| 1 | //! Durable task, gate, and PR-attempt tools. |
| 2 | |
| 3 | use std::path::{Path, PathBuf}; |
| 4 | use std::process::Stdio; |
| 5 | use std::time::Instant; |
| 6 | |
| 7 | use async_trait::async_trait; |
| 8 | use chrono::Utc; |
| 9 | use serde_json::{Value, json}; |
| 10 | use tokio::process::Command; |
| 11 | use uuid::Uuid; |
| 12 | |
| 13 | use crate::dependencies::ExternalTool; |
| 14 | use crate::task_manager::{ |
| 15 | NewTaskRequest, TaskArtifactRef, TaskAttemptRecord, TaskCancelDisposition, TaskGateRecord, |
| 16 | TaskRecord, |
| 17 | }; |
| 18 | use crate::tools::shell::BashTool; |
| 19 | use crate::tools::spec::{ |
| 20 | ApprovalRequirement, ToolCapability, ToolContext, ToolError, ToolResult, ToolSpec, |
| 21 | optional_bool, optional_bool_opt, optional_str, optional_u64, required_str, |
| 22 | }; |
| 23 | use crate::work_graph::{ |
| 24 | CancelOutcome, OperationIntent, OperationObservation, OperationOwnerSnapshot, OwnerState, |
| 25 | task_owner_snapshot, |
| 26 | }; |
| 27 | use codewhale_execpolicy::command_safety::{SafetyLevel, analyze_command}; |
| 28 | |
| 29 | const MAX_SUMMARY_CHARS: usize = 900; |
| 30 | const DEFAULT_GATE_TIMEOUT_MS: u64 = 120_000; |
| 31 | const MAX_GATE_TIMEOUT_MS: u64 = 600_000; |
| 32 | |
| 33 | fn build_gate_command_parts(command: &str) -> (String, Vec<String>) { |
| 34 | ( |
| 35 | "/bin/sh".to_string(), |
| 36 | vec!["-lc".to_string(), command.to_string()], |
| 37 | ) |
| 38 | } |
| 39 | |
| 40 | fn build_gate_command(command: &str, cwd: &Path) -> Command { |
| 41 | let (program, args) = build_gate_command_parts(command); |
| 42 | let mut cmd = Command::new(program); |
| 43 | cmd.args(args) |
| 44 | .current_dir(cwd) |
| 45 | .stdout(Stdio::piped()) |
| 46 | .stderr(Stdio::piped()); |
| 47 | cmd |
| 48 | } |
| 49 | |
| 50 | fn task_shell_wait_input(mut input: Value) -> Value { |
| 51 | if input.get("wait").is_none_or(Value::is_null) |
| 52 | && input.get("block").is_none_or(Value::is_null) |
| 53 | && let Some(object) = input.as_object_mut() |
| 54 | { |
| 55 | object.insert("wait".to_string(), Value::Bool(false)); |
| 56 | } |
| 57 | input |
| 58 | } |
| 59 | |
| 60 | /// Unified durable-task tool (piagent phase B). |
| 61 | /// |
| 62 | /// The model sees one tool, `tasks`, with an `action` parameter routing to |
| 63 | /// the per-action logic below. The per-action `task_*` / `pr_attempt_*` |
| 64 | /// execution aliases were removed in v0.9.3. |
| 65 | /// |
| 66 | /// `TaskShellStartTool` / `TaskShellWaitTool` stay separate: the registry |
| 67 | /// gates them behind `allow_shell` (see `with_runtime_task_shell_tools`), |
| 68 | /// which differs from every other action in this family. |
| 69 | pub struct TasksTool { |
| 70 | name: &'static str, |
| 71 | forced_action: Option<&'static str>, |
| 72 | read_only: bool, |
| 73 | } |
| 74 | |
| 75 | pub struct TaskShellStartTool; |
| 76 | pub struct TaskShellWaitTool; |
| 77 | |
| 78 | /// Actions the Plan-mode read-only surface exposes. |
| 79 | const READ_ACTIONS: &[&str] = &["list", "read", "pr_attempt_list", "pr_attempt_read"]; |
| 80 | const ALL_ACTIONS: &[&str] = &[ |
| 81 | "create", |
| 82 | "list", |
| 83 | "read", |
| 84 | "cancel", |
| 85 | "gate_run", |
| 86 | "pr_attempt_record", |
| 87 | "pr_attempt_list", |
| 88 | "pr_attempt_read", |
| 89 | "pr_attempt_preflight", |
| 90 | ]; |
| 91 | |
| 92 | impl TasksTool { |
| 93 | pub const fn new(name: &'static str) -> Self { |
| 94 | Self { |
| 95 | name, |
| 96 | forced_action: None, |
| 97 | read_only: false, |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | /// Plan-mode variant: only the read-only actions are advertised and routed. |
| 102 | pub const fn read_only(name: &'static str) -> Self { |
| 103 | Self { |
| 104 | name, |
| 105 | forced_action: None, |
| 106 | read_only: true, |
| 107 | } |
| 108 | } |
| 109 | |
| 110 | #[cfg(test)] |
| 111 | pub const fn alias(name: &'static str, action: &'static str) -> Self { |
| 112 | Self { |
| 113 | name, |
| 114 | forced_action: Some(action), |
| 115 | read_only: false, |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | fn allowed_actions(&self) -> &'static [&'static str] { |
| 120 | if self.read_only { |
| 121 | READ_ACTIONS |
| 122 | } else { |
| 123 | ALL_ACTIONS |
| 124 | } |
| 125 | } |
| 126 | |
| 127 | fn resolve_action<'a>(&'a self, input: &'a Value) -> Result<&'a str, ToolError> { |
| 128 | let action = match self.forced_action { |
| 129 | Some(action) => action, |
| 130 | None => input.get("action").and_then(Value::as_str).ok_or_else(|| { |
| 131 | ToolError::invalid_input(format!( |
| 132 | "tasks: missing `action` (one of: {})", |
| 133 | self.allowed_actions().join(", ") |
| 134 | )) |
| 135 | })?, |
| 136 | }; |
| 137 | if self.allowed_actions().contains(&action) { |
| 138 | Ok(action) |
| 139 | } else { |
| 140 | Err(ToolError::invalid_input(format!( |
| 141 | "tasks: invalid action `{action}` (one of: {})", |
| 142 | self.allowed_actions().join(", ") |
| 143 | ))) |
| 144 | } |
| 145 | } |
| 146 | |
| 147 | fn action_is_read(action: &str) -> bool { |
| 148 | READ_ACTIONS.contains(&action) |
| 149 | } |
| 150 | |
| 151 | /// Whether this action executes code (drives static capabilities and the |
| 152 | /// Plan-mode "no ExecutesCode tools" invariant). |
| 153 | fn action_executes_code(action: &str) -> bool { |
| 154 | action == "gate_run" |
| 155 | } |
| 156 | |
| 157 | fn action_requires_approval(action: &str) -> bool { |
| 158 | !Self::action_is_read(action) |
| 159 | } |
| 160 | } |
| 161 | |
| 162 | #[async_trait] |
| 163 | impl ToolSpec for TasksTool { |
| 164 | fn name(&self) -> &'static str { |
| 165 | self.name |
| 166 | } |
| 167 | |
| 168 | fn model_visible(&self) -> bool { |
| 169 | self.forced_action.is_none() |
| 170 | } |
| 171 | |
| 172 | fn description(&self) -> &'static str { |
| 173 | match self.forced_action { |
| 174 | Some("create") => { |
| 175 | "Create/enqueue a durable background task through TaskManager. Durable tasks are restart-aware executable work, distinct from sub-agents." |
| 176 | } |
| 177 | Some("list") => { |
| 178 | "List recent durable tasks with status, linked thread/turn ids, and concise summaries." |
| 179 | } |
| 180 | Some("read") => { |
| 181 | "Read durable task detail including timeline, checklist, gate evidence, artifacts, and PR attempts." |
| 182 | } |
| 183 | Some("cancel") => { |
| 184 | "Cancel a queued or running durable task through TaskManager. Requires approval because it changes work state." |
| 185 | } |
| 186 | Some("gate_run") => { |
| 187 | "Run an approved verification gate command and return structured evidence. When inside a durable task, the gate result and log artifact are attached to that task." |
| 188 | } |
| 189 | Some("pr_attempt_record") => { |
| 190 | "Capture current git diff as a durable PR work attempt with patch artifact, changed files, and verification notes." |
| 191 | } |
| 192 | Some("pr_attempt_list") => "List PR attempts recorded on a durable task.", |
| 193 | Some("pr_attempt_read") => { |
| 194 | "Read one recorded PR attempt and its patch artifact reference." |
| 195 | } |
| 196 | Some("pr_attempt_preflight") => { |
| 197 | "Run `git apply --check` for a recorded attempt patch. This is a no-mutation preflight; actual apply remains explicit and approval-gated elsewhere." |
| 198 | } |
| 199 | _ if self.read_only => { |
| 200 | "Inspect durable tasks and their PR attempts. Actions: \"list\", \"read\", \"pr_attempt_list\", \"pr_attempt_read\"." |
| 201 | } |
| 202 | _ => { |
| 203 | "Manage durable background tasks through TaskManager. Durable tasks are restart-aware executable work, distinct from sub-agents. Actions: \"create\" (enqueue; approval), \"list\", \"read\", \"cancel\" (approval), \"gate_run\" (run an approved verification gate command and return structured evidence; approval), \"pr_attempt_record\", \"pr_attempt_list\", \"pr_attempt_read\", \"pr_attempt_preflight\". Use task_shell_start for long-running shell work." |
| 204 | } |
| 205 | } |
| 206 | } |
| 207 | |
| 208 | fn input_schema(&self) -> Value { |
| 209 | if let Some(action) = self.forced_action { |
| 210 | return legacy_action_schema(action); |
| 211 | } |
| 212 | let actions: Vec<&str> = self.allowed_actions().to_vec(); |
| 213 | let mut properties = serde_json::Map::new(); |
| 214 | properties.insert( |
| 215 | "action".to_string(), |
| 216 | json!({ |
| 217 | "type": "string", |
| 218 | "enum": actions, |
| 219 | "description": "Action to perform." |
| 220 | }), |
| 221 | ); |
| 222 | if !self.read_only { |
| 223 | properties.insert( |
| 224 | "prompt".to_string(), |
| 225 | json!({ "type": "string", "description": "Work prompt for the durable task (action=create)." }), |
| 226 | ); |
| 227 | properties.insert( |
| 228 | "name".to_string(), |
| 229 | json!({ "type": "string", "description": "Short run name shown in queues; omit to derive from the prompt. (action=create)" }), |
| 230 | ); |
| 231 | properties.insert( |
| 232 | "model_provider".to_string(), |
| 233 | json!({ "type": "string", "description": "Provider kind for the pinned model. Omit to inherit the configured provider." }), |
| 234 | ); |
| 235 | properties.insert( |
| 236 | "model_provider_id".to_string(), |
| 237 | json!({ "type": "string", "description": "Exact configured provider id, including named custom routes. Keeps the model on that route." }), |
| 238 | ); |
| 239 | properties.insert( |
| 240 | "model".to_string(), |
| 241 | json!({ "type": "string", "description": "(action=create)" }), |
| 242 | ); |
| 243 | properties.insert( |
| 244 | "workspace".to_string(), |
| 245 | json!({ "type": "string", "description": "Workspace path; defaults to current workspace. (action=create)" }), |
| 246 | ); |
| 247 | properties.insert( |
| 248 | "mode".to_string(), |
| 249 | json!({ "type": "string", "enum": ["agent", "plan", "operate"], "description": "(action=create)" }), |
| 250 | ); |
| 251 | properties.insert( |
| 252 | "allow_shell".to_string(), |
| 253 | json!({ "type": "boolean", "description": "(action=create)" }), |
| 254 | ); |
| 255 | properties.insert( |
| 256 | "trust_mode".to_string(), |
| 257 | json!({ "type": "boolean", "description": "(action=create)" }), |
| 258 | ); |
| 259 | properties.insert( |
| 260 | "auto_approve".to_string(), |
| 261 | json!({ "type": "boolean", "description": "(action=create)" }), |
| 262 | ); |
| 263 | properties.insert( |
| 264 | "gate".to_string(), |
| 265 | json!({ |
| 266 | "type": "string", |
| 267 | "enum": ["fmt", "check", "clippy", "test", "custom"], |
| 268 | "description": "Gate category. (action=gate_run)" |
| 269 | }), |
| 270 | ); |
| 271 | properties.insert( |
| 272 | "command".to_string(), |
| 273 | json!({ "type": "string", "description": "Command to run. (action=gate_run)" }), |
| 274 | ); |
| 275 | properties.insert( |
| 276 | "cwd".to_string(), |
| 277 | json!({ "type": "string", "description": "Optional working directory within the workspace. (action=gate_run)" }), |
| 278 | ); |
| 279 | properties.insert( |
| 280 | "timeout_ms".to_string(), |
| 281 | json!({ "type": "integer", "minimum": 1000, "maximum": 600000, "description": "(action=gate_run)" }), |
| 282 | ); |
| 283 | properties.insert( |
| 284 | "attempt_group_id".to_string(), |
| 285 | json!({ "type": "string", "description": "(action=pr_attempt_record)" }), |
| 286 | ); |
| 287 | properties.insert( |
| 288 | "attempt_index".to_string(), |
| 289 | json!({ "type": "integer", "minimum": 1, "description": "(action=pr_attempt_record)" }), |
| 290 | ); |
| 291 | properties.insert( |
| 292 | "attempt_count".to_string(), |
| 293 | json!({ "type": "integer", "minimum": 1, "description": "(action=pr_attempt_record)" }), |
| 294 | ); |
| 295 | properties.insert( |
| 296 | "summary".to_string(), |
| 297 | json!({ "type": "string", "description": "Attempt summary (action=pr_attempt_record)." }), |
| 298 | ); |
| 299 | properties.insert( |
| 300 | "verification".to_string(), |
| 301 | json!({ "type": "array", "items": { "type": "string" }, "description": "(action=pr_attempt_record)" }), |
| 302 | ); |
| 303 | } |
| 304 | properties.insert( |
| 305 | "attempt_id".to_string(), |
| 306 | json!({ "type": "string", "description": "(action=pr_attempt_read/preflight)" }), |
| 307 | ); |
| 308 | properties.insert( |
| 309 | "task_id".to_string(), |
| 310 | json!({ "type": "string", "description": "Full task id or unambiguous prefix (action=read/cancel); task id, defaults to active task (action=pr_attempt_*)." }), |
| 311 | ); |
| 312 | properties.insert( |
| 313 | "limit".to_string(), |
| 314 | json!({ "type": "integer", "minimum": 1, "maximum": 100, "default": 20, "description": "(action=list)" }), |
| 315 | ); |
| 316 | json!({ |
| 317 | "type": "object", |
| 318 | "properties": properties, |
| 319 | "additionalProperties": false |
| 320 | }) |
| 321 | } |
| 322 | |
| 323 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 324 | match self.forced_action { |
| 325 | Some(action) if Self::action_executes_code(action) => { |
| 326 | vec![ |
| 327 | ToolCapability::ExecutesCode, |
| 328 | ToolCapability::RequiresApproval, |
| 329 | ] |
| 330 | } |
| 331 | Some(action) if Self::action_is_read(action) => vec![ToolCapability::ReadOnly], |
| 332 | Some(_) => vec![ToolCapability::RequiresApproval], |
| 333 | None if self.read_only => vec![ToolCapability::ReadOnly], |
| 334 | None => vec![ |
| 335 | ToolCapability::ExecutesCode, |
| 336 | ToolCapability::RequiresApproval, |
| 337 | ], |
| 338 | } |
| 339 | } |
| 340 | |
| 341 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 342 | match self.forced_action { |
| 343 | Some(action) if Self::action_requires_approval(action) => ApprovalRequirement::Required, |
| 344 | Some(_) => ApprovalRequirement::Auto, |
| 345 | None if self.read_only => ApprovalRequirement::Auto, |
| 346 | None => ApprovalRequirement::Required, |
| 347 | } |
| 348 | } |
| 349 | |
| 350 | fn approval_requirement_for(&self, input: &Value) -> ApprovalRequirement { |
| 351 | match self.resolve_action(input) { |
| 352 | Ok(action) if Self::action_requires_approval(action) => ApprovalRequirement::Required, |
| 353 | Ok(_) => ApprovalRequirement::Auto, |
| 354 | Err(_) => self.approval_requirement(), |
| 355 | } |
| 356 | } |
| 357 | |
| 358 | fn is_read_only_for(&self, input: &Value) -> bool { |
| 359 | match self.resolve_action(input) { |
| 360 | Ok(action) => Self::action_is_read(action), |
| 361 | Err(_) => self.is_read_only(), |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 366 | crate::core::engine::tool_catalog::enforce_tool_denial( |
| 367 | context, |
| 368 | self.name(), |
| 369 | &json!({"action": self.resolve_action(&input)?}), |
| 370 | )?; |
| 371 | match self.resolve_action(&input)? { |
| 372 | "create" => self.execute_create(&input, context).await, |
| 373 | "list" => self.execute_list(&input, context).await, |
| 374 | "read" => self.execute_read(&input, context).await, |
| 375 | "cancel" => self.execute_cancel(&input, context).await, |
| 376 | "gate_run" => self.execute_gate_run(&input, context).await, |
| 377 | "pr_attempt_record" => self.execute_pr_attempt_record(&input, context).await, |
| 378 | "pr_attempt_list" => self.execute_pr_attempt_list(&input, context).await, |
| 379 | "pr_attempt_read" => self.execute_pr_attempt_read(&input, context).await, |
| 380 | "pr_attempt_preflight" => self.execute_pr_attempt_preflight(&input, context).await, |
| 381 | action => Err(ToolError::invalid_input(format!( |
| 382 | "tasks: invalid action `{action}`" |
| 383 | ))), |
| 384 | } |
| 385 | } |
| 386 | } |
| 387 | |
| 388 | /// The exact schema the legacy per-action tool exposed, kept so hidden alias |
| 389 | /// registrations report an identical contract to the pre-unification tools. |
| 390 | fn legacy_action_schema(action: &str) -> Value { |
| 391 | match action { |
| 392 | "create" => json!({ |
| 393 | "type": "object", |
| 394 | "properties": { |
| 395 | "prompt": { "type": "string", "description": "Work prompt for the durable task." }, |
| 396 | "model": { "type": "string" }, |
| 397 | "model_provider": { "type": "string", "description": "Provider kind for the pinned model." }, |
| 398 | "model_provider_id": { "type": "string", "description": "Exact configured provider id." }, |
| 399 | "workspace": { "type": "string", "description": "Workspace path; defaults to current workspace." }, |
| 400 | "mode": { "type": "string", "enum": ["agent", "plan", "operate"] }, |
| 401 | "allow_shell": { "type": "boolean" }, |
| 402 | "trust_mode": { "type": "boolean" }, |
| 403 | "auto_approve": { "type": "boolean" } |
| 404 | }, |
| 405 | "required": ["prompt"], |
| 406 | "additionalProperties": false |
| 407 | }), |
| 408 | "list" => json!({ |
| 409 | "type": "object", |
| 410 | "properties": { |
| 411 | "limit": { "type": "integer", "minimum": 1, "maximum": 100, "default": 20 } |
| 412 | }, |
| 413 | "additionalProperties": false |
| 414 | }), |
| 415 | "read" | "cancel" => json!({ |
| 416 | "type": "object", |
| 417 | "properties": { |
| 418 | "task_id": { "type": "string", "description": "Full task id or unambiguous prefix." } |
| 419 | }, |
| 420 | "required": ["task_id"], |
| 421 | "additionalProperties": false |
| 422 | }), |
| 423 | "gate_run" => json!({ |
| 424 | "type": "object", |
| 425 | "properties": { |
| 426 | "gate": { |
| 427 | "type": "string", |
| 428 | "enum": ["fmt", "check", "clippy", "test", "custom"], |
| 429 | "description": "Gate category." |
| 430 | }, |
| 431 | "command": { "type": "string", "description": "Command to run." }, |
| 432 | "cwd": { "type": "string", "description": "Optional working directory within the workspace." }, |
| 433 | "timeout_ms": { "type": "integer", "minimum": 1000, "maximum": 600000 } |
| 434 | }, |
| 435 | "required": ["gate", "command"], |
| 436 | "additionalProperties": false |
| 437 | }), |
| 438 | "pr_attempt_record" => json!({ |
| 439 | "type": "object", |
| 440 | "properties": { |
| 441 | "task_id": { "type": "string", "description": "Task to attach to; defaults to active task." }, |
| 442 | "attempt_group_id": { "type": "string" }, |
| 443 | "attempt_index": { "type": "integer", "minimum": 1 }, |
| 444 | "attempt_count": { "type": "integer", "minimum": 1 }, |
| 445 | "summary": { "type": "string" }, |
| 446 | "verification": { "type": "array", "items": { "type": "string" } } |
| 447 | }, |
| 448 | "required": ["summary"], |
| 449 | "additionalProperties": false |
| 450 | }), |
| 451 | "pr_attempt_list" => task_id_schema(), |
| 452 | // pr_attempt_read / pr_attempt_preflight share the attempt-id schema. |
| 453 | _ => json!({ |
| 454 | "type": "object", |
| 455 | "properties": { |
| 456 | "task_id": { "type": "string", "description": "Task id; defaults to active task." }, |
| 457 | "attempt_id": { "type": "string" } |
| 458 | }, |
| 459 | "required": ["attempt_id"], |
| 460 | "additionalProperties": false |
| 461 | }), |
| 462 | } |
| 463 | } |
| 464 | |
| 465 | impl TasksTool { |
| 466 | async fn execute_create( |
| 467 | &self, |
| 468 | input: &Value, |
| 469 | context: &ToolContext, |
| 470 | ) -> Result<ToolResult, ToolError> { |
| 471 | let manager = context |
| 472 | .runtime |
| 473 | .task_manager |
| 474 | .as_ref() |
| 475 | .ok_or_else(|| ToolError::not_available("TaskManager is not attached"))?; |
| 476 | let workspace = optional_str(input, "workspace")? |
| 477 | .map(PathBuf::from) |
| 478 | .unwrap_or_else(|| context.workspace.clone()); |
| 479 | let prompt = required_str(input, "prompt")?.to_string(); |
| 480 | let req = NewTaskRequest { |
| 481 | prompt: prompt.clone(), |
| 482 | name: optional_str(input, "name")?.map(ToString::to_string), |
| 483 | model: optional_str(input, "model")?.map(ToString::to_string), |
| 484 | model_provider: optional_str(input, "model_provider")?.map(ToString::to_string), |
| 485 | model_provider_id: optional_str(input, "model_provider_id")?.map(ToString::to_string), |
| 486 | workspace: Some(workspace), |
| 487 | mode: optional_str(input, "mode")?.map(ToString::to_string), |
| 488 | // Authority declarations: read strictly. A malformed value that |
| 489 | // silently reads as "unset" is a restriction that evaporates. |
| 490 | allow_shell: optional_bool_opt(input, "allow_shell")?, |
| 491 | trust_mode: optional_bool_opt(input, "trust_mode")?, |
| 492 | auto_approve: optional_bool_opt(input, "auto_approve")?, |
| 493 | owner_session_id: Some(context.state_namespace.clone()), |
| 494 | }; |
| 495 | let task_id = crate::task_manager::TaskManager::new_task_id(); |
| 496 | if let Some(work) = context.runtime.work.as_ref() |
| 497 | && let Err(err) = work.register_operation( |
| 498 | &context.state_namespace, |
| 499 | OperationIntent::new( |
| 500 | format!("task:{task_id}"), |
| 501 | prompt, |
| 502 | true, |
| 503 | "task_create", |
| 504 | &task_id, |
| 505 | ), |
| 506 | ) |
| 507 | { |
| 508 | // Bookkeeping must not veto the task: every later reconcile is |
| 509 | // guarded by `has_operation_binding`, so an unbound task merely |
| 510 | // goes unreported on the Work surface. |
| 511 | tracing::warn!( |
| 512 | task_id = %task_id, |
| 513 | error = %err, |
| 514 | "task work-graph registration skipped; running unbound" |
| 515 | ); |
| 516 | } |
| 517 | let task = match manager.add_task_with_id(req, task_id.clone()).await { |
| 518 | Ok(task) => task, |
| 519 | Err(err) => { |
| 520 | if let Some(work) = context.runtime.work.as_ref() { |
| 521 | let _ = work.reconcile_operation( |
| 522 | &context.state_namespace, |
| 523 | OperationOwnerSnapshot::new( |
| 524 | format!("task:{task_id}"), |
| 525 | OwnerState::Failed, |
| 526 | 1, |
| 527 | Utc::now().timestamp_millis(), |
| 528 | ), |
| 529 | ); |
| 530 | } |
| 531 | return Err(ToolError::execution_failed(err.to_string())); |
| 532 | } |
| 533 | }; |
| 534 | let lifecycle_warning = reconcile_task_record(context, &task).err().map(|err| { |
| 535 | tracing::warn!(task_id = %task.id, error = %err, "task was created but Work lifecycle reconciliation failed"); |
| 536 | err.to_string() |
| 537 | }); |
| 538 | task_result_with_lifecycle_warning("task_create", &task, lifecycle_warning.as_deref()) |
| 539 | } |
| 540 | |
| 541 | async fn execute_list( |
| 542 | &self, |
| 543 | input: &Value, |
| 544 | context: &ToolContext, |
| 545 | ) -> Result<ToolResult, ToolError> { |
| 546 | let manager = context |
| 547 | .runtime |
| 548 | .task_manager |
| 549 | .as_ref() |
| 550 | .ok_or_else(|| ToolError::not_available("TaskManager is not attached"))?; |
| 551 | let limit = optional_u64(input, "limit", 20)?.clamp(1, 100) as usize; |
| 552 | let tasks = manager |
| 553 | .list_tasks_for_owner(Some(limit), None, &context.state_namespace) |
| 554 | .await |
| 555 | .map_err(|error| ToolError::execution_failed(error.to_string()))?; |
| 556 | ToolResult::json(&json!({ |
| 557 | "summary": format!("{} durable task(s)", tasks.len()), |
| 558 | "tasks": tasks, |
| 559 | })) |
| 560 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 561 | } |
| 562 | |
| 563 | async fn execute_read( |
| 564 | &self, |
| 565 | input: &Value, |
| 566 | context: &ToolContext, |
| 567 | ) -> Result<ToolResult, ToolError> { |
| 568 | let task = read_task_for_input(input, context).await?; |
| 569 | task_result("task_read", &task) |
| 570 | } |
| 571 | |
| 572 | async fn execute_cancel( |
| 573 | &self, |
| 574 | input: &Value, |
| 575 | context: &ToolContext, |
| 576 | ) -> Result<ToolResult, ToolError> { |
| 577 | let manager = context |
| 578 | .runtime |
| 579 | .task_manager |
| 580 | .as_ref() |
| 581 | .ok_or_else(|| ToolError::not_available("TaskManager is not attached"))?; |
| 582 | let task_id = required_str(input, "task_id")?; |
| 583 | let cancellation = if context.runtime.active_task_id.as_deref() == Some(task_id) { |
| 584 | // `active_task_id` is stamped from the immutable runtime thread |
| 585 | // record, not model input. Preserve self-cancel for a running task, |
| 586 | // but fail closed for ownerless legacy records. |
| 587 | manager.cancel_task_for_active_runtime(task_id).await |
| 588 | } else { |
| 589 | manager |
| 590 | .cancel_task_for_owner(task_id, &context.state_namespace) |
| 591 | .await |
| 592 | } |
| 593 | .map_err(|e| ToolError::execution_failed(e.to_string()))?; |
| 594 | let task = cancellation.task; |
| 595 | let cancel_outcome = match cancellation.disposition { |
| 596 | TaskCancelDisposition::Forced => CancelOutcome::Forced, |
| 597 | TaskCancelDisposition::Requested => CancelOutcome::Requested, |
| 598 | TaskCancelDisposition::AlreadyFinished => CancelOutcome::AlreadyFinished, |
| 599 | }; |
| 600 | let mut lifecycle_warnings = Vec::new(); |
| 601 | if let Some(work) = context.runtime.work.as_ref() { |
| 602 | let external = format!("task:{}", task.id); |
| 603 | if work.has_operation_binding(Some(&context.state_namespace), &external) |
| 604 | && let Err(err) = work.reconcile_observation( |
| 605 | &context.state_namespace, |
| 606 | &external, |
| 607 | OperationObservation::CancelUpdate { |
| 608 | outcome: cancel_outcome, |
| 609 | at: Utc::now().timestamp_millis(), |
| 610 | }, |
| 611 | ) |
| 612 | { |
| 613 | tracing::warn!(task_id = %task.id, error = %err, "task was cancelled but Work cancel reconciliation failed"); |
| 614 | lifecycle_warnings.push(err); |
| 615 | } |
| 616 | } |
| 617 | if let Err(err) = reconcile_task_record(context, &task) { |
| 618 | tracing::warn!(task_id = %task.id, error = %err, "task cancellation succeeded but owner-state reconciliation failed"); |
| 619 | lifecycle_warnings.push(err.to_string()); |
| 620 | } |
| 621 | let lifecycle_warning = |
| 622 | (!lifecycle_warnings.is_empty()).then(|| lifecycle_warnings.join("; ")); |
| 623 | task_result_with_lifecycle_warning("task_cancel", &task, lifecycle_warning.as_deref()) |
| 624 | } |
| 625 | |
| 626 | async fn execute_gate_run( |
| 627 | &self, |
| 628 | input: &Value, |
| 629 | context: &ToolContext, |
| 630 | ) -> Result<ToolResult, ToolError> { |
| 631 | crate::core::engine::tool_catalog::enforce_tool_denial(context, "task_gate_run", input)?; |
| 632 | if context.shell_policy != crate::worker_profile::ShellPolicy::Full { |
| 633 | return Err(ToolError::permission_denied( |
| 634 | "Gate commands require full shell permission.", |
| 635 | )); |
| 636 | } |
| 637 | let gate = required_str(input, "gate")?.to_string(); |
| 638 | let command = required_str(input, "command")?.to_string(); |
| 639 | let timeout_ms = optional_u64(input, "timeout_ms", DEFAULT_GATE_TIMEOUT_MS)? |
| 640 | .clamp(1_000, MAX_GATE_TIMEOUT_MS); |
| 641 | let cwd = resolve_cwd(context, optional_str(input, "cwd")?)?; |
| 642 | |
| 643 | let safety = analyze_command(&command); |
| 644 | if !context.auto_approve && matches!(safety.level, SafetyLevel::Dangerous) { |
| 645 | return Ok(ToolResult::error(format!( |
| 646 | "BLOCKED: gate command classified dangerous: {}", |
| 647 | safety.reasons.join("; ") |
| 648 | )) |
| 649 | .with_metadata(json!({ |
| 650 | "safety_level": "dangerous", |
| 651 | "blocked": true, |
| 652 | "reasons": safety.reasons, |
| 653 | }))); |
| 654 | } |
| 655 | |
| 656 | let started = Instant::now(); |
| 657 | let mut cmd = build_gate_command(&command, &cwd); |
| 658 | let output = |
| 659 | tokio::time::timeout(std::time::Duration::from_millis(timeout_ms), cmd.output()).await; |
| 660 | |
| 661 | let duration_ms = u64::try_from(started.elapsed().as_millis()).unwrap_or(u64::MAX); |
| 662 | let (exit_code, stdout, stderr, timed_out, spawn_error) = match output { |
| 663 | Ok(Ok(out)) => ( |
| 664 | out.status.code(), |
| 665 | String::from_utf8_lossy(&out.stdout).to_string(), |
| 666 | String::from_utf8_lossy(&out.stderr).to_string(), |
| 667 | false, |
| 668 | None, |
| 669 | ), |
| 670 | Ok(Err(err)) => ( |
| 671 | None, |
| 672 | String::new(), |
| 673 | String::new(), |
| 674 | false, |
| 675 | Some(err.to_string()), |
| 676 | ), |
| 677 | Err(_) => (None, String::new(), String::new(), true, None), |
| 678 | }; |
| 679 | |
| 680 | let full_log = format!( |
| 681 | "$ {command}\n\n[stdout]\n{stdout}\n\n[stderr]\n{stderr}\n{}", |
| 682 | spawn_error |
| 683 | .as_ref() |
| 684 | .map(|e| format!("\n[spawn_error]\n{e}\n")) |
| 685 | .unwrap_or_default() |
| 686 | ); |
| 687 | let summary_source = if !stderr.trim().is_empty() { |
| 688 | stderr.as_str() |
| 689 | } else if !stdout.trim().is_empty() { |
| 690 | stdout.as_str() |
| 691 | } else { |
| 692 | spawn_error.as_deref().unwrap_or("(no output)") |
| 693 | }; |
| 694 | let summary = summarize(summary_source, MAX_SUMMARY_CHARS); |
| 695 | let status = if timed_out { |
| 696 | "timeout" |
| 697 | } else if spawn_error.is_some() { |
| 698 | "failed" |
| 699 | } else if exit_code == Some(0) { |
| 700 | "passed" |
| 701 | } else { |
| 702 | "failed" |
| 703 | }; |
| 704 | let classification = classify_gate_failure(&gate, status, timed_out, &stderr, &stdout); |
| 705 | let log_path = write_runtime_artifact(context, "gate", &full_log).await?; |
| 706 | let gate_record = TaskGateRecord { |
| 707 | id: format!("gate_{}", &Uuid::new_v4().to_string()[..8]), |
| 708 | gate: gate.clone(), |
| 709 | command: command.clone(), |
| 710 | cwd: cwd.clone(), |
| 711 | exit_code, |
| 712 | status: status.to_string(), |
| 713 | classification, |
| 714 | duration_ms, |
| 715 | summary: summary.clone(), |
| 716 | log_path: log_path.clone(), |
| 717 | recorded_at: Utc::now(), |
| 718 | }; |
| 719 | |
| 720 | let content = json!({ |
| 721 | "gate": gate_record, |
| 722 | "stdout_summary": summarize(&stdout, MAX_SUMMARY_CHARS), |
| 723 | "stderr_summary": summarize(&stderr, MAX_SUMMARY_CHARS), |
| 724 | }); |
| 725 | let mut metadata = json!({ |
| 726 | "command": command, |
| 727 | "cwd": cwd, |
| 728 | "exit_code": exit_code, |
| 729 | "duration_ms": duration_ms, |
| 730 | "timed_out": timed_out, |
| 731 | "task_updates": { |
| 732 | "gate": gate_record, |
| 733 | "artifacts": artifact_updates("gate_log", log_path.clone(), &summary) |
| 734 | } |
| 735 | }); |
| 736 | if let Some(path) = log_path { |
| 737 | metadata["artifact_path"] = json!(path); |
| 738 | } |
| 739 | Ok(ToolResult::json(&content) |
| 740 | .map_err(|e| ToolError::execution_failed(e.to_string()))? |
| 741 | .with_metadata(metadata)) |
| 742 | } |
| 743 | |
| 744 | async fn execute_pr_attempt_record( |
| 745 | &self, |
| 746 | input: &Value, |
| 747 | context: &ToolContext, |
| 748 | ) -> Result<ToolResult, ToolError> { |
| 749 | let task_id = read_task_for_input(input, context).await?.id; |
| 750 | let base_sha = git_output(&context.workspace, &["rev-parse", "HEAD"]) |
| 751 | .await |
| 752 | .ok(); |
| 753 | let head_sha = base_sha.clone(); |
| 754 | let branch = git_output(&context.workspace, &["rev-parse", "--abbrev-ref", "HEAD"]) |
| 755 | .await |
| 756 | .ok(); |
| 757 | let diff = git_output(&context.workspace, &["diff", "--binary", "--no-color"]).await?; |
| 758 | if diff.trim().is_empty() { |
| 759 | return Ok(ToolResult::error( |
| 760 | "No working-tree diff to record as an attempt.", |
| 761 | )); |
| 762 | } |
| 763 | let changed_files = git_output(&context.workspace, &["diff", "--name-only"]) |
| 764 | .await? |
| 765 | .lines() |
| 766 | .filter(|line| !line.trim().is_empty()) |
| 767 | .map(ToString::to_string) |
| 768 | .collect::<Vec<_>>(); |
| 769 | let patch_path = write_task_artifact_for(context, &task_id, "attempt_patch", &diff).await?; |
| 770 | let attempt = TaskAttemptRecord { |
| 771 | id: format!("attempt_{}", &Uuid::new_v4().to_string()[..8]), |
| 772 | attempt_group_id: optional_str(input, "attempt_group_id")? |
| 773 | .map(ToString::to_string) |
| 774 | .unwrap_or_else(|| format!("attempt_group_{}", &Uuid::new_v4().to_string()[..8])), |
| 775 | attempt_index: optional_u64(input, "attempt_index", 1)?.max(1) as u32, |
| 776 | attempt_count: optional_u64(input, "attempt_count", 1)?.max(1) as u32, |
| 777 | base_ref: branch.clone(), |
| 778 | base_sha, |
| 779 | head_ref: branch, |
| 780 | head_sha, |
| 781 | summary: required_str(input, "summary")?.to_string(), |
| 782 | changed_files, |
| 783 | patch_path: patch_path.clone(), |
| 784 | verification: input |
| 785 | .get("verification") |
| 786 | .and_then(Value::as_array) |
| 787 | .map(|items| { |
| 788 | items |
| 789 | .iter() |
| 790 | .filter_map(Value::as_str) |
| 791 | .map(ToString::to_string) |
| 792 | .collect() |
| 793 | }) |
| 794 | .unwrap_or_default(), |
| 795 | selected: false, |
| 796 | recorded_at: Utc::now(), |
| 797 | }; |
| 798 | let metadata = json!({ |
| 799 | "task_id": task_id, |
| 800 | "task_updates": { |
| 801 | "attempt": attempt, |
| 802 | "artifacts": artifact_updates("attempt_patch", patch_path.clone(), "Captured git diff for PR attempt") |
| 803 | } |
| 804 | }); |
| 805 | if context.runtime.active_task_id.as_deref() != Some(task_id.as_str()) |
| 806 | && let Some(manager) = context.runtime.task_manager.as_ref() |
| 807 | { |
| 808 | manager |
| 809 | .record_tool_metadata(&task_id, &metadata) |
| 810 | .await |
| 811 | .map_err(|e| ToolError::execution_failed(e.to_string()))?; |
| 812 | } |
| 813 | Ok(ToolResult::json(&metadata) |
| 814 | .map_err(|e| ToolError::execution_failed(e.to_string()))? |
| 815 | .with_metadata(metadata)) |
| 816 | } |
| 817 | |
| 818 | async fn execute_pr_attempt_list( |
| 819 | &self, |
| 820 | input: &Value, |
| 821 | context: &ToolContext, |
| 822 | ) -> Result<ToolResult, ToolError> { |
| 823 | let task = read_task_for_input(input, context).await?; |
| 824 | ToolResult::json(&json!({ "task_id": task.id, "attempts": task.attempts })) |
| 825 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 826 | } |
| 827 | |
| 828 | async fn execute_pr_attempt_read( |
| 829 | &self, |
| 830 | input: &Value, |
| 831 | context: &ToolContext, |
| 832 | ) -> Result<ToolResult, ToolError> { |
| 833 | let task = read_task_for_input(input, context).await?; |
| 834 | let attempt_id = required_str(input, "attempt_id")?; |
| 835 | let attempt = task |
| 836 | .attempts |
| 837 | .iter() |
| 838 | .find(|attempt| attempt.id == attempt_id) |
| 839 | .ok_or_else(|| ToolError::invalid_input(format!("Attempt not found: {attempt_id}")))?; |
| 840 | ToolResult::json(attempt).map_err(|e| ToolError::execution_failed(e.to_string())) |
| 841 | } |
| 842 | |
| 843 | async fn execute_pr_attempt_preflight( |
| 844 | &self, |
| 845 | input: &Value, |
| 846 | context: &ToolContext, |
| 847 | ) -> Result<ToolResult, ToolError> { |
| 848 | let manager = context |
| 849 | .runtime |
| 850 | .task_manager |
| 851 | .as_ref() |
| 852 | .ok_or_else(|| ToolError::not_available("TaskManager is not attached"))?; |
| 853 | let task = read_task_for_input(input, context).await?; |
| 854 | let attempt_id = required_str(input, "attempt_id")?; |
| 855 | let attempt = task |
| 856 | .attempts |
| 857 | .iter() |
| 858 | .find(|attempt| attempt.id == attempt_id) |
| 859 | .ok_or_else(|| ToolError::invalid_input(format!("Attempt not found: {attempt_id}")))?; |
| 860 | let patch_ref = attempt |
| 861 | .patch_path |
| 862 | .as_ref() |
| 863 | .ok_or_else(|| ToolError::invalid_input("Attempt has no patch artifact"))?; |
| 864 | let patch_path = manager.artifact_absolute_path(patch_ref); |
| 865 | let workspace = context.workspace.clone(); |
| 866 | let out = tokio::task::spawn_blocking(move || { |
| 867 | crate::dependencies::Git::command() |
| 868 | .ok_or_else(|| std::io::Error::new(std::io::ErrorKind::NotFound, "git not found"))? |
| 869 | .args(["apply", "--check"]) |
| 870 | .arg(&patch_path) |
| 871 | .current_dir(&workspace) |
| 872 | .output() |
| 873 | }) |
| 874 | .await |
| 875 | .map_err(|join_err| { |
| 876 | // Surface the otherwise-discarded join error for debugging; the |
| 877 | // returned ToolError (and thus user-facing behavior) is unchanged. |
| 878 | tracing::debug!(error = %join_err, "git apply --check spawn_blocking task failed to join"); |
| 879 | ToolError::execution_failed(format!("git apply --check panicked: {join_err}")) |
| 880 | })? |
| 881 | .map_err(|e| ToolError::execution_failed(format!("git apply --check failed: {e}")))?; |
| 882 | let stdout = String::from_utf8_lossy(&out.stdout).to_string(); |
| 883 | let stderr = String::from_utf8_lossy(&out.stderr).to_string(); |
| 884 | ToolResult::json(&json!({ |
| 885 | "attempt_id": attempt_id, |
| 886 | "patch_path": patch_ref, |
| 887 | "would_apply": out.status.success(), |
| 888 | "exit_code": out.status.code(), |
| 889 | "stdout_summary": summarize(&stdout, MAX_SUMMARY_CHARS), |
| 890 | "stderr_summary": summarize(&stderr, MAX_SUMMARY_CHARS), |
| 891 | "mutated_worktree": false |
| 892 | })) |
| 893 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 894 | } |
| 895 | } |
| 896 | |
| 897 | #[async_trait] |
| 898 | impl ToolSpec for TaskShellStartTool { |
| 899 | fn name(&self) -> &'static str { |
| 900 | "task_shell_start" |
| 901 | } |
| 902 | |
| 903 | fn description(&self) -> &'static str { |
| 904 | "Start a long-running shell command in the background and return a shell task_id immediately. Completion is delivered automatically as an internal runtime event and remains visible in the task/status surface; use task_shell_wait only for early output, explicit barriers, or gate evidence on the active durable task." |
| 905 | } |
| 906 | |
| 907 | fn input_schema(&self) -> Value { |
| 908 | json!({ |
| 909 | "type": "object", |
| 910 | "properties": { |
| 911 | "command": { "type": "string" }, |
| 912 | "cwd": { "type": "string", "description": "Optional working directory within the workspace." }, |
| 913 | "timeout_ms": { "type": "integer", "minimum": 1000, "maximum": 600000 }, |
| 914 | "stdin": { "type": "string" }, |
| 915 | "tty": { "type": "boolean" } |
| 916 | }, |
| 917 | "required": ["command"], |
| 918 | "additionalProperties": false |
| 919 | }) |
| 920 | } |
| 921 | |
| 922 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 923 | vec![ |
| 924 | ToolCapability::ExecutesCode, |
| 925 | ToolCapability::RequiresApproval, |
| 926 | ] |
| 927 | } |
| 928 | |
| 929 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 930 | ApprovalRequirement::Required |
| 931 | } |
| 932 | |
| 933 | fn starts_detached_for(&self, input: &Value) -> bool { |
| 934 | input.get("command").and_then(Value::as_str).is_some() |
| 935 | } |
| 936 | |
| 937 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 938 | crate::core::engine::tool_catalog::enforce_tool_denial(context, self.name(), &input)?; |
| 939 | let mut shell_input = json!({ |
| 940 | "command": required_str(&input, "command")?, |
| 941 | "background": true, |
| 942 | "timeout_ms": optional_u64(&input, "timeout_ms", DEFAULT_GATE_TIMEOUT_MS)? |
| 943 | .clamp(1_000, MAX_GATE_TIMEOUT_MS), |
| 944 | }); |
| 945 | if let Some(cwd) = optional_str(&input, "cwd")? { |
| 946 | let cwd = resolve_cwd(context, Some(cwd))?; |
| 947 | shell_input["cwd"] = json!(cwd); |
| 948 | } |
| 949 | if let Some(stdin) = optional_str(&input, "stdin")? { |
| 950 | shell_input["stdin"] = json!(stdin); |
| 951 | } |
| 952 | if optional_bool(&input, "tty", false)? { |
| 953 | shell_input["tty"] = json!(true); |
| 954 | } |
| 955 | let mut result = BashTool::new("Bash").execute(shell_input, context).await?; |
| 956 | if let Some(metadata) = result.metadata.as_mut() { |
| 957 | metadata["background"] = json!(true); |
| 958 | metadata["task_shell"] = json!(true); |
| 959 | } |
| 960 | Ok(result) |
| 961 | } |
| 962 | } |
| 963 | |
| 964 | #[async_trait] |
| 965 | impl ToolSpec for TaskShellWaitTool { |
| 966 | fn name(&self) -> &'static str { |
| 967 | "task_shell_wait" |
| 968 | } |
| 969 | |
| 970 | fn description(&self) -> &'static str { |
| 971 | "Poll a background shell task without blocking the agent indefinitely. Completion is delivered automatically; use this only for early output, explicit barriers, or gate evidence. If `gate` is supplied and the shell task has completed, records structured gate evidence on the active durable task." |
| 972 | } |
| 973 | |
| 974 | fn input_schema(&self) -> Value { |
| 975 | json!({ |
| 976 | "type": "object", |
| 977 | "properties": { |
| 978 | "task_id": { "type": "string", "description": "Background shell task id returned by task_shell_start or `Bash`." }, |
| 979 | "wait": { "type": "boolean", "default": false }, |
| 980 | "timeout_ms": { "type": "integer", "minimum": 1000, "maximum": 600000 }, |
| 981 | "gate": { "type": "string", "enum": ["fmt", "check", "clippy", "test", "custom"] }, |
| 982 | "command": { "type": "string", "description": "Original command, used when recording gate evidence." } |
| 983 | }, |
| 984 | "required": ["task_id"], |
| 985 | "additionalProperties": false |
| 986 | }) |
| 987 | } |
| 988 | |
| 989 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 990 | vec![ToolCapability::ReadOnly] |
| 991 | } |
| 992 | |
| 993 | fn approval_requirement(&self) -> ApprovalRequirement { |
| 994 | ApprovalRequirement::Auto |
| 995 | } |
| 996 | |
| 997 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 998 | crate::core::engine::tool_catalog::enforce_tool_denial(context, self.name(), &input)?; |
| 999 | let shell_input = task_shell_wait_input(input.clone()); |
| 1000 | let result = BashTool::alias("exec_shell_wait", "wait") |
| 1001 | .execute(shell_input, context) |
| 1002 | .await?; |
| 1003 | let Some(gate) = optional_str(&input, "gate")? else { |
| 1004 | return Ok(result); |
| 1005 | }; |
| 1006 | let status = result |
| 1007 | .metadata |
| 1008 | .as_ref() |
| 1009 | .and_then(|m| m.get("status")) |
| 1010 | .and_then(Value::as_str) |
| 1011 | .unwrap_or("Running"); |
| 1012 | if status == "Running" { |
| 1013 | return Ok(result); |
| 1014 | } |
| 1015 | let exit_code = result |
| 1016 | .metadata |
| 1017 | .as_ref() |
| 1018 | .and_then(|m| m.get("exit_code")) |
| 1019 | .and_then(Value::as_i64) |
| 1020 | .and_then(|v| i32::try_from(v).ok()); |
| 1021 | let duration_ms = result |
| 1022 | .metadata |
| 1023 | .as_ref() |
| 1024 | .and_then(|m| m.get("duration_ms")) |
| 1025 | .and_then(Value::as_u64) |
| 1026 | .unwrap_or_default(); |
| 1027 | let command = optional_str(&input, "command")?.unwrap_or("(background shell)"); |
| 1028 | let log_path = write_runtime_artifact(context, "background_gate", &result.content).await?; |
| 1029 | let gate_status = if exit_code == Some(0) { |
| 1030 | "passed" |
| 1031 | } else if status == "TimedOut" { |
| 1032 | "timeout" |
| 1033 | } else { |
| 1034 | "failed" |
| 1035 | }; |
| 1036 | let gate_record = TaskGateRecord { |
| 1037 | id: format!("gate_{}", &Uuid::new_v4().to_string()[..8]), |
| 1038 | gate: gate.to_string(), |
| 1039 | command: command.to_string(), |
| 1040 | cwd: context.workspace.clone(), |
| 1041 | exit_code, |
| 1042 | status: gate_status.to_string(), |
| 1043 | classification: classify_gate_failure( |
| 1044 | gate, |
| 1045 | gate_status, |
| 1046 | status == "TimedOut", |
| 1047 | &result.content, |
| 1048 | "", |
| 1049 | ), |
| 1050 | duration_ms, |
| 1051 | summary: summarize(&result.content, MAX_SUMMARY_CHARS), |
| 1052 | log_path: log_path.clone(), |
| 1053 | recorded_at: Utc::now(), |
| 1054 | }; |
| 1055 | let mut metadata = result.metadata.clone().unwrap_or_else(|| json!({})); |
| 1056 | metadata["background"] = json!(true); |
| 1057 | metadata["task_updates"] = json!({ |
| 1058 | "gate": gate_record, |
| 1059 | "artifacts": artifact_updates("background_gate_log", log_path, "Background shell gate output") |
| 1060 | }); |
| 1061 | Ok(result.with_metadata(metadata)) |
| 1062 | } |
| 1063 | } |
| 1064 | |
| 1065 | fn reconcile_task_record(context: &ToolContext, task: &TaskRecord) -> Result<(), ToolError> { |
| 1066 | let Some(work) = context.runtime.work.as_ref() else { |
| 1067 | return Ok(()); |
| 1068 | }; |
| 1069 | let external = format!("task:{}", task.id); |
| 1070 | if !work.has_operation_binding(Some(&context.state_namespace), &external) { |
| 1071 | return Ok(()); |
| 1072 | } |
| 1073 | work.reconcile_operation( |
| 1074 | &context.state_namespace, |
| 1075 | task_owner_snapshot( |
| 1076 | &task.id, |
| 1077 | task.status, |
| 1078 | task.lifecycle_seq, |
| 1079 | task.created_at, |
| 1080 | task.started_at, |
| 1081 | task.ended_at, |
| 1082 | ), |
| 1083 | ) |
| 1084 | .map(|_| ()) |
| 1085 | .map_err(ToolError::execution_failed) |
| 1086 | } |
| 1087 | |
| 1088 | fn task_result(label: &str, task: &TaskRecord) -> Result<ToolResult, ToolError> { |
| 1089 | task_result_with_lifecycle_warning(label, task, None) |
| 1090 | } |
| 1091 | |
| 1092 | fn task_result_with_lifecycle_warning( |
| 1093 | label: &str, |
| 1094 | task: &TaskRecord, |
| 1095 | lifecycle_warning: Option<&str>, |
| 1096 | ) -> Result<ToolResult, ToolError> { |
| 1097 | ToolResult::json(&json!({ |
| 1098 | "summary": format!("{label}: {} ({:?})", task.id, task.status), |
| 1099 | "task": task, |
| 1100 | "lifecycle_warning": lifecycle_warning, |
| 1101 | "execution_ownership": if task.execution_scope.is_some() { "scope_bound" } else { "unverified" }, |
| 1102 | })) |
| 1103 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 1104 | } |
| 1105 | |
| 1106 | fn resolve_cwd(context: &ToolContext, raw: Option<&str>) -> Result<PathBuf, ToolError> { |
| 1107 | match raw { |
| 1108 | Some(path) => { |
| 1109 | let resolved = context.resolve_path(path)?; |
| 1110 | if resolved.is_dir() { |
| 1111 | Ok(resolved) |
| 1112 | } else { |
| 1113 | Err(ToolError::invalid_input(format!( |
| 1114 | "cwd must be a directory: {path}" |
| 1115 | ))) |
| 1116 | } |
| 1117 | } |
| 1118 | None => Ok(context.workspace.clone()), |
| 1119 | } |
| 1120 | } |
| 1121 | |
| 1122 | async fn write_runtime_artifact( |
| 1123 | context: &ToolContext, |
| 1124 | label: &str, |
| 1125 | content: &str, |
| 1126 | ) -> Result<Option<PathBuf>, ToolError> { |
| 1127 | let Some(task_id) = context.runtime.active_task_id.as_deref() else { |
| 1128 | return Ok(None); |
| 1129 | }; |
| 1130 | let manager = context.runtime.task_manager.as_ref(); |
| 1131 | if let Some(manager) = manager { |
| 1132 | return manager |
| 1133 | .write_task_artifact(task_id, label, content) |
| 1134 | .map(Some) |
| 1135 | .map_err(|e| ToolError::execution_failed(e.to_string())); |
| 1136 | } |
| 1137 | let Some(data_dir) = context.runtime.task_data_dir.as_ref() else { |
| 1138 | return Ok(None); |
| 1139 | }; |
| 1140 | let artifact_dir = data_dir.join("artifacts").join(task_id); |
| 1141 | let filename = format!( |
| 1142 | "{}_{}.txt", |
| 1143 | Utc::now().format("%Y%m%dT%H%M%S%.3fZ"), |
| 1144 | sanitize_filename(label) |
| 1145 | ); |
| 1146 | let absolute = artifact_dir.join(filename); |
| 1147 | let content_owned = content.to_owned(); |
| 1148 | let abs = absolute.clone(); |
| 1149 | tokio::task::spawn_blocking(move || { |
| 1150 | std::fs::create_dir_all(&artifact_dir)?; |
| 1151 | std::fs::write(&abs, content_owned)?; |
| 1152 | Ok::<(), std::io::Error>(()) |
| 1153 | }) |
| 1154 | .await |
| 1155 | .map_err(|e| { |
| 1156 | // Surface the otherwise-discarded join error for debugging; the |
| 1157 | // returned ToolError (and thus user-facing behavior) is unchanged. |
| 1158 | tracing::debug!(error = %e, "artifact write spawn_blocking task failed to join"); |
| 1159 | ToolError::execution_failed(format!("artifact write task panicked: {e}")) |
| 1160 | })? |
| 1161 | .map_err(|e| ToolError::execution_failed(format!("write artifact: {e}")))?; |
| 1162 | Ok(Some( |
| 1163 | absolute |
| 1164 | .strip_prefix(data_dir) |
| 1165 | .map(PathBuf::from) |
| 1166 | .unwrap_or(absolute), |
| 1167 | )) |
| 1168 | } |
| 1169 | |
| 1170 | async fn write_task_artifact_for( |
| 1171 | context: &ToolContext, |
| 1172 | task_id: &str, |
| 1173 | label: &str, |
| 1174 | content: &str, |
| 1175 | ) -> Result<Option<PathBuf>, ToolError> { |
| 1176 | if let Some(manager) = context.runtime.task_manager.as_ref() { |
| 1177 | return manager |
| 1178 | .write_task_artifact(task_id, label, content) |
| 1179 | .map(Some) |
| 1180 | .map_err(|e| ToolError::execution_failed(e.to_string())); |
| 1181 | } |
| 1182 | if context.runtime.active_task_id.as_deref() != Some(task_id) { |
| 1183 | return Ok(None); |
| 1184 | } |
| 1185 | write_runtime_artifact(context, label, content).await |
| 1186 | } |
| 1187 | |
| 1188 | fn artifact_updates(label: &str, path: Option<PathBuf>, summary: &str) -> Value { |
| 1189 | match path { |
| 1190 | Some(path) => json!([TaskArtifactRef { |
| 1191 | label: label.to_string(), |
| 1192 | path, |
| 1193 | summary: summarize(summary, 240), |
| 1194 | created_at: Utc::now(), |
| 1195 | }]), |
| 1196 | None => json!([]), |
| 1197 | } |
| 1198 | } |
| 1199 | |
| 1200 | async fn read_task_for_input( |
| 1201 | input: &Value, |
| 1202 | context: &ToolContext, |
| 1203 | ) -> Result<TaskRecord, ToolError> { |
| 1204 | let manager = context |
| 1205 | .runtime |
| 1206 | .task_manager |
| 1207 | .as_ref() |
| 1208 | .ok_or_else(|| ToolError::not_available("TaskManager is not attached"))?; |
| 1209 | let task_id = task_id_from_input_or_context(input, context)?; |
| 1210 | if context.runtime.active_task_id.as_deref() == Some(task_id.as_str()) { |
| 1211 | // Runtime thread construction stamps `active_task_id` from its durable |
| 1212 | // thread record. It is not a tool parameter, so an owned task may read |
| 1213 | // itself even though its execution engine has a distinct namespace. |
| 1214 | manager |
| 1215 | .get_task_for_active_runtime(&task_id) |
| 1216 | .await |
| 1217 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 1218 | } else { |
| 1219 | manager |
| 1220 | .get_task_for_owner(&task_id, &context.state_namespace) |
| 1221 | .await |
| 1222 | .map_err(|e| ToolError::execution_failed(e.to_string())) |
| 1223 | } |
| 1224 | } |
| 1225 | |
| 1226 | fn task_id_from_input_or_context( |
| 1227 | input: &Value, |
| 1228 | context: &ToolContext, |
| 1229 | ) -> Result<String, ToolError> { |
| 1230 | optional_str(input, "task_id")? |
| 1231 | .map(ToString::to_string) |
| 1232 | .or_else(|| context.runtime.active_task_id.clone()) |
| 1233 | .ok_or_else(|| { |
| 1234 | ToolError::invalid_input("task_id is required when no durable task is active") |
| 1235 | }) |
| 1236 | } |
| 1237 | |
| 1238 | fn task_id_schema() -> Value { |
| 1239 | json!({ |
| 1240 | "type": "object", |
| 1241 | "properties": { |
| 1242 | "task_id": { "type": "string", "description": "Task id; defaults to active task." } |
| 1243 | }, |
| 1244 | "additionalProperties": false |
| 1245 | }) |
| 1246 | } |
| 1247 | |
| 1248 | async fn git_output(workspace: &Path, args: &[&str]) -> Result<String, ToolError> { |
| 1249 | let args_owned: Vec<String> = args.iter().map(|s| (*s).to_owned()).collect(); |
| 1250 | let cwd = workspace.to_path_buf(); |
| 1251 | let out = tokio::task::spawn_blocking(move || { |
| 1252 | let arg_refs: Vec<&str> = args_owned.iter().map(String::as_str).collect(); |
| 1253 | crate::dependencies::Git::output(&arg_refs, &cwd) |
| 1254 | }) |
| 1255 | .await |
| 1256 | .map_err(|e| { |
| 1257 | // Surface the otherwise-discarded join error for debugging; the |
| 1258 | // returned ToolError (and thus user-facing behavior) is unchanged. |
| 1259 | tracing::debug!(error = %e, "git spawn_blocking task failed to join"); |
| 1260 | ToolError::execution_failed(format!("git task panicked: {e}")) |
| 1261 | })? |
| 1262 | .map_err(|e| ToolError::execution_failed(format!("failed to run git: {e}")))?; |
| 1263 | if !out.status.success() { |
| 1264 | return Err(ToolError::execution_failed(format!( |
| 1265 | "git {} failed: {}", |
| 1266 | args.join(" "), |
| 1267 | String::from_utf8_lossy(&out.stderr).trim() |
| 1268 | ))); |
| 1269 | } |
| 1270 | Ok(String::from_utf8_lossy(&out.stdout).trim_end().to_string()) |
| 1271 | } |
| 1272 | |
| 1273 | fn classify_gate_failure( |
| 1274 | gate: &str, |
| 1275 | status: &str, |
| 1276 | timed_out: bool, |
| 1277 | stderr: &str, |
| 1278 | stdout: &str, |
| 1279 | ) -> String { |
| 1280 | if timed_out { |
| 1281 | return "timeout".to_string(); |
| 1282 | } |
| 1283 | if status == "passed" { |
| 1284 | return "passed".to_string(); |
| 1285 | } |
| 1286 | let haystack = format!("{stderr}\n{stdout}").to_ascii_lowercase(); |
| 1287 | if haystack.contains("address already in use") || haystack.contains("port") { |
| 1288 | "environment_port_binding".to_string() |
| 1289 | } else if gate == "clippy" || haystack.contains("warning:") { |
| 1290 | "lint_failure".to_string() |
| 1291 | } else if gate == "test" || haystack.contains("test result: failed") { |
| 1292 | "test_failure".to_string() |
| 1293 | } else if haystack.contains("error: could not compile") |
| 1294 | || haystack.contains("compilation failed") |
| 1295 | { |
| 1296 | "compile_error".to_string() |
| 1297 | } else { |
| 1298 | "environment_or_tooling_failure".to_string() |
| 1299 | } |
| 1300 | } |
| 1301 | |
| 1302 | fn summarize(text: &str, limit: usize) -> String { |
| 1303 | let mut out = String::new(); |
| 1304 | for (idx, ch) in text.chars().enumerate() { |
| 1305 | if idx >= limit.saturating_sub(3) { |
| 1306 | out.push_str("..."); |
| 1307 | return out; |
| 1308 | } |
| 1309 | if ch.is_control() && ch != '\n' && ch != '\t' { |
| 1310 | continue; |
| 1311 | } |
| 1312 | out.push(ch); |
| 1313 | } |
| 1314 | if out.trim().is_empty() { |
| 1315 | "(no output)".to_string() |
| 1316 | } else { |
| 1317 | out |
| 1318 | } |
| 1319 | } |
| 1320 | |
| 1321 | fn sanitize_filename(input: &str) -> String { |
| 1322 | let mut out = String::new(); |
| 1323 | for ch in input.chars() { |
| 1324 | if ch.is_ascii_alphanumeric() || ch == '_' || ch == '-' { |
| 1325 | out.push(ch); |
| 1326 | } else { |
| 1327 | out.push('_'); |
| 1328 | } |
| 1329 | } |
| 1330 | if out.is_empty() { |
| 1331 | "artifact".to_string() |
| 1332 | } else { |
| 1333 | out |
| 1334 | } |
| 1335 | } |
| 1336 | |
| 1337 | #[cfg(test)] |
| 1338 | mod tests { |
| 1339 | use super::*; |
| 1340 | use crate::tools::spec::ToolSpec; |
| 1341 | |
| 1342 | #[test] |
| 1343 | fn durable_task_schema_requires_prompt() { |
| 1344 | let schema = TasksTool::alias("task_create", "create").input_schema(); |
| 1345 | assert_eq!(schema["required"][0], "prompt"); |
| 1346 | assert!(schema["properties"]["prompt"].is_object()); |
| 1347 | } |
| 1348 | |
| 1349 | #[test] |
| 1350 | fn create_mode_enum_advertises_operate_not_yolo() { |
| 1351 | let create = TasksTool::alias("task_create", "create").input_schema(); |
| 1352 | assert_eq!( |
| 1353 | create["properties"]["mode"]["enum"], |
| 1354 | json!(["agent", "plan", "operate"]) |
| 1355 | ); |
| 1356 | let canonical = TasksTool::new("tasks").input_schema(); |
| 1357 | assert_eq!( |
| 1358 | canonical["properties"]["mode"]["enum"], |
| 1359 | json!(["agent", "plan", "operate"]) |
| 1360 | ); |
| 1361 | } |
| 1362 | |
| 1363 | #[test] |
| 1364 | fn gate_classifier_detects_timeout() { |
| 1365 | assert_eq!( |
| 1366 | classify_gate_failure("test", "timeout", true, "", ""), |
| 1367 | "timeout" |
| 1368 | ); |
| 1369 | } |
| 1370 | |
| 1371 | #[test] |
| 1372 | fn canonical_schema_lists_all_actions_and_union_fields() { |
| 1373 | let schema = TasksTool::new("tasks").input_schema(); |
| 1374 | let actions = schema["properties"]["action"]["enum"] |
| 1375 | .as_array() |
| 1376 | .expect("action enum"); |
| 1377 | for action in [ |
| 1378 | "create", |
| 1379 | "list", |
| 1380 | "read", |
| 1381 | "cancel", |
| 1382 | "gate_run", |
| 1383 | "pr_attempt_record", |
| 1384 | "pr_attempt_list", |
| 1385 | "pr_attempt_read", |
| 1386 | "pr_attempt_preflight", |
| 1387 | ] { |
| 1388 | assert!( |
| 1389 | actions.iter().any(|value| value.as_str() == Some(action)), |
| 1390 | "canonical schema must offer action {action}" |
| 1391 | ); |
| 1392 | } |
| 1393 | for field in [ |
| 1394 | "prompt", |
| 1395 | "task_id", |
| 1396 | "gate", |
| 1397 | "command", |
| 1398 | "attempt_id", |
| 1399 | "limit", |
| 1400 | ] { |
| 1401 | assert!( |
| 1402 | schema["properties"][field].is_object(), |
| 1403 | "canonical schema must carry union field {field}" |
| 1404 | ); |
| 1405 | } |
| 1406 | assert_eq!(schema["additionalProperties"], json!(false)); |
| 1407 | } |
| 1408 | |
| 1409 | #[test] |
| 1410 | fn read_only_variant_only_offers_read_actions() { |
| 1411 | let tool = TasksTool::read_only("tasks"); |
| 1412 | let schema = tool.input_schema(); |
| 1413 | assert_eq!( |
| 1414 | schema["properties"]["action"]["enum"], |
| 1415 | json!(["list", "read", "pr_attempt_list", "pr_attempt_read"]) |
| 1416 | ); |
| 1417 | assert!(!schema["properties"]["prompt"].is_object()); |
| 1418 | assert!(!schema["properties"]["gate"].is_object()); |
| 1419 | // pr_attempt_read is a read action: its id field must be advertised |
| 1420 | // on the read-only surface too. |
| 1421 | assert!(schema["properties"]["attempt_id"].is_object()); |
| 1422 | assert!(schema["properties"]["task_id"].is_object()); |
| 1423 | assert_eq!(tool.approval_requirement(), ApprovalRequirement::Auto); |
| 1424 | assert!(tool.is_read_only()); |
| 1425 | assert_eq!(tool.capabilities(), vec![ToolCapability::ReadOnly]); |
| 1426 | } |
| 1427 | |
| 1428 | #[test] |
| 1429 | fn aliases_hide_from_model_and_force_action() { |
| 1430 | let create = TasksTool::alias("task_create", "create"); |
| 1431 | assert!(!create.model_visible()); |
| 1432 | assert_eq!(create.name(), "task_create"); |
| 1433 | assert_eq!(create.approval_requirement(), ApprovalRequirement::Required); |
| 1434 | |
| 1435 | let gate = TasksTool::alias("task_gate_run", "gate_run"); |
| 1436 | assert_eq!(gate.approval_requirement(), ApprovalRequirement::Required); |
| 1437 | assert!(gate.capabilities().contains(&ToolCapability::ExecutesCode)); |
| 1438 | |
| 1439 | let list = TasksTool::alias("task_list", "list"); |
| 1440 | assert_eq!(list.approval_requirement(), ApprovalRequirement::Auto); |
| 1441 | assert!(list.is_read_only_for(&json!({}))); |
| 1442 | |
| 1443 | let canonical = TasksTool::new("tasks"); |
| 1444 | assert!(canonical.model_visible()); |
| 1445 | assert_eq!( |
| 1446 | canonical.approval_requirement_for(&json!({"action": "list"})), |
| 1447 | ApprovalRequirement::Auto |
| 1448 | ); |
| 1449 | assert_eq!( |
| 1450 | canonical.approval_requirement_for(&json!({"action": "cancel"})), |
| 1451 | ApprovalRequirement::Required |
| 1452 | ); |
| 1453 | assert_eq!( |
| 1454 | canonical.approval_requirement_for(&json!({"action": "gate_run"})), |
| 1455 | ApprovalRequirement::Required |
| 1456 | ); |
| 1457 | assert!(canonical.is_read_only_for(&json!({"action": "pr_attempt_read"}))); |
| 1458 | assert!(!canonical.is_read_only_for(&json!({"action": "create"}))); |
| 1459 | } |
| 1460 | |
| 1461 | #[test] |
| 1462 | fn canonical_rejects_unknown_or_missing_action() { |
| 1463 | let tool = TasksTool::new("tasks"); |
| 1464 | let err = tool |
| 1465 | .resolve_action(&json!({})) |
| 1466 | .expect_err("missing action must fail"); |
| 1467 | assert!(err.to_string().contains("missing `action`")); |
| 1468 | let err = tool |
| 1469 | .resolve_action(&json!({"action": "explode"})) |
| 1470 | .expect_err("unknown action must fail"); |
| 1471 | assert!(err.to_string().contains("invalid action")); |
| 1472 | |
| 1473 | let read_only = TasksTool::read_only("tasks"); |
| 1474 | let err = read_only |
| 1475 | .resolve_action(&json!({"action": "gate_run"})) |
| 1476 | .expect_err("read-only surface must reject exec actions"); |
| 1477 | assert!(err.to_string().contains("invalid action")); |
| 1478 | } |
| 1479 | |
| 1480 | #[test] |
| 1481 | fn background_shell_schema_is_explicit() { |
| 1482 | let schema = TaskShellStartTool.input_schema(); |
| 1483 | assert_eq!(schema["required"][0], "command"); |
| 1484 | assert_eq!(schema["properties"]["timeout_ms"]["maximum"], 600000); |
| 1485 | |
| 1486 | let wait_schema = TaskShellWaitTool.input_schema(); |
| 1487 | assert_eq!(wait_schema["required"][0], "task_id"); |
| 1488 | assert!(wait_schema["properties"]["gate"].is_object()); |
| 1489 | } |
| 1490 | |
| 1491 | #[test] |
| 1492 | fn task_shell_wait_keeps_its_documented_nonblocking_default() { |
| 1493 | assert_eq!( |
| 1494 | task_shell_wait_input(json!({"task_id": "shell_1"}))["wait"], |
| 1495 | false |
| 1496 | ); |
| 1497 | assert_eq!( |
| 1498 | task_shell_wait_input(json!({"task_id": "shell_1", "wait": true}))["wait"], |
| 1499 | true |
| 1500 | ); |
| 1501 | assert_eq!( |
| 1502 | task_shell_wait_input(json!({"task_id": "shell_1", "block": true}))["block"], |
| 1503 | true |
| 1504 | ); |
| 1505 | assert_eq!( |
| 1506 | task_shell_wait_input(json!({"task_id": "shell_1", "wait": null}))["wait"], |
| 1507 | false |
| 1508 | ); |
| 1509 | assert_eq!( |
| 1510 | task_shell_wait_input(json!({"task_id": "shell_1", "block": null}))["wait"], |
| 1511 | false |
| 1512 | ); |
| 1513 | } |
| 1514 | |
| 1515 | #[tokio::test] |
| 1516 | async fn task_shell_wait_null_is_a_nonblocking_snapshot() { |
| 1517 | let workspace = tempfile::tempdir().expect("workspace"); |
| 1518 | let context = ToolContext::new(workspace.path()); |
| 1519 | let started = TaskShellStartTool |
| 1520 | .execute(json!({"command": "sleep 2", "timeout_ms": 5_000}), &context) |
| 1521 | .await |
| 1522 | .expect("start background shell"); |
| 1523 | let task_id = started |
| 1524 | .metadata |
| 1525 | .as_ref() |
| 1526 | .and_then(|metadata| metadata.get("task_id")) |
| 1527 | .and_then(Value::as_str) |
| 1528 | .expect("task id") |
| 1529 | .to_string(); |
| 1530 | |
| 1531 | let before = std::time::Instant::now(); |
| 1532 | let snapshot = TaskShellWaitTool |
| 1533 | .execute( |
| 1534 | json!({"task_id": task_id, "wait": null, "timeout_ms": 5_000}), |
| 1535 | &context, |
| 1536 | ) |
| 1537 | .await |
| 1538 | .expect("poll background shell"); |
| 1539 | assert!( |
| 1540 | before.elapsed() < std::time::Duration::from_secs(1), |
| 1541 | "task_shell_wait wait:null must preserve the nonblocking default" |
| 1542 | ); |
| 1543 | assert_eq!( |
| 1544 | snapshot |
| 1545 | .metadata |
| 1546 | .as_ref() |
| 1547 | .and_then(|metadata| metadata.get("status")) |
| 1548 | .and_then(Value::as_str), |
| 1549 | Some("Running") |
| 1550 | ); |
| 1551 | |
| 1552 | BashTool::alias("exec_shell_cancel", "cancel") |
| 1553 | .execute(json!({"task_id": task_id}), &context) |
| 1554 | .await |
| 1555 | .expect("cancel background shell"); |
| 1556 | } |
| 1557 | |
| 1558 | #[test] |
| 1559 | fn gate_command_uses_login_shell_invocation() { |
| 1560 | let (program, args) = build_gate_command_parts("echo hello"); |
| 1561 | assert_eq!(program, "/bin/sh"); |
| 1562 | assert_eq!(args, vec!["-lc".to_string(), "echo hello".to_string()]); |
| 1563 | } |
| 1564 | } |
| 1565 |