| 1 | //! Narrow model-facing agent coordination tools. |
| 2 | //! |
| 3 | //! Keeps `agent` as the creation surface. These five tools wrap existing |
| 4 | //! SubAgentManager / mailbox / checkpoint machinery without restoring the |
| 5 | //! retired lifecycle theater (`agent_open` / `agent_eval` / …). |
| 6 | |
| 7 | use std::sync::Arc; |
| 8 | use std::time::{Duration, Instant}; |
| 9 | |
| 10 | use async_trait::async_trait; |
| 11 | use serde_json::{Value, json}; |
| 12 | |
| 13 | use super::{ |
| 14 | COMPLETED_AGENT_RETENTION, ParentMailReceipt, SharedSubAgentManager, SubAgentRuntime, |
| 15 | SubAgentStatus, parse_agent_ref, subagent_session_projection, subagent_status_name, |
| 16 | wait_for_subagents_from_input, |
| 17 | }; |
| 18 | use crate::tools::registry::ToolRegistryBuilder; |
| 19 | use crate::tools::spec::{ |
| 20 | ApprovalRequirement, ToolCapability, ToolContext, ToolError, ToolResult, ToolSpec, |
| 21 | }; |
| 22 | |
| 23 | /// Bounds for `agents/wait`. Short on purpose: a blocked wait makes the |
| 24 | /// session deaf to typed input, and settled children already report back as |
| 25 | /// `<codewhale:subagent.done>` sentinels that start a fresh turn (#4097). |
| 26 | const COORD_WAIT_DEFAULT_TIMEOUT_SECS: u64 = 30; |
| 27 | const COORD_WAIT_MIN_TIMEOUT_SECS: u64 = 1; |
| 28 | const COORD_WAIT_MAX_TIMEOUT_SECS: u64 = 120; |
| 29 | const COORD_WAIT_CHECK_INTERVAL: Duration = Duration::from_millis(250); |
| 30 | const RECENT_PROGRESS_LIMIT: usize = 8; |
| 31 | pub(super) const COORDINATION_RECORD_LIMIT: usize = 128; |
| 32 | const COORDINATION_INSPECT_LIMIT: usize = 24; |
| 33 | pub(super) const COORDINATION_PROJECTION_DECISION_LIMIT: usize = 8; |
| 34 | pub(super) const COORDINATION_PROJECTION_BYTE_LIMIT: usize = 4096; |
| 35 | |
| 36 | mod ledger; |
| 37 | |
| 38 | // The ledger types moved to `ledger` unchanged and are re-published here, so |
| 39 | // `crate::tools::subagent::coord::{DecisionRecord, …}` still resolves for every |
| 40 | // consumer that never had reason to know where the definitions sit — the point |
| 41 | // of the split was to shorten two files, not to make four other files import |
| 42 | // differently. |
| 43 | // |
| 44 | // A glob, not a list: several of these types are named only from `cfg(test)` |
| 45 | // code in other modules, and an explicit `pub use` of those reads as an unused |
| 46 | // import in a release build. The glob also keeps each item's own visibility, so |
| 47 | // `MAX_RECONCILIATION_RETRIES` stays reachable here without becoming part of |
| 48 | // the module's public surface. |
| 49 | pub use ledger::*; |
| 50 | |
| 51 | // ── agents/list ────────────────────────────────────────────────────────── |
| 52 | |
| 53 | pub struct AgentsListTool { |
| 54 | manager: SharedSubAgentManager, |
| 55 | } |
| 56 | |
| 57 | impl AgentsListTool { |
| 58 | #[must_use] |
| 59 | pub fn new(manager: SharedSubAgentManager) -> Self { |
| 60 | Self { manager } |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | #[async_trait] |
| 65 | impl ToolSpec for AgentsListTool { |
| 66 | fn model_visible(&self) -> bool { |
| 67 | // #5462: `agent` is the sole model-facing sub-agent surface. These |
| 68 | // narrow tools stay registered and executable by name so a persisted |
| 69 | // transcript replays byte-for-byte, but they are never advertised in |
| 70 | // the catalog and can never be returned by `tool_search` — the same |
| 71 | // shape `rlm` and `exec_shell` already use. |
| 72 | false |
| 73 | } |
| 74 | |
| 75 | fn name(&self) -> &'static str { |
| 76 | "agents/list" |
| 77 | } |
| 78 | |
| 79 | fn description(&self) -> &'static str { |
| 80 | "List child agents: ids, parent hierarchy, state, bounded recent progress, and token budget. Read-only coordination view — does not spawn or wake workers." |
| 81 | } |
| 82 | |
| 83 | fn input_schema(&self) -> Value { |
| 84 | json!({ |
| 85 | "type": "object", |
| 86 | "properties": { |
| 87 | "include_archived": { |
| 88 | "type": "boolean", |
| 89 | "description": "Include prior-session / archived agents. Default false." |
| 90 | }, |
| 91 | "agent_id": { |
| 92 | "type": "string", |
| 93 | "description": "Optional single agent id or session name to inspect." |
| 94 | } |
| 95 | }, |
| 96 | "required": [] |
| 97 | }) |
| 98 | } |
| 99 | |
| 100 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 101 | vec![ToolCapability::ReadOnly] |
| 102 | } |
| 103 | |
| 104 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 105 | ApprovalRequirement::Auto |
| 106 | } |
| 107 | |
| 108 | fn is_read_only_for(&self, _input: &Value) -> bool { |
| 109 | true |
| 110 | } |
| 111 | |
| 112 | fn supports_parallel_for(&self, _input: &Value) -> bool { |
| 113 | true |
| 114 | } |
| 115 | |
| 116 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 117 | let include_archived = input |
| 118 | .get("include_archived") |
| 119 | .and_then(Value::as_bool) |
| 120 | .unwrap_or(false); |
| 121 | let agent_ref = parse_agent_ref(&input)?; |
| 122 | |
| 123 | let mut manager = self.manager.write().await; |
| 124 | manager.cleanup_for_session(&context.state_namespace, COMPLETED_AGENT_RETENTION); |
| 125 | let summaries = if let Some(agent_ref) = agent_ref { |
| 126 | let summary = manager |
| 127 | .coordination_summary_for_session( |
| 128 | &context.state_namespace, |
| 129 | &agent_ref, |
| 130 | RECENT_PROGRESS_LIMIT, |
| 131 | ) |
| 132 | .map_err(|err| ToolError::invalid_input(err.to_string()))?; |
| 133 | vec![summary] |
| 134 | } else { |
| 135 | manager.list_coordination_summaries_for_session( |
| 136 | &context.state_namespace, |
| 137 | include_archived, |
| 138 | RECENT_PROGRESS_LIMIT, |
| 139 | ) |
| 140 | }; |
| 141 | drop(manager); |
| 142 | |
| 143 | let payload = json!({ |
| 144 | "action": "list", |
| 145 | "count": summaries.len(), |
| 146 | "agents": summaries, |
| 147 | }); |
| 148 | let mut tool_result = ToolResult::json(&payload) |
| 149 | .map_err(|err| ToolError::execution_failed(err.to_string()))?; |
| 150 | tool_result.metadata = Some(json!({ |
| 151 | "action": "list", |
| 152 | "count": summaries.len(), |
| 153 | })); |
| 154 | Ok(tool_result) |
| 155 | } |
| 156 | } |
| 157 | |
| 158 | // ── agents/message ─────────────────────────────────────────────────────── |
| 159 | |
| 160 | pub struct AgentsMessageTool { |
| 161 | manager: SharedSubAgentManager, |
| 162 | caller_agent_id: Option<String>, |
| 163 | } |
| 164 | |
| 165 | impl AgentsMessageTool { |
| 166 | #[must_use] |
| 167 | pub fn new(manager: SharedSubAgentManager) -> Self { |
| 168 | Self { |
| 169 | manager, |
| 170 | caller_agent_id: None, |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | #[must_use] |
| 175 | pub(crate) fn with_optional_caller(mut self, caller_agent_id: Option<String>) -> Self { |
| 176 | self.caller_agent_id = caller_agent_id; |
| 177 | self |
| 178 | } |
| 179 | } |
| 180 | |
| 181 | #[async_trait] |
| 182 | impl ToolSpec for AgentsMessageTool { |
| 183 | fn model_visible(&self) -> bool { |
| 184 | // #5462: `agent` is the sole model-facing sub-agent surface. These |
| 185 | // narrow tools stay registered and executable by name so a persisted |
| 186 | // transcript replays byte-for-byte, but they are never advertised in |
| 187 | // the catalog and can never be returned by `tool_search` — the same |
| 188 | // shape `rlm` and `exec_shell` already use. |
| 189 | false |
| 190 | } |
| 191 | |
| 192 | fn name(&self) -> &'static str { |
| 193 | "agents/message" |
| 194 | } |
| 195 | |
| 196 | fn description(&self) -> &'static str { |
| 197 | "Queue a parent message onto a running child without waking it. The message stays queued until a later agents/followup delivers it through the child's live input channel. Use agents/followup directly when you want immediate delivery." |
| 198 | } |
| 199 | |
| 200 | fn input_schema(&self) -> Value { |
| 201 | json!({ |
| 202 | "type": "object", |
| 203 | "properties": { |
| 204 | "agent_id": { |
| 205 | "type": "string", |
| 206 | "description": "Target child agent id or session name." |
| 207 | }, |
| 208 | "message": { |
| 209 | "type": "string", |
| 210 | "description": "Message text to queue." |
| 211 | } |
| 212 | }, |
| 213 | "required": ["agent_id", "message"] |
| 214 | }) |
| 215 | } |
| 216 | |
| 217 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 218 | vec![ToolCapability::RequiresApproval] |
| 219 | } |
| 220 | |
| 221 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 222 | ApprovalRequirement::Required |
| 223 | } |
| 224 | |
| 225 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 226 | let agent_ref = |
| 227 | parse_agent_ref(&input)?.ok_or_else(|| ToolError::missing_field("agent_id"))?; |
| 228 | let message = input |
| 229 | .get("message") |
| 230 | .or_else(|| input.get("text")) |
| 231 | .and_then(Value::as_str) |
| 232 | .map(str::trim) |
| 233 | .filter(|s| !s.is_empty()) |
| 234 | .ok_or_else(|| ToolError::missing_field("message"))? |
| 235 | .to_string(); |
| 236 | |
| 237 | let receipt = { |
| 238 | let mut manager = self.manager.write().await; |
| 239 | manager |
| 240 | .ensure_caller_controls_descendant_for_session( |
| 241 | &context.state_namespace, |
| 242 | &agent_ref, |
| 243 | self.caller_agent_id.as_deref(), |
| 244 | "agents/message", |
| 245 | ) |
| 246 | .map_err(|err| ToolError::invalid_input(err.to_string()))?; |
| 247 | manager |
| 248 | .queue_running_parent_message_for_session( |
| 249 | &context.state_namespace, |
| 250 | &agent_ref, |
| 251 | message, |
| 252 | ) |
| 253 | .map_err(|err| ToolError::invalid_input(err.to_string()))? |
| 254 | }; |
| 255 | |
| 256 | let payload = json!({ |
| 257 | "action": "message", |
| 258 | "agent_id": receipt.agent_id, |
| 259 | "queued": true, |
| 260 | "woke": false, |
| 261 | "queue_depth": receipt.queue_depth, |
| 262 | "status": receipt.status, |
| 263 | "note": "Message queued without waking the child.", |
| 264 | }); |
| 265 | let mut tool_result = ToolResult::json(&payload) |
| 266 | .map_err(|err| ToolError::execution_failed(err.to_string()))?; |
| 267 | tool_result.metadata = Some(json!({ |
| 268 | "action": "message", |
| 269 | "agent_id": receipt.agent_id, |
| 270 | "woke": false, |
| 271 | "queue_depth": receipt.queue_depth, |
| 272 | })); |
| 273 | Ok(tool_result) |
| 274 | } |
| 275 | } |
| 276 | |
| 277 | // ── agents/followup ────────────────────────────────────────────────────── |
| 278 | |
| 279 | pub struct AgentsFollowupTool { |
| 280 | manager: SharedSubAgentManager, |
| 281 | caller_agent_id: Option<String>, |
| 282 | /// Runtime for checkpoint resume. `None` (legacy/test construction) |
| 283 | /// keeps the queue-only followup behavior. |
| 284 | runtime: Option<SubAgentRuntime>, |
| 285 | } |
| 286 | |
| 287 | impl AgentsFollowupTool { |
| 288 | #[must_use] |
| 289 | pub fn new(manager: SharedSubAgentManager) -> Self { |
| 290 | Self { |
| 291 | manager, |
| 292 | caller_agent_id: None, |
| 293 | runtime: None, |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | #[must_use] |
| 298 | pub fn with_runtime(mut self, runtime: SubAgentRuntime) -> Self { |
| 299 | self.runtime = Some(runtime); |
| 300 | self |
| 301 | } |
| 302 | |
| 303 | #[must_use] |
| 304 | pub(crate) fn with_optional_caller(mut self, caller_agent_id: Option<String>) -> Self { |
| 305 | self.caller_agent_id = caller_agent_id; |
| 306 | self |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | impl AgentsFollowupTool { |
| 311 | async fn followup_one( |
| 312 | &self, |
| 313 | agent_ref: &str, |
| 314 | message: &str, |
| 315 | context: &ToolContext, |
| 316 | ) -> Result<Value, ToolError> { |
| 317 | let mut manager = self.manager.write().await; |
| 318 | let (source, target) = manager |
| 319 | .continuation_target_for_caller( |
| 320 | &context.state_namespace, |
| 321 | agent_ref, |
| 322 | self.caller_agent_id.as_deref(), |
| 323 | "agents/followup", |
| 324 | ) |
| 325 | .map_err(|error| ToolError::invalid_input(error.to_string()))?; |
| 326 | let snapshot = manager |
| 327 | .get_result(&target) |
| 328 | .map_err(|error| ToolError::invalid_input(error.to_string()))?; |
| 329 | let resumed_already = source != target; |
| 330 | let receipt = if super::subagent_checkpoint_is_continuable(&snapshot) |
| 331 | && self.runtime.is_some() |
| 332 | { |
| 333 | let snapshot = manager |
| 334 | .resume_from_checkpoint_for_session( |
| 335 | &context.state_namespace, |
| 336 | Arc::clone(&self.manager), |
| 337 | self.runtime.clone().expect("runtime checked"), |
| 338 | &target, |
| 339 | message, |
| 340 | ) |
| 341 | .map_err(|error| ToolError::execution_failed(error.to_string()))?; |
| 342 | ParentMailReceipt { |
| 343 | agent_id: snapshot.agent_id.clone(), |
| 344 | status: subagent_status_name(&snapshot.status).to_string(), |
| 345 | queue_depth: 0, |
| 346 | woke: true, |
| 347 | continued_from_checkpoint: true, |
| 348 | continuation_handle: None, |
| 349 | note: format!( |
| 350 | "resumed from checkpoint {source} as {}; original receipt retained", |
| 351 | snapshot.agent_id |
| 352 | ), |
| 353 | } |
| 354 | } else if resumed_already && snapshot.status != SubAgentStatus::Running { |
| 355 | ParentMailReceipt { |
| 356 | agent_id: target, status: subagent_status_name(&snapshot.status).to_string(), |
| 357 | queue_depth: 0, woke: false, continued_from_checkpoint: true, |
| 358 | continuation_handle: None, |
| 359 | note: "Existing continuation has settled; no duplicate worker was started and no message was delivered.".to_string(), |
| 360 | } |
| 361 | } else { |
| 362 | manager |
| 363 | .followup_child_for_session(&context.state_namespace, &target, message.to_string()) |
| 364 | .map_err(|error| ToolError::invalid_input(error.to_string()))? |
| 365 | }; |
| 366 | let child_route = manager |
| 367 | .get_worker_record_for_session(&context.state_namespace, &receipt.agent_id) |
| 368 | .and_then(|record| record.spec.child_route); |
| 369 | Ok(json!({ |
| 370 | "action": "followup", "from": source, "to": receipt.agent_id, |
| 371 | "agent_id": receipt.agent_id, "queued": receipt.woke || receipt.queue_depth > 0, |
| 372 | "woke": receipt.woke, "queue_depth": receipt.queue_depth, "status": receipt.status, |
| 373 | "continued_from_checkpoint": receipt.continued_from_checkpoint || resumed_already, |
| 374 | "continuation_handle": receipt.continuation_handle, "note": receipt.note, |
| 375 | "child_route": child_route, |
| 376 | })) |
| 377 | } |
| 378 | } |
| 379 | |
| 380 | #[async_trait] |
| 381 | impl ToolSpec for AgentsFollowupTool { |
| 382 | fn model_visible(&self) -> bool { |
| 383 | // #5462: `agent` is the sole model-facing sub-agent surface. These |
| 384 | // narrow tools stay registered and executable by name so a persisted |
| 385 | // transcript replays byte-for-byte, but they are never advertised in |
| 386 | // the catalog and can never be returned by `tool_search` — the same |
| 387 | // shape `rlm` and `exec_shell` already use. |
| 388 | false |
| 389 | } |
| 390 | |
| 391 | fn name(&self) -> &'static str { |
| 392 | "agents/followup" |
| 393 | } |
| 394 | |
| 395 | fn description(&self) -> &'static str { |
| 396 | "Queue a message and attempt to resume an idle or interrupted child. Running children receive the message on their next step; interrupted_continuable children are resumed from their checkpoint into a fresh agent loop (new agent id, original prompt plus prior conversation tail) when a runtime is attached, and otherwise keep queue-only semantics with the continuation_handle returned." |
| 397 | } |
| 398 | |
| 399 | fn input_schema(&self) -> Value { |
| 400 | json!({ |
| 401 | "type": "object", |
| 402 | "properties": { |
| 403 | "agent_id": {"type": "string"}, |
| 404 | "agent_ids": {"type": "array", "items": {"type": "string", "minLength": 1}, "minItems": 1, "maxItems": 32}, |
| 405 | "all_parked": {"type": "boolean", "description": "Continue every owned parked child, up to 32."}, |
| 406 | "message": {"type": "string", "minLength": 1} |
| 407 | }, |
| 408 | "required": ["message"], |
| 409 | "oneOf": [{"required": ["agent_id"]}, {"required": ["agent_ids"]}, {"required": ["all_parked"]}] |
| 410 | }) |
| 411 | } |
| 412 | |
| 413 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 414 | vec![ToolCapability::RequiresApproval] |
| 415 | } |
| 416 | |
| 417 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 418 | ApprovalRequirement::Required |
| 419 | } |
| 420 | |
| 421 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 422 | let message = input |
| 423 | .get("message") |
| 424 | .or_else(|| input.get("text")) |
| 425 | .and_then(Value::as_str) |
| 426 | .map(str::trim) |
| 427 | .filter(|value| !value.is_empty()) |
| 428 | .ok_or_else(|| ToolError::missing_field("message"))?; |
| 429 | let single = parse_agent_ref(&input)?; |
| 430 | let all_parked = super::parse_optional_bool(&input, &["all_parked"])?.unwrap_or(false); |
| 431 | let batch = input.get("agent_ids"); |
| 432 | if usize::from(single.is_some()) + usize::from(batch.is_some()) + usize::from(all_parked) |
| 433 | != 1 |
| 434 | { |
| 435 | return Err(ToolError::invalid_input( |
| 436 | "followup requires exactly one of agent_id, agent_ids, or all_parked=true", |
| 437 | )); |
| 438 | } |
| 439 | let mut targets = if let Some(ids) = batch { |
| 440 | let ids = ids |
| 441 | .as_array() |
| 442 | .filter(|ids| !ids.is_empty() && ids.len() <= 32) |
| 443 | .ok_or_else(|| { |
| 444 | ToolError::invalid_input("agent_ids must contain 1..32 nonempty strings") |
| 445 | })?; |
| 446 | ids.iter() |
| 447 | .map(|id| { |
| 448 | id.as_str() |
| 449 | .map(str::trim) |
| 450 | .filter(|id| !id.is_empty()) |
| 451 | .map(str::to_string) |
| 452 | .ok_or_else(|| { |
| 453 | ToolError::invalid_input("agent_ids must contain nonempty strings") |
| 454 | }) |
| 455 | }) |
| 456 | .collect::<Result<Vec<_>, _>>()? |
| 457 | } else if let Some(id) = single.as_ref() { |
| 458 | vec![id.clone()] |
| 459 | } else { |
| 460 | let manager = self.manager.read().await; |
| 461 | let mut ids = manager |
| 462 | .agents |
| 463 | .values() |
| 464 | .filter(|agent| { |
| 465 | manager.agent_is_owned_by_session(agent, &context.state_namespace) |
| 466 | && agent |
| 467 | .checkpoint |
| 468 | .as_ref() |
| 469 | .is_some_and(|checkpoint| checkpoint.parked_at_turn_end) |
| 470 | && matches!(agent.status, SubAgentStatus::Interrupted(_)) |
| 471 | && manager |
| 472 | .ensure_caller_controls_descendant( |
| 473 | &agent.id, |
| 474 | self.caller_agent_id.as_deref(), |
| 475 | "agents/followup", |
| 476 | ) |
| 477 | .is_ok() |
| 478 | && manager |
| 479 | .continuation_target(&agent.id) |
| 480 | .is_ok_and(|target| target == agent.id) |
| 481 | }) |
| 482 | .map(|agent| agent.id.clone()) |
| 483 | .collect::<Vec<_>>(); |
| 484 | ids.sort(); |
| 485 | if ids.len() > 32 { |
| 486 | return Err(ToolError::invalid_input( |
| 487 | "More than 32 parked children; use explicit agent_ids batches", |
| 488 | )); |
| 489 | } |
| 490 | ids |
| 491 | }; |
| 492 | let mut seen = std::collections::HashSet::new(); |
| 493 | targets.retain(|target| seen.insert(target.clone())); |
| 494 | let mut results = Vec::new(); |
| 495 | let mut errors = Vec::new(); |
| 496 | // Each mutation and both hierarchy checks share the manager write lock. |
| 497 | // A target failure cannot erase successful results from another target. |
| 498 | for target in targets { |
| 499 | match self.followup_one(&target, message, context).await { |
| 500 | Ok(payload) => results.push(payload), |
| 501 | Err(error) if single.is_some() => return Err(error), |
| 502 | Err(error) => errors.push(json!({"from": target, "error": error.to_string()})), |
| 503 | } |
| 504 | } |
| 505 | let payload = if single.is_some() { |
| 506 | results.pop().expect("single target returned a result") |
| 507 | } else { |
| 508 | json!({"action": "followup", "results": results, "errors": errors}) |
| 509 | }; |
| 510 | let mut result = ToolResult::json(&payload) |
| 511 | .map_err(|error| ToolError::execution_failed(error.to_string()))?; |
| 512 | result.metadata = Some(payload.clone()); |
| 513 | Ok(result) |
| 514 | } |
| 515 | } |
| 516 | |
| 517 | // ── agents/interrupt ───────────────────────────────────────────────────── |
| 518 | |
| 519 | pub struct AgentsInterruptTool { |
| 520 | manager: SharedSubAgentManager, |
| 521 | /// Optional caller identity for fail-closed self-interrupt checks. |
| 522 | caller_agent_id: Option<String>, |
| 523 | } |
| 524 | |
| 525 | impl AgentsInterruptTool { |
| 526 | #[must_use] |
| 527 | pub fn new(manager: SharedSubAgentManager) -> Self { |
| 528 | Self { |
| 529 | manager, |
| 530 | caller_agent_id: None, |
| 531 | } |
| 532 | } |
| 533 | |
| 534 | #[must_use] |
| 535 | #[allow(dead_code)] // arms self-interrupt fail-closed when child registries thread caller (P1.2) |
| 536 | pub fn with_caller(mut self, caller_agent_id: impl Into<String>) -> Self { |
| 537 | self.caller_agent_id = Some(caller_agent_id.into()); |
| 538 | self |
| 539 | } |
| 540 | |
| 541 | #[must_use] |
| 542 | pub(crate) fn with_optional_caller(mut self, caller_agent_id: Option<String>) -> Self { |
| 543 | self.caller_agent_id = caller_agent_id; |
| 544 | self |
| 545 | } |
| 546 | } |
| 547 | |
| 548 | #[async_trait] |
| 549 | impl ToolSpec for AgentsInterruptTool { |
| 550 | fn model_visible(&self) -> bool { |
| 551 | // #5462: `agent` is the sole model-facing sub-agent surface. These |
| 552 | // narrow tools stay registered and executable by name so a persisted |
| 553 | // transcript replays byte-for-byte, but they are never advertised in |
| 554 | // the catalog and can never be returned by `tool_search` — the same |
| 555 | // shape `rlm` and `exec_shell` already use. |
| 556 | false |
| 557 | } |
| 558 | |
| 559 | fn name(&self) -> &'static str { |
| 560 | "agents/interrupt" |
| 561 | } |
| 562 | |
| 563 | fn description(&self) -> &'static str { |
| 564 | "Interrupt a running child agent, preserve its checkpoint, and return the prior state. Fails closed on root or self targets. Prefer this over cancel when you may resume later." |
| 565 | } |
| 566 | |
| 567 | fn input_schema(&self) -> Value { |
| 568 | json!({ |
| 569 | "type": "object", |
| 570 | "properties": { |
| 571 | "agent_id": { |
| 572 | "type": "string", |
| 573 | "description": "Child agent id or session name to interrupt." |
| 574 | }, |
| 575 | "reason": { |
| 576 | "type": "string", |
| 577 | "description": "Optional interrupt reason recorded on the checkpoint." |
| 578 | } |
| 579 | }, |
| 580 | "required": ["agent_id"] |
| 581 | }) |
| 582 | } |
| 583 | |
| 584 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 585 | vec![ToolCapability::RequiresApproval] |
| 586 | } |
| 587 | |
| 588 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 589 | ApprovalRequirement::Required |
| 590 | } |
| 591 | |
| 592 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 593 | let agent_ref = |
| 594 | parse_agent_ref(&input)?.ok_or_else(|| ToolError::missing_field("agent_id"))?; |
| 595 | let reason = input |
| 596 | .get("reason") |
| 597 | .and_then(Value::as_str) |
| 598 | .map(str::trim) |
| 599 | .filter(|s| !s.is_empty()) |
| 600 | .unwrap_or("interrupted by parent via agents/interrupt") |
| 601 | .to_string(); |
| 602 | |
| 603 | let (prior, snapshot) = { |
| 604 | let mut manager = self.manager.write().await; |
| 605 | manager |
| 606 | .interrupt_child_for_session( |
| 607 | &context.state_namespace, |
| 608 | &agent_ref, |
| 609 | self.caller_agent_id.as_deref(), |
| 610 | reason, |
| 611 | ) |
| 612 | .map_err(|err| ToolError::invalid_input(err.to_string()))? |
| 613 | }; |
| 614 | |
| 615 | let worker_record = { |
| 616 | let manager = self.manager.read().await; |
| 617 | manager.get_worker_record_for_session(&context.state_namespace, &snapshot.agent_id) |
| 618 | }; |
| 619 | let projection = |
| 620 | subagent_session_projection(&self.manager, snapshot, false, context, worker_record) |
| 621 | .await; |
| 622 | let payload = json!({ |
| 623 | "action": "interrupt", |
| 624 | "agent_id": projection.agent_id, |
| 625 | "prior_status": subagent_status_name(&prior.status), |
| 626 | "prior_steps_taken": prior.steps_taken, |
| 627 | "status": projection.status, |
| 628 | "checkpoint_preserved": projection.checkpoint.is_some(), |
| 629 | "continuable": projection.continuable, |
| 630 | "projection": projection, |
| 631 | "child_route": projection.child_route, |
| 632 | }); |
| 633 | let mut tool_result = ToolResult::json(&payload) |
| 634 | .map_err(|err| ToolError::execution_failed(err.to_string()))?; |
| 635 | tool_result.metadata = Some(json!({ |
| 636 | "action": "interrupt", |
| 637 | "agent_id": payload["agent_id"], |
| 638 | "checkpoint_preserved": payload["checkpoint_preserved"], |
| 639 | "child_route": payload["child_route"], |
| 640 | })); |
| 641 | Ok(tool_result) |
| 642 | } |
| 643 | } |
| 644 | |
| 645 | // ── agents/wait ────────────────────────────────────────────────────────── |
| 646 | |
| 647 | pub struct AgentsWaitTool { |
| 648 | manager: SharedSubAgentManager, |
| 649 | } |
| 650 | |
| 651 | impl AgentsWaitTool { |
| 652 | #[must_use] |
| 653 | pub fn new(manager: SharedSubAgentManager) -> Self { |
| 654 | Self { manager } |
| 655 | } |
| 656 | } |
| 657 | |
| 658 | #[async_trait] |
| 659 | impl ToolSpec for AgentsWaitTool { |
| 660 | fn model_visible(&self) -> bool { |
| 661 | // #5462: `agent` is the sole model-facing sub-agent surface. These |
| 662 | // narrow tools stay registered and executable by name so a persisted |
| 663 | // transcript replays byte-for-byte, but they are never advertised in |
| 664 | // the catalog and can never be returned by `tool_search` — the same |
| 665 | // shape `rlm` and `exec_shell` already use. |
| 666 | false |
| 667 | } |
| 668 | |
| 669 | fn name(&self) -> &'static str { |
| 670 | "agents/wait" |
| 671 | } |
| 672 | |
| 673 | fn description(&self) -> &'static str { |
| 674 | "Block briefly until watched children settle or the timeout elapses. Keep waits short: on timeout, end your turn — settled children wake you automatically as completion sentinels; polling agents/list in a loop is not the right shape either. until=all is the fan-out join: it returns only when every child running at call time has left running, with each child's outcome. until=completion (default) returns as soon as any one child settles. until=activity also returns on progress." |
| 675 | } |
| 676 | |
| 677 | fn input_schema(&self) -> Value { |
| 678 | json!({ |
| 679 | "type": "object", |
| 680 | "properties": { |
| 681 | "agent_id": { |
| 682 | "type": "string", |
| 683 | "description": "Optional specific child. When omitted, watches every child running at call time." |
| 684 | }, |
| 685 | "timeout_secs": { |
| 686 | "type": "integer", |
| 687 | "minimum": 1, |
| 688 | "maximum": 120, |
| 689 | "description": "Maximum seconds to block. Default 30. Keep it short — on timeout, end your turn; settled children report back as completion sentinels." |
| 690 | }, |
| 691 | "until": { |
| 692 | "type": "string", |
| 693 | "enum": ["completion", "all", "activity"], |
| 694 | "description": "completion (default): return when any one child leaves running. all: return only when every watched child has left running — use this after a fan-out so one wait covers the whole batch. activity: also return when recent progress changes. Children spawned after the call are not watched; no children means an immediate return." |
| 695 | } |
| 696 | }, |
| 697 | "required": [] |
| 698 | }) |
| 699 | } |
| 700 | |
| 701 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 702 | vec![ToolCapability::ReadOnly] |
| 703 | } |
| 704 | |
| 705 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 706 | ApprovalRequirement::Auto |
| 707 | } |
| 708 | |
| 709 | fn is_read_only_for(&self, _input: &Value) -> bool { |
| 710 | true |
| 711 | } |
| 712 | |
| 713 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 714 | dispatch_wait(&input, Arc::clone(&self.manager), context).await |
| 715 | } |
| 716 | } |
| 717 | |
| 718 | /// Single entry point for every blocking wait, shared by `agents/wait` and |
| 719 | /// `agent(action="wait")` so the two surfaces cannot drift. |
| 720 | /// |
| 721 | /// `until` selects the join shape: |
| 722 | /// - `completion` (default) — return as soon as any one watched child settles. |
| 723 | /// - `all` — return only when every watched child has settled (the fan-out |
| 724 | /// join the parent should use after dispatching a batch). |
| 725 | /// - `activity` — also return when a running child makes visible progress. |
| 726 | pub(super) async fn dispatch_wait( |
| 727 | input: &Value, |
| 728 | manager: SharedSubAgentManager, |
| 729 | context: &ToolContext, |
| 730 | ) -> Result<ToolResult, ToolError> { |
| 731 | let until = input |
| 732 | .get("until") |
| 733 | .and_then(Value::as_str) |
| 734 | .unwrap_or("completion") |
| 735 | .trim() |
| 736 | .to_ascii_lowercase(); |
| 737 | |
| 738 | match until.as_str() { |
| 739 | "" | "completion" => { |
| 740 | let mut wait_input = input.clone(); |
| 741 | if wait_input.get("action").is_none() { |
| 742 | wait_input["action"] = json!("wait"); |
| 743 | } |
| 744 | wait_for_subagents_from_input(&wait_input, manager, context).await |
| 745 | } |
| 746 | "all" => wait_for_all_children(input, manager, context).await, |
| 747 | "activity" => wait_for_activity(input, manager, context).await, |
| 748 | other => Err(ToolError::invalid_input(format!( |
| 749 | "Invalid until '{other}'. Use completion, all, or activity." |
| 750 | ))), |
| 751 | } |
| 752 | } |
| 753 | |
| 754 | /// `until=all`: block until every child that was running when the call was |
| 755 | /// made has left `Running`. |
| 756 | /// |
| 757 | /// The watch set is fixed at call time. A child spawned while this wait is |
| 758 | /// blocked is deliberately **not** joined — the parent asked to join the batch |
| 759 | /// it had just dispatched, and silently extending the set would make the call |
| 760 | /// unbounded in a way the caller never asked for. Callers that fan out again |
| 761 | /// simply issue another wait. |
| 762 | /// |
| 763 | /// Cancel-safe (no lock is held across an await), honours `timeout_secs`, and |
| 764 | /// returns immediately with `all_settled: true` when nothing is running. |
| 765 | async fn wait_for_all_children( |
| 766 | input: &Value, |
| 767 | manager: SharedSubAgentManager, |
| 768 | context: &ToolContext, |
| 769 | ) -> Result<ToolResult, ToolError> { |
| 770 | let timeout_secs = input |
| 771 | .get("timeout_secs") |
| 772 | .or_else(|| input.get("timeout")) |
| 773 | .and_then(Value::as_u64) |
| 774 | .unwrap_or(COORD_WAIT_DEFAULT_TIMEOUT_SECS) |
| 775 | .clamp(COORD_WAIT_MIN_TIMEOUT_SECS, COORD_WAIT_MAX_TIMEOUT_SECS); |
| 776 | let timeout = Duration::from_secs(timeout_secs); |
| 777 | let agent_ref = parse_agent_ref(input)?; |
| 778 | |
| 779 | // Resolve the watch set up front so a bad reference fails immediately |
| 780 | // rather than blocking for the whole timeout. |
| 781 | let watched: Vec<String> = { |
| 782 | let manager = manager.read().await; |
| 783 | if let Some(agent_ref) = &agent_ref { |
| 784 | let snapshot = manager |
| 785 | .get_result_by_ref_for_session(&context.state_namespace, agent_ref) |
| 786 | .map_err(|err| ToolError::invalid_input(err.to_string()))?; |
| 787 | if snapshot.status != SubAgentStatus::Running { |
| 788 | // Already settled: hand back its outcome rather than an empty |
| 789 | // "nothing to join" that hides what the caller asked about. |
| 790 | let settled = json!({ |
| 791 | "agent_id": snapshot.agent_id, |
| 792 | "name": snapshot.name, |
| 793 | "status": subagent_status_name(&snapshot.status), |
| 794 | "steps_taken": snapshot.steps_taken, |
| 795 | }); |
| 796 | drop(manager); |
| 797 | return wait_all_payload(&[settled], &[], 0, false); |
| 798 | } |
| 799 | vec![snapshot.agent_id] |
| 800 | } else { |
| 801 | manager |
| 802 | .list_filtered_for_session(&context.state_namespace, false) |
| 803 | .into_iter() |
| 804 | .filter(|snapshot| snapshot.status == SubAgentStatus::Running) |
| 805 | .map(|snapshot| snapshot.agent_id) |
| 806 | .collect() |
| 807 | } |
| 808 | }; |
| 809 | |
| 810 | // Zero children is an immediate return, never a hang. |
| 811 | if watched.is_empty() { |
| 812 | return wait_all_payload(&[], &[], 0, false); |
| 813 | } |
| 814 | |
| 815 | let started = Instant::now(); |
| 816 | let cancelled = async { |
| 817 | match &context.cancel_token { |
| 818 | Some(token) => token.cancelled().await, |
| 819 | None => std::future::pending().await, |
| 820 | } |
| 821 | }; |
| 822 | tokio::pin!(cancelled); |
| 823 | |
| 824 | loop { |
| 825 | let (settled, still_running) = { |
| 826 | let manager = manager.read().await; |
| 827 | let mut settled = Vec::new(); |
| 828 | let mut still_running = Vec::new(); |
| 829 | for agent_id in &watched { |
| 830 | match manager.get_result_by_ref_for_session(&context.state_namespace, agent_id) { |
| 831 | Ok(snapshot) if snapshot.status == SubAgentStatus::Running => { |
| 832 | still_running.push(json!({ |
| 833 | "agent_id": snapshot.agent_id, |
| 834 | "name": snapshot.name, |
| 835 | "status": "running", |
| 836 | })); |
| 837 | } |
| 838 | Ok(snapshot) => settled.push(json!({ |
| 839 | "agent_id": snapshot.agent_id, |
| 840 | "name": snapshot.name, |
| 841 | "status": subagent_status_name(&snapshot.status), |
| 842 | "steps_taken": snapshot.steps_taken, |
| 843 | })), |
| 844 | // A watched child that vanished from the ledger (retention |
| 845 | // cleanup) is no longer running; report it rather than |
| 846 | // blocking on a record that will never settle. |
| 847 | Err(_) => settled.push(json!({ |
| 848 | "agent_id": agent_id, |
| 849 | "status": "gone", |
| 850 | })), |
| 851 | } |
| 852 | } |
| 853 | (settled, still_running) |
| 854 | }; |
| 855 | |
| 856 | if still_running.is_empty() { |
| 857 | return wait_all_payload(&settled, &[], started.elapsed().as_millis(), false); |
| 858 | } |
| 859 | if started.elapsed() >= timeout { |
| 860 | return wait_all_payload( |
| 861 | &settled, |
| 862 | &still_running, |
| 863 | started.elapsed().as_millis(), |
| 864 | true, |
| 865 | ); |
| 866 | } |
| 867 | |
| 868 | tokio::select! { |
| 869 | biased; |
| 870 | () = &mut cancelled => { |
| 871 | return Err(ToolError::cancelled( |
| 872 | "Wait interrupted by user cancellation before every child settled.".to_string(), |
| 873 | )); |
| 874 | } |
| 875 | () = tokio::time::sleep(COORD_WAIT_CHECK_INTERVAL) => {} |
| 876 | } |
| 877 | } |
| 878 | } |
| 879 | |
| 880 | /// `until=all` result: every watched child with its own outcome, so the parent |
| 881 | /// can synthesize from one return instead of re-inspecting each child. |
| 882 | fn wait_all_payload( |
| 883 | settled: &[Value], |
| 884 | still_running: &[Value], |
| 885 | waited_ms: u128, |
| 886 | timed_out: bool, |
| 887 | ) -> Result<ToolResult, ToolError> { |
| 888 | let note = if timed_out { |
| 889 | "The wait interval ended; the children are still running. You may answer the user or continue other work. Ordinary turn completion keeps them running; results arrive as <codewhale:subagent.done> sentinels. Use followup only when a child actually needs continuation." |
| 890 | } else if settled.is_empty() { |
| 891 | "No sub-agents were running; nothing to join." |
| 892 | } else { |
| 893 | "Every watched child has settled. Full results arrive as <codewhale:subagent.done> sentinels — synthesize from those." |
| 894 | }; |
| 895 | let payload = json!({ |
| 896 | "action": "wait", |
| 897 | "until": "all", |
| 898 | "all_settled": still_running.is_empty(), |
| 899 | "settled": settled, |
| 900 | "still_running": still_running, |
| 901 | "waited_ms": u64::try_from(waited_ms).unwrap_or(u64::MAX), |
| 902 | "timed_out": timed_out, |
| 903 | "note": note, |
| 904 | }); |
| 905 | let mut tool_result = |
| 906 | ToolResult::json(&payload).map_err(|err| ToolError::execution_failed(err.to_string()))?; |
| 907 | tool_result.metadata = Some(json!({ |
| 908 | "action": "wait", |
| 909 | "until": "all", |
| 910 | "all_settled": still_running.is_empty(), |
| 911 | "settled": settled.len(), |
| 912 | "running": still_running.len(), |
| 913 | "timed_out": timed_out, |
| 914 | })); |
| 915 | Ok(tool_result) |
| 916 | } |
| 917 | |
| 918 | async fn wait_for_activity( |
| 919 | input: &Value, |
| 920 | manager: SharedSubAgentManager, |
| 921 | context: &ToolContext, |
| 922 | ) -> Result<ToolResult, ToolError> { |
| 923 | let timeout_secs = input |
| 924 | .get("timeout_secs") |
| 925 | .or_else(|| input.get("timeout")) |
| 926 | .and_then(Value::as_u64) |
| 927 | .unwrap_or(COORD_WAIT_DEFAULT_TIMEOUT_SECS) |
| 928 | .clamp(COORD_WAIT_MIN_TIMEOUT_SECS, COORD_WAIT_MAX_TIMEOUT_SECS); |
| 929 | let timeout = Duration::from_secs(timeout_secs); |
| 930 | let agent_ref = parse_agent_ref(input)?; |
| 931 | |
| 932 | let (watched, baseline): (Vec<String>, Vec<(String, u64)>) = { |
| 933 | let manager = manager.read().await; |
| 934 | if let Some(agent_ref) = &agent_ref { |
| 935 | let snap = manager |
| 936 | .get_result_by_ref_for_session(&context.state_namespace, agent_ref) |
| 937 | .map_err(|err| ToolError::invalid_input(err.to_string()))?; |
| 938 | let fp = manager.activity_fingerprint(&snap.agent_id).unwrap_or(0); |
| 939 | if snap.status != SubAgentStatus::Running { |
| 940 | let payload = json!({ |
| 941 | "action": "wait", |
| 942 | "until": "activity", |
| 943 | "reason": "already_settled", |
| 944 | "timed_out": false, |
| 945 | "agent_id": snap.agent_id, |
| 946 | "status": subagent_status_name(&snap.status), |
| 947 | }); |
| 948 | let mut tool_result = ToolResult::json(&payload) |
| 949 | .map_err(|err| ToolError::execution_failed(err.to_string()))?; |
| 950 | tool_result.metadata = Some(json!({ "action": "wait", "timed_out": false })); |
| 951 | return Ok(tool_result); |
| 952 | } |
| 953 | (vec![snap.agent_id.clone()], vec![(snap.agent_id, fp)]) |
| 954 | } else { |
| 955 | let running = manager |
| 956 | .list_filtered_for_session(&context.state_namespace, false) |
| 957 | .into_iter() |
| 958 | .filter(|s| s.status == SubAgentStatus::Running) |
| 959 | .map(|s| s.agent_id) |
| 960 | .collect::<Vec<_>>(); |
| 961 | let baseline = running |
| 962 | .iter() |
| 963 | .map(|id| { |
| 964 | let fp = manager.activity_fingerprint(id).unwrap_or(0); |
| 965 | (id.clone(), fp) |
| 966 | }) |
| 967 | .collect(); |
| 968 | (running, baseline) |
| 969 | } |
| 970 | }; |
| 971 | |
| 972 | if watched.is_empty() { |
| 973 | let payload = json!({ |
| 974 | "action": "wait", |
| 975 | "until": "activity", |
| 976 | "note": "No running sub-agents; nothing to wait for.", |
| 977 | "timed_out": false, |
| 978 | }); |
| 979 | let mut tool_result = ToolResult::json(&payload) |
| 980 | .map_err(|err| ToolError::execution_failed(err.to_string()))?; |
| 981 | tool_result.metadata = Some(json!({ "action": "wait", "timed_out": false })); |
| 982 | return Ok(tool_result); |
| 983 | } |
| 984 | |
| 985 | let started = Instant::now(); |
| 986 | let cancelled = async { |
| 987 | match &context.cancel_token { |
| 988 | Some(token) => token.cancelled().await, |
| 989 | None => std::future::pending().await, |
| 990 | } |
| 991 | }; |
| 992 | tokio::pin!(cancelled); |
| 993 | |
| 994 | loop { |
| 995 | let outcome = { |
| 996 | let manager = manager.read().await; |
| 997 | let mut settled = Vec::new(); |
| 998 | let mut activity = Vec::new(); |
| 999 | for (id, base_fp) in &baseline { |
| 1000 | if let Ok(snap) = |
| 1001 | manager.get_result_by_ref_for_session(&context.state_namespace, id) |
| 1002 | { |
| 1003 | if snap.status != SubAgentStatus::Running { |
| 1004 | settled.push(snap); |
| 1005 | continue; |
| 1006 | } |
| 1007 | let fp = manager.activity_fingerprint(id).unwrap_or(0); |
| 1008 | if fp != *base_fp { |
| 1009 | activity.push(json!({ |
| 1010 | "agent_id": id, |
| 1011 | "status": "running", |
| 1012 | "activity_fingerprint": fp, |
| 1013 | })); |
| 1014 | } |
| 1015 | } |
| 1016 | } |
| 1017 | ( |
| 1018 | settled, |
| 1019 | activity, |
| 1020 | manager.running_count_for_session(&context.state_namespace), |
| 1021 | ) |
| 1022 | }; |
| 1023 | |
| 1024 | if !outcome.0.is_empty() || !outcome.1.is_empty() { |
| 1025 | let payload = json!({ |
| 1026 | "action": "wait", |
| 1027 | "until": "activity", |
| 1028 | "settled": outcome.0.iter().map(|s| json!({ |
| 1029 | "agent_id": s.agent_id, |
| 1030 | "status": subagent_status_name(&s.status), |
| 1031 | })).collect::<Vec<_>>(), |
| 1032 | "activity": outcome.1, |
| 1033 | "running": outcome.2, |
| 1034 | "elapsed_ms": started.elapsed().as_millis(), |
| 1035 | "timed_out": false, |
| 1036 | }); |
| 1037 | let mut tool_result = ToolResult::json(&payload) |
| 1038 | .map_err(|err| ToolError::execution_failed(err.to_string()))?; |
| 1039 | tool_result.metadata = Some(json!({ |
| 1040 | "action": "wait", |
| 1041 | "timed_out": false, |
| 1042 | "settled": outcome.0.len(), |
| 1043 | "activity": outcome.1.len(), |
| 1044 | })); |
| 1045 | return Ok(tool_result); |
| 1046 | } |
| 1047 | |
| 1048 | if started.elapsed() >= timeout { |
| 1049 | let payload = json!({ |
| 1050 | "action": "wait", |
| 1051 | "until": "activity", |
| 1052 | "settled": [], |
| 1053 | "activity": [], |
| 1054 | "running": outcome.2, |
| 1055 | "elapsed_ms": started.elapsed().as_millis(), |
| 1056 | "timed_out": true, |
| 1057 | "note": "The wait interval ended without new child activity. Children keep running after ordinary turn completion and report through <codewhale:subagent.done> sentinels.", |
| 1058 | }); |
| 1059 | let mut tool_result = ToolResult::json(&payload) |
| 1060 | .map_err(|err| ToolError::execution_failed(err.to_string()))?; |
| 1061 | tool_result.metadata = Some(json!({ "action": "wait", "timed_out": true })); |
| 1062 | return Ok(tool_result); |
| 1063 | } |
| 1064 | |
| 1065 | tokio::select! { |
| 1066 | biased; |
| 1067 | () = &mut cancelled => { |
| 1068 | return Err(ToolError::cancelled( |
| 1069 | "Wait interrupted by user cancellation before child activity.".to_string(), |
| 1070 | )); |
| 1071 | } |
| 1072 | () = tokio::time::sleep(COORD_WAIT_CHECK_INTERVAL) => {} |
| 1073 | } |
| 1074 | } |
| 1075 | } |
| 1076 | |
| 1077 | /// Register the narrow coordination tools alongside `agent`. |
| 1078 | pub fn register_coordination_tools( |
| 1079 | builder: ToolRegistryBuilder, |
| 1080 | manager: SharedSubAgentManager, |
| 1081 | runtime: SubAgentRuntime, |
| 1082 | ) -> ToolRegistryBuilder { |
| 1083 | // `runtime.parent_agent_id` is the identity of the agent this registry is |
| 1084 | // being built FOR: `runtime_for_nested_agent_tools` stamps the child's own |
| 1085 | // id there before `new_with_owner` registers tools, so anything that agent |
| 1086 | // spawns records it as parent. Thread that identity through every mutating |
| 1087 | // hierarchy tool: a child may control only its own descendants, while the |
| 1088 | // root registry (`None`) may control any child (TUI-DOG-017). |
| 1089 | let caller = runtime.parent_agent_id.clone(); |
| 1090 | let message = AgentsMessageTool::new(Arc::clone(&manager)).with_optional_caller(caller.clone()); |
| 1091 | let followup = AgentsFollowupTool::new(Arc::clone(&manager)) |
| 1092 | .with_optional_caller(caller.clone()) |
| 1093 | .with_runtime(runtime.clone()); |
| 1094 | let interrupt = |
| 1095 | AgentsInterruptTool::new(Arc::clone(&manager)).with_optional_caller(caller.clone()); |
| 1096 | let coordinate = AgentsCoordinateTool::new(Arc::clone(&manager), caller); |
| 1097 | builder |
| 1098 | .with_tool(Arc::new(AgentsListTool::new(Arc::clone(&manager)))) |
| 1099 | .with_tool(Arc::new(message)) |
| 1100 | .with_tool(Arc::new(followup)) |
| 1101 | .with_tool(Arc::new(interrupt)) |
| 1102 | .with_tool(Arc::new(coordinate)) |
| 1103 | .with_tool(Arc::new(AgentsWaitTool::new(manager))) |
| 1104 | } |
| 1105 | |
| 1106 | pub struct AgentsCoordinateTool { |
| 1107 | manager: SharedSubAgentManager, |
| 1108 | caller: Option<String>, |
| 1109 | } |
| 1110 | |
| 1111 | impl AgentsCoordinateTool { |
| 1112 | #[must_use] |
| 1113 | pub fn new(manager: SharedSubAgentManager, caller: Option<String>) -> Self { |
| 1114 | Self { manager, caller } |
| 1115 | } |
| 1116 | } |
| 1117 | |
| 1118 | #[async_trait] |
| 1119 | impl ToolSpec for AgentsCoordinateTool { |
| 1120 | fn model_visible(&self) -> bool { |
| 1121 | // #5462: `agent` is the sole model-facing sub-agent surface. These |
| 1122 | // narrow tools stay registered and executable by name so a persisted |
| 1123 | // transcript replays byte-for-byte, but they are never advertised in |
| 1124 | // the catalog and can never be returned by `tool_search` — the same |
| 1125 | // shape `rlm` and `exec_shell` already use. |
| 1126 | false |
| 1127 | } |
| 1128 | |
| 1129 | fn name(&self) -> &'static str { |
| 1130 | "agents/coordinate" |
| 1131 | } |
| 1132 | |
| 1133 | fn description(&self) -> &'static str { |
| 1134 | "Record or inspect bounded coordination state: propose/accept/supersede decisions, expand the caller's write claim before mutation, reconcile multiple decision records into one neutral fan-in receipt, or release stale write-claims whose owner is no longer running." |
| 1135 | } |
| 1136 | |
| 1137 | fn input_schema(&self) -> Value { |
| 1138 | json!({ |
| 1139 | "type": "object", |
| 1140 | "properties": { |
| 1141 | "action": { "type": "string", "enum": ["inspect", "propose", "accept", "supersede", "claim", "reconcile", "release"] }, |
| 1142 | "decision_id": { "type": "string" }, |
| 1143 | "subject": { "type": "string" }, |
| 1144 | "expected_version": { "type": "integer", "minimum": 1 }, |
| 1145 | "scope": { "type": "array", "items": { "type": "string" } }, |
| 1146 | "constraints": { "type": "array", "items": { "type": "string" } }, |
| 1147 | "evidence_handles": { "type": "array", "items": { "type": "string" } }, |
| 1148 | "roots": { "type": "array", "items": { "type": "string" } }, |
| 1149 | "exact_files": { "type": "array", "items": { "type": "string" } }, |
| 1150 | "contracts": { "type": "array", "items": { "type": "string" } }, |
| 1151 | "owner": { "type": "string" }, |
| 1152 | "input_decisions": { "type": "array", "items": { "type": "string" } }, |
| 1153 | "outcome": { "type": "string" }, |
| 1154 | "candidate_handles": { "type": "array", "items": { "type": "string" } }, |
| 1155 | "retry_count": { "type": "integer", "minimum": 0, "maximum": 3 }, |
| 1156 | "retry_limit": { "type": "integer", "minimum": 1, "maximum": 3 }, |
| 1157 | "reviewer_evidence_handles": { "type": "array", "items": { "type": "string" } }, |
| 1158 | "verifier_evidence_handles": { "type": "array", "items": { "type": "string" } }, |
| 1159 | "verification_outcome": { "type": "string" }, |
| 1160 | "limit": { "type": "integer", "minimum": 1, "maximum": 24 } |
| 1161 | }, |
| 1162 | "required": ["action"] |
| 1163 | }) |
| 1164 | } |
| 1165 | |
| 1166 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 1167 | // #5123-class: this tool mutates the coordination ledger and expands |
| 1168 | // the caller's write claim (actions propose/accept/supersede/claim/ |
| 1169 | // reconcile) — declaring ReadOnly was a lie that let policy layers |
| 1170 | // treat a mutating call as a safe read. Only `inspect` is read-only, |
| 1171 | // which is what is_read_only_for reports. |
| 1172 | vec![ToolCapability::WritesFiles] |
| 1173 | } |
| 1174 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 1175 | // Stays Auto: coordination records are session-scoped in-memory |
| 1176 | // state, and gating them would deadlock autonomous sub-agent fan-in. |
| 1177 | ApprovalRequirement::Auto |
| 1178 | } |
| 1179 | fn is_read_only_for(&self, input: &Value) -> bool { |
| 1180 | input.get("action").and_then(Value::as_str) == Some("inspect") |
| 1181 | } |
| 1182 | |
| 1183 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 1184 | let action = input |
| 1185 | .get("action") |
| 1186 | .and_then(Value::as_str) |
| 1187 | .unwrap_or("inspect"); |
| 1188 | let bounded_text = |key: &str| { |
| 1189 | input |
| 1190 | .get(key) |
| 1191 | .and_then(Value::as_str) |
| 1192 | .map(|value| value.chars().take(512).collect::<String>()) |
| 1193 | }; |
| 1194 | // Tool authority is the runtime caller identity. Root cannot supply an |
| 1195 | // arbitrary child owner and mutate that child's decisions/claim. |
| 1196 | let owner = self.caller.clone().unwrap_or_else(|| "root".to_string()); |
| 1197 | let strings = |key: &str| { |
| 1198 | input |
| 1199 | .get(key) |
| 1200 | .and_then(Value::as_array) |
| 1201 | .map(|items| { |
| 1202 | items |
| 1203 | .iter() |
| 1204 | .take(24) |
| 1205 | .filter_map(Value::as_str) |
| 1206 | .map(|value| value.chars().take(512).collect::<String>()) |
| 1207 | .collect::<Vec<_>>() |
| 1208 | }) |
| 1209 | .unwrap_or_default() |
| 1210 | }; |
| 1211 | if action == "inspect" { |
| 1212 | let manager = self.manager.read().await; |
| 1213 | let value = manager.inspect_coordination_for_session( |
| 1214 | &context.state_namespace, |
| 1215 | bounded_text("subject").as_deref(), |
| 1216 | input |
| 1217 | .get("limit") |
| 1218 | .and_then(Value::as_u64) |
| 1219 | .unwrap_or(COORDINATION_INSPECT_LIMIT as u64) as usize, |
| 1220 | ); |
| 1221 | return ToolResult::json(&value) |
| 1222 | .map_err(|e| ToolError::execution_failed(e.to_string())); |
| 1223 | } |
| 1224 | if !matches!( |
| 1225 | action, |
| 1226 | "propose" | "accept" | "supersede" | "claim" | "reconcile" | "release" |
| 1227 | ) { |
| 1228 | return Err(ToolError::invalid_input(format!( |
| 1229 | "unknown coordination action '{action}'" |
| 1230 | ))); |
| 1231 | } |
| 1232 | |
| 1233 | let mut manager = self.manager.write().await; |
| 1234 | if let Some(caller) = self.caller.as_deref() { |
| 1235 | manager |
| 1236 | .get_result_by_ref_for_session(&context.state_namespace, caller) |
| 1237 | .map_err(|_| { |
| 1238 | ToolError::invalid_input("Agent not found in the active session".to_string()) |
| 1239 | })?; |
| 1240 | } |
| 1241 | if matches!(action, "accept" | "supersede") { |
| 1242 | let decision_id = bounded_text("decision_id").unwrap_or_default(); |
| 1243 | if !manager |
| 1244 | .coordination_decision_is_owned_by_session(&context.state_namespace, &decision_id) |
| 1245 | { |
| 1246 | return Err(ToolError::invalid_input( |
| 1247 | "Coordination decision not found in the active session".to_string(), |
| 1248 | )); |
| 1249 | } |
| 1250 | } |
| 1251 | if action == "reconcile" |
| 1252 | && strings("input_decisions").iter().any(|decision_id| { |
| 1253 | !manager.coordination_decision_is_owned_by_session( |
| 1254 | &context.state_namespace, |
| 1255 | decision_id, |
| 1256 | ) |
| 1257 | }) |
| 1258 | { |
| 1259 | return Err(ToolError::invalid_input( |
| 1260 | "One or more coordination decisions were not found in the active session" |
| 1261 | .to_string(), |
| 1262 | )); |
| 1263 | } |
| 1264 | let coordination_before = manager.coordination.clone(); |
| 1265 | let mutation = match action { |
| 1266 | "propose" => manager |
| 1267 | .record_coordination_decision(DecisionRecord { |
| 1268 | decision_id: bounded_text("decision_id").unwrap_or_default(), |
| 1269 | subject: bounded_text("subject").unwrap_or_default(), |
| 1270 | status: DecisionStatus::Proposed, |
| 1271 | owner, |
| 1272 | scope: strings("scope"), |
| 1273 | constraints: strings("constraints"), |
| 1274 | evidence_handles: strings("evidence_handles"), |
| 1275 | version: 1, |
| 1276 | sequence: 0, |
| 1277 | }) |
| 1278 | .map_err(ToolError::invalid_input) |
| 1279 | .and_then(|record| { |
| 1280 | serde_json::to_value(record) |
| 1281 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 1282 | }), |
| 1283 | "accept" | "supersede" => input |
| 1284 | .get("expected_version") |
| 1285 | .and_then(Value::as_u64) |
| 1286 | .and_then(|value| u32::try_from(value).ok()) |
| 1287 | .ok_or_else(|| { |
| 1288 | ToolError::invalid_input( |
| 1289 | "accept/supersede requires expected_version".to_string(), |
| 1290 | ) |
| 1291 | }) |
| 1292 | .and_then(|expected_version| { |
| 1293 | manager |
| 1294 | .update_coordination_decision( |
| 1295 | &bounded_text("decision_id").unwrap_or_default(), |
| 1296 | if action == "accept" { |
| 1297 | DecisionStatus::Accepted |
| 1298 | } else { |
| 1299 | DecisionStatus::Superseded |
| 1300 | }, |
| 1301 | &owner, |
| 1302 | expected_version, |
| 1303 | ) |
| 1304 | .map_err(ToolError::invalid_input) |
| 1305 | }) |
| 1306 | .and_then(|record| { |
| 1307 | serde_json::to_value(record) |
| 1308 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 1309 | }), |
| 1310 | "claim" => manager |
| 1311 | .expand_write_claim( |
| 1312 | &owner, |
| 1313 | strings("roots"), |
| 1314 | strings("exact_files"), |
| 1315 | strings("contracts"), |
| 1316 | ) |
| 1317 | .map_err(ToolError::invalid_input) |
| 1318 | .and_then(|claim| { |
| 1319 | serde_json::to_value(claim) |
| 1320 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 1321 | }), |
| 1322 | "reconcile" => manager |
| 1323 | .reconcile_coordination( |
| 1324 | bounded_text("subject").unwrap_or_default(), |
| 1325 | owner, |
| 1326 | strings("input_decisions"), |
| 1327 | bounded_text("outcome").unwrap_or_default(), |
| 1328 | strings("evidence_handles"), |
| 1329 | strings("candidate_handles"), |
| 1330 | input |
| 1331 | .get("retry_count") |
| 1332 | .and_then(Value::as_u64) |
| 1333 | .and_then(|value| u32::try_from(value).ok()) |
| 1334 | .unwrap_or_default(), |
| 1335 | input |
| 1336 | .get("retry_limit") |
| 1337 | .and_then(Value::as_u64) |
| 1338 | .and_then(|value| u32::try_from(value).ok()) |
| 1339 | .unwrap_or(MAX_RECONCILIATION_RETRIES), |
| 1340 | strings("reviewer_evidence_handles"), |
| 1341 | strings("verifier_evidence_handles"), |
| 1342 | bounded_text("verification_outcome").unwrap_or_default(), |
| 1343 | ) |
| 1344 | .map_err(ToolError::invalid_input) |
| 1345 | .and_then(|receipt| { |
| 1346 | serde_json::to_value(receipt) |
| 1347 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 1348 | }), |
| 1349 | "release" => { |
| 1350 | let owner = input |
| 1351 | .get("owner") |
| 1352 | .and_then(Value::as_str) |
| 1353 | .map(|value| value.to_string()); |
| 1354 | let released = manager |
| 1355 | .release_stale_write_claims(owner) |
| 1356 | .map_err(ToolError::invalid_input)?; |
| 1357 | let sequence = manager.coordination.sequence; |
| 1358 | serde_json::to_value(json!({ |
| 1359 | "released": released.len(), |
| 1360 | "owners": released, |
| 1361 | "sequence": sequence |
| 1362 | })) |
| 1363 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 1364 | } |
| 1365 | _ => unreachable!("coordination action validated above"), |
| 1366 | }; |
| 1367 | let value = match mutation { |
| 1368 | Ok(value) => value, |
| 1369 | Err(error) => { |
| 1370 | // Contention failures deliberately append a durable receipt. |
| 1371 | // Stamp and persist every sequence allocated by the failed |
| 1372 | // action before returning its error; validation failures that |
| 1373 | // did not mutate the ledger allocate nothing. |
| 1374 | let first_new_sequence = coordination_before.sequence.saturating_add(1); |
| 1375 | let last_new_sequence = manager.coordination.sequence; |
| 1376 | for sequence in first_new_sequence..=last_new_sequence { |
| 1377 | if let Err(stamp_error) = manager |
| 1378 | .stamp_coordination_sequence_for_session(sequence, &context.state_namespace) |
| 1379 | { |
| 1380 | manager.coordination = coordination_before; |
| 1381 | return Err(ToolError::execution_failed(format!( |
| 1382 | "{error}; additionally failed to stamp coordination receipt: {stamp_error}" |
| 1383 | ))); |
| 1384 | } |
| 1385 | } |
| 1386 | if last_new_sequence >= first_new_sequence |
| 1387 | && let Err(persist_error) = manager.persist_state_synchronously() |
| 1388 | { |
| 1389 | manager.coordination = coordination_before; |
| 1390 | return Err(ToolError::execution_failed(format!( |
| 1391 | "{error}; additionally failed to persist coordination receipt: {persist_error}" |
| 1392 | ))); |
| 1393 | } |
| 1394 | return Err(error); |
| 1395 | } |
| 1396 | }; |
| 1397 | let Some(sequence) = value.get("sequence").and_then(Value::as_u64) else { |
| 1398 | manager.coordination = coordination_before; |
| 1399 | return Err(ToolError::execution_failed(format!( |
| 1400 | "coordination action '{action}' produced no durable sequence" |
| 1401 | ))); |
| 1402 | }; |
| 1403 | if let Err(error) = |
| 1404 | manager.stamp_coordination_sequence_for_session(sequence, &context.state_namespace) |
| 1405 | { |
| 1406 | manager.coordination = coordination_before; |
| 1407 | return Err(ToolError::execution_failed(error)); |
| 1408 | } |
| 1409 | if let Err(error) = manager.persist_state_synchronously() { |
| 1410 | manager.coordination = coordination_before; |
| 1411 | return Err(ToolError::execution_failed(format!( |
| 1412 | "failed to persist coordination action '{action}': {error}" |
| 1413 | ))); |
| 1414 | } |
| 1415 | ToolResult::json(&value).map_err(|e| ToolError::execution_failed(e.to_string())) |
| 1416 | } |
| 1417 | } |
| 1418 | |
| 1419 | #[cfg(test)] |
| 1420 | mod tests { |
| 1421 | use super::*; |
| 1422 | use crate::tools::spec::ToolContext; |
| 1423 | use codewhale_models::Role; |
| 1424 | use std::collections::BTreeSet; |
| 1425 | use tempfile::tempdir; |
| 1426 | |
| 1427 | #[test] |
| 1428 | fn coordinate_tool_does_not_declare_read_only() { |
| 1429 | // #5123-class: the tool mutates the coordination ledger and expands |
| 1430 | // write claims; its declared capabilities must not say ReadOnly. |
| 1431 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 1432 | super::super::SubAgentManager::new(std::path::PathBuf::from("."), 1), |
| 1433 | )); |
| 1434 | let tool = AgentsCoordinateTool::new(manager, None); |
| 1435 | let capabilities = ToolSpec::capabilities(&tool); |
| 1436 | assert!( |
| 1437 | !capabilities.contains(&ToolCapability::ReadOnly), |
| 1438 | "agents/coordinate mutates the ledger — ReadOnly is a lie: {capabilities:?}" |
| 1439 | ); |
| 1440 | // …but the dynamic check still marks inspect as read-only. |
| 1441 | assert!(tool.is_read_only_for(&json!({"action": "inspect"}))); |
| 1442 | assert!(!tool.is_read_only_for(&json!({"action": "propose"}))); |
| 1443 | } |
| 1444 | |
| 1445 | #[test] |
| 1446 | fn coordination_descriptions_match_implemented_resume_behavior() { |
| 1447 | // Checkpoint resume is implemented (#5242): the descriptions must |
| 1448 | // describe the real behavior, including the honest queue-only |
| 1449 | // fallback when no runtime is attached. |
| 1450 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 1451 | super::super::SubAgentManager::new(std::env::temp_dir(), 1), |
| 1452 | )); |
| 1453 | let message = AgentsMessageTool::new(Arc::clone(&manager)); |
| 1454 | let followup = AgentsFollowupTool::new(manager); |
| 1455 | |
| 1456 | assert!(!message.description().contains("natural resume")); |
| 1457 | assert!(message.description().contains("stays queued")); |
| 1458 | assert!(followup.description().contains("attempt to resume")); |
| 1459 | assert!( |
| 1460 | followup |
| 1461 | .description() |
| 1462 | .contains("resumed from their checkpoint") |
| 1463 | ); |
| 1464 | assert!(followup.description().contains("queue-only semantics")); |
| 1465 | } |
| 1466 | |
| 1467 | async fn manager_with_running_child( |
| 1468 | workspace: &std::path::Path, |
| 1469 | ) -> (SharedSubAgentManager, String) { |
| 1470 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 1471 | super::super::SubAgentManager::new(workspace.to_path_buf(), 4), |
| 1472 | )); |
| 1473 | let agent_id = { |
| 1474 | let mut guard = manager.write().await; |
| 1475 | guard.insert_test_running_agent("coord_child", workspace) |
| 1476 | }; |
| 1477 | (manager, agent_id) |
| 1478 | } |
| 1479 | |
| 1480 | async fn manager_with_agent_hierarchy( |
| 1481 | workspace: &std::path::Path, |
| 1482 | ) -> (SharedSubAgentManager, String, String, String) { |
| 1483 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 1484 | super::super::SubAgentManager::new(workspace.to_path_buf(), 8), |
| 1485 | )); |
| 1486 | let (parent, child, sibling) = { |
| 1487 | let mut guard = manager.write().await; |
| 1488 | let parent = guard.insert_test_running_agent("hierarchy_parent", workspace); |
| 1489 | let child = guard.insert_test_running_agent("hierarchy_child", workspace); |
| 1490 | let sibling = guard.insert_test_running_agent("hierarchy_sibling", workspace); |
| 1491 | for (agent_id, parent_id) in [ |
| 1492 | (&parent, "root"), |
| 1493 | (&child, parent.as_str()), |
| 1494 | (&sibling, "root"), |
| 1495 | ] { |
| 1496 | let record = guard |
| 1497 | .worker_records |
| 1498 | .get_mut(agent_id) |
| 1499 | .expect("hierarchy worker record"); |
| 1500 | record.parent_run_id = Some(parent_id.to_string()); |
| 1501 | record.spec.parent_run_id = Some(parent_id.to_string()); |
| 1502 | } |
| 1503 | (parent, child, sibling) |
| 1504 | }; |
| 1505 | (manager, parent, child, sibling) |
| 1506 | } |
| 1507 | |
| 1508 | #[tokio::test] |
| 1509 | async fn message_queues_without_waking() { |
| 1510 | let tmp = tempdir().unwrap(); |
| 1511 | let (manager, agent_id) = manager_with_running_child(tmp.path()).await; |
| 1512 | let tool = AgentsMessageTool::new(Arc::clone(&manager)); |
| 1513 | let result = tool |
| 1514 | .execute( |
| 1515 | json!({ "agent_id": agent_id, "message": "hold this" }), |
| 1516 | &ToolContext::new(tmp.path()), |
| 1517 | ) |
| 1518 | .await |
| 1519 | .expect("message ok"); |
| 1520 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 1521 | assert_eq!(body["woke"], json!(false)); |
| 1522 | assert_eq!(body["queued"], json!(true)); |
| 1523 | assert_eq!(body["queue_depth"], json!(1)); |
| 1524 | |
| 1525 | let guard = manager.read().await; |
| 1526 | let depth = guard.queued_mail_depth(&agent_id).unwrap(); |
| 1527 | assert_eq!(depth, 1); |
| 1528 | assert!(!guard.child_was_woken(&agent_id)); |
| 1529 | } |
| 1530 | |
| 1531 | #[tokio::test] |
| 1532 | async fn followup_does_not_claim_wake_when_live_channel_is_closed() { |
| 1533 | let tmp = tempdir().unwrap(); |
| 1534 | let (manager, agent_id) = manager_with_running_child(tmp.path()).await; |
| 1535 | let result = AgentsFollowupTool::new(Arc::clone(&manager)) |
| 1536 | .execute( |
| 1537 | json!({ "agent_id": agent_id, "message": "try to wake" }), |
| 1538 | &ToolContext::new(tmp.path()), |
| 1539 | ) |
| 1540 | .await |
| 1541 | .expect("truthful closed-channel receipt"); |
| 1542 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 1543 | assert_eq!(body["woke"], json!(false)); |
| 1544 | assert_eq!(body["queue_depth"], json!(1)); |
| 1545 | assert!( |
| 1546 | body["note"].as_str().unwrap_or_default().contains("closed"), |
| 1547 | "{body}" |
| 1548 | ); |
| 1549 | |
| 1550 | let guard = manager.read().await; |
| 1551 | assert_eq!(guard.queued_mail_depth(&agent_id), Some(1)); |
| 1552 | assert!(!guard.child_was_woken(&agent_id)); |
| 1553 | } |
| 1554 | |
| 1555 | #[tokio::test] |
| 1556 | async fn hierarchy_mutations_allow_own_descendants_and_deny_siblings_or_ancestors() { |
| 1557 | let tmp = tempdir().unwrap(); |
| 1558 | let (manager, parent, child, sibling) = manager_with_agent_hierarchy(tmp.path()).await; |
| 1559 | let context = ToolContext::new(tmp.path()); |
| 1560 | |
| 1561 | AgentsMessageTool::new(Arc::clone(&manager)) |
| 1562 | .with_optional_caller(Some(parent.clone())) |
| 1563 | .execute( |
| 1564 | json!({ "agent_id": child, "message": "bounded parent note" }), |
| 1565 | &context, |
| 1566 | ) |
| 1567 | .await |
| 1568 | .expect("parent may message its own child"); |
| 1569 | AgentsFollowupTool::new(Arc::clone(&manager)) |
| 1570 | .with_optional_caller(Some(parent.clone())) |
| 1571 | .execute( |
| 1572 | json!({ "agent_id": child, "message": "resume own child" }), |
| 1573 | &context, |
| 1574 | ) |
| 1575 | .await |
| 1576 | .expect("parent may follow up its own child"); |
| 1577 | |
| 1578 | let sibling_message = AgentsMessageTool::new(Arc::clone(&manager)) |
| 1579 | .with_optional_caller(Some(parent.clone())) |
| 1580 | .execute( |
| 1581 | json!({ "agent_id": sibling, "message": "cross branch" }), |
| 1582 | &context, |
| 1583 | ) |
| 1584 | .await |
| 1585 | .expect_err("sibling message must fail closed") |
| 1586 | .to_string(); |
| 1587 | assert!( |
| 1588 | sibling_message.contains("own descendants"), |
| 1589 | "{sibling_message}" |
| 1590 | ); |
| 1591 | |
| 1592 | let ancestor_followup = AgentsFollowupTool::new(Arc::clone(&manager)) |
| 1593 | .with_optional_caller(Some(child.clone())) |
| 1594 | .execute( |
| 1595 | json!({ "agent_id": parent, "message": "wake ancestor" }), |
| 1596 | &context, |
| 1597 | ) |
| 1598 | .await |
| 1599 | .expect_err("ancestor followup must fail closed") |
| 1600 | .to_string(); |
| 1601 | assert!( |
| 1602 | ancestor_followup.contains("own descendants"), |
| 1603 | "{ancestor_followup}" |
| 1604 | ); |
| 1605 | |
| 1606 | let sibling_interrupt = AgentsInterruptTool::new(Arc::clone(&manager)) |
| 1607 | .with_optional_caller(Some(parent.clone())) |
| 1608 | .execute(json!({ "agent_id": sibling }), &context) |
| 1609 | .await |
| 1610 | .expect_err("sibling interrupt must fail closed") |
| 1611 | .to_string(); |
| 1612 | assert!( |
| 1613 | sibling_interrupt.contains("own descendants"), |
| 1614 | "{sibling_interrupt}" |
| 1615 | ); |
| 1616 | |
| 1617 | let interrupted = AgentsInterruptTool::new(Arc::clone(&manager)) |
| 1618 | .with_optional_caller(Some(parent)) |
| 1619 | .execute(json!({ "agent_id": child }), &context) |
| 1620 | .await |
| 1621 | .expect("parent may interrupt its own child"); |
| 1622 | let body: Value = serde_json::from_str(&interrupted.content).unwrap(); |
| 1623 | assert_eq!(body["status"], json!("interrupted")); |
| 1624 | } |
| 1625 | |
| 1626 | #[tokio::test] |
| 1627 | async fn coordinate_inspect_is_side_effect_free_and_mutations_are_synchronously_durable() { |
| 1628 | let tmp = tempdir().unwrap(); |
| 1629 | let blocked_state_path = tmp.path().join("blocked-state"); |
| 1630 | std::fs::create_dir(&blocked_state_path).unwrap(); |
| 1631 | let blocked_manager = Arc::new(tokio::sync::RwLock::new( |
| 1632 | super::super::SubAgentManager::new(tmp.path().to_path_buf(), 4) |
| 1633 | .with_state_path(blocked_state_path), |
| 1634 | )); |
| 1635 | let blocked_tool = AgentsCoordinateTool::new(Arc::clone(&blocked_manager), None); |
| 1636 | |
| 1637 | blocked_tool |
| 1638 | .execute( |
| 1639 | json!({ "action": "inspect" }), |
| 1640 | &ToolContext::new(tmp.path()), |
| 1641 | ) |
| 1642 | .await |
| 1643 | .expect("read-only inspect must not attempt persistence"); |
| 1644 | let error = blocked_tool |
| 1645 | .execute( |
| 1646 | json!({ |
| 1647 | "action": "propose", |
| 1648 | "decision_id": "durable-decision", |
| 1649 | "subject": "durability", |
| 1650 | "constraints": ["persist before acknowledgement"] |
| 1651 | }), |
| 1652 | &ToolContext::new(tmp.path()), |
| 1653 | ) |
| 1654 | .await |
| 1655 | .expect_err("mutation must fail when its receipt cannot persist") |
| 1656 | .to_string(); |
| 1657 | assert!(error.contains("failed to persist"), "{error}"); |
| 1658 | assert!( |
| 1659 | blocked_manager |
| 1660 | .read() |
| 1661 | .await |
| 1662 | .coordination |
| 1663 | .decisions |
| 1664 | .is_empty(), |
| 1665 | "failed persistence must roll the in-memory decision back" |
| 1666 | ); |
| 1667 | |
| 1668 | let durable_workspace = tempdir().unwrap(); |
| 1669 | let state_path = durable_workspace.path().join("subagents.v1.json"); |
| 1670 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 1671 | super::super::SubAgentManager::new(durable_workspace.path().to_path_buf(), 4) |
| 1672 | .with_state_path(state_path.clone()), |
| 1673 | )); |
| 1674 | AgentsCoordinateTool::new(Arc::clone(&manager), None) |
| 1675 | .execute( |
| 1676 | json!({ |
| 1677 | "action": "propose", |
| 1678 | "decision_id": "durable-decision", |
| 1679 | "subject": "durability", |
| 1680 | "constraints": ["persist before acknowledgement"] |
| 1681 | }), |
| 1682 | &ToolContext::new(durable_workspace.path()), |
| 1683 | ) |
| 1684 | .await |
| 1685 | .expect("durable mutation"); |
| 1686 | let mut replayed = |
| 1687 | super::super::SubAgentManager::new(durable_workspace.path().to_path_buf(), 4) |
| 1688 | .with_state_path(state_path); |
| 1689 | replayed.load_state().expect("reload durable action"); |
| 1690 | assert_eq!(replayed.coordination.decisions.len(), 1); |
| 1691 | assert_eq!( |
| 1692 | replayed.coordination.decisions[0].decision_id, |
| 1693 | "durable-decision" |
| 1694 | ); |
| 1695 | } |
| 1696 | |
| 1697 | #[tokio::test] |
| 1698 | async fn release_action_clears_only_stale_write_claims_and_persists() { |
| 1699 | let tmp = tempdir().unwrap(); |
| 1700 | let state_path = tmp.path().join("subagents.v1.json"); |
| 1701 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 1702 | super::super::SubAgentManager::new(tmp.path().to_path_buf(), 4) |
| 1703 | .with_state_path(state_path.clone()), |
| 1704 | )); |
| 1705 | let live = { |
| 1706 | let mut guard = manager.write().await; |
| 1707 | let live = guard.insert_test_running_agent("live-builder", tmp.path()); |
| 1708 | let active = [live.clone()].into_iter().collect::<BTreeSet<_>>(); |
| 1709 | for claim in [ |
| 1710 | WriteScopeClaim { |
| 1711 | owner: live.clone(), |
| 1712 | roots: vec!["src/live".into()], |
| 1713 | exact_files: Vec::new(), |
| 1714 | contracts: Vec::new(), |
| 1715 | }, |
| 1716 | WriteScopeClaim { |
| 1717 | owner: "zombie-builder".into(), |
| 1718 | roots: vec!["src/zombie".into()], |
| 1719 | exact_files: Vec::new(), |
| 1720 | contracts: Vec::new(), |
| 1721 | }, |
| 1722 | ] { |
| 1723 | guard |
| 1724 | .coordination |
| 1725 | .register_claim(claim, false, |candidate| active.contains(candidate)) |
| 1726 | .expect("initial claim"); |
| 1727 | } |
| 1728 | let _ = guard.persist_state_synchronously(); |
| 1729 | live |
| 1730 | }; |
| 1731 | |
| 1732 | let result = AgentsCoordinateTool::new(Arc::clone(&manager), None) |
| 1733 | .execute( |
| 1734 | json!({ "action": "release" }), |
| 1735 | &ToolContext::new(tmp.path()), |
| 1736 | ) |
| 1737 | .await |
| 1738 | .expect("release sweeps stale claims"); |
| 1739 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 1740 | assert_eq!(body["released"], json!(1)); |
| 1741 | assert_eq!(body["owners"], json!(["zombie-builder"])); |
| 1742 | |
| 1743 | // The sweep is durable: reloading the ledger keeps only the live claim. |
| 1744 | let mut replayed = super::super::SubAgentManager::new(tmp.path().to_path_buf(), 4) |
| 1745 | .with_state_path(state_path); |
| 1746 | replayed.load_state().expect("reload released ledger"); |
| 1747 | assert_eq!(replayed.coordination.write_claims.len(), 1); |
| 1748 | assert_eq!(replayed.coordination.write_claims[0].claim.owner, live); |
| 1749 | |
| 1750 | // Named-owner release of a live claimant is a no-op and still succeeds. |
| 1751 | let noop = AgentsCoordinateTool::new(Arc::clone(&manager), None) |
| 1752 | .execute( |
| 1753 | json!({ "action": "release", "owner": live }), |
| 1754 | &ToolContext::new(tmp.path()), |
| 1755 | ) |
| 1756 | .await |
| 1757 | .expect("live release is a no-op"); |
| 1758 | let body: Value = serde_json::from_str(&noop.content).unwrap(); |
| 1759 | assert_eq!(body["released"], json!(0)); |
| 1760 | } |
| 1761 | |
| 1762 | #[tokio::test] |
| 1763 | async fn rejected_claim_contention_is_persisted_before_returning_the_error() { |
| 1764 | let tmp = tempdir().unwrap(); |
| 1765 | let state_path = tmp.path().join("subagents.v1.json"); |
| 1766 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 1767 | super::super::SubAgentManager::new(tmp.path().to_path_buf(), 4) |
| 1768 | .with_state_path(state_path.clone()), |
| 1769 | )); |
| 1770 | let (claimant, owner) = { |
| 1771 | let mut guard = manager.write().await; |
| 1772 | let claimant = guard.insert_test_running_agent("claimant", tmp.path()); |
| 1773 | let owner = guard.insert_test_running_agent("owner", tmp.path()); |
| 1774 | let active = [claimant.clone(), owner.clone()] |
| 1775 | .into_iter() |
| 1776 | .collect::<BTreeSet<_>>(); |
| 1777 | for claim in [ |
| 1778 | WriteScopeClaim { |
| 1779 | owner: claimant.clone(), |
| 1780 | roots: vec!["src/claimant".into()], |
| 1781 | exact_files: Vec::new(), |
| 1782 | contracts: Vec::new(), |
| 1783 | }, |
| 1784 | WriteScopeClaim { |
| 1785 | owner: owner.clone(), |
| 1786 | roots: vec!["src/shared".into()], |
| 1787 | exact_files: Vec::new(), |
| 1788 | contracts: Vec::new(), |
| 1789 | }, |
| 1790 | ] { |
| 1791 | guard |
| 1792 | .coordination |
| 1793 | .register_claim(claim, false, |candidate| active.contains(candidate)) |
| 1794 | .expect("initial non-overlapping claim"); |
| 1795 | } |
| 1796 | (claimant, owner) |
| 1797 | }; |
| 1798 | |
| 1799 | let error = AgentsCoordinateTool::new(Arc::clone(&manager), Some(claimant.clone())) |
| 1800 | .execute( |
| 1801 | json!({ "action": "claim", "roots": ["src/shared/nested"] }), |
| 1802 | &ToolContext::new(tmp.path()), |
| 1803 | ) |
| 1804 | .await |
| 1805 | .expect_err("overlap must block") |
| 1806 | .to_string(); |
| 1807 | assert!( |
| 1808 | error.contains(&owner) && error.contains("contention"), |
| 1809 | "{error}" |
| 1810 | ); |
| 1811 | |
| 1812 | let mut replayed = super::super::SubAgentManager::new(tmp.path().to_path_buf(), 4) |
| 1813 | .with_state_path(state_path); |
| 1814 | replayed.load_state().expect("reload contention receipt"); |
| 1815 | assert_eq!(replayed.coordination.contentions.len(), 1); |
| 1816 | assert_eq!(replayed.coordination.contentions[0].claimant, claimant); |
| 1817 | assert_eq!( |
| 1818 | replayed.coordination.contentions[0].conflicting_owner, |
| 1819 | owner |
| 1820 | ); |
| 1821 | } |
| 1822 | |
| 1823 | #[tokio::test] |
| 1824 | async fn coordination_resolution_survives_reload_and_resolving_claim_eviction() { |
| 1825 | let tmp = tempdir().unwrap(); |
| 1826 | let state_path = tmp.path().join("subagents.v1.json"); |
| 1827 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 1828 | super::super::SubAgentManager::new(tmp.path().to_path_buf(), 4) |
| 1829 | .with_state_path(state_path.clone()), |
| 1830 | )); |
| 1831 | let claimant = { |
| 1832 | let mut guard = manager.write().await; |
| 1833 | let claimant = guard.insert_test_running_agent("claimant", tmp.path()); |
| 1834 | let owner = guard.insert_test_running_agent("owner", tmp.path()); |
| 1835 | let active = [claimant.clone(), owner.clone()] |
| 1836 | .into_iter() |
| 1837 | .collect::<BTreeSet<_>>(); |
| 1838 | for claim in [ |
| 1839 | WriteScopeClaim { |
| 1840 | owner: claimant.clone(), |
| 1841 | roots: vec!["src/claimant".into()], |
| 1842 | exact_files: Vec::new(), |
| 1843 | contracts: Vec::new(), |
| 1844 | }, |
| 1845 | WriteScopeClaim { |
| 1846 | owner: owner.clone(), |
| 1847 | roots: vec!["src/shared".into()], |
| 1848 | exact_files: Vec::new(), |
| 1849 | contracts: Vec::new(), |
| 1850 | }, |
| 1851 | ] { |
| 1852 | guard |
| 1853 | .coordination |
| 1854 | .register_claim(claim, false, |candidate| active.contains(candidate)) |
| 1855 | .expect("initial non-overlapping claim"); |
| 1856 | } |
| 1857 | claimant |
| 1858 | }; |
| 1859 | |
| 1860 | AgentsCoordinateTool::new(Arc::clone(&manager), Some(claimant.clone())) |
| 1861 | .execute( |
| 1862 | json!({ "action": "claim", "roots": ["src/shared/nested"] }), |
| 1863 | &ToolContext::new(tmp.path()), |
| 1864 | ) |
| 1865 | .await |
| 1866 | .expect_err("overlap must block and persist its receipt"); |
| 1867 | |
| 1868 | let resolution_sequence = { |
| 1869 | let mut guard = manager.write().await; |
| 1870 | let record = guard |
| 1871 | .coordination |
| 1872 | .register_claim( |
| 1873 | WriteScopeClaim { |
| 1874 | owner: claimant.clone(), |
| 1875 | roots: vec!["src/isolated".into()], |
| 1876 | exact_files: Vec::new(), |
| 1877 | contracts: Vec::new(), |
| 1878 | }, |
| 1879 | true, |
| 1880 | |_| true, |
| 1881 | ) |
| 1882 | .expect("later isolated claim resolves contention"); |
| 1883 | guard |
| 1884 | .persist_state_synchronously() |
| 1885 | .expect("persist resolved contention"); |
| 1886 | record.sequence |
| 1887 | }; |
| 1888 | |
| 1889 | let mut replayed = super::super::SubAgentManager::new(tmp.path().to_path_buf(), 4) |
| 1890 | .with_state_path(state_path.clone()); |
| 1891 | replayed.load_state().expect("reload resolved contention"); |
| 1892 | assert_eq!(replayed.coordination.contentions.len(), 1); |
| 1893 | assert_eq!( |
| 1894 | replayed.coordination.contentions[0].disposition, |
| 1895 | WriteContentionDisposition::ResolvedBySuccessfulClaim |
| 1896 | ); |
| 1897 | assert_eq!( |
| 1898 | replayed.coordination.contentions[0].resolution_sequence, |
| 1899 | Some(resolution_sequence) |
| 1900 | ); |
| 1901 | |
| 1902 | let slots = COORDINATION_RECORD_LIMIT - replayed.coordination.write_claims.len(); |
| 1903 | for index in 0..slots { |
| 1904 | replayed |
| 1905 | .coordination |
| 1906 | .register_claim( |
| 1907 | WriteScopeClaim { |
| 1908 | owner: format!("inactive-fill-{index:03}"), |
| 1909 | roots: vec![format!("pkg/fill-{index:03}")], |
| 1910 | exact_files: Vec::new(), |
| 1911 | contracts: Vec::new(), |
| 1912 | }, |
| 1913 | true, |
| 1914 | |_| false, |
| 1915 | ) |
| 1916 | .expect("fill inactive claim capacity"); |
| 1917 | } |
| 1918 | for index in 0..2 { |
| 1919 | replayed |
| 1920 | .coordination |
| 1921 | .register_claim( |
| 1922 | WriteScopeClaim { |
| 1923 | owner: format!("inactive-overflow-{index}"), |
| 1924 | roots: vec![format!("pkg/overflow-{index}")], |
| 1925 | exact_files: Vec::new(), |
| 1926 | contracts: Vec::new(), |
| 1927 | }, |
| 1928 | true, |
| 1929 | |_| false, |
| 1930 | ) |
| 1931 | .expect("evict oldest inactive claim at capacity"); |
| 1932 | } |
| 1933 | assert!( |
| 1934 | !replayed |
| 1935 | .coordination |
| 1936 | .write_claims |
| 1937 | .iter() |
| 1938 | .any(|claim| claim.claim.owner == claimant), |
| 1939 | "the resolving claimant claim must be evicted for the durability regression" |
| 1940 | ); |
| 1941 | replayed |
| 1942 | .persist_state_synchronously() |
| 1943 | .expect("persist after inactive claim eviction"); |
| 1944 | |
| 1945 | let mut final_replay = super::super::SubAgentManager::new(tmp.path().to_path_buf(), 4) |
| 1946 | .with_state_path(state_path); |
| 1947 | final_replay |
| 1948 | .load_state() |
| 1949 | .expect("reload after resolving claim eviction"); |
| 1950 | let projection = final_replay.coordination_detail_projection(None, 24); |
| 1951 | assert!( |
| 1952 | !projection |
| 1953 | .write_claims |
| 1954 | .iter() |
| 1955 | .any(|claim| claim.claim.owner == claimant) |
| 1956 | ); |
| 1957 | assert_eq!(projection.contentions.len(), 1); |
| 1958 | assert_eq!( |
| 1959 | projection.contentions[0].disposition, |
| 1960 | WriteContentionDisposition::ResolvedBySuccessfulClaim |
| 1961 | ); |
| 1962 | assert_eq!( |
| 1963 | projection.contentions[0].resolution_sequence, |
| 1964 | Some(resolution_sequence) |
| 1965 | ); |
| 1966 | assert!(!crate::tui::coordination_detail::needs_attention( |
| 1967 | &projection |
| 1968 | )); |
| 1969 | let pager = crate::tui::coordination_detail::format( |
| 1970 | codewhale_localization::Locale::En, |
| 1971 | &projection, |
| 1972 | ); |
| 1973 | assert!( |
| 1974 | pager.contains("disposition resolved_by_successful_claim"), |
| 1975 | "{pager}" |
| 1976 | ); |
| 1977 | assert!(!pager.contains("disposition blocked_pending"), "{pager}"); |
| 1978 | } |
| 1979 | |
| 1980 | #[tokio::test] |
| 1981 | async fn interrupt_fails_closed_on_self() { |
| 1982 | let tmp = tempdir().unwrap(); |
| 1983 | let (manager, agent_id) = manager_with_running_child(tmp.path()).await; |
| 1984 | let tool = AgentsInterruptTool::new(Arc::clone(&manager)).with_caller(agent_id.clone()); |
| 1985 | let err = tool |
| 1986 | .execute( |
| 1987 | json!({ "agent_id": agent_id }), |
| 1988 | &ToolContext::new(tmp.path()), |
| 1989 | ) |
| 1990 | .await |
| 1991 | .expect_err("self interrupt must fail"); |
| 1992 | let msg = err.to_string().to_ascii_lowercase(); |
| 1993 | assert!( |
| 1994 | msg.contains("self") || msg.contains("own"), |
| 1995 | "unexpected error: {err}" |
| 1996 | ); |
| 1997 | } |
| 1998 | |
| 1999 | #[tokio::test] |
| 2000 | async fn interrupt_fails_closed_on_missing_target() { |
| 2001 | let tmp = tempdir().unwrap(); |
| 2002 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 2003 | super::super::SubAgentManager::new(tmp.path().to_path_buf(), 2), |
| 2004 | )); |
| 2005 | let tool = AgentsInterruptTool::new(manager); |
| 2006 | let err = tool |
| 2007 | .execute( |
| 2008 | json!({ "agent_id": "agent_missing" }), |
| 2009 | &ToolContext::new(tmp.path()), |
| 2010 | ) |
| 2011 | .await |
| 2012 | .expect_err("missing target"); |
| 2013 | assert!(err.to_string().contains("not found") || err.to_string().contains("Agent")); |
| 2014 | } |
| 2015 | |
| 2016 | #[tokio::test] |
| 2017 | async fn wait_times_out_when_child_stays_running() { |
| 2018 | let tmp = tempdir().unwrap(); |
| 2019 | let (manager, agent_id) = manager_with_running_child(tmp.path()).await; |
| 2020 | let tool = AgentsWaitTool::new(manager); |
| 2021 | let result = tool |
| 2022 | .execute( |
| 2023 | json!({ |
| 2024 | "agent_id": agent_id, |
| 2025 | "timeout_secs": 1, |
| 2026 | "until": "activity" |
| 2027 | }), |
| 2028 | &ToolContext::new(tmp.path()), |
| 2029 | ) |
| 2030 | .await |
| 2031 | .expect("wait returns"); |
| 2032 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 2033 | assert_eq!(body["timed_out"], json!(true)); |
| 2034 | } |
| 2035 | |
| 2036 | #[tokio::test] |
| 2037 | async fn list_resolves_target_and_reports_queue() { |
| 2038 | let tmp = tempdir().unwrap(); |
| 2039 | let (manager, agent_id) = manager_with_running_child(tmp.path()).await; |
| 2040 | { |
| 2041 | let mut guard = manager.write().await; |
| 2042 | guard |
| 2043 | .queue_parent_message(&agent_id, "note".into(), false) |
| 2044 | .unwrap(); |
| 2045 | } |
| 2046 | let tool = AgentsListTool::new(manager); |
| 2047 | let result = tool |
| 2048 | .execute( |
| 2049 | json!({ "agent_id": agent_id }), |
| 2050 | &ToolContext::new(tmp.path()), |
| 2051 | ) |
| 2052 | .await |
| 2053 | .expect("list ok"); |
| 2054 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 2055 | assert_eq!(body["count"], json!(1)); |
| 2056 | assert_eq!(body["agents"][0]["agent_id"], json!(agent_id)); |
| 2057 | assert!(body["agents"][0]["queued_mail"].as_u64().unwrap_or(0) >= 1); |
| 2058 | } |
| 2059 | |
| 2060 | #[tokio::test] |
| 2061 | async fn followup_interrupted_continuable_without_runtime_queues_honestly() { |
| 2062 | let tmp = tempdir().unwrap(); |
| 2063 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 2064 | super::super::SubAgentManager::new(tmp.path().to_path_buf(), 4), |
| 2065 | )); |
| 2066 | let (agent_id, handle) = { |
| 2067 | let mut guard = manager.write().await; |
| 2068 | guard.insert_test_interrupted_continuable_agent( |
| 2069 | "paused_child", |
| 2070 | tmp.path(), |
| 2071 | vec![codewhale_models::Message { |
| 2072 | role: Role::User, |
| 2073 | content: vec![codewhale_models::ContentBlock::Text { |
| 2074 | text: "prior work".to_string(), |
| 2075 | cache_control: None, |
| 2076 | }], |
| 2077 | }], |
| 2078 | ) |
| 2079 | }; |
| 2080 | // No runtime attached: checkpoint resume is unavailable, so followup |
| 2081 | // keeps the honest queue-only semantics with the continuation handle. |
| 2082 | let tool = AgentsFollowupTool::new(Arc::clone(&manager)); |
| 2083 | let result = tool |
| 2084 | .execute( |
| 2085 | json!({ "agent_id": agent_id, "message": "please continue" }), |
| 2086 | &ToolContext::new(tmp.path()), |
| 2087 | ) |
| 2088 | .await |
| 2089 | .expect("followup ok"); |
| 2090 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 2091 | assert_eq!(body["queued"], json!(true)); |
| 2092 | assert_eq!(body["woke"], json!(false)); |
| 2093 | assert_eq!(body["continued_from_checkpoint"], json!(false)); |
| 2094 | assert_eq!(body["continuation_handle"], json!(handle)); |
| 2095 | let note = body["note"].as_str().unwrap_or_default(); |
| 2096 | assert!( |
| 2097 | note.contains("attach a runtime") && note.contains(&handle), |
| 2098 | "note must point at the resume path with the continuation handle: {note}" |
| 2099 | ); |
| 2100 | |
| 2101 | let guard = manager.read().await; |
| 2102 | assert_eq!(guard.queued_mail_depth(&agent_id).unwrap(), 1); |
| 2103 | assert!(!guard.child_was_woken(&agent_id)); |
| 2104 | } |
| 2105 | |
| 2106 | // === until="all": the fan-out join =================================== |
| 2107 | // |
| 2108 | // Before this existed a parent with five children had to issue five |
| 2109 | // waits — while the prompt told it not to poll. These lock the join in. |
| 2110 | |
| 2111 | fn empty_manager(workspace: &std::path::Path) -> SharedSubAgentManager { |
| 2112 | Arc::new(tokio::sync::RwLock::new( |
| 2113 | super::super::SubAgentManager::new(workspace.to_path_buf(), 8), |
| 2114 | )) |
| 2115 | } |
| 2116 | |
| 2117 | async fn settle(manager: &SharedSubAgentManager, agent_id: &str, status: SubAgentStatus) { |
| 2118 | let mut guard = manager.write().await; |
| 2119 | if let Some(agent) = guard.agents.get_mut(agent_id) { |
| 2120 | agent.status = status; |
| 2121 | } |
| 2122 | } |
| 2123 | |
| 2124 | #[test] |
| 2125 | fn wait_schema_offers_all_as_a_first_class_until() { |
| 2126 | let tmp = tempdir().unwrap(); |
| 2127 | let tool = AgentsWaitTool::new(empty_manager(tmp.path())); |
| 2128 | let schema = tool.input_schema(); |
| 2129 | let until = &schema["properties"]["until"]; |
| 2130 | assert_eq!( |
| 2131 | until["enum"], |
| 2132 | json!(["completion", "all", "activity"]), |
| 2133 | "until must expose all alongside completion/activity: {schema}" |
| 2134 | ); |
| 2135 | let described = until["description"].as_str().unwrap_or_default(); |
| 2136 | assert!( |
| 2137 | described.contains("every watched child") && described.contains("any one child"), |
| 2138 | "the schema must make completion vs all unmistakable: {described}" |
| 2139 | ); |
| 2140 | } |
| 2141 | |
| 2142 | #[tokio::test] |
| 2143 | async fn wait_until_all_on_an_already_settled_child_reports_its_outcome() { |
| 2144 | let tmp = tempdir().unwrap(); |
| 2145 | let manager = empty_manager(tmp.path()); |
| 2146 | let agent_id = { |
| 2147 | let mut guard = manager.write().await; |
| 2148 | guard.insert_test_running_agent("all_already_done", tmp.path()) |
| 2149 | }; |
| 2150 | settle(&manager, &agent_id, SubAgentStatus::Completed).await; |
| 2151 | |
| 2152 | let result = dispatch_wait( |
| 2153 | &json!({ "until": "all", "agent_id": agent_id, "timeout_secs": 60 }), |
| 2154 | Arc::clone(&manager), |
| 2155 | &ToolContext::new(tmp.path()), |
| 2156 | ) |
| 2157 | .await |
| 2158 | .expect("a settled child is an immediate return"); |
| 2159 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 2160 | assert_eq!(body["all_settled"], json!(true), "{body}"); |
| 2161 | let settled = body["settled"].as_array().unwrap(); |
| 2162 | assert_eq!(settled.len(), 1, "{body}"); |
| 2163 | assert_eq!(settled[0]["status"], json!("completed"), "{body}"); |
| 2164 | } |
| 2165 | |
| 2166 | #[tokio::test] |
| 2167 | async fn foreign_session_is_excluded_from_default_and_explicit_list_waits() { |
| 2168 | let tmp = tempdir().unwrap(); |
| 2169 | let manager = empty_manager(tmp.path()); |
| 2170 | let agent_a = { |
| 2171 | let mut guard = manager.write().await; |
| 2172 | let agent_id = guard.insert_test_running_agent("foreign_wait_a", tmp.path()); |
| 2173 | guard.assign_test_session_owner(&agent_id, "session-a"); |
| 2174 | agent_id |
| 2175 | }; |
| 2176 | let context_b = ToolContext::new(tmp.path()).with_state_namespace("session-b"); |
| 2177 | |
| 2178 | let listed = AgentsListTool::new(Arc::clone(&manager)) |
| 2179 | .execute(json!({}), &context_b) |
| 2180 | .await |
| 2181 | .expect("B list"); |
| 2182 | let listed: Value = serde_json::from_str(&listed.content).unwrap(); |
| 2183 | assert_eq!(listed["count"], json!(0)); |
| 2184 | |
| 2185 | let default_wait = dispatch_wait( |
| 2186 | &json!({ "until": "all", "timeout_secs": 60 }), |
| 2187 | Arc::clone(&manager), |
| 2188 | &context_b, |
| 2189 | ) |
| 2190 | .await |
| 2191 | .expect("B default wait has no visible children"); |
| 2192 | let default_wait: Value = serde_json::from_str(&default_wait.content).unwrap(); |
| 2193 | assert!(default_wait["settled"].as_array().unwrap().is_empty()); |
| 2194 | |
| 2195 | let error = dispatch_wait( |
| 2196 | &json!({ "until": "all", "agent_id": agent_a, "timeout_secs": 60 }), |
| 2197 | manager, |
| 2198 | &context_b, |
| 2199 | ) |
| 2200 | .await |
| 2201 | .expect_err("B explicit wait must reject A") |
| 2202 | .to_string(); |
| 2203 | assert!( |
| 2204 | error.contains("Agent not found in the active session"), |
| 2205 | "{error}" |
| 2206 | ); |
| 2207 | } |
| 2208 | |
| 2209 | #[tokio::test] |
| 2210 | async fn wait_until_all_returns_immediately_with_no_children() { |
| 2211 | let tmp = tempdir().unwrap(); |
| 2212 | let started = Instant::now(); |
| 2213 | let result = dispatch_wait( |
| 2214 | &json!({ "until": "all", "timeout_secs": 60 }), |
| 2215 | empty_manager(tmp.path()), |
| 2216 | &ToolContext::new(tmp.path()), |
| 2217 | ) |
| 2218 | .await |
| 2219 | .expect("wait-for-all with zero children must return, not hang"); |
| 2220 | assert!( |
| 2221 | started.elapsed() < Duration::from_secs(5), |
| 2222 | "zero children must not burn the timeout" |
| 2223 | ); |
| 2224 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 2225 | assert_eq!(body["all_settled"], json!(true)); |
| 2226 | assert_eq!(body["timed_out"], json!(false)); |
| 2227 | assert!(body["settled"].as_array().unwrap().is_empty(), "{body}"); |
| 2228 | } |
| 2229 | |
| 2230 | #[tokio::test] |
| 2231 | async fn wait_until_all_blocks_until_every_child_settles() { |
| 2232 | let tmp = tempdir().unwrap(); |
| 2233 | let manager = empty_manager(tmp.path()); |
| 2234 | let (first, second, third) = { |
| 2235 | let mut guard = manager.write().await; |
| 2236 | ( |
| 2237 | guard.insert_test_running_agent("all_first", tmp.path()), |
| 2238 | guard.insert_test_running_agent("all_second", tmp.path()), |
| 2239 | guard.insert_test_running_agent("all_third", tmp.path()), |
| 2240 | ) |
| 2241 | }; |
| 2242 | |
| 2243 | // Staggered settles: an `until=completion` wait would return after the |
| 2244 | // first one. `until=all` must stay blocked through the last. |
| 2245 | let flip = Arc::clone(&manager); |
| 2246 | let (a, b, c) = (first.clone(), second.clone(), third.clone()); |
| 2247 | tokio::spawn(async move { |
| 2248 | tokio::time::sleep(Duration::from_millis(50)).await; |
| 2249 | settle(&flip, &a, SubAgentStatus::Completed).await; |
| 2250 | tokio::time::sleep(Duration::from_millis(150)).await; |
| 2251 | settle(&flip, &b, SubAgentStatus::Failed("boom".to_string())).await; |
| 2252 | tokio::time::sleep(Duration::from_millis(150)).await; |
| 2253 | settle(&flip, &c, SubAgentStatus::Cancelled).await; |
| 2254 | }); |
| 2255 | |
| 2256 | let result = dispatch_wait( |
| 2257 | &json!({ "until": "all", "timeout_secs": 30 }), |
| 2258 | Arc::clone(&manager), |
| 2259 | &ToolContext::new(tmp.path()), |
| 2260 | ) |
| 2261 | .await |
| 2262 | .expect("wait-for-all should succeed"); |
| 2263 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 2264 | assert_eq!(body["all_settled"], json!(true), "{body}"); |
| 2265 | assert_eq!(body["timed_out"], json!(false), "{body}"); |
| 2266 | assert!( |
| 2267 | body["still_running"].as_array().unwrap().is_empty(), |
| 2268 | "{body}" |
| 2269 | ); |
| 2270 | |
| 2271 | // Per-child outcomes come back on the single return. |
| 2272 | let settled = body["settled"].as_array().unwrap(); |
| 2273 | assert_eq!(settled.len(), 3, "{body}"); |
| 2274 | let outcomes: std::collections::BTreeMap<&str, &str> = settled |
| 2275 | .iter() |
| 2276 | .map(|entry| { |
| 2277 | ( |
| 2278 | entry["agent_id"].as_str().unwrap(), |
| 2279 | entry["status"].as_str().unwrap(), |
| 2280 | ) |
| 2281 | }) |
| 2282 | .collect(); |
| 2283 | assert_eq!(outcomes.get(first.as_str()), Some(&"completed"), "{body}"); |
| 2284 | assert_eq!(outcomes.get(second.as_str()), Some(&"failed"), "{body}"); |
| 2285 | assert_eq!(outcomes.get(third.as_str()), Some(&"cancelled"), "{body}"); |
| 2286 | } |
| 2287 | |
| 2288 | #[tokio::test] |
| 2289 | async fn wait_until_all_times_out_reporting_settled_and_still_running() { |
| 2290 | let tmp = tempdir().unwrap(); |
| 2291 | let manager = empty_manager(tmp.path()); |
| 2292 | let (done, stuck) = { |
| 2293 | let mut guard = manager.write().await; |
| 2294 | ( |
| 2295 | guard.insert_test_running_agent("all_done", tmp.path()), |
| 2296 | guard.insert_test_running_agent("all_stuck", tmp.path()), |
| 2297 | ) |
| 2298 | }; |
| 2299 | |
| 2300 | let request = json!({ "until": "all", "timeout_secs": 1 }); |
| 2301 | let context = ToolContext::new(tmp.path()); |
| 2302 | let wait = dispatch_wait(&request, Arc::clone(&manager), &context); |
| 2303 | tokio::pin!(wait); |
| 2304 | // Capture both running children before completing one. The timeout |
| 2305 | // receipt must not depend on a background task winning a 50ms race. |
| 2306 | assert!(futures_util::poll!(wait.as_mut()).is_pending()); |
| 2307 | settle(&manager, &done, SubAgentStatus::Completed).await; |
| 2308 | |
| 2309 | let result = wait |
| 2310 | .await |
| 2311 | .expect("a timeout is a partial receipt, not an error"); |
| 2312 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 2313 | assert_eq!(body["timed_out"], json!(true), "{body}"); |
| 2314 | assert_eq!(body["all_settled"], json!(false), "{body}"); |
| 2315 | |
| 2316 | let settled = body["settled"].as_array().unwrap(); |
| 2317 | assert_eq!(settled.len(), 1, "{body}"); |
| 2318 | assert_eq!(settled[0]["agent_id"], json!(done), "{body}"); |
| 2319 | assert_eq!(settled[0]["status"], json!("completed"), "{body}"); |
| 2320 | |
| 2321 | let running = body["still_running"].as_array().unwrap(); |
| 2322 | assert_eq!(running.len(), 1, "{body}"); |
| 2323 | assert_eq!(running[0]["agent_id"], json!(stuck), "{body}"); |
| 2324 | } |
| 2325 | |
| 2326 | #[tokio::test] |
| 2327 | async fn wait_until_all_ignores_children_spawned_mid_wait() { |
| 2328 | let tmp = tempdir().unwrap(); |
| 2329 | let manager = empty_manager(tmp.path()); |
| 2330 | let original = { |
| 2331 | let mut guard = manager.write().await; |
| 2332 | guard.insert_test_running_agent("all_original", tmp.path()) |
| 2333 | }; |
| 2334 | |
| 2335 | let flip = Arc::clone(&manager); |
| 2336 | let tmp_path = tmp.path().to_path_buf(); |
| 2337 | let original_id = original.clone(); |
| 2338 | tokio::spawn(async move { |
| 2339 | tokio::time::sleep(Duration::from_millis(50)).await; |
| 2340 | { |
| 2341 | let mut guard = flip.write().await; |
| 2342 | guard.insert_test_running_agent("all_latecomer", &tmp_path); |
| 2343 | } |
| 2344 | settle(&flip, &original_id, SubAgentStatus::Completed).await; |
| 2345 | }); |
| 2346 | |
| 2347 | let result = dispatch_wait( |
| 2348 | &json!({ "until": "all", "timeout_secs": 30 }), |
| 2349 | Arc::clone(&manager), |
| 2350 | &ToolContext::new(tmp.path()), |
| 2351 | ) |
| 2352 | .await |
| 2353 | .expect("wait-for-all should succeed"); |
| 2354 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 2355 | // The watch set is the batch as of call time: the latecomer must not |
| 2356 | // extend a wait the caller never asked to include it in. |
| 2357 | assert_eq!(body["all_settled"], json!(true), "{body}"); |
| 2358 | assert_eq!(body["timed_out"], json!(false), "{body}"); |
| 2359 | let settled = body["settled"].as_array().unwrap(); |
| 2360 | assert_eq!(settled.len(), 1, "{body}"); |
| 2361 | assert_eq!(settled[0]["agent_id"], json!(original), "{body}"); |
| 2362 | } |
| 2363 | |
| 2364 | #[tokio::test] |
| 2365 | async fn wait_rejects_unknown_until_naming_every_supported_mode() { |
| 2366 | let tmp = tempdir().unwrap(); |
| 2367 | let error = dispatch_wait( |
| 2368 | &json!({ "until": "forever" }), |
| 2369 | empty_manager(tmp.path()), |
| 2370 | &ToolContext::new(tmp.path()), |
| 2371 | ) |
| 2372 | .await |
| 2373 | .expect_err("an unknown until must fail loudly"); |
| 2374 | let message = error.to_string(); |
| 2375 | for mode in ["completion", "all", "activity"] { |
| 2376 | assert!(message.contains(mode), "{message}"); |
| 2377 | } |
| 2378 | } |
| 2379 | |
| 2380 | #[tokio::test] |
| 2381 | async fn followup_interrupted_continuable_resumes_with_runtime() { |
| 2382 | let tmp = tempdir().unwrap(); |
| 2383 | let manager = Arc::new(tokio::sync::RwLock::new( |
| 2384 | super::super::SubAgentManager::new(tmp.path().to_path_buf(), 4), |
| 2385 | )); |
| 2386 | let (agent_id, _handle) = { |
| 2387 | let mut guard = manager.write().await; |
| 2388 | guard.insert_test_interrupted_continuable_agent( |
| 2389 | "paused_child", |
| 2390 | tmp.path(), |
| 2391 | vec![codewhale_models::Message { |
| 2392 | role: Role::User, |
| 2393 | content: vec![codewhale_models::ContentBlock::Text { |
| 2394 | text: "prior work".to_string(), |
| 2395 | cache_control: None, |
| 2396 | }], |
| 2397 | }], |
| 2398 | ) |
| 2399 | }; |
| 2400 | let mut runtime = super::super::tests::stub_runtime(); |
| 2401 | runtime.manager = Arc::clone(&manager); |
| 2402 | let tool = AgentsFollowupTool::new(Arc::clone(&manager)).with_runtime(runtime); |
| 2403 | let result = tool |
| 2404 | .execute( |
| 2405 | json!({ "agent_id": agent_id, "message": "please continue" }), |
| 2406 | &ToolContext::new(tmp.path()), |
| 2407 | ) |
| 2408 | .await |
| 2409 | .expect("followup ok"); |
| 2410 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 2411 | assert_eq!(body["queued"], json!(true)); |
| 2412 | assert_eq!(body["woke"], json!(true)); |
| 2413 | assert_eq!(body["continued_from_checkpoint"], json!(true)); |
| 2414 | let note = body["note"].as_str().unwrap_or_default(); |
| 2415 | assert!(note.contains("resumed from checkpoint"), "{note}"); |
| 2416 | let resumed_id = body["agent_id"].as_str().unwrap_or_default(); |
| 2417 | assert_ne!( |
| 2418 | resumed_id, agent_id, |
| 2419 | "resume re-dispatches under a new agent id" |
| 2420 | ); |
| 2421 | |
| 2422 | // A fresh record exists for the resumed session; the prior terminal |
| 2423 | // record stays immutable (receipts are never rewritten). |
| 2424 | let guard = manager.read().await; |
| 2425 | guard.get_result(resumed_id).expect("resumed agent exists"); |
| 2426 | let prior = guard.get_result(&agent_id).expect("prior record"); |
| 2427 | assert!(matches!(prior.status, SubAgentStatus::Interrupted(_))); |
| 2428 | } |
| 2429 | } |
| 2430 |