| 1 | //! Sub-agent and background-task routing helpers for the TUI loop. |
| 2 | |
| 3 | use std::time::{Duration, Instant}; |
| 4 | |
| 5 | use crate::task_manager::{TaskRecord, TaskStatus, TaskSummary}; |
| 6 | use crate::tools::subagent::{ |
| 7 | AgentWorkerStatus, MailboxMessage, SubAgentResult, SubAgentStatus, |
| 8 | subagent_progress_tool_display_name, |
| 9 | }; |
| 10 | use crate::tui::app::{ |
| 11 | AgentCurrentActivity, AgentCurrentActivityStatus, AgentProgressMeta, AgentRecentAction, App, |
| 12 | MAX_AGENT_RECENT_ACTIONS, TaskPanelEntry, TaskPanelEntryKind, bound_agent_activity_text, |
| 13 | }; |
| 14 | use crate::tui::history::{HistoryCell, SubAgentCell, summarize_tool_output}; |
| 15 | use crate::tui::pager::PagerView; |
| 16 | use crate::tui::tool_routing::refreshes_workspace_context_on_completion; |
| 17 | use crate::tui::widgets::agent_card::{ |
| 18 | AgentLifecycle, DelegateCard, FanoutCard, apply_to_delegate, apply_to_fanout, |
| 19 | }; |
| 20 | use crate::tui::workspace_context; |
| 21 | use codewhale_config::AppMode; |
| 22 | |
| 23 | /// Keep settled cards visible briefly, then archive them from the compact |
| 24 | /// live projection. Their transcript card and persisted agent record remain |
| 25 | /// reachable through the Agents register. |
| 26 | const SUBAGENT_TERMINAL_CARD_TTL: Duration = Duration::from_secs(45); |
| 27 | const SUBAGENT_TERMINAL_CARD_MAX_RETAINED: usize = 24; |
| 28 | |
| 29 | pub(super) fn running_agent_count(app: &App) -> usize { |
| 30 | let mut ids: std::collections::HashSet<&str> = |
| 31 | app.agent_progress.keys().map(String::as_str).collect(); |
| 32 | for agent in app |
| 33 | .subagent_cache |
| 34 | .iter() |
| 35 | .filter(|agent| matches!(agent.status, SubAgentStatus::Running)) |
| 36 | { |
| 37 | ids.insert(agent.agent_id.as_str()); |
| 38 | } |
| 39 | ids.len() |
| 40 | } |
| 41 | |
| 42 | /// Describe detached workers that deliberately survive a parent-turn stop. |
| 43 | /// |
| 44 | /// `agent` starts are detached from the turn cancellation token, so a plain |
| 45 | /// "Request cancelled" receipt is incomplete whenever live workers remain. |
| 46 | /// Use stable UI labels where available and raw ids as a lossless fallback; |
| 47 | /// sorting keeps the receipt deterministic across HashMap iteration order. |
| 48 | pub(super) fn parent_stop_status(app: &App, base: &str) -> String { |
| 49 | let mut ids = std::collections::BTreeSet::new(); |
| 50 | ids.extend(app.agent_progress.keys().cloned()); |
| 51 | ids.extend( |
| 52 | app.subagent_cache |
| 53 | .iter() |
| 54 | .filter(|agent| matches!(agent.status, SubAgentStatus::Running)) |
| 55 | .map(|agent| agent.agent_id.clone()), |
| 56 | ); |
| 57 | if ids.is_empty() { |
| 58 | return base.to_string(); |
| 59 | } |
| 60 | |
| 61 | let labels = ids |
| 62 | .into_iter() |
| 63 | .map(|id| { |
| 64 | app.agent_label_map |
| 65 | .get(&id) |
| 66 | .filter(|label| !label.trim().is_empty()) |
| 67 | .cloned() |
| 68 | .unwrap_or(id) |
| 69 | }) |
| 70 | .collect::<Vec<_>>(); |
| 71 | format!( |
| 72 | "{base}; detached workers continue (none canceled): {}", |
| 73 | labels.join(", ") |
| 74 | ) |
| 75 | } |
| 76 | |
| 77 | pub(super) fn active_fanout_counts(app: &App) -> Option<(usize, usize)> { |
| 78 | // Read running count from the canonical slot states on the active |
| 79 | // FanoutCard, if one exists. Used by `rlm` and any future multi-child |
| 80 | // dispatch the parent agent makes via repeated `agent`. |
| 81 | if let Some(idx) = app.last_fanout_card_index |
| 82 | && let Some(HistoryCell::SubAgent(SubAgentCell::Fanout(card))) = app.history.get(idx) |
| 83 | { |
| 84 | let running = card |
| 85 | .workers |
| 86 | .iter() |
| 87 | .filter(|slot| matches!(slot.status, AgentLifecycle::Running)) |
| 88 | .count(); |
| 89 | return Some((running, card.worker_count())); |
| 90 | } |
| 91 | None |
| 92 | } |
| 93 | |
| 94 | /// True when this child settled by being *parked* at the parent's turn end |
| 95 | /// rather than by asking anyone anything (#5906). |
| 96 | /// |
| 97 | /// The runtime parks a turn-owned child with a `needs_input` note that reads |
| 98 | /// like a question ("Resume this parked child with ..."), so every surface |
| 99 | /// that keys off `needs_input` used to render a parked husk with the same |
| 100 | /// "waiting for input" label as a child a user can actually answer. The |
| 101 | /// checkpoint records which of the two it is; this is the one place the UI |
| 102 | /// asks, so no surface has to sniff the reason string. |
| 103 | pub(crate) fn subagent_is_parked(agent: &SubAgentResult) -> bool { |
| 104 | agent |
| 105 | .checkpoint |
| 106 | .as_ref() |
| 107 | .is_some_and(|checkpoint| checkpoint.parked_at_turn_end) |
| 108 | } |
| 109 | |
| 110 | /// The one-line recovery a parked row shows instead of a pending question: |
| 111 | /// nobody will answer it, so it names the two ways out that actually exist. |
| 112 | pub(crate) fn parked_recovery_detail(app: &App) -> String { |
| 113 | codewhale_localization::tr( |
| 114 | app.ui_locale, |
| 115 | codewhale_localization::MessageId::AgentStatusParkedRecovery, |
| 116 | ) |
| 117 | .into_owned() |
| 118 | } |
| 119 | |
| 120 | pub(super) fn reconcile_subagent_activity_state(app: &mut App) { |
| 121 | reconcile_subagent_activity_state_at(app, Instant::now()); |
| 122 | } |
| 123 | |
| 124 | pub(super) fn apply_subagent_terminal_projection( |
| 125 | app: &mut App, |
| 126 | agent_id: &str, |
| 127 | status: SubAgentStatus, |
| 128 | result: Option<String>, |
| 129 | ) -> bool { |
| 130 | app.agent_progress.remove(agent_id); |
| 131 | |
| 132 | let worker_status = worker_status_for_terminal_projection(&status); |
| 133 | let safe_result = result.map(|result| bound_agent_activity_text(&result)); |
| 134 | let meta = app |
| 135 | .agent_progress_meta |
| 136 | .entry(agent_id.to_string()) |
| 137 | .or_default(); |
| 138 | // A parked child projects as `Interrupted` here. Interrupted-over-Waiting |
| 139 | // and interrupted-over-Parked are both "the settled state the surface |
| 140 | // already established, restated" — do not downgrade either (#5906). |
| 141 | let sticky = meta |
| 142 | .current_activity |
| 143 | .as_ref() |
| 144 | .map(|activity| activity.status) |
| 145 | .filter(|status| { |
| 146 | matches!( |
| 147 | status, |
| 148 | AgentCurrentActivityStatus::Waiting | AgentCurrentActivityStatus::Parked |
| 149 | ) |
| 150 | }); |
| 151 | let activity_status = match sticky { |
| 152 | Some(status) if worker_status == AgentWorkerStatus::Interrupted => status, |
| 153 | _ => worker_status.into(), |
| 154 | }; |
| 155 | let step = meta |
| 156 | .current_activity |
| 157 | .as_ref() |
| 158 | .and_then(|activity| activity.step); |
| 159 | meta.current_activity = Some(AgentCurrentActivity::bounded( |
| 160 | activity_status, |
| 161 | safe_result.clone(), |
| 162 | None, |
| 163 | step, |
| 164 | )); |
| 165 | meta.current_tool = None; |
| 166 | |
| 167 | let Some(agent) = app |
| 168 | .subagent_cache |
| 169 | .iter_mut() |
| 170 | .find(|agent| agent.agent_id == agent_id) |
| 171 | else { |
| 172 | reconcile_subagent_activity_state(app); |
| 173 | return false; |
| 174 | }; |
| 175 | |
| 176 | agent.worker_status = Some(worker_status); |
| 177 | agent.status = status; |
| 178 | if let Some(result) = safe_result { |
| 179 | agent.result = Some(result); |
| 180 | } |
| 181 | reconcile_subagent_activity_state(app); |
| 182 | true |
| 183 | } |
| 184 | |
| 185 | fn worker_status_for_terminal_projection(status: &SubAgentStatus) -> AgentWorkerStatus { |
| 186 | match status { |
| 187 | SubAgentStatus::Running => AgentWorkerStatus::Running, |
| 188 | SubAgentStatus::Completed => AgentWorkerStatus::Completed, |
| 189 | SubAgentStatus::Interrupted(_) => AgentWorkerStatus::Interrupted, |
| 190 | SubAgentStatus::Failed(_) | SubAgentStatus::BudgetExhausted => AgentWorkerStatus::Failed, |
| 191 | SubAgentStatus::Cancelled => AgentWorkerStatus::Cancelled, |
| 192 | } |
| 193 | } |
| 194 | |
| 195 | pub(super) fn reconcile_subagent_activity_state_at(app: &mut App, now: Instant) { |
| 196 | reconcile_terminal_subagent_card_retention(app, now); |
| 197 | |
| 198 | let cached_agents = app.subagent_cache.clone(); |
| 199 | let running_agents: Vec<(String, String)> = cached_agents |
| 200 | .iter() |
| 201 | .filter(|agent| matches!(agent.status, SubAgentStatus::Running)) |
| 202 | .map(|agent| { |
| 203 | ( |
| 204 | agent.agent_id.clone(), |
| 205 | summarize_tool_output(&agent.assignment.objective), |
| 206 | ) |
| 207 | }) |
| 208 | .collect(); |
| 209 | |
| 210 | let running_ids: std::collections::HashSet<String> = |
| 211 | running_agents.iter().map(|(id, _)| id.clone()).collect(); |
| 212 | // Evict a progress row only when the authoritative cache actually knows |
| 213 | // the agent and reports it non-running. A progress-only entry — an agent |
| 214 | // whose AgentSpawned/AgentList delivery was dropped under channel |
| 215 | // pressure so the cache has never seen it — must survive until the cache |
| 216 | // supersedes it, or spawned agents flicker in and out of the sidebar. |
| 217 | let cached_ids: std::collections::HashSet<String> = cached_agents |
| 218 | .iter() |
| 219 | .map(|agent| agent.agent_id.clone()) |
| 220 | .collect(); |
| 221 | app.agent_progress |
| 222 | .retain(|id, _| running_ids.contains(id) || !cached_ids.contains(id)); |
| 223 | let progress_ids: std::collections::HashSet<String> = |
| 224 | app.agent_progress.keys().cloned().collect(); |
| 225 | app.agent_progress_meta |
| 226 | .retain(|id, _| cached_ids.contains(id) || progress_ids.contains(id)); |
| 227 | |
| 228 | for (id, objective) in &running_agents { |
| 229 | app.agent_progress |
| 230 | .entry(id.clone()) |
| 231 | .or_insert_with(|| objective.clone()); |
| 232 | } |
| 233 | |
| 234 | let recovery_detail = parked_recovery_detail(app); |
| 235 | for agent in &cached_agents { |
| 236 | let meta = app |
| 237 | .agent_progress_meta |
| 238 | .entry(agent.agent_id.clone()) |
| 239 | .or_insert_with(|| AgentProgressMeta { |
| 240 | parent_run_id: agent.parent_run_id.clone(), |
| 241 | spawn_depth: agent.spawn_depth, |
| 242 | ..AgentProgressMeta::default() |
| 243 | }); |
| 244 | meta.parent_run_id = agent.parent_run_id.clone(); |
| 245 | meta.spawn_depth = agent.spawn_depth; |
| 246 | |
| 247 | let existing = meta.current_activity.clone(); |
| 248 | let parked = subagent_is_parked(agent); |
| 249 | // Parked outranks `needs_input`: the park note *is* phrased as a |
| 250 | // question, and reading it as one is exactly the bug (#5906). |
| 251 | let mut structured_status = if parked { |
| 252 | AgentCurrentActivityStatus::Parked |
| 253 | } else if agent.needs_input.is_some() { |
| 254 | AgentCurrentActivityStatus::Waiting |
| 255 | } else if let Some(worker_status) = agent.worker_status { |
| 256 | worker_status.into() |
| 257 | } else if matches!(agent.status, SubAgentStatus::Running) { |
| 258 | existing |
| 259 | .as_ref() |
| 260 | .map(|activity| activity.status) |
| 261 | .unwrap_or(AgentCurrentActivityStatus::Running) |
| 262 | } else { |
| 263 | worker_status_for_terminal_projection(&agent.status).into() |
| 264 | }; |
| 265 | if structured_status == AgentCurrentActivityStatus::Interrupted |
| 266 | && existing |
| 267 | .as_ref() |
| 268 | .is_some_and(|activity| activity.status == AgentCurrentActivityStatus::Waiting) |
| 269 | { |
| 270 | structured_status = AgentCurrentActivityStatus::Waiting; |
| 271 | } |
| 272 | |
| 273 | let detail = parked |
| 274 | .then(|| recovery_detail.clone()) |
| 275 | .or_else(|| { |
| 276 | agent |
| 277 | .needs_input |
| 278 | .as_ref() |
| 279 | .map(|needs_input| needs_input.question.clone()) |
| 280 | }) |
| 281 | .or_else(|| { |
| 282 | existing |
| 283 | .as_ref() |
| 284 | .filter(|activity| activity.status == structured_status) |
| 285 | .and_then(|activity| activity.detail.clone()) |
| 286 | }) |
| 287 | .or_else(|| agent.result.clone()); |
| 288 | let current_tool = existing |
| 289 | .as_ref() |
| 290 | .filter(|_| structured_status == AgentCurrentActivityStatus::RunningTool) |
| 291 | .and_then(|activity| activity.current_tool.clone()); |
| 292 | let step = (agent.steps_taken > 0) |
| 293 | .then_some(agent.steps_taken) |
| 294 | .or_else(|| existing.as_ref().and_then(|activity| activity.step)); |
| 295 | meta.current_activity = Some(AgentCurrentActivity::bounded( |
| 296 | structured_status, |
| 297 | detail, |
| 298 | current_tool.clone(), |
| 299 | step, |
| 300 | )); |
| 301 | meta.current_tool = current_tool; |
| 302 | } |
| 303 | |
| 304 | if running_ids.is_empty() { |
| 305 | app.agent_activity_started_at = None; |
| 306 | } else if app.agent_activity_started_at.is_none() { |
| 307 | app.agent_activity_started_at = Some(Instant::now()); |
| 308 | } |
| 309 | |
| 310 | reconcile_cards_with_snapshots(app); |
| 311 | } |
| 312 | |
| 313 | fn reconcile_terminal_subagent_card_retention(app: &mut App, now: Instant) { |
| 314 | let current_ids: std::collections::HashSet<String> = app |
| 315 | .subagent_cache |
| 316 | .iter() |
| 317 | .map(|agent| agent.agent_id.clone()) |
| 318 | .collect(); |
| 319 | app.subagent_terminal_seen_at |
| 320 | .retain(|id, _| current_ids.contains(id)); |
| 321 | |
| 322 | for agent in &app.subagent_cache { |
| 323 | if matches!(agent.status, SubAgentStatus::Running) { |
| 324 | app.subagent_terminal_seen_at.remove(&agent.agent_id); |
| 325 | } else { |
| 326 | app.subagent_terminal_seen_at |
| 327 | .entry(agent.agent_id.clone()) |
| 328 | .or_insert(now); |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | app.subagent_cache.retain(|agent| { |
| 333 | if matches!(agent.status, SubAgentStatus::Running) { |
| 334 | return true; |
| 335 | } |
| 336 | app.subagent_terminal_seen_at |
| 337 | .get(&agent.agent_id) |
| 338 | .and_then(|seen_at| now.checked_duration_since(*seen_at)) |
| 339 | .is_none_or(|age| age <= SUBAGENT_TERMINAL_CARD_TTL) |
| 340 | }); |
| 341 | |
| 342 | let mut terminal_seen: Vec<(String, Instant)> = app |
| 343 | .subagent_cache |
| 344 | .iter() |
| 345 | .filter(|agent| !matches!(agent.status, SubAgentStatus::Running)) |
| 346 | .filter_map(|agent| { |
| 347 | app.subagent_terminal_seen_at |
| 348 | .get(&agent.agent_id) |
| 349 | .map(|seen_at| (agent.agent_id.clone(), *seen_at)) |
| 350 | }) |
| 351 | .collect(); |
| 352 | terminal_seen.sort_by(|a, b| b.1.cmp(&a.1).then_with(|| a.0.cmp(&b.0))); |
| 353 | let keep_terminal_ids: std::collections::HashSet<String> = terminal_seen |
| 354 | .into_iter() |
| 355 | .take(SUBAGENT_TERMINAL_CARD_MAX_RETAINED) |
| 356 | .map(|(id, _)| id) |
| 357 | .collect(); |
| 358 | app.subagent_cache.retain(|agent| { |
| 359 | matches!(agent.status, SubAgentStatus::Running) |
| 360 | || keep_terminal_ids.contains(agent.agent_id.as_str()) |
| 361 | }); |
| 362 | |
| 363 | let kept_ids: std::collections::HashSet<String> = app |
| 364 | .subagent_cache |
| 365 | .iter() |
| 366 | .map(|agent| agent.agent_id.clone()) |
| 367 | .collect(); |
| 368 | app.subagent_terminal_seen_at |
| 369 | .retain(|id, _| kept_ids.contains(id)); |
| 370 | } |
| 371 | |
| 372 | /// Sync in-transcript card slots that still render as running against the |
| 373 | /// canonical manager snapshot statuses. A card can miss its terminal mailbox |
| 374 | /// envelope (e.g. API-timeout interruption observed only via `AgentList`), |
| 375 | /// which would otherwise leave the fanout/delegate UI counting the agent as |
| 376 | /// running indefinitely. |
| 377 | fn reconcile_cards_with_snapshots(app: &mut App) { |
| 378 | let non_running: Vec<(String, AgentLifecycle)> = app |
| 379 | .subagent_cache |
| 380 | .iter() |
| 381 | .filter_map(|agent| { |
| 382 | let lifecycle = match &agent.status { |
| 383 | SubAgentStatus::Running => return None, |
| 384 | SubAgentStatus::Interrupted(_) => AgentLifecycle::Interrupted, |
| 385 | SubAgentStatus::Completed => AgentLifecycle::Completed, |
| 386 | SubAgentStatus::Failed(_) => AgentLifecycle::Failed, |
| 387 | SubAgentStatus::Cancelled => AgentLifecycle::Cancelled, |
| 388 | SubAgentStatus::BudgetExhausted => AgentLifecycle::Failed, |
| 389 | }; |
| 390 | Some((agent.agent_id.clone(), lifecycle)) |
| 391 | }) |
| 392 | .collect(); |
| 393 | for (agent_id, lifecycle) in non_running { |
| 394 | let Some(&idx) = app.subagent_card_index.get(&agent_id) else { |
| 395 | continue; |
| 396 | }; |
| 397 | let updated = match app.history.get_mut(idx) { |
| 398 | Some(HistoryCell::SubAgent(SubAgentCell::Delegate(card))) |
| 399 | if card.agent_id == agent_id |
| 400 | && matches!( |
| 401 | card.status, |
| 402 | AgentLifecycle::Pending | AgentLifecycle::Running |
| 403 | ) => |
| 404 | { |
| 405 | card.status = lifecycle; |
| 406 | true |
| 407 | } |
| 408 | Some(HistoryCell::SubAgent(SubAgentCell::Fanout(card))) => { |
| 409 | match card.workers.iter_mut().find(|slot| { |
| 410 | slot.agent_id == agent_id |
| 411 | && matches!( |
| 412 | slot.status, |
| 413 | AgentLifecycle::Pending | AgentLifecycle::Running |
| 414 | ) |
| 415 | }) { |
| 416 | Some(slot) => { |
| 417 | slot.status = lifecycle; |
| 418 | true |
| 419 | } |
| 420 | None => false, |
| 421 | } |
| 422 | } |
| 423 | _ => false, |
| 424 | }; |
| 425 | if updated { |
| 426 | app.bump_history_cell(idx); |
| 427 | } |
| 428 | } |
| 429 | } |
| 430 | |
| 431 | fn subagent_status_rank(status: &SubAgentStatus) -> u8 { |
| 432 | match status { |
| 433 | SubAgentStatus::Running => 0, |
| 434 | SubAgentStatus::Interrupted(_) => 1, |
| 435 | SubAgentStatus::Failed(_) => 2, |
| 436 | SubAgentStatus::Completed => 3, |
| 437 | SubAgentStatus::Cancelled => 4, |
| 438 | SubAgentStatus::BudgetExhausted => 2, |
| 439 | } |
| 440 | } |
| 441 | |
| 442 | pub(super) fn sort_subagents_in_place(agents: &mut [SubAgentResult]) { |
| 443 | agents.sort_by(|a, b| { |
| 444 | subagent_status_rank(&a.status) |
| 445 | .cmp(&subagent_status_rank(&b.status)) |
| 446 | .then_with(|| a.agent_type.as_str().cmp(b.agent_type.as_str())) |
| 447 | .then_with(|| a.agent_id.cmp(&b.agent_id)) |
| 448 | }); |
| 449 | } |
| 450 | |
| 451 | pub(super) fn subagent_message_refreshes_workspace_context(message: &MailboxMessage) -> bool { |
| 452 | matches!( |
| 453 | message, |
| 454 | MailboxMessage::ToolCallCompleted { tool_name, .. } |
| 455 | if refreshes_workspace_context_on_completion(tool_name) |
| 456 | ) |
| 457 | } |
| 458 | |
| 459 | /// Route a `MailboxMessage` envelope to the matching in-transcript card, |
| 460 | /// allocating a `DelegateCard` or `FanoutCard` on first sight (issue #128). |
| 461 | pub(super) fn handle_subagent_mailbox_for_turn( |
| 462 | app: &mut App, |
| 463 | turn_id: &str, |
| 464 | seq: u64, |
| 465 | message: &MailboxMessage, |
| 466 | ) -> bool { |
| 467 | // Accumulate sub-agent token costs for the real-time footer counter (#166). |
| 468 | if let MailboxMessage::TokenUsage { |
| 469 | agent_id: _, |
| 470 | source_id, |
| 471 | route, |
| 472 | usage, |
| 473 | } = message |
| 474 | { |
| 475 | // The child's own route truth always wins and is never guessed from |
| 476 | // provider identity: `route` is the immutable envelope its client was |
| 477 | // frozen with at construction, so its billing mode, billing surface |
| 478 | // and endpoint fingerprint are the child's dispatch receipt. A child |
| 479 | // whose route froze as Unknown stays Unknown. |
| 480 | // |
| 481 | // Sub-agent spend joins the parent total, so it also joins the |
| 482 | // completeness counters `/cost` reports against that total. |
| 483 | let stable_source = if source_id.trim().is_empty() { |
| 484 | format!("mailbox:{turn_id}:{seq}") |
| 485 | } else { |
| 486 | source_id.clone() |
| 487 | }; |
| 488 | let source_fingerprint = crate::cost_status::usage_source_fingerprint(&stable_source); |
| 489 | if crate::cost_status::usage_source_seen(&stable_source) { |
| 490 | // The runtime sink runs synchronously before mailbox publication. |
| 491 | // Fold its identity and money into App together, before a following |
| 492 | // TurnComplete can snapshot either half. A post-seal response has |
| 493 | // no mailbox event and is swept by the event loop instead. |
| 494 | let first_live_delivery = !app |
| 495 | .session |
| 496 | .subagent_usage_sources |
| 497 | .contains(&source_fingerprint); |
| 498 | let pending_runtime_cost = crate::cost_status::drain(); |
| 499 | if !pending_runtime_cost.is_empty() { |
| 500 | app.absorb_pending_background_cost(&pending_runtime_cost); |
| 501 | } |
| 502 | if first_live_delivery |
| 503 | && app |
| 504 | .session |
| 505 | .subagent_usage_sources |
| 506 | .contains(&source_fingerprint) |
| 507 | { |
| 508 | record_agent_current_activity(app, message); |
| 509 | } |
| 510 | return false; |
| 511 | } |
| 512 | if app |
| 513 | .session |
| 514 | .subagent_usage_sources |
| 515 | .insert(source_fingerprint) |
| 516 | { |
| 517 | // Update the Work-row receipt under the same replay guard as |
| 518 | // cost. A mailbox replay must not make either total grow twice. |
| 519 | // Usage confirms or replaces the child's frozen spawn model; the |
| 520 | // parent's configured/default route is never child evidence. |
| 521 | record_agent_current_activity(app, message); |
| 522 | let audit = route.audit(usage); |
| 523 | app.record_turn_cost_audit(&audit); |
| 524 | app.record_turn_cost_route_receipt(route.receipt(&audit)); |
| 525 | if let Some(cost) = audit.estimate { |
| 526 | app.accrue_subagent_cost_estimate(cost); |
| 527 | } |
| 528 | } |
| 529 | // No transcript card changed. The event-loop redraw still repaints |
| 530 | // the Work row and footer from the updated receipt state. |
| 531 | return false; |
| 532 | } |
| 533 | |
| 534 | // Resolve (or allocate) the target cell for this envelope. ChildSpawned |
| 535 | // is special — it always belongs to the active fanout card if one |
| 536 | // exists; otherwise it seeds a new one. |
| 537 | let display_message = bounded_mailbox_message(message); |
| 538 | let agent_id = display_message.agent_id().to_string(); |
| 539 | record_agent_current_activity(app, message); |
| 540 | if subagent_message_refreshes_workspace_context(message) { |
| 541 | workspace_context::refresh_now(app, Instant::now()); |
| 542 | } |
| 543 | |
| 544 | if matches!(message, MailboxMessage::ChildSpawned { .. }) |
| 545 | && let Some(idx) = app.last_fanout_card_index |
| 546 | && let Some(HistoryCell::SubAgent(SubAgentCell::Fanout(card))) = app.history.get_mut(idx) |
| 547 | { |
| 548 | let updated = apply_to_fanout(card, &display_message); |
| 549 | app.subagent_card_index.insert(agent_id, idx); |
| 550 | if updated { |
| 551 | app.bump_history_cell(idx); |
| 552 | } |
| 553 | return updated; |
| 554 | } |
| 555 | |
| 556 | // Existing card for this agent_id? Mutate in place. |
| 557 | if let Some(&idx) = app.subagent_card_index.get(&agent_id) { |
| 558 | let updated = match app.history.get_mut(idx) { |
| 559 | Some(HistoryCell::SubAgent(SubAgentCell::Delegate(card))) => { |
| 560 | apply_to_delegate(card, &display_message) |
| 561 | } |
| 562 | Some(HistoryCell::SubAgent(SubAgentCell::Fanout(card))) => { |
| 563 | apply_to_fanout(card, &display_message) |
| 564 | } |
| 565 | _ => false, |
| 566 | }; |
| 567 | if updated { |
| 568 | // idx is already in scope from the outer |
| 569 | // `if let Some(&idx) = app.subagent_card_index.get(&agent_id)`. |
| 570 | app.bump_history_cell(idx); |
| 571 | } |
| 572 | return updated; |
| 573 | } |
| 574 | |
| 575 | // No existing card — only `Started` reasonably opens one. Anything else |
| 576 | // for an unknown agent_id is dropped (likely arrived after the cell was |
| 577 | // cleared, e.g. session-resume edge cases). |
| 578 | let agent_type = match &display_message { |
| 579 | MailboxMessage::Started { agent_type, .. } => agent_type.clone(), |
| 580 | MailboxMessage::Completed { .. } |
| 581 | | MailboxMessage::Failed { .. } |
| 582 | | MailboxMessage::Interrupted { .. } |
| 583 | | MailboxMessage::Cancelled { .. } => "unknown".to_string(), |
| 584 | _ => return false, |
| 585 | }; |
| 586 | |
| 587 | let dispatch_kind = app.pending_subagent_dispatch.as_deref(); |
| 588 | let is_fanout = matches!(dispatch_kind, Some("rlm_open" | "rlm_eval" | "rlm")); |
| 589 | |
| 590 | if is_fanout { |
| 591 | // Reuse the active fanout card for sibling spawns; otherwise create |
| 592 | // one anchored at this position so subsequent siblings join it. |
| 593 | if let Some(idx) = app.last_fanout_card_index |
| 594 | && let Some(HistoryCell::SubAgent(SubAgentCell::Fanout(card))) = |
| 595 | app.history.get_mut(idx) |
| 596 | { |
| 597 | let updated = card.claim_pending_worker(&agent_id, AgentLifecycle::Running); |
| 598 | app.subagent_card_index.insert(agent_id, idx); |
| 599 | if updated { |
| 600 | app.bump_history_cell(idx); |
| 601 | } |
| 602 | updated |
| 603 | } else { |
| 604 | let mut card = FanoutCard::new(dispatch_kind.unwrap_or("rlm_eval").to_string()); |
| 605 | card.upsert_worker(&agent_id, AgentLifecycle::Running); |
| 606 | app.add_message(HistoryCell::SubAgent(SubAgentCell::Fanout(card))); |
| 607 | let idx = app.history.len().saturating_sub(1); |
| 608 | app.last_fanout_card_index = Some(idx); |
| 609 | app.subagent_card_index.insert(agent_id, idx); |
| 610 | app.bump_history_cell(idx); |
| 611 | true |
| 612 | } |
| 613 | } else { |
| 614 | let mut card = DelegateCard::new(agent_id.clone(), agent_type.clone()); |
| 615 | apply_to_delegate(&mut card, &display_message); |
| 616 | app.add_message(HistoryCell::SubAgent(SubAgentCell::Delegate(card))); |
| 617 | let idx = app.history.len().saturating_sub(1); |
| 618 | app.subagent_card_index.insert(agent_id.clone(), idx); |
| 619 | // Single delegate consumes the pending dispatch label so a follow-on |
| 620 | // tool call doesn't accidentally inherit it. |
| 621 | app.pending_subagent_dispatch = None; |
| 622 | // idx was just inserted on the line above — no need to re-query. |
| 623 | app.bump_history_cell(idx); |
| 624 | true |
| 625 | } |
| 626 | } |
| 627 | |
| 628 | #[cfg(test)] |
| 629 | pub(super) fn handle_subagent_mailbox(app: &mut App, seq: u64, message: &MailboxMessage) -> bool { |
| 630 | handle_subagent_mailbox_for_turn(app, "test-turn", seq, message) |
| 631 | } |
| 632 | |
| 633 | fn bounded_mailbox_message(message: &MailboxMessage) -> MailboxMessage { |
| 634 | match message { |
| 635 | MailboxMessage::Progress { agent_id, status } => MailboxMessage::Progress { |
| 636 | agent_id: agent_id.clone(), |
| 637 | status: bound_agent_activity_text(status), |
| 638 | }, |
| 639 | MailboxMessage::ToolCallStarted { |
| 640 | agent_id, |
| 641 | tool_name, |
| 642 | step, |
| 643 | } => MailboxMessage::ToolCallStarted { |
| 644 | agent_id: agent_id.clone(), |
| 645 | tool_name: bound_agent_activity_text(subagent_progress_tool_display_name(tool_name)), |
| 646 | step: *step, |
| 647 | }, |
| 648 | MailboxMessage::ToolCallCompleted { |
| 649 | agent_id, |
| 650 | tool_name, |
| 651 | step, |
| 652 | ok, |
| 653 | } => MailboxMessage::ToolCallCompleted { |
| 654 | agent_id: agent_id.clone(), |
| 655 | tool_name: bound_agent_activity_text(subagent_progress_tool_display_name(tool_name)), |
| 656 | step: *step, |
| 657 | ok: *ok, |
| 658 | }, |
| 659 | MailboxMessage::Completed { agent_id, summary } => MailboxMessage::Completed { |
| 660 | agent_id: agent_id.clone(), |
| 661 | summary: bound_agent_activity_text(summary), |
| 662 | }, |
| 663 | MailboxMessage::Failed { agent_id, error } => MailboxMessage::Failed { |
| 664 | agent_id: agent_id.clone(), |
| 665 | error: bound_agent_activity_text(error), |
| 666 | }, |
| 667 | MailboxMessage::Interrupted { agent_id, reason } => MailboxMessage::Interrupted { |
| 668 | agent_id: agent_id.clone(), |
| 669 | reason: bound_agent_activity_text(reason), |
| 670 | }, |
| 671 | // Item text is model-authored and reaches the transcript, so it gets |
| 672 | // the same redaction/bounding every other displayed child string gets. |
| 673 | // Ids, statuses, and counts are preserved exactly — bounding must not |
| 674 | // change what the ledger says. |
| 675 | MailboxMessage::WorkState { agent_id, todo } => MailboxMessage::WorkState { |
| 676 | agent_id: agent_id.clone(), |
| 677 | todo: crate::tools::todo::TodoListSnapshot { |
| 678 | items: todo |
| 679 | .items |
| 680 | .iter() |
| 681 | .map(|item| crate::tools::todo::TodoItem { |
| 682 | content: bound_agent_activity_text(&item.content), |
| 683 | ..item.clone() |
| 684 | }) |
| 685 | .collect(), |
| 686 | ..todo.clone() |
| 687 | }, |
| 688 | }, |
| 689 | _ => message.clone(), |
| 690 | } |
| 691 | } |
| 692 | |
| 693 | fn record_agent_current_activity(app: &mut App, message: &MailboxMessage) { |
| 694 | let agent_id = message.agent_id().to_string(); |
| 695 | let meta = app.agent_progress_meta.entry(agent_id).or_default(); |
| 696 | if let MailboxMessage::TokenUsage { route, usage, .. } = message { |
| 697 | // The child's own used-token tally (input + output), matching the |
| 698 | // worker budget's `usage_total_tokens`. Counting only completions made |
| 699 | // the strip look "stuck" on tiny numbers while the child was burning |
| 700 | // context. Absent until a real envelope lands, so an agent with no |
| 701 | // reported usage shows no number instead of a zero. |
| 702 | let turn_total = |
| 703 | u64::from(usage.input_tokens).saturating_add(u64::from(usage.output_tokens)); |
| 704 | meta.received_tokens = Some(meta.received_tokens.unwrap_or(0).saturating_add(turn_total)); |
| 705 | meta.resolved_provider = Some(route.provider.as_str().to_string()); |
| 706 | meta.resolved_model = Some(bound_agent_activity_text( |
| 707 | &crate::cost_status::sanitize_persisted_route_label(&route.model), |
| 708 | )) |
| 709 | .filter(|model| !model.trim().is_empty()); |
| 710 | return; |
| 711 | } |
| 712 | if let MailboxMessage::WorkState { todo, .. } = message { |
| 713 | // Work state is a separate fact from what the agent is doing right |
| 714 | // now — publishing a ledger update must not invent an activity |
| 715 | // transition. It does update the remaining-to-do chip for the strip. |
| 716 | if todo.is_empty() { |
| 717 | meta.todos_remaining = None; |
| 718 | } else { |
| 719 | let remaining = todo |
| 720 | .items |
| 721 | .iter() |
| 722 | .filter(|item| !item.status.is_settled()) |
| 723 | .count(); |
| 724 | meta.todos_remaining = Some(u32::try_from(remaining).unwrap_or(u32::MAX)); |
| 725 | } |
| 726 | return; |
| 727 | } |
| 728 | if let MailboxMessage::ToolCallCompleted { |
| 729 | tool_name, |
| 730 | step, |
| 731 | ok, |
| 732 | .. |
| 733 | } = message |
| 734 | { |
| 735 | if meta.recent_actions.len() == MAX_AGENT_RECENT_ACTIONS { |
| 736 | meta.recent_actions.pop_front(); |
| 737 | } |
| 738 | meta.recent_actions.push_back(AgentRecentAction::bounded( |
| 739 | subagent_progress_tool_display_name(tool_name), |
| 740 | *step, |
| 741 | *ok, |
| 742 | )); |
| 743 | } |
| 744 | let previous = meta.current_activity.clone(); |
| 745 | |
| 746 | let (status, detail, current_tool, step) = match message { |
| 747 | MailboxMessage::Started { agent_type, .. } => ( |
| 748 | AgentCurrentActivityStatus::Running, |
| 749 | Some(format!("started {agent_type}")), |
| 750 | None, |
| 751 | None, |
| 752 | ), |
| 753 | MailboxMessage::Progress { status, .. } => ( |
| 754 | previous |
| 755 | .as_ref() |
| 756 | .map(|activity| activity.status) |
| 757 | .unwrap_or(AgentCurrentActivityStatus::Running), |
| 758 | Some(status.clone()), |
| 759 | previous |
| 760 | .as_ref() |
| 761 | .and_then(|activity| activity.current_tool.clone()), |
| 762 | previous.as_ref().and_then(|activity| activity.step), |
| 763 | ), |
| 764 | MailboxMessage::ToolCallStarted { |
| 765 | tool_name, step, .. |
| 766 | } => ( |
| 767 | AgentCurrentActivityStatus::RunningTool, |
| 768 | None, |
| 769 | Some(subagent_progress_tool_display_name(tool_name).to_string()), |
| 770 | Some(*step), |
| 771 | ), |
| 772 | MailboxMessage::ToolCallCompleted { |
| 773 | tool_name, |
| 774 | step, |
| 775 | ok, |
| 776 | .. |
| 777 | } => ( |
| 778 | AgentCurrentActivityStatus::Running, |
| 779 | Some(format!( |
| 780 | "{} {}", |
| 781 | subagent_progress_tool_display_name(tool_name), |
| 782 | if *ok { "completed" } else { "failed" } |
| 783 | )), |
| 784 | None, |
| 785 | Some(*step), |
| 786 | ), |
| 787 | MailboxMessage::ChildSpawned { parent_id, .. } => ( |
| 788 | AgentCurrentActivityStatus::Starting, |
| 789 | Some(format!("spawned by {parent_id}")), |
| 790 | None, |
| 791 | None, |
| 792 | ), |
| 793 | MailboxMessage::Completed { summary, .. } => ( |
| 794 | AgentCurrentActivityStatus::Done, |
| 795 | Some(summary.clone()), |
| 796 | None, |
| 797 | previous.as_ref().and_then(|activity| activity.step), |
| 798 | ), |
| 799 | MailboxMessage::Failed { error, .. } => ( |
| 800 | AgentCurrentActivityStatus::Failed, |
| 801 | Some(error.clone()), |
| 802 | None, |
| 803 | previous.as_ref().and_then(|activity| activity.step), |
| 804 | ), |
| 805 | MailboxMessage::Interrupted { reason, .. } => ( |
| 806 | AgentCurrentActivityStatus::Waiting, |
| 807 | Some(reason.clone()), |
| 808 | None, |
| 809 | previous.as_ref().and_then(|activity| activity.step), |
| 810 | ), |
| 811 | MailboxMessage::Cancelled { .. } => ( |
| 812 | AgentCurrentActivityStatus::Canceled, |
| 813 | None, |
| 814 | None, |
| 815 | previous.as_ref().and_then(|activity| activity.step), |
| 816 | ), |
| 817 | MailboxMessage::TokenUsage { .. } => unreachable!("token usage handled above"), |
| 818 | MailboxMessage::WorkState { .. } => unreachable!("work state handled above"), |
| 819 | }; |
| 820 | |
| 821 | meta.current_activity = Some(AgentCurrentActivity::bounded( |
| 822 | status, |
| 823 | detail, |
| 824 | current_tool.clone(), |
| 825 | step, |
| 826 | )); |
| 827 | meta.current_tool = current_tool; |
| 828 | if let MailboxMessage::ToolCallCompleted { |
| 829 | tool_name, |
| 830 | ok: true, |
| 831 | .. |
| 832 | } = message |
| 833 | && is_file_mutation_tool(tool_name) |
| 834 | { |
| 835 | meta.files_touched = meta.files_touched.saturating_add(1); |
| 836 | } |
| 837 | } |
| 838 | |
| 839 | fn is_file_mutation_tool(name: &str) -> bool { |
| 840 | matches!( |
| 841 | name, |
| 842 | "write_file" | "edit_file" | "apply_patch" | "fim_edit" | "Write" | "Edit" |
| 843 | ) |
| 844 | } |
| 845 | |
| 846 | pub(super) fn task_mode_label(mode: AppMode) -> &'static str { |
| 847 | mode.as_setting() |
| 848 | } |
| 849 | |
| 850 | pub(super) fn task_summary_to_panel_entry(summary: TaskSummary) -> TaskPanelEntry { |
| 851 | TaskPanelEntry { |
| 852 | id: summary.id, |
| 853 | status: task_status_label(summary.status).to_string(), |
| 854 | prompt_summary: summary.prompt_summary, |
| 855 | duration_ms: summary.duration_ms, |
| 856 | kind: TaskPanelEntryKind::Background, |
| 857 | stale: false, |
| 858 | elapsed_since_output_ms: None, |
| 859 | owner_agent_id: None, |
| 860 | owner_agent_name: None, |
| 861 | current_tool: None, |
| 862 | role: None, |
| 863 | files_touched: 0, |
| 864 | } |
| 865 | } |
| 866 | |
| 867 | fn task_status_label(status: TaskStatus) -> &'static str { |
| 868 | match status { |
| 869 | TaskStatus::Queued => "queued", |
| 870 | TaskStatus::Running => "running", |
| 871 | TaskStatus::Completed => "completed", |
| 872 | TaskStatus::Failed => "failed", |
| 873 | TaskStatus::Canceled => "canceled", |
| 874 | } |
| 875 | } |
| 876 | |
| 877 | pub(super) fn format_task_list(tasks: &[TaskSummary]) -> String { |
| 878 | if tasks.is_empty() { |
| 879 | return "No tasks found.".to_string(); |
| 880 | } |
| 881 | |
| 882 | let show_session = tasks.iter().any(|task| task.owner_session_id.is_some()); |
| 883 | let mut lines = vec![format!("Tasks ({})", tasks.len())]; |
| 884 | // Build headers with the same format strings as the rows so the ID |
| 885 | // column (21-char `task_` ids) can never drift out of alignment again. |
| 886 | if show_session { |
| 887 | lines.push(format!( |
| 888 | "{:<21} {:<9} {:<12} {:>8} {}", |
| 889 | "ID", "Status", "Session", "Time", "Title" |
| 890 | )); |
| 891 | } else { |
| 892 | lines.push(format!( |
| 893 | "{:<21} {:<9} {:>8} {}", |
| 894 | "ID", "Status", "Time", "Title" |
| 895 | )); |
| 896 | } |
| 897 | lines.push("------------------------------------------------------------".to_string()); |
| 898 | for task in tasks { |
| 899 | let duration = task |
| 900 | .duration_ms |
| 901 | .map(crate::elapsed::format_elapsed_ms) |
| 902 | .unwrap_or_else(|| "-".to_string()); |
| 903 | let owner_session = task.owner_session_id.as_deref().unwrap_or("-"); |
| 904 | let owner_session = if owner_session.chars().count() > 12 { |
| 905 | format!("{}…", owner_session.chars().take(11).collect::<String>()) |
| 906 | } else { |
| 907 | owner_session.to_string() |
| 908 | }; |
| 909 | if show_session { |
| 910 | lines.push(format!( |
| 911 | "{:<21} {:<9} {:<12} {:>8} {}", |
| 912 | task.id, |
| 913 | task_status_label(task.status), |
| 914 | owner_session, |
| 915 | duration, |
| 916 | task.prompt_summary |
| 917 | )); |
| 918 | } else { |
| 919 | lines.push(format!( |
| 920 | "{:<21} {:<9} {:>8} {}", |
| 921 | task.id, |
| 922 | task_status_label(task.status), |
| 923 | duration, |
| 924 | task.prompt_summary |
| 925 | )); |
| 926 | } |
| 927 | } |
| 928 | lines.push("Use /task show <id> for timeline details.".to_string()); |
| 929 | lines.join("\n") |
| 930 | } |
| 931 | |
| 932 | pub(super) fn open_task_pager(app: &mut App, task: &TaskRecord) { |
| 933 | let width = app |
| 934 | .viewport |
| 935 | .last_transcript_area |
| 936 | .map(|area| area.width) |
| 937 | .unwrap_or(100) |
| 938 | .saturating_sub(4); |
| 939 | app.view_stack.push(PagerView::from_text( |
| 940 | format!("Task {}", task.id), |
| 941 | &format_task_detail(task), |
| 942 | width.max(60), |
| 943 | )); |
| 944 | } |
| 945 | |
| 946 | fn format_task_detail(task: &TaskRecord) -> String { |
| 947 | let mut lines = Vec::new(); |
| 948 | lines.push(format!("Task: {}", task.id)); |
| 949 | lines.push(format!("Status: {}", task_status_label(task.status))); |
| 950 | lines.push(format!("Mode: {}", task.mode)); |
| 951 | lines.push(format!("Model: {}", task.model)); |
| 952 | lines.push(format!( |
| 953 | "Workspace: {}", |
| 954 | crate::utils::display_path(&task.workspace) |
| 955 | )); |
| 956 | if let Some(owner_session_id) = task.owner_session_id.as_deref() { |
| 957 | lines.push(format!("Owning Session: {owner_session_id}")); |
| 958 | } |
| 959 | if let Some(thread_id) = task.thread_id.as_ref() { |
| 960 | lines.push(format!("Runtime Thread: {thread_id}")); |
| 961 | } |
| 962 | if let Some(turn_id) = task.turn_id.as_ref() { |
| 963 | lines.push(format!("Runtime Turn: {turn_id}")); |
| 964 | } |
| 965 | if task.runtime_event_count > 0 { |
| 966 | lines.push(format!("Runtime Events: {}", task.runtime_event_count)); |
| 967 | } |
| 968 | lines.push(format!("Created: {}", task.created_at)); |
| 969 | if let Some(started_at) = task.started_at { |
| 970 | lines.push(format!("Started: {started_at}")); |
| 971 | } |
| 972 | if let Some(ended_at) = task.ended_at { |
| 973 | lines.push(format!("Ended: {ended_at}")); |
| 974 | } |
| 975 | if let Some(duration) = task.duration_ms { |
| 976 | lines.push(format!( |
| 977 | "Duration: {}", |
| 978 | crate::elapsed::format_elapsed_ms(duration) |
| 979 | )); |
| 980 | } |
| 981 | lines.push(String::new()); |
| 982 | lines.push("Prompt:".to_string()); |
| 983 | lines.push(task.prompt.clone()); |
| 984 | |
| 985 | if let Some(summary) = task.result_summary.as_ref() { |
| 986 | lines.push(String::new()); |
| 987 | lines.push("Result Summary:".to_string()); |
| 988 | lines.push(summary.clone()); |
| 989 | } |
| 990 | if let Some(path) = task.result_detail_path.as_ref() { |
| 991 | lines.push(format!("Result Artifact: {}", path.display())); |
| 992 | } |
| 993 | if let Some(error) = task.error.as_ref() { |
| 994 | lines.push(String::new()); |
| 995 | lines.push(format!("Error: {error}")); |
| 996 | } |
| 997 | |
| 998 | lines.push(String::new()); |
| 999 | lines.push("Tool Calls:".to_string()); |
| 1000 | if task.tool_calls.is_empty() { |
| 1001 | lines.push("- (none)".to_string()); |
| 1002 | } else { |
| 1003 | for tool in &task.tool_calls { |
| 1004 | let status = match tool.status { |
| 1005 | crate::task_manager::TaskToolStatus::Running => "running", |
| 1006 | crate::task_manager::TaskToolStatus::Success => "success", |
| 1007 | crate::task_manager::TaskToolStatus::Failed => "failed", |
| 1008 | crate::task_manager::TaskToolStatus::Canceled => "canceled", |
| 1009 | }; |
| 1010 | let mut line = format!( |
| 1011 | "- {} [{}] {}", |
| 1012 | tool.name, |
| 1013 | status, |
| 1014 | tool.output_summary.as_deref().unwrap_or("(no summary)") |
| 1015 | ); |
| 1016 | if let Some(duration) = tool.duration_ms { |
| 1017 | line.push_str(&format!(" ({:.2}s)", duration as f64 / 1000.0)); |
| 1018 | } |
| 1019 | lines.push(line); |
| 1020 | if let Some(path) = tool.detail_path.as_ref() { |
| 1021 | lines.push(format!(" detail: {}", path.display())); |
| 1022 | } |
| 1023 | if let Some(path) = tool.patch_ref.as_ref() { |
| 1024 | lines.push(format!(" patch: {}", path.display())); |
| 1025 | } |
| 1026 | } |
| 1027 | } |
| 1028 | |
| 1029 | lines.push(String::new()); |
| 1030 | lines.push("Timeline:".to_string()); |
| 1031 | if task.timeline.is_empty() { |
| 1032 | lines.push("- (none)".to_string()); |
| 1033 | } else { |
| 1034 | for entry in &task.timeline { |
| 1035 | lines.push(format!( |
| 1036 | "- [{}] {}: {}", |
| 1037 | entry.timestamp, entry.kind, entry.summary |
| 1038 | )); |
| 1039 | if let Some(path) = entry.detail_path.as_ref() { |
| 1040 | lines.push(format!(" detail: {}", path.display())); |
| 1041 | } |
| 1042 | } |
| 1043 | } |
| 1044 | |
| 1045 | lines.join("\n") |
| 1046 | } |
| 1047 | |
| 1048 | #[cfg(test)] |
| 1049 | mod tests { |
| 1050 | use super::*; |
| 1051 | use crate::config::Config; |
| 1052 | use crate::task_manager::{TaskStatus, TaskSummary}; |
| 1053 | use crate::tools::subagent::{FleetRole, SubAgentAssignment}; |
| 1054 | use crate::tui::app::{InitialInput, TuiOptions}; |
| 1055 | use crate::tui::widgets::agent_card::AgentLifecycle; |
| 1056 | use chrono::Utc; |
| 1057 | use std::path::PathBuf; |
| 1058 | |
| 1059 | fn test_options() -> TuiOptions { |
| 1060 | TuiOptions { |
| 1061 | model: "test-model".to_string(), |
| 1062 | allow_shell: true, |
| 1063 | max_subagents: 4, |
| 1064 | start_in_agent_mode: true, |
| 1065 | initial_input: None::<InitialInput>, |
| 1066 | startup_notice: None, |
| 1067 | ..crate::test_support::test_tui_options(PathBuf::from(".")) |
| 1068 | } |
| 1069 | } |
| 1070 | |
| 1071 | fn test_route( |
| 1072 | provider: crate::config::ApiProvider, |
| 1073 | model: &str, |
| 1074 | ) -> crate::cost_status::EffectiveRouteEnvelope { |
| 1075 | crate::cost_status::EffectiveRouteEnvelope::capture( |
| 1076 | None, |
| 1077 | provider, |
| 1078 | provider.as_str(), |
| 1079 | model, |
| 1080 | Some(provider.default_base_url()), |
| 1081 | Utc::now(), |
| 1082 | ) |
| 1083 | } |
| 1084 | |
| 1085 | fn task_summary(id: &str, status: TaskStatus, duration_ms: Option<u64>) -> TaskSummary { |
| 1086 | TaskSummary { |
| 1087 | execution_binding_known: true, |
| 1088 | id: id.to_string(), |
| 1089 | status, |
| 1090 | prompt_summary: "Fix task list output".to_string(), |
| 1091 | name: None, |
| 1092 | model: "deepseek-v4-pro".to_string(), |
| 1093 | model_provider: None, |
| 1094 | model_provider_id: None, |
| 1095 | mode: "agent".to_string(), |
| 1096 | workspace: PathBuf::from("/tmp"), |
| 1097 | created_at: Utc::now(), |
| 1098 | started_at: None, |
| 1099 | ended_at: None, |
| 1100 | duration_ms, |
| 1101 | lifecycle_seq: 1, |
| 1102 | error: None, |
| 1103 | terminal_reason: None, |
| 1104 | thread_id: None, |
| 1105 | turn_id: None, |
| 1106 | owner_session_id: None, |
| 1107 | } |
| 1108 | } |
| 1109 | |
| 1110 | fn subagent_result(id: &str, status: SubAgentStatus) -> SubAgentResult { |
| 1111 | SubAgentResult { |
| 1112 | usage: None, |
| 1113 | name: id.to_string(), |
| 1114 | agent_id: id.to_string(), |
| 1115 | context_mode: "fresh".to_string(), |
| 1116 | fork_context: false, |
| 1117 | workspace: None, |
| 1118 | git_branch: None, |
| 1119 | agent_type: FleetRole::Worker, |
| 1120 | assignment: SubAgentAssignment { |
| 1121 | objective: format!("objective-{id}"), |
| 1122 | role: Some("worker".to_string()), |
| 1123 | }, |
| 1124 | model: "deepseek-v4-flash".to_string(), |
| 1125 | nickname: None, |
| 1126 | status, |
| 1127 | worker_status: None, |
| 1128 | runtime_permissions: None, |
| 1129 | parent_run_id: None, |
| 1130 | spawn_depth: 0, |
| 1131 | child_route: None, |
| 1132 | result: None, |
| 1133 | steps_taken: 0, |
| 1134 | checkpoint: None, |
| 1135 | needs_input: None, |
| 1136 | duration_ms: 0, |
| 1137 | started_at: None, |
| 1138 | from_prior_session: false, |
| 1139 | } |
| 1140 | } |
| 1141 | |
| 1142 | #[test] |
| 1143 | fn task_list_includes_title_header_and_time_column() { |
| 1144 | let output = format_task_list(&[ |
| 1145 | task_summary("task_12345678", TaskStatus::Running, None), |
| 1146 | task_summary("task_abcdef12", TaskStatus::Completed, Some(1234)), |
| 1147 | ]); |
| 1148 | |
| 1149 | assert!(output.contains(&format!( |
| 1150 | "{:<21} {:<9} {:>8} {}", |
| 1151 | "ID", "Status", "Time", "Title" |
| 1152 | ))); |
| 1153 | assert!(output.contains(&format!( |
| 1154 | "{:<21} {:<9} {:>8} {}", |
| 1155 | "task_12345678", "running", "-", "Fix task list output" |
| 1156 | ))); |
| 1157 | assert!(output.contains(&format!( |
| 1158 | "{:<21} {:<9} {:>8} {}", |
| 1159 | "task_abcdef12", "completed", "1s", "Fix task list output" |
| 1160 | ))); |
| 1161 | } |
| 1162 | |
| 1163 | #[test] |
| 1164 | fn task_list_shows_owner_session_when_present() { |
| 1165 | let mut task = task_summary("task_owned", TaskStatus::Running, None); |
| 1166 | task.owner_session_id = Some("session-123456".to_string()); |
| 1167 | |
| 1168 | let output = format_task_list(&[task]); |
| 1169 | |
| 1170 | assert!(output.contains(&format!( |
| 1171 | "{:<21} {:<9} {:<12} {:>8} {}", |
| 1172 | "ID", "Status", "Session", "Time", "Title" |
| 1173 | ))); |
| 1174 | // Owner ids are truncated to 11 chars + '…' so the rendered value |
| 1175 | // stays inside the 12-wide Session column and cannot drift the |
| 1176 | // remaining columns out of alignment. |
| 1177 | assert!(output.contains("session-123…"), "{output}"); |
| 1178 | } |
| 1179 | |
| 1180 | #[test] |
| 1181 | fn mailbox_progress_reports_transcript_change_only_for_visible_card_updates() { |
| 1182 | let mut app = App::new(test_options(), &Config::default()); |
| 1183 | let started = MailboxMessage::started("agent_live", FleetRole::Worker); |
| 1184 | assert!( |
| 1185 | handle_subagent_mailbox(&mut app, 1, &started), |
| 1186 | "first started envelope creates a visible card" |
| 1187 | ); |
| 1188 | |
| 1189 | let progress = |
| 1190 | MailboxMessage::progress("agent_live", "step 1/100: requesting model response"); |
| 1191 | assert!( |
| 1192 | !handle_subagent_mailbox(&mut app, 2, &progress), |
| 1193 | "low-signal progress for an already-running card is a no-op" |
| 1194 | ); |
| 1195 | |
| 1196 | let tool = MailboxMessage::ToolCallStarted { |
| 1197 | agent_id: "agent_live".to_string(), |
| 1198 | tool_name: "read_file".to_string(), |
| 1199 | step: 1, |
| 1200 | }; |
| 1201 | assert!( |
| 1202 | handle_subagent_mailbox(&mut app, 3, &tool), |
| 1203 | "tool progress still updates the visible transcript card" |
| 1204 | ); |
| 1205 | assert_eq!( |
| 1206 | app.agent_progress_meta["agent_live"] |
| 1207 | .current_tool |
| 1208 | .as_deref(), |
| 1209 | Some("read_file") |
| 1210 | ); |
| 1211 | |
| 1212 | let completed = MailboxMessage::ToolCallCompleted { |
| 1213 | agent_id: "agent_live".to_string(), |
| 1214 | tool_name: "read_file".to_string(), |
| 1215 | step: 1, |
| 1216 | ok: true, |
| 1217 | }; |
| 1218 | assert!(handle_subagent_mailbox(&mut app, 4, &completed)); |
| 1219 | assert_eq!(app.agent_progress_meta["agent_live"].current_tool, None); |
| 1220 | |
| 1221 | let wrote = MailboxMessage::ToolCallCompleted { |
| 1222 | agent_id: "agent_live".to_string(), |
| 1223 | tool_name: "apply_patch".to_string(), |
| 1224 | step: 2, |
| 1225 | ok: true, |
| 1226 | }; |
| 1227 | assert!(handle_subagent_mailbox(&mut app, 5, &wrote)); |
| 1228 | assert_eq!(app.agent_progress_meta["agent_live"].files_touched, 1); |
| 1229 | } |
| 1230 | |
| 1231 | #[test] |
| 1232 | fn canonical_child_file_activity_counts_only_successful_mutations() { |
| 1233 | let mut app = App::new(test_options(), &Config::default()); |
| 1234 | |
| 1235 | for (step, tool_name) in ["read_file", "list_dir", "file_search", "grep_files"] |
| 1236 | .into_iter() |
| 1237 | .enumerate() |
| 1238 | { |
| 1239 | record_agent_current_activity( |
| 1240 | &mut app, |
| 1241 | &MailboxMessage::ToolCallCompleted { |
| 1242 | agent_id: "agent_files".to_string(), |
| 1243 | tool_name: tool_name.to_string(), |
| 1244 | step: step as u32, |
| 1245 | ok: true, |
| 1246 | }, |
| 1247 | ); |
| 1248 | } |
| 1249 | assert_eq!(app.agent_progress_meta["agent_files"].files_touched, 0); |
| 1250 | |
| 1251 | for (step, tool_name) in ["write_file", "edit_file", "apply_patch"] |
| 1252 | .into_iter() |
| 1253 | .enumerate() |
| 1254 | { |
| 1255 | record_agent_current_activity( |
| 1256 | &mut app, |
| 1257 | &MailboxMessage::ToolCallCompleted { |
| 1258 | agent_id: "agent_files".to_string(), |
| 1259 | tool_name: tool_name.to_string(), |
| 1260 | step: (step + 10) as u32, |
| 1261 | ok: true, |
| 1262 | }, |
| 1263 | ); |
| 1264 | } |
| 1265 | assert_eq!(app.agent_progress_meta["agent_files"].files_touched, 3); |
| 1266 | |
| 1267 | record_agent_current_activity( |
| 1268 | &mut app, |
| 1269 | &MailboxMessage::ToolCallCompleted { |
| 1270 | agent_id: "agent_files".to_string(), |
| 1271 | tool_name: "write_file".to_string(), |
| 1272 | step: 20, |
| 1273 | ok: false, |
| 1274 | }, |
| 1275 | ); |
| 1276 | assert_eq!(app.agent_progress_meta["agent_files"].files_touched, 3); |
| 1277 | } |
| 1278 | |
| 1279 | #[test] |
| 1280 | fn recent_actions_are_three_bounded_structured_tool_outcomes() { |
| 1281 | let mut app = App::new(test_options(), &Config::default()); |
| 1282 | let agent_id = "agent_recent"; |
| 1283 | for step in 1..=5 { |
| 1284 | record_agent_current_activity( |
| 1285 | &mut app, |
| 1286 | &MailboxMessage::ToolCallCompleted { |
| 1287 | agent_id: agent_id.to_string(), |
| 1288 | tool_name: format!("\u{1b}[31mtool_{step}\u{1b}[0m"), |
| 1289 | step, |
| 1290 | ok: step != 4, |
| 1291 | }, |
| 1292 | ); |
| 1293 | } |
| 1294 | record_agent_current_activity( |
| 1295 | &mut app, |
| 1296 | &MailboxMessage::Progress { |
| 1297 | agent_id: agent_id.to_string(), |
| 1298 | status: "tool_99 completed".to_string(), |
| 1299 | }, |
| 1300 | ); |
| 1301 | |
| 1302 | let actions = &app.agent_progress_meta[agent_id].recent_actions; |
| 1303 | assert_eq!(actions.len(), MAX_AGENT_RECENT_ACTIONS); |
| 1304 | assert_eq!( |
| 1305 | actions.iter().map(|action| action.step).collect::<Vec<_>>(), |
| 1306 | vec![3, 4, 5] |
| 1307 | ); |
| 1308 | assert!(!actions.iter().any(|action| action.step == 99)); |
| 1309 | assert!(actions.iter().all(|action| !action.tool.contains('\u{1b}'))); |
| 1310 | assert!(!actions[1].ok); |
| 1311 | } |
| 1312 | |
| 1313 | #[test] |
| 1314 | fn token_usage_records_only_the_effective_child_route_facts() { |
| 1315 | let mut app = App::new(test_options(), &Config::default()); |
| 1316 | let changed = handle_subagent_mailbox( |
| 1317 | &mut app, |
| 1318 | 91, |
| 1319 | &MailboxMessage::TokenUsage { |
| 1320 | agent_id: "agent_route".to_string(), |
| 1321 | source_id: "response-route".to_string(), |
| 1322 | route: Box::new(test_route( |
| 1323 | crate::config::ApiProvider::Openrouter, |
| 1324 | "vendor/model-real", |
| 1325 | )), |
| 1326 | usage: codewhale_models::Usage::default(), |
| 1327 | }, |
| 1328 | ); |
| 1329 | |
| 1330 | assert!(!changed, "route facts do not allocate a transcript card"); |
| 1331 | let meta = &app.agent_progress_meta["agent_route"]; |
| 1332 | assert_eq!(meta.resolved_provider.as_deref(), Some("openrouter")); |
| 1333 | assert_eq!(meta.resolved_model.as_deref(), Some("vendor/model-real")); |
| 1334 | assert!(meta.current_activity.is_none()); |
| 1335 | } |
| 1336 | |
| 1337 | #[test] |
| 1338 | fn token_usage_accumulates_input_plus_output_across_child_turns() { |
| 1339 | let mut app = App::new(test_options(), &Config::default()); |
| 1340 | let route = test_route(crate::config::ApiProvider::Deepseek, "deepseek-v4-flash"); |
| 1341 | handle_subagent_mailbox( |
| 1342 | &mut app, |
| 1343 | 1, |
| 1344 | &MailboxMessage::TokenUsage { |
| 1345 | agent_id: "agent_spend".to_string(), |
| 1346 | source_id: "response-1".to_string(), |
| 1347 | route: Box::new(route.clone()), |
| 1348 | usage: codewhale_models::Usage { |
| 1349 | input_tokens: 1_000, |
| 1350 | output_tokens: 40, |
| 1351 | ..Default::default() |
| 1352 | }, |
| 1353 | }, |
| 1354 | ); |
| 1355 | handle_subagent_mailbox( |
| 1356 | &mut app, |
| 1357 | 2, |
| 1358 | &MailboxMessage::TokenUsage { |
| 1359 | agent_id: "agent_spend".to_string(), |
| 1360 | source_id: "response-2".to_string(), |
| 1361 | route: Box::new(route), |
| 1362 | usage: codewhale_models::Usage { |
| 1363 | input_tokens: 2_000, |
| 1364 | output_tokens: 60, |
| 1365 | ..Default::default() |
| 1366 | }, |
| 1367 | }, |
| 1368 | ); |
| 1369 | |
| 1370 | assert_eq!( |
| 1371 | app.agent_progress_meta["agent_spend"].received_tokens, |
| 1372 | Some(3_100), |
| 1373 | "work-bar tally must match worker budget total (input+output)" |
| 1374 | ); |
| 1375 | } |
| 1376 | |
| 1377 | #[test] |
| 1378 | fn typed_mailbox_lifecycle_projects_running_waiting_failed_and_done() { |
| 1379 | let mut app = App::new(test_options(), &Config::default()); |
| 1380 | |
| 1381 | assert!(handle_subagent_mailbox( |
| 1382 | &mut app, |
| 1383 | 1, |
| 1384 | &MailboxMessage::started("agent_running", FleetRole::Worker), |
| 1385 | )); |
| 1386 | assert_eq!( |
| 1387 | app.agent_progress_meta["agent_running"] |
| 1388 | .current_activity |
| 1389 | .as_ref() |
| 1390 | .map(|activity| activity.status), |
| 1391 | Some(AgentCurrentActivityStatus::Running) |
| 1392 | ); |
| 1393 | |
| 1394 | assert!(handle_subagent_mailbox( |
| 1395 | &mut app, |
| 1396 | 2, |
| 1397 | &MailboxMessage::ToolCallStarted { |
| 1398 | agent_id: "agent_running".to_string(), |
| 1399 | tool_name: "read_file".to_string(), |
| 1400 | step: 3, |
| 1401 | }, |
| 1402 | )); |
| 1403 | let running = app.agent_progress_meta["agent_running"] |
| 1404 | .current_activity |
| 1405 | .as_ref() |
| 1406 | .expect("running tool projection"); |
| 1407 | assert_eq!(running.status, AgentCurrentActivityStatus::RunningTool); |
| 1408 | assert_eq!(running.current_tool.as_deref(), Some("read_file")); |
| 1409 | assert_eq!(running.step, Some(3)); |
| 1410 | |
| 1411 | assert!(handle_subagent_mailbox( |
| 1412 | &mut app, |
| 1413 | 3, |
| 1414 | &MailboxMessage::Interrupted { |
| 1415 | agent_id: "agent_running".to_string(), |
| 1416 | reason: "approval needed".to_string(), |
| 1417 | }, |
| 1418 | )); |
| 1419 | let waiting = app.agent_progress_meta["agent_running"] |
| 1420 | .current_activity |
| 1421 | .as_ref() |
| 1422 | .expect("waiting projection"); |
| 1423 | assert_eq!(waiting.status, AgentCurrentActivityStatus::Waiting); |
| 1424 | assert_eq!(waiting.detail.as_deref(), Some("approval needed")); |
| 1425 | |
| 1426 | for (seq, agent_id, terminal, expected) in [ |
| 1427 | ( |
| 1428 | 4, |
| 1429 | "agent_failed", |
| 1430 | MailboxMessage::Failed { |
| 1431 | agent_id: "agent_failed".to_string(), |
| 1432 | error: "verification failed".to_string(), |
| 1433 | }, |
| 1434 | AgentCurrentActivityStatus::Failed, |
| 1435 | ), |
| 1436 | ( |
| 1437 | 5, |
| 1438 | "agent_done", |
| 1439 | MailboxMessage::Completed { |
| 1440 | agent_id: "agent_done".to_string(), |
| 1441 | summary: "verification complete".to_string(), |
| 1442 | }, |
| 1443 | AgentCurrentActivityStatus::Done, |
| 1444 | ), |
| 1445 | ] { |
| 1446 | assert!(handle_subagent_mailbox(&mut app, seq, &terminal)); |
| 1447 | assert_eq!( |
| 1448 | app.agent_progress_meta[agent_id] |
| 1449 | .current_activity |
| 1450 | .as_ref() |
| 1451 | .map(|activity| activity.status), |
| 1452 | Some(expected) |
| 1453 | ); |
| 1454 | } |
| 1455 | } |
| 1456 | |
| 1457 | #[test] |
| 1458 | fn reconcile_projects_typed_status_when_activity_detail_is_missing() { |
| 1459 | let mut app = App::new(test_options(), &Config::default()); |
| 1460 | let mut agent = subagent_result("agent_model_wait", SubAgentStatus::Running); |
| 1461 | agent.worker_status = Some(AgentWorkerStatus::ModelWait); |
| 1462 | app.subagent_cache.push(agent); |
| 1463 | |
| 1464 | reconcile_subagent_activity_state_at(&mut app, Instant::now()); |
| 1465 | |
| 1466 | let activity = app.agent_progress_meta["agent_model_wait"] |
| 1467 | .current_activity |
| 1468 | .as_ref() |
| 1469 | .expect("typed activity fallback"); |
| 1470 | assert_eq!(activity.status, AgentCurrentActivityStatus::ModelWait); |
| 1471 | assert_eq!(activity.detail, None); |
| 1472 | } |
| 1473 | |
| 1474 | #[test] |
| 1475 | fn mailbox_compact_projection_redacts_secrets_and_control_sequences() { |
| 1476 | let mut app = App::new(test_options(), &Config::default()); |
| 1477 | let agent_id = "agent_safe_projection"; |
| 1478 | assert!(handle_subagent_mailbox( |
| 1479 | &mut app, |
| 1480 | 1, |
| 1481 | &MailboxMessage::started(agent_id, FleetRole::Worker), |
| 1482 | )); |
| 1483 | let secret = "sk-mailbox-secret-1234567890"; |
| 1484 | let raw = format!( |
| 1485 | "\u{1b}[31mrunning\u{1b}[0m\napi_key={secret}\n\u{1b}]8;;https://example.invalid\u{7}details\u{1b}]8;;\u{7}\u{1}" |
| 1486 | ); |
| 1487 | assert!(handle_subagent_mailbox( |
| 1488 | &mut app, |
| 1489 | 2, |
| 1490 | &MailboxMessage::progress(agent_id, raw.clone()), |
| 1491 | )); |
| 1492 | |
| 1493 | let activity = app.agent_progress_meta[agent_id] |
| 1494 | .current_activity |
| 1495 | .as_ref() |
| 1496 | .expect("safe activity projection"); |
| 1497 | let detail = activity.detail.as_deref().expect("safe detail"); |
| 1498 | assert!(detail.contains("[redacted]"), "{detail:?}"); |
| 1499 | assert!(!detail.contains(secret), "{detail:?}"); |
| 1500 | assert!(!detail.contains('\u{1b}'), "{detail:?}"); |
| 1501 | assert!(!detail.contains("example.invalid"), "{detail:?}"); |
| 1502 | |
| 1503 | let card_index = app.subagent_card_index[agent_id]; |
| 1504 | let HistoryCell::SubAgent(SubAgentCell::Delegate(card)) = &app.history[card_index] else { |
| 1505 | panic!("expected delegate card"); |
| 1506 | }; |
| 1507 | let rendered = card |
| 1508 | .render_lines(120, &codewhale_palette::UI_THEME) |
| 1509 | .into_iter() |
| 1510 | .flat_map(|line| line.spans.into_iter().map(|span| span.content.into_owned())) |
| 1511 | .collect::<String>(); |
| 1512 | assert!(rendered.contains("[redacted]"), "{rendered:?}"); |
| 1513 | assert!(!rendered.contains(secret), "{rendered:?}"); |
| 1514 | assert!(!rendered.contains('\u{1b}'), "{rendered:?}"); |
| 1515 | assert!(!rendered.contains("example.invalid"), "{rendered:?}"); |
| 1516 | assert!( |
| 1517 | raw.contains(secret), |
| 1518 | "source mailbox payload stays untouched" |
| 1519 | ); |
| 1520 | assert!( |
| 1521 | raw.contains('\u{1b}'), |
| 1522 | "source mailbox payload stays untouched" |
| 1523 | ); |
| 1524 | } |
| 1525 | |
| 1526 | #[test] |
| 1527 | fn reconcile_keeps_progress_only_rows_until_cache_knows_the_agent() { |
| 1528 | let mut app = App::new(test_options(), &Config::default()); |
| 1529 | |
| 1530 | // A progress-first agent: its AgentSpawned/AgentList delivery was |
| 1531 | // dropped under channel pressure, so the authoritative cache has |
| 1532 | // never seen it. Its sidebar row must survive reconciliation. |
| 1533 | app.agent_progress |
| 1534 | .insert("agent_orphan".to_string(), "step 2/10".to_string()); |
| 1535 | app.agent_progress_meta.insert( |
| 1536 | "agent_orphan".to_string(), |
| 1537 | AgentProgressMeta { |
| 1538 | parent_run_id: None, |
| 1539 | spawn_depth: 0, |
| 1540 | ..AgentProgressMeta::default() |
| 1541 | }, |
| 1542 | ); |
| 1543 | |
| 1544 | // A terminal agent the cache DOES know about: its stale progress row |
| 1545 | // must still be evicted. |
| 1546 | app.subagent_cache |
| 1547 | .push(subagent_result("agent_done", SubAgentStatus::Completed)); |
| 1548 | app.agent_progress |
| 1549 | .insert("agent_done".to_string(), "step 9/10".to_string()); |
| 1550 | app.agent_progress_meta.insert( |
| 1551 | "agent_done".to_string(), |
| 1552 | AgentProgressMeta { |
| 1553 | parent_run_id: None, |
| 1554 | spawn_depth: 0, |
| 1555 | ..AgentProgressMeta::default() |
| 1556 | }, |
| 1557 | ); |
| 1558 | |
| 1559 | reconcile_subagent_activity_state_at(&mut app, Instant::now()); |
| 1560 | |
| 1561 | assert!( |
| 1562 | app.agent_progress.contains_key("agent_orphan"), |
| 1563 | "progress-only agent unknown to the cache must survive reconcile" |
| 1564 | ); |
| 1565 | assert!( |
| 1566 | app.agent_progress_meta.contains_key("agent_orphan"), |
| 1567 | "progress-only meta unknown to the cache must survive reconcile" |
| 1568 | ); |
| 1569 | assert!( |
| 1570 | !app.agent_progress.contains_key("agent_done"), |
| 1571 | "cache-known terminal agent progress must still be evicted" |
| 1572 | ); |
| 1573 | assert_eq!( |
| 1574 | app.agent_progress_meta["agent_done"] |
| 1575 | .current_activity |
| 1576 | .as_ref() |
| 1577 | .map(|activity| activity.status), |
| 1578 | Some(AgentCurrentActivityStatus::Done), |
| 1579 | "cache-known terminal agents retain a bounded terminal projection" |
| 1580 | ); |
| 1581 | |
| 1582 | // Once the authoritative cache reports the orphan as terminal, the |
| 1583 | // normal eviction applies and the row is released. |
| 1584 | app.subagent_cache |
| 1585 | .push(subagent_result("agent_orphan", SubAgentStatus::Completed)); |
| 1586 | reconcile_subagent_activity_state_at(&mut app, Instant::now()); |
| 1587 | assert!( |
| 1588 | !app.agent_progress.contains_key("agent_orphan"), |
| 1589 | "cache supersedes the progress-only row once it knows the agent" |
| 1590 | ); |
| 1591 | assert_eq!( |
| 1592 | app.agent_progress_meta["agent_orphan"] |
| 1593 | .current_activity |
| 1594 | .as_ref() |
| 1595 | .map(|activity| activity.status), |
| 1596 | Some(AgentCurrentActivityStatus::Done) |
| 1597 | ); |
| 1598 | } |
| 1599 | |
| 1600 | #[test] |
| 1601 | fn terminal_cards_archive_after_forty_five_seconds_without_losing_history() { |
| 1602 | let mut app = App::new(test_options(), &Config::default()); |
| 1603 | let agent_id = "agent_archive"; |
| 1604 | assert!(handle_subagent_mailbox( |
| 1605 | &mut app, |
| 1606 | 1, |
| 1607 | &MailboxMessage::started(agent_id, FleetRole::Worker), |
| 1608 | )); |
| 1609 | app.subagent_cache |
| 1610 | .push(subagent_result(agent_id, SubAgentStatus::Completed)); |
| 1611 | |
| 1612 | let observed = Instant::now(); |
| 1613 | reconcile_subagent_activity_state_at(&mut app, observed); |
| 1614 | reconcile_subagent_activity_state_at(&mut app, observed + SUBAGENT_TERMINAL_CARD_TTL); |
| 1615 | assert!( |
| 1616 | app.subagent_cache |
| 1617 | .iter() |
| 1618 | .any(|agent| agent.agent_id == agent_id), |
| 1619 | "the terminal card stays visible through its full 45-second grace period" |
| 1620 | ); |
| 1621 | |
| 1622 | reconcile_subagent_activity_state_at( |
| 1623 | &mut app, |
| 1624 | observed + SUBAGENT_TERMINAL_CARD_TTL + Duration::from_millis(1), |
| 1625 | ); |
| 1626 | assert!( |
| 1627 | !app.subagent_cache |
| 1628 | .iter() |
| 1629 | .any(|agent| agent.agent_id == agent_id), |
| 1630 | "the compact live cache must archive the settled card after the grace period" |
| 1631 | ); |
| 1632 | assert!( |
| 1633 | app.subagent_card_index.contains_key(agent_id), |
| 1634 | "archiving a compact card must not delete its transcript record" |
| 1635 | ); |
| 1636 | } |
| 1637 | |
| 1638 | #[test] |
| 1639 | fn apply_subagent_terminal_projection_clears_live_progress_and_card_state() { |
| 1640 | let mut app = App::new(test_options(), &Config::default()); |
| 1641 | let started = MailboxMessage::started("agent_done", FleetRole::Worker); |
| 1642 | assert!(handle_subagent_mailbox(&mut app, 1, &started)); |
| 1643 | let card_idx = app.subagent_card_index["agent_done"]; |
| 1644 | let initial_revision = app.history_revisions[card_idx]; |
| 1645 | |
| 1646 | app.subagent_cache |
| 1647 | .push(subagent_result("agent_done", SubAgentStatus::Running)); |
| 1648 | app.agent_progress |
| 1649 | .insert("agent_done".to_string(), "step 4/10".to_string()); |
| 1650 | app.agent_progress_meta.insert( |
| 1651 | "agent_done".to_string(), |
| 1652 | AgentProgressMeta { |
| 1653 | parent_run_id: None, |
| 1654 | spawn_depth: 0, |
| 1655 | ..AgentProgressMeta::default() |
| 1656 | }, |
| 1657 | ); |
| 1658 | |
| 1659 | assert!(apply_subagent_terminal_projection( |
| 1660 | &mut app, |
| 1661 | "agent_done", |
| 1662 | SubAgentStatus::Cancelled, |
| 1663 | Some("cancelled by user".to_string()) |
| 1664 | )); |
| 1665 | |
| 1666 | assert!(!app.agent_progress.contains_key("agent_done")); |
| 1667 | assert_eq!( |
| 1668 | app.agent_progress_meta["agent_done"] |
| 1669 | .current_activity |
| 1670 | .as_ref() |
| 1671 | .map(|activity| activity.status), |
| 1672 | Some(AgentCurrentActivityStatus::Canceled) |
| 1673 | ); |
| 1674 | let agent = app |
| 1675 | .subagent_cache |
| 1676 | .iter() |
| 1677 | .find(|agent| agent.agent_id == "agent_done") |
| 1678 | .expect("projected agent remains cached"); |
| 1679 | assert_eq!(agent.status, SubAgentStatus::Cancelled); |
| 1680 | assert_eq!(agent.worker_status, Some(AgentWorkerStatus::Cancelled)); |
| 1681 | assert_eq!(agent.result.as_deref(), Some("cancelled by user")); |
| 1682 | assert_eq!(running_agent_count(&app), 0); |
| 1683 | assert_ne!( |
| 1684 | app.history_revisions[card_idx], initial_revision, |
| 1685 | "terminal projection should invalidate the stale running card" |
| 1686 | ); |
| 1687 | match &app.history[card_idx] { |
| 1688 | HistoryCell::SubAgent(SubAgentCell::Delegate(card)) => { |
| 1689 | assert_eq!(card.status, AgentLifecycle::Cancelled); |
| 1690 | } |
| 1691 | cell => panic!("expected delegate card, got {cell:?}"), |
| 1692 | } |
| 1693 | } |
| 1694 | |
| 1695 | #[test] |
| 1696 | fn parent_stop_status_names_only_workers_that_continue_detached() { |
| 1697 | let mut app = App::new(test_options(), &Config::default()); |
| 1698 | app.subagent_cache |
| 1699 | .push(subagent_result("agent_b", SubAgentStatus::Running)); |
| 1700 | app.subagent_cache |
| 1701 | .push(subagent_result("agent_done", SubAgentStatus::Completed)); |
| 1702 | app.agent_progress |
| 1703 | .insert("agent_a".to_string(), "running tool".to_string()); |
| 1704 | app.agent_label_map |
| 1705 | .insert("agent_a".to_string(), "Agent 1".to_string()); |
| 1706 | app.agent_label_map |
| 1707 | .insert("agent_b".to_string(), "Southern Right".to_string()); |
| 1708 | app.agent_label_map |
| 1709 | .insert("agent_done".to_string(), "Finished worker".to_string()); |
| 1710 | |
| 1711 | let status = parent_stop_status(&app, "Request cancelled"); |
| 1712 | assert_eq!( |
| 1713 | status, |
| 1714 | "Request cancelled; detached workers continue (none canceled): Agent 1, Southern Right" |
| 1715 | ); |
| 1716 | assert!(!status.contains("Finished worker")); |
| 1717 | } |
| 1718 | |
| 1719 | #[test] |
| 1720 | fn parent_stop_status_is_unchanged_without_detached_workers() { |
| 1721 | let app = App::new(test_options(), &Config::default()); |
| 1722 | assert_eq!( |
| 1723 | parent_stop_status(&app, "Request cancelled"), |
| 1724 | "Request cancelled" |
| 1725 | ); |
| 1726 | } |
| 1727 | |
| 1728 | #[test] |
| 1729 | fn completion_before_started_allocates_recovery_delegate_card() { |
| 1730 | let mut app = App::new(test_options(), &Config::default()); |
| 1731 | let completed = MailboxMessage::Completed { |
| 1732 | agent_id: "agent_early".to_string(), |
| 1733 | summary: "recovered after early completion".to_string(), |
| 1734 | }; |
| 1735 | assert!( |
| 1736 | handle_subagent_mailbox(&mut app, 1, &completed), |
| 1737 | "completion-first delivery must still open a card" |
| 1738 | ); |
| 1739 | assert!(app.subagent_card_index.contains_key("agent_early")); |
| 1740 | |
| 1741 | let started = MailboxMessage::started("agent_early", FleetRole::Worker); |
| 1742 | assert!(handle_subagent_mailbox(&mut app, 2, &started)); |
| 1743 | match app.history.last() { |
| 1744 | Some(HistoryCell::SubAgent(SubAgentCell::Delegate(card))) => { |
| 1745 | assert_eq!(card.agent_id, "agent_early"); |
| 1746 | assert_ne!(card.agent_type, "…"); |
| 1747 | } |
| 1748 | other => panic!("expected delegate card, got {other:?}"), |
| 1749 | } |
| 1750 | } |
| 1751 | |
| 1752 | /// #4810: each child card carries that child's own ledger — never the |
| 1753 | /// parent's, never a sibling's, and never as transcript messages. |
| 1754 | #[test] |
| 1755 | fn sibling_child_cards_render_disjoint_todo_lists() { |
| 1756 | use crate::tools::todo::{TodoItem, TodoListSnapshot, TodoStatus}; |
| 1757 | |
| 1758 | fn snapshot(id: u32, content: &str) -> TodoListSnapshot { |
| 1759 | TodoListSnapshot { |
| 1760 | items: vec![TodoItem { |
| 1761 | id, |
| 1762 | content: content.to_string(), |
| 1763 | status: TodoStatus::InProgress, |
| 1764 | }], |
| 1765 | completion_pct: 0, |
| 1766 | in_progress_id: Some(id), |
| 1767 | } |
| 1768 | } |
| 1769 | |
| 1770 | fn card_text(app: &App, agent_id: &str) -> String { |
| 1771 | let idx = app.subagent_card_index[agent_id]; |
| 1772 | let HistoryCell::SubAgent(SubAgentCell::Delegate(card)) = &app.history[idx] else { |
| 1773 | panic!("expected a delegate card for {agent_id}"); |
| 1774 | }; |
| 1775 | assert_eq!(card.agent_id, agent_id); |
| 1776 | card.render_lines(120, &app.ui_theme) |
| 1777 | .into_iter() |
| 1778 | .flat_map(|line| line.spans.into_iter().map(|span| span.content.into_owned())) |
| 1779 | .collect() |
| 1780 | } |
| 1781 | |
| 1782 | let mut app = App::new(test_options(), &Config::default()); |
| 1783 | // The parent's own ledger exists and must never surface on a child. |
| 1784 | let parent_item = "PARENT ONLY: ship the release"; |
| 1785 | |
| 1786 | for (seq, id) in ["agent_left", "agent_right"].into_iter().enumerate() { |
| 1787 | assert!(handle_subagent_mailbox( |
| 1788 | &mut app, |
| 1789 | seq as u64 + 1, |
| 1790 | &MailboxMessage::started(id, FleetRole::Worker), |
| 1791 | )); |
| 1792 | } |
| 1793 | assert!(handle_subagent_mailbox( |
| 1794 | &mut app, |
| 1795 | 3, |
| 1796 | &MailboxMessage::WorkState { |
| 1797 | agent_id: "agent_left".to_string(), |
| 1798 | todo: snapshot(1, "LEFT: map the call sites"), |
| 1799 | }, |
| 1800 | )); |
| 1801 | assert!(handle_subagent_mailbox( |
| 1802 | &mut app, |
| 1803 | 4, |
| 1804 | &MailboxMessage::WorkState { |
| 1805 | agent_id: "agent_right".to_string(), |
| 1806 | todo: snapshot(1, "RIGHT: write the migration"), |
| 1807 | }, |
| 1808 | )); |
| 1809 | |
| 1810 | let left = card_text(&app, "agent_left"); |
| 1811 | let right = card_text(&app, "agent_right"); |
| 1812 | assert!(left.contains("LEFT: map the call sites"), "{left}"); |
| 1813 | assert!(!left.contains("RIGHT:"), "sibling leak: {left}"); |
| 1814 | assert!(!left.contains(parent_item), "parent leak: {left}"); |
| 1815 | assert!(right.contains("RIGHT: write the migration"), "{right}"); |
| 1816 | assert!(!right.contains("LEFT:"), "sibling leak: {right}"); |
| 1817 | assert!(!right.contains(parent_item), "parent leak: {right}"); |
| 1818 | |
| 1819 | // A nested child of agent_left gets its own card and its own ledger; |
| 1820 | // neither its parent's card nor its uncle's card absorbs it. |
| 1821 | assert!(handle_subagent_mailbox( |
| 1822 | &mut app, |
| 1823 | 5, |
| 1824 | &MailboxMessage::started("agent_nested", FleetRole::Worker), |
| 1825 | )); |
| 1826 | assert!(handle_subagent_mailbox( |
| 1827 | &mut app, |
| 1828 | 6, |
| 1829 | &MailboxMessage::WorkState { |
| 1830 | agent_id: "agent_nested".to_string(), |
| 1831 | todo: snapshot(1, "NESTED: read one file"), |
| 1832 | }, |
| 1833 | )); |
| 1834 | assert!(card_text(&app, "agent_nested").contains("NESTED: read one file")); |
| 1835 | assert!(!card_text(&app, "agent_left").contains("NESTED:")); |
| 1836 | assert!(!card_text(&app, "agent_right").contains("NESTED:")); |
| 1837 | |
| 1838 | // Nothing entered the transcript as an ordinary message: every To-do |
| 1839 | // row lives inside a sub-agent card. |
| 1840 | assert!( |
| 1841 | !app.history.iter().any(|cell| { |
| 1842 | !matches!(cell, HistoryCell::SubAgent(_)) |
| 1843 | && format!("{cell:?}").contains("map the call sites") |
| 1844 | }), |
| 1845 | "child To-do must not become an ordinary transcript message" |
| 1846 | ); |
| 1847 | } |
| 1848 | |
| 1849 | /// Work state is a ledger fact, not an activity transition: it must not |
| 1850 | /// invent or overwrite what the agent is currently doing. |
| 1851 | #[test] |
| 1852 | fn work_state_updates_todos_remaining_without_rewriting_activity() { |
| 1853 | use crate::tools::todo::{TodoItem, TodoListSnapshot, TodoStatus}; |
| 1854 | |
| 1855 | let mut app = App::new(test_options(), &Config::default()); |
| 1856 | assert!(handle_subagent_mailbox( |
| 1857 | &mut app, |
| 1858 | 1, |
| 1859 | &MailboxMessage::started("agent_todos", FleetRole::Worker), |
| 1860 | )); |
| 1861 | assert!(handle_subagent_mailbox( |
| 1862 | &mut app, |
| 1863 | 2, |
| 1864 | &MailboxMessage::ToolCallStarted { |
| 1865 | agent_id: "agent_todos".to_string(), |
| 1866 | tool_name: "read_file".to_string(), |
| 1867 | step: 1, |
| 1868 | }, |
| 1869 | )); |
| 1870 | assert!(handle_subagent_mailbox( |
| 1871 | &mut app, |
| 1872 | 3, |
| 1873 | &MailboxMessage::WorkState { |
| 1874 | agent_id: "agent_todos".to_string(), |
| 1875 | todo: TodoListSnapshot { |
| 1876 | items: vec![ |
| 1877 | TodoItem { |
| 1878 | id: 1, |
| 1879 | content: "done already".to_string(), |
| 1880 | status: TodoStatus::Completed, |
| 1881 | }, |
| 1882 | TodoItem { |
| 1883 | id: 2, |
| 1884 | content: "still cooking".to_string(), |
| 1885 | status: TodoStatus::InProgress, |
| 1886 | }, |
| 1887 | TodoItem { |
| 1888 | id: 3, |
| 1889 | content: "not yet".to_string(), |
| 1890 | status: TodoStatus::Pending, |
| 1891 | }, |
| 1892 | ], |
| 1893 | completion_pct: 33, |
| 1894 | in_progress_id: Some(2), |
| 1895 | }, |
| 1896 | }, |
| 1897 | )); |
| 1898 | |
| 1899 | let meta = &app.agent_progress_meta["agent_todos"]; |
| 1900 | assert_eq!(meta.todos_remaining, Some(2)); |
| 1901 | let activity = meta.current_activity.as_ref().expect("activity"); |
| 1902 | assert_eq!(activity.status, AgentCurrentActivityStatus::RunningTool); |
| 1903 | assert_eq!(activity.current_tool.as_deref(), Some("read_file")); |
| 1904 | |
| 1905 | // Empty publish clears the chip source (no list → no figure). |
| 1906 | assert!(handle_subagent_mailbox( |
| 1907 | &mut app, |
| 1908 | 4, |
| 1909 | &MailboxMessage::WorkState { |
| 1910 | agent_id: "agent_todos".to_string(), |
| 1911 | todo: TodoListSnapshot::default(), |
| 1912 | }, |
| 1913 | )); |
| 1914 | assert_eq!(app.agent_progress_meta["agent_todos"].todos_remaining, None); |
| 1915 | } |
| 1916 | |
| 1917 | #[test] |
| 1918 | fn work_state_envelope_does_not_rewrite_current_activity() { |
| 1919 | let mut app = App::new(test_options(), &Config::default()); |
| 1920 | assert!(handle_subagent_mailbox( |
| 1921 | &mut app, |
| 1922 | 1, |
| 1923 | &MailboxMessage::started("agent_x", FleetRole::Worker), |
| 1924 | )); |
| 1925 | assert!(handle_subagent_mailbox( |
| 1926 | &mut app, |
| 1927 | 2, |
| 1928 | &MailboxMessage::ToolCallStarted { |
| 1929 | agent_id: "agent_x".to_string(), |
| 1930 | tool_name: "read_file".to_string(), |
| 1931 | step: 2, |
| 1932 | }, |
| 1933 | )); |
| 1934 | assert!(handle_subagent_mailbox( |
| 1935 | &mut app, |
| 1936 | 3, |
| 1937 | &MailboxMessage::WorkState { |
| 1938 | agent_id: "agent_x".to_string(), |
| 1939 | todo: crate::tools::todo::TodoListSnapshot { |
| 1940 | items: vec![crate::tools::todo::TodoItem { |
| 1941 | id: 1, |
| 1942 | content: "keep reading".to_string(), |
| 1943 | status: crate::tools::todo::TodoStatus::InProgress, |
| 1944 | }], |
| 1945 | completion_pct: 0, |
| 1946 | in_progress_id: Some(1), |
| 1947 | }, |
| 1948 | }, |
| 1949 | )); |
| 1950 | |
| 1951 | let activity = app.agent_progress_meta["agent_x"] |
| 1952 | .current_activity |
| 1953 | .as_ref() |
| 1954 | .expect("activity"); |
| 1955 | assert_eq!(activity.status, AgentCurrentActivityStatus::RunningTool); |
| 1956 | assert_eq!(activity.current_tool.as_deref(), Some("read_file")); |
| 1957 | } |
| 1958 | |
| 1959 | /// Displayed child ledger text goes through the same redaction the rest of |
| 1960 | /// the child's displayed strings do. |
| 1961 | #[test] |
| 1962 | fn work_state_item_text_is_redacted_before_it_reaches_the_card() { |
| 1963 | let mut app = App::new(test_options(), &Config::default()); |
| 1964 | assert!(handle_subagent_mailbox( |
| 1965 | &mut app, |
| 1966 | 1, |
| 1967 | &MailboxMessage::started("agent_secret", FleetRole::Worker), |
| 1968 | )); |
| 1969 | let secret = "sk-mailbox-secret-1234567890"; |
| 1970 | assert!(handle_subagent_mailbox( |
| 1971 | &mut app, |
| 1972 | 2, |
| 1973 | &MailboxMessage::WorkState { |
| 1974 | agent_id: "agent_secret".to_string(), |
| 1975 | todo: crate::tools::todo::TodoListSnapshot { |
| 1976 | items: vec![crate::tools::todo::TodoItem { |
| 1977 | id: 4, |
| 1978 | content: format!("rotate api_key={secret}"), |
| 1979 | status: crate::tools::todo::TodoStatus::InProgress, |
| 1980 | }], |
| 1981 | completion_pct: 0, |
| 1982 | in_progress_id: Some(4), |
| 1983 | }, |
| 1984 | }, |
| 1985 | )); |
| 1986 | |
| 1987 | let idx = app.subagent_card_index["agent_secret"]; |
| 1988 | let HistoryCell::SubAgent(SubAgentCell::Delegate(card)) = &app.history[idx] else { |
| 1989 | panic!("expected delegate card"); |
| 1990 | }; |
| 1991 | let rendered: String = card |
| 1992 | .render_lines(120, &codewhale_palette::UI_THEME) |
| 1993 | .into_iter() |
| 1994 | .flat_map(|line| line.spans.into_iter().map(|span| span.content.into_owned())) |
| 1995 | .collect(); |
| 1996 | assert!(!rendered.contains(secret), "{rendered}"); |
| 1997 | assert!(rendered.contains("[redacted]"), "{rendered}"); |
| 1998 | assert!( |
| 1999 | rendered.contains("#4"), |
| 2000 | "item identity is preserved: {rendered}" |
| 2001 | ); |
| 2002 | } |
| 2003 | |
| 2004 | #[test] |
| 2005 | fn fanout_completion_burst_preserves_started_to_done_ordering() { |
| 2006 | let mut app = App::new(test_options(), &Config::default()); |
| 2007 | app.pending_subagent_dispatch = Some("rlm_eval".to_string()); |
| 2008 | for (seq, id) in ["agent_a", "agent_b"].into_iter().enumerate() { |
| 2009 | assert!(handle_subagent_mailbox( |
| 2010 | &mut app, |
| 2011 | seq as u64 + 1, |
| 2012 | &MailboxMessage::started(id, FleetRole::Scout), |
| 2013 | )); |
| 2014 | } |
| 2015 | assert!(handle_subagent_mailbox( |
| 2016 | &mut app, |
| 2017 | 3, |
| 2018 | &MailboxMessage::Completed { |
| 2019 | agent_id: "agent_a".to_string(), |
| 2020 | summary: "a done".to_string(), |
| 2021 | }, |
| 2022 | )); |
| 2023 | let Some(HistoryCell::SubAgent(SubAgentCell::Fanout(card))) = app.history.last() else { |
| 2024 | panic!("expected fanout card"); |
| 2025 | }; |
| 2026 | assert_eq!(card.workers.len(), 2); |
| 2027 | assert_eq!(card.workers[0].status, AgentLifecycle::Completed); |
| 2028 | assert_eq!(card.workers[1].status, AgentLifecycle::Running); |
| 2029 | } |
| 2030 | } |
| 2031 |