 | } |
| 30 | |
| 31 | impl fmt::Display for SessionFailureClass { |
| 32 | fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { |
| 33 | let label = match self { |
| 34 | Self::CommandExit => "command_exit", |
| 35 | Self::Network => "network", |
| 36 | Self::SandboxApproval => "sandbox_approval", |
| 37 | Self::MissingDependency => "missing_dependency", |
| 38 | Self::Timeout => "timeout", |
| 39 | Self::BackgroundJob => "background_job", |
| 40 | Self::ToolSchema => "tool_schema", |
| 41 | Self::Model => "model", |
| 42 | Self::Unknown => "unknown", |
| 43 | Self::UnclosedTurn => "unclosed_turn", |
| 44 | }; |
| 45 | f.write_str(label) |
| 46 | } |
| 47 | } |
| 48 | |
| 49 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 50 | pub(crate) struct SessionFailureSource { |
| 51 | pub line: usize, |
| 52 | #[serde(skip_serializing_if = "Option::is_none")] |
| 53 | pub event: Option<String>, |
| 54 | #[serde(skip_serializing_if = "Option::is_none")] |
| 55 | pub turn_ref: Option<String>, |
| 56 | #[serde(skip_serializing_if = "Option::is_none")] |
| 57 | pub tool_name: Option<String>, |
| 58 | #[serde(skip_serializing_if = "Option::is_none")] |
| 59 | pub timestamp: Option<String>, |
| 60 | } |
| 61 | |
| 62 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 63 | pub(crate) struct SessionFailureSummary { |
| 64 | pub total_lines: usize, |
| 65 | pub malformed_lines: usize, |
| 66 | pub counts: BTreeMap<SessionFailureClass, usize>, |
| 67 | pub sources: BTreeMap<SessionFailureClass, Vec<SessionFailureSource>>, |
| 68 | } |
| 69 | |
| 70 | impl SessionFailureSummary { |
| 71 | #[must_use] |
| 72 | pub(crate) fn count(&self, class: SessionFailureClass) -> usize { |
| 73 | self.counts.get(&class).copied().unwrap_or(0) |
| 74 | } |
| 75 | |
| 76 | fn record(&mut self, class: SessionFailureClass, source: SessionFailureSource) { |
| 77 | *self.counts.entry(class).or_insert(0) += 1; |
| 78 | self.sources.entry(class).or_default().push(source); |
| 79 | } |
| 80 | } |
| 81 | |
| 82 | #[must_use] |
| 83 | pub(crate) fn analyze_session_failure_jsonl(jsonl: &str) -> SessionFailureSummary { |
| 84 | let mut summary = SessionFailureSummary { |
| 85 | total_lines: 0, |
| 86 | malformed_lines: 0, |
| 87 | counts: BTreeMap::new(), |
| 88 | sources: BTreeMap::new(), |
| 89 | }; |
| 90 | let mut open_turns: BTreeMap<String, SessionFailureSource> = BTreeMap::new(); |
| 91 | |
| 92 | for (idx, raw_line) in jsonl.lines().enumerate() { |
| 93 | let line_no = idx + 1; |
| 94 | let trimmed = raw_line.trim(); |
| 95 | if trimmed.is_empty() { |
| 96 | continue; |
| 97 | } |
| 98 | summary.total_lines += 1; |
| 99 | let Ok(value) = serde_json::from_str::<Value>(trimmed) else { |
| 100 | summary.malformed_lines += 1; |
| 101 | continue; |
| 102 | }; |
| 103 | |
| 104 | let event = event_name(&value); |
| 105 | let turn_id = string_field_any(&value, &["turn_id", "turnId", "run_id"]); |
| 106 | let source = source_handle(line_no, &value, event.clone(), turn_id.as_deref()); |
| 107 | let failure_signal = has_failure_signal(&value); |
| 108 | |
| 109 | if event_matches( |
| 110 | event.as_deref(), |
| 111 | &["turn_started", "turnstarted", "turn_start"], |
| 112 | ) { |
| 113 | if let Some(turn_id) = turn_id { |
| 114 | open_turns.insert(turn_id, source); |
| 115 | } |
| 116 | continue; |
| 117 | } |
| 118 | if event_matches( |
| 119 | event.as_deref(), |
| 120 | &[ |
| 121 | "turn_complete", |
| 122 | "turncompleted", |
| 123 | "turn_finished", |
| 124 | "turnfinished", |
| 125 | ], |
| 126 | ) { |
| 127 | if let Some(turn_id) = turn_id.as_ref() { |
| 128 | open_turns.remove(turn_id); |
| 129 | } |
| 130 | if failure_signal { |
| 131 | let class = classify_failure_signal(&value); |
| 132 | summary.record(class, source); |
| 133 | } |
| 134 | continue; |
| 135 | } |
| 136 | |
| 137 | if failure_signal { |
| 138 | let class = classify_failure_signal(&value); |
| 139 | summary.record(class, source); |
| 140 | } |
| 141 | } |
| 142 | |
| 143 | for (_, source) in open_turns { |
| 144 | summary.record(SessionFailureClass::UnclosedTurn, source); |
| 145 | } |
| 146 | |
| 147 | summary |
| 148 | } |
| 149 | |
| 150 | #[must_use] |
| 151 | pub(crate) fn format_redacted_failure_summary(summary: &SessionFailureSummary) -> String { |
| 152 | if summary.counts.is_empty() { |
| 153 | return "No session failure signals detected.".to_string(); |
| 154 | } |
| 155 | let mut lines = vec![format!( |
| 156 | "Session failure diagnostics: {} JSONL lines inspected, {} malformed skipped.", |
| 157 | summary.total_lines, summary.malformed_lines |
| 158 | )]; |
| 159 | for (class, count) in &summary.counts { |
| 160 | let sample = summary |
| 161 | .sources |
| 162 | .get(class) |
| 163 | .and_then(|sources| sources.first()) |
| 164 | .map(format_source) |
| 165 | .unwrap_or_else(|| "no source".to_string()); |
| 166 | lines.push(format!("- {class}: {count} (sample: {sample})")); |
| 167 | } |
| 168 | lines.join("\n") |
| 169 | } |
| 170 | |
| 171 | fn source_handle( |
| 172 | line: usize, |
| 173 | value: &Value, |
| 174 | event: Option<String>, |
| 175 | turn_id: Option<&str>, |
| 176 | ) -> SessionFailureSource { |
| 177 | let tool_name = string_field_any(value, &["tool_name", "toolName", "tool"]) |
| 178 | .filter(|name| event.as_deref().is_none_or(|event| event != name)); |
| 179 | SessionFailureSource { |
| 180 | line, |
| 181 | event, |
| 182 | turn_ref: turn_id.map(crate::utils::redacted_identifier_for_log), |
| 183 | tool_name, |
| 184 | timestamp: string_field_any(value, &["timestamp", "ts", "created_at", "createdAt"]), |
| 185 | } |
| 186 | } |
| 187 | |
| 188 | fn format_source(source: &SessionFailureSource) -> String { |
| 189 | let mut parts = vec![format!("line {}", source.line)]; |
| 190 | if let Some(event) = source.event.as_deref() { |
| 191 | parts.push(format!("event={event}")); |
| 192 | } |
| 193 | if let Some(turn_ref) = source.turn_ref.as_deref() { |
| 194 | parts.push(format!("turn={turn_ref}")); |
| 195 | } |
| 196 | if let Some(tool_name) = source.tool_name.as_deref() { |
| 197 | parts.push(format!("tool={tool_name}")); |
| 198 | } |
| 199 | if let Some(timestamp) = source.timestamp.as_deref() { |
| 200 | parts.push(format!("ts={timestamp}")); |
| 201 | } |
| 202 | parts.join(" ") |
| 203 | } |
| 204 | |
| 205 | fn has_failure_signal(value: &Value) -> bool { |
| 206 | numeric_field_any(value, &["exit_code", "exitCode"]).is_some_and(|code| code != 0) |
| 207 | || bool_field_any(value, &["success"]).is_some_and(|success| !success) |
| 208 | || bool_field_any(value, &["is_error", "isError"]).unwrap_or(false) |
| 209 | || failure_status(value).is_some() |
| 210 | || string_field_any(value, &["error", "stderr"]).is_some_and(|text| !text.is_empty()) |
| 211 | } |
| 212 | |
| 213 | fn classify_failure_signal(value: &Value) -> SessionFailureClass { |
| 214 | if let Some(message) = diagnostic_message(value) { |
| 215 | return classify_session_failure(value, &message); |
| 216 | } |
| 217 | if let Some(status) = failure_status(value) { |
| 218 | let lower = status.to_ascii_lowercase(); |
| 219 | if lower.contains("timeout") || lower.contains("timed_out") { |
| 220 | return SessionFailureClass::Timeout; |
| 221 | } |
| 222 | if lower.contains("cancel") || lower.contains("background") || lower.contains("stale") { |
| 223 | return SessionFailureClass::BackgroundJob; |
| 224 | } |
| 225 | } |
| 226 | if numeric_field_any(value, &["exit_code", "exitCode"]).is_some_and(|code| code != 0) { |
| 227 | return SessionFailureClass::CommandExit; |
| 228 | } |
| 229 | SessionFailureClass::Unknown |
| 230 | } |
| 231 | |
| 232 | fn classify_session_failure(value: &Value, message: &str) -> SessionFailureClass { |
| 233 | let lower = message.to_ascii_lowercase(); |
| 234 | if lower.contains("background") |
| 235 | || lower.contains("task_shell") |
| 236 | || lower.contains("job timed out") |
| 237 | || lower.contains("job cancelled") |
| 238 | || lower.contains("stale job") |
| 239 | { |
| 240 | return SessionFailureClass::BackgroundJob; |
| 241 | } |
| 242 | if lower.contains("sandbox") |
| 243 | || lower.contains("approval") |
| 244 | || lower.contains("permission denied") |
| 245 | || lower.contains("operation not permitted") |
| 246 | || lower.contains("read-only") |
| 247 | || lower.contains("access is denied") |
| 248 | { |
| 249 | return SessionFailureClass::SandboxApproval; |
| 250 | } |
| 251 | if lower.contains("command not found") |
| 252 | || lower.contains("no such file or directory") |
| 253 | || lower.contains("missing binary") |
| 254 | || lower.contains("enoent") |
| 255 | || lower.contains("not installed") |
| 256 | { |
| 257 | return SessionFailureClass::MissingDependency; |
| 258 | } |
| 259 | if lower.contains("missing field") |
| 260 | || lower.contains("invalid tool") |
| 261 | || lower.contains("invalid input") |
| 262 | || lower.contains("schema") |
| 263 | || lower.contains("tool arguments") |
| 264 | { |
| 265 | return SessionFailureClass::ToolSchema; |
| 266 | } |
| 267 | if numeric_field_any(value, &["exit_code", "exitCode"]).is_some_and(|code| code != 0) |
| 268 | || lower.contains("non-zero") |
| 269 | || lower.contains("exit status") |
| 270 | || lower.contains("exit code") |
| 271 | { |
| 272 | return SessionFailureClass::CommandExit; |
| 273 | } |
| 274 | match classify_error_message(message) { |
| 275 | ErrorCategory::Network | ErrorCategory::RateLimit => SessionFailureClass::Network, |
| 276 | ErrorCategory::Timeout => SessionFailureClass::Timeout, |
| 277 | ErrorCategory::Authorization => SessionFailureClass::SandboxApproval, |
| 278 | ErrorCategory::Authentication => SessionFailureClass::Model, |
| 279 | ErrorCategory::State => SessionFailureClass::MissingDependency, |
| 280 | ErrorCategory::InvalidInput | ErrorCategory::Parse => SessionFailureClass::ToolSchema, |
| 281 | ErrorCategory::Tool => SessionFailureClass::CommandExit, |
| 282 | ErrorCategory::Internal if lower.contains("model") => SessionFailureClass::Model, |
| 283 | ErrorCategory::Internal => SessionFailureClass::Unknown, |
| 284 | } |
| 285 | } |
| 286 | |
| 287 | fn diagnostic_message(value: &Value) -> Option<String> { |
| 288 | let mut parts = Vec::new(); |
| 289 | collect_string_fields( |
| 290 | value, |
| 291 | &mut parts, |
| 292 | &[ |
| 293 | "error", "message", "stderr", "reason", "result", "output", "content", |
| 294 | ], |
| 295 | 0, |
| 296 | ); |
| 297 | let mut seen = BTreeSet::new(); |
| 298 | let deduped = parts |
| 299 | .into_iter() |
| 300 | .filter(|part| !part.trim().is_empty()) |
| 301 | .filter(|part| seen.insert(part.clone())) |
| 302 | .collect::<Vec<_>>(); |
| 303 | (!deduped.is_empty()).then(|| deduped.join(" ")) |
| 304 | } |
| 305 | |
| 306 | fn collect_string_fields(value: &Value, out: &mut Vec<String>, keys: &[&str], depth: usize) { |
| 307 | if depth > 4 { |
| 308 | return; |
| 309 | } |
| 310 | match value { |
| 311 | Value::Object(map) => { |
| 312 | for (key, value) in map { |
| 313 | if keys |
| 314 | .iter() |
| 315 | .any(|candidate| key.eq_ignore_ascii_case(candidate)) |
| 316 | && let Some(text) = value.as_str() |
| 317 | { |
| 318 | out.push(text.to_string()); |
| 319 | } |
| 320 | collect_string_fields(value, out, keys, depth + 1); |
| 321 | } |
| 322 | } |
| 323 | Value::Array(items) => { |
| 324 | for item in items { |
| 325 | collect_string_fields(item, out, keys, depth + 1); |
| 326 | } |
| 327 | } |
| 328 | _ => {} |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | fn event_name(value: &Value) -> Option<String> { |
| 333 | string_field_any(value, &["event", "type", "kind"]).map(|event| normalize_event(&event)) |
| 334 | } |
| 335 | |
| 336 | fn normalize_event(event: &str) -> String { |
| 337 | event |
| 338 | .trim() |
| 339 | .trim_matches('"') |
| 340 | .replace(['-', ' ', '.'], "_") |
| 341 | .to_ascii_lowercase() |
| 342 | } |
| 343 | |
| 344 | fn event_matches(event: Option<&str>, aliases: &[&str]) -> bool { |
| 345 | event.is_some_and(|event| aliases.contains(&event)) |
| 346 | } |
| 347 | |
| 348 | fn failure_status(value: &Value) -> Option<String> { |
| 349 | string_field_any(value, &["status", "state", "outcome"]).filter(|status| { |
| 350 | let normalized = normalize_event(status); |
| 351 | matches!( |
| 352 | normalized.as_str(), |
| 353 | "failed" |
| 354 | | "failure" |
| 355 | | "error" |
| 356 | | "errored" |
| 357 | | "cancelled" |
| 358 | | "canceled" |
| 359 | | "timeout" |
| 360 | | "timed_out" |
| 361 | | "stale" |
| 362 | ) |
| 363 | }) |
| 364 | } |
| 365 | |
| 366 | fn string_field_any(value: &Value, keys: &[&str]) -> Option<String> { |
| 367 | string_field_any_at(value, keys, 0) |
| 368 | } |
| 369 | |
| 370 | fn string_field_any_at(value: &Value, keys: &[&str], depth: usize) -> Option<String> { |
| 371 | if depth > 4 { |
| 372 | return None; |
| 373 | } |
| 374 | match value { |
| 375 | Value::Object(map) => { |
| 376 | for key in keys { |
| 377 | if let Some(value) = map.iter().find_map(|(candidate, value)| { |
| 378 | candidate.eq_ignore_ascii_case(key).then_some(value) |
| 379 | }) && let Some(text) = value.as_str() |
| 380 | { |
| 381 | return Some(text.to_string()); |
| 382 | } |
| 383 | } |
| 384 | for child in map.values() { |
| 385 | if let Some(found) = string_field_any_at(child, keys, depth + 1) { |
| 386 | return Some(found); |
| 387 | } |
| 388 | } |
| 389 | None |
| 390 | } |
| 391 | Value::Array(items) => items |
| 392 | .iter() |
| 393 | .find_map(|item| string_field_any_at(item, keys, depth + 1)), |
| 394 | _ => None, |
| 395 | } |
| 396 | } |
| 397 | |
| 398 | fn numeric_field_any(value: &Value, keys: &[&str]) -> Option<i64> { |
| 399 | numeric_field_any_at(value, keys, 0) |
| 400 | } |
| 401 | |
| 402 | fn numeric_field_any_at(value: &Value, keys: &[&str], depth: usize) -> Option<i64> { |
| 403 | if depth > 4 { |
| 404 | return None; |
| 405 | } |
| 406 | match value { |
| 407 | Value::Object(map) => { |
| 408 | for key in keys { |
| 409 | if let Some(value) = map.iter().find_map(|(candidate, value)| { |
| 410 | candidate.eq_ignore_ascii_case(key).then_some(value) |
| 411 | }) && let Some(number) = value.as_i64() |
| 412 | { |
| 413 | return Some(number); |
| 414 | } |
| 415 | } |
| 416 | for child in map.values() { |
| 417 | if let Some(found) = numeric_field_any_at(child, keys, depth + 1) { |
| 418 | return Some(found); |
| 419 | } |
| 420 | } |
| 421 | None |
| 422 | } |
| 423 | Value::Array(items) => items |
| 424 | .iter() |
| 425 | .find_map(|item| numeric_field_any_at(item, keys, depth + 1)), |
| 426 | _ => None, |
| 427 | } |
| 428 | } |
| 429 | |
| 430 | fn bool_field_any(value: &Value, keys: &[&str]) -> Option<bool> { |
| 431 | bool_field_any_at(value, keys, 0) |
| 432 | } |
| 433 | |
| 434 | fn bool_field_any_at(value: &Value, keys: &[&str], depth: usize) -> Option<bool> { |
| 435 | if depth > 4 { |
| 436 | return None; |
| 437 | } |
| 438 | match value { |
| 439 | Value::Object(map) => { |
| 440 | for key in keys { |
| 441 | if let Some(value) = map.iter().find_map(|(candidate, value)| { |
| 442 | candidate.eq_ignore_ascii_case(key).then_some(value) |
| 443 | }) && let Some(flag) = value.as_bool() |
| 444 | { |
| 445 | return Some(flag); |
| 446 | } |
| 447 | } |
| 448 | for child in map.values() { |
| 449 | if let Some(found) = bool_field_any_at(child, keys, depth + 1) { |
| 450 | return Some(found); |
| 451 | } |
| 452 | } |
| 453 | None |
| 454 | } |
| 455 | Value::Array(items) => items |
| 456 | .iter() |
| 457 | .find_map(|item| bool_field_any_at(item, keys, depth + 1)), |
| 458 | _ => None, |
| 459 | } |
| 460 | } |
| 461 | |
| 462 | #[cfg(test)] |
| 463 | mod tests { |
| 464 | use super::*; |
| 465 | |
| 466 | #[test] |
| 467 | fn synthetic_jsonl_classifies_environment_and_tool_failures() { |
| 468 | let jsonl = r#" |
| 469 | {"event":"turn_started","turn_id":"turn-secret-1","timestamp":"2026-06-25T12:00:00Z"} |
| 470 | {"event":"tool_call_complete","turn_id":"turn-secret-1","tool_name":"exec_shell","exit_code":127,"stderr":"bash: rg: command not found"} |
| 471 | {"event":"tool_call_complete","turn_id":"turn-secret-1","tool_name":"exec_shell","exit_code":2,"stderr":"command failed with exit code 2"} |
| 472 | {"event":"tool_call_complete","turn_id":"turn-secret-1","tool_name":"web_search","error":"DNS resolution failed for api.example.test"} |
| 473 | {"event":"tool_call_complete","turn_id":"turn-secret-1","tool_name":"apply_patch","error":"Permission denied by sandbox: read-only filesystem"} |
| 474 | {"event":"tool_call_complete","turn_id":"turn-secret-1","tool_name":"exec_shell","error":"request timed out after 30s"} |
| 475 | {"event":"tool_call_complete","turn_id":"turn-secret-1","tool_name":"task_shell_wait","error":"background job timed out"} |
| 476 | {"event":"tool_call_complete","turn_id":"turn-secret-1","tool_name":"mcp_tool","error":"missing field: tool arguments"} |
| 477 | {"event":"turn_complete","turn_id":"turn-secret-1","status":"completed"} |
| 478 | {"event":"turn_started","turn_id":"turn-secret-2"} |
| 479 | not json at all |
| 480 | "#; |
| 481 | |
| 482 | let summary = analyze_session_failure_jsonl(jsonl); |
| 483 | |
| 484 | assert_eq!(summary.malformed_lines, 1); |
| 485 | assert_eq!(summary.count(SessionFailureClass::MissingDependency), 1); |
| 486 | assert_eq!(summary.count(SessionFailureClass::CommandExit), 1); |
| 487 | assert_eq!(summary.count(SessionFailureClass::Network), 1); |
| 488 | assert_eq!(summary.count(SessionFailureClass::SandboxApproval), 1); |
| 489 | assert_eq!(summary.count(SessionFailureClass::Timeout), 1); |
| 490 | assert_eq!(summary.count(SessionFailureClass::BackgroundJob), 1); |
| 491 | assert_eq!(summary.count(SessionFailureClass::ToolSchema), 1); |
| 492 | assert_eq!(summary.count(SessionFailureClass::UnclosedTurn), 1); |
| 493 | |
| 494 | let sources = summary |
| 495 | .sources |
| 496 | .get(&SessionFailureClass::MissingDependency) |
| 497 | .expect("missing-dependency source"); |
| 498 | assert_eq!(sources[0].tool_name.as_deref(), Some("exec_shell")); |
| 499 | assert!( |
| 500 | sources[0] |
| 501 | .turn_ref |
| 502 | .as_deref() |
| 503 | .is_some_and(|turn| turn.starts_with("<redacted:")), |
| 504 | "turn ids must be redacted: {sources:?}" |
| 505 | ); |
| 506 | } |
| 507 | |
| 508 | #[test] |
| 509 | fn redacted_summary_omits_raw_messages_and_paths() { |
| 510 | let jsonl = r#" |
| 511 | {"event":"turn_started","turn_id":"turn-secret-1"} |
| 512 | {"event":"tool_call_complete","turn_id":"turn-secret-1","tool_name":"read_file","timestamp":"2026-06-25T12:34:56Z","error":"No such file or directory: /Users/alice/secret/project/.env"} |
| 513 | "#; |
| 514 | |
| 515 | let summary = analyze_session_failure_jsonl(jsonl); |
| 516 | let rendered = format_redacted_failure_summary(&summary); |
| 517 | |
| 518 | assert!(rendered.contains("missing_dependency")); |
| 519 | assert!(rendered.contains("line 3")); |
| 520 | assert!(rendered.contains("tool=read_file")); |
| 521 | assert!(rendered.contains("ts=2026-06-25T12:34:56Z")); |
| 522 | assert!(!rendered.contains("alice")); |
| 523 | assert!(!rendered.contains(".env")); |
| 524 | assert!(!rendered.contains("turn-secret-1")); |
| 525 | } |
| 526 | |
| 527 | #[test] |
| 528 | fn successful_content_does_not_create_unknown_failure() { |
| 529 | let jsonl = r#" |
| 530 | {"event":"turn_started","turn_id":"turn-secret-1"} |
| 531 | {"event":"tool_call_complete","turn_id":"turn-secret-1","tool_name":"exec_shell","success":true,"content":"command output mentioning error budgets is still normal content"} |
| 532 | {"event":"turn_complete","turn_id":"turn-secret-1","status":"completed","message":"done"} |
| 533 | "#; |
| 534 | |
| 535 | let summary = analyze_session_failure_jsonl(jsonl); |
| 536 | |
| 537 | assert_eq!(summary.count(SessionFailureClass::Unknown), 0); |
| 538 | assert_eq!(summary.count(SessionFailureClass::CommandExit), 0); |
| 539 | assert_eq!(summary.count(SessionFailureClass::UnclosedTurn), 0); |
| 540 | assert!( |
| 541 | summary.counts.is_empty(), |
| 542 | "summary should be empty: {summary:?}" |
| 543 | ); |
| 544 | } |
| 545 | |
| 546 | #[test] |
| 547 | fn empty_error_field_is_not_a_failure_signal() { |
| 548 | let jsonl = r#" |
| 549 | {"event":"tool_call_complete","success":true,"error":"","stderr":""} |
| 550 | "#; |
| 551 | |
| 552 | let summary = analyze_session_failure_jsonl(jsonl); |
| 553 | |
| 554 | assert!( |
| 555 | summary.counts.is_empty(), |
| 556 | "empty error/stderr sentinels should not signal failure: {summary:?}" |
| 557 | ); |
| 558 | } |
| 559 | |
| 560 | #[test] |
| 561 | fn nested_generic_fields_do_not_shadow_source_handles() { |
| 562 | let jsonl = r#" |
| 563 | {"event":"tool_call_complete","turn_id":"turn-real","tool_name":"exec_shell","success":false,"payload":{"id":"nested-id","name":"nested-name","error":"nested error"}} |
| 564 | "#; |
| 565 | |
| 566 | let summary = analyze_session_failure_jsonl(jsonl); |
| 567 | let source = summary |
| 568 | .sources |
| 569 | .values() |
| 570 | .flat_map(|sources| sources.iter()) |
| 571 | .next() |
| 572 | .expect("failure source"); |
| 573 | |
| 574 | assert_eq!(source.tool_name.as_deref(), Some("exec_shell")); |
| 575 | assert!( |
| 576 | source |
| 577 | .turn_ref |
| 578 | .as_deref() |
| 579 | .is_some_and(|turn| turn.starts_with("<redacted:")), |
| 580 | "top-level turn id should be redacted and used: {source:?}" |
| 581 | ); |
| 582 | } |
| 583 | |
| 584 | #[test] |
| 585 | fn deeply_nested_error_field_is_ignored() { |
| 586 | let jsonl = r#" |
| 587 | {"event":"tool_call_complete","payload":{"a":{"b":{"c":{"d":{"e":{"error":"too deep"}}}}}}} |
| 588 | "#; |
| 589 | |
| 590 | let summary = analyze_session_failure_jsonl(jsonl); |
| 591 | |
| 592 | assert!( |
| 593 | summary.counts.is_empty(), |
| 594 | "deeply nested error fields should not signal failure: {summary:?}" |
| 595 | ); |
| 596 | } |
| 597 | |
| 598 | #[test] |
| 599 | fn authentication_failures_are_model_failures_not_sandbox_approval() { |
| 600 | let jsonl = r#" |
| 601 | {"event":"model_response","success":false,"error":"Authentication failed: invalid API key"} |
| 602 | "#; |
| 603 | |
| 604 | let summary = analyze_session_failure_jsonl(jsonl); |
| 605 | |
| 606 | assert_eq!(summary.count(SessionFailureClass::Model), 1); |
| 607 | assert_eq!(summary.count(SessionFailureClass::SandboxApproval), 0); |
| 608 | } |
| 609 | } |
| 610 |