| 1 | //! Elevated Workflow plan assessment for approval cards (#4126). |
| 2 | //! |
| 3 | //! Pure, UI-free analysis of a [`WorkflowSpec`] (and optional planner risk |
| 4 | //! string) so callers can decide whether an operator approval card is required |
| 5 | //! and what fields that card should show. |
| 6 | |
| 7 | use serde::{Deserialize, Serialize}; |
| 8 | |
| 9 | use crate::{ |
| 10 | IsolationMode, LeafSpec, PermissionSpec, TaskMode, WorkflowNode, WorkflowSpec, |
| 11 | leaf_is_write_capable, leaf_wants_worktree, |
| 12 | }; |
| 13 | |
| 14 | /// Default soft token budget from product config (`[workflow].default_token_budget`). |
| 15 | /// Plans requesting more than this are treated as high-budget. |
| 16 | pub const DEFAULT_HIGH_BUDGET_THRESHOLD: u64 = 120_000; |
| 17 | |
| 18 | /// Options that refine elevation assessment beyond the IR itself. |
| 19 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 20 | pub struct ElevationOptions { |
| 21 | /// Token budget declared on the tool call (may outrank `spec.budget`). |
| 22 | pub token_budget: Option<u64>, |
| 23 | /// Threshold above which a token budget is considered high. |
| 24 | pub high_budget_threshold: u64, |
| 25 | /// Whether the parent session currently allows writes. |
| 26 | pub parent_allows_write: bool, |
| 27 | /// Whether the parent session currently allows network. |
| 28 | pub parent_allows_network: bool, |
| 29 | } |
| 30 | |
| 31 | impl Default for ElevationOptions { |
| 32 | fn default() -> Self { |
| 33 | Self { |
| 34 | token_budget: None, |
| 35 | high_budget_threshold: DEFAULT_HIGH_BUDGET_THRESHOLD, |
| 36 | // Assume Act/read-write parent unless callers narrow posture. |
| 37 | parent_allows_write: true, |
| 38 | parent_allows_network: true, |
| 39 | } |
| 40 | } |
| 41 | } |
| 42 | |
| 43 | /// Summary of why a Workflow plan needs (or does not need) elevated approval. |
| 44 | #[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)] |
| 45 | pub struct WorkflowPlanElevation { |
| 46 | pub elevated: bool, |
| 47 | pub goal: String, |
| 48 | pub child_count: usize, |
| 49 | pub child_summary: String, |
| 50 | pub writes: bool, |
| 51 | pub shell: bool, |
| 52 | pub network: bool, |
| 53 | pub secrets: bool, |
| 54 | pub worktree: bool, |
| 55 | pub high_budget: bool, |
| 56 | pub broader_authority: bool, |
| 57 | /// Human-readable budget line for the approval card. |
| 58 | pub budget_label: String, |
| 59 | /// Distinct elevation reasons (for audit / impact lines). |
| 60 | pub reasons: Vec<String>, |
| 61 | } |
| 62 | |
| 63 | impl WorkflowPlanElevation { |
| 64 | /// Card field labels/values used by the TUI approval modal (#4126). |
| 65 | #[must_use] |
| 66 | pub fn card_fields(&self) -> Vec<(&'static str, String)> { |
| 67 | vec![ |
| 68 | ("Goal", self.goal.clone()), |
| 69 | ("Children", self.child_summary.clone()), |
| 70 | ("Writes", yes_no(self.writes)), |
| 71 | ("Shell", yes_no(self.shell)), |
| 72 | ("Network", yes_no(self.network)), |
| 73 | ("Budget", self.budget_label.clone()), |
| 74 | ] |
| 75 | } |
| 76 | |
| 77 | /// True when the plan is fully inside the read-only envelope. |
| 78 | #[must_use] |
| 79 | pub fn is_read_only_envelope(&self) -> bool { |
| 80 | !self.elevated |
| 81 | && !self.writes |
| 82 | && !self.shell |
| 83 | && !self.network |
| 84 | && !self.secrets |
| 85 | && !self.worktree |
| 86 | && !self.high_budget |
| 87 | && !self.broader_authority |
| 88 | } |
| 89 | } |
| 90 | |
| 91 | fn yes_no(flag: bool) -> String { |
| 92 | if flag { |
| 93 | "yes".to_string() |
| 94 | } else { |
| 95 | "no".to_string() |
| 96 | } |
| 97 | } |
| 98 | |
| 99 | /// Assess elevation for a compiled [`WorkflowSpec`]. |
| 100 | #[must_use] |
| 101 | pub fn assess_workflow_elevation( |
| 102 | spec: &WorkflowSpec, |
| 103 | options: ElevationOptions, |
| 104 | ) -> WorkflowPlanElevation { |
| 105 | let mut child_ids = Vec::new(); |
| 106 | let mut writes = false; |
| 107 | let mut shell = false; |
| 108 | let mut network = false; |
| 109 | let mut secrets = false; |
| 110 | let mut worktree = false; |
| 111 | |
| 112 | walk_nodes( |
| 113 | &spec.nodes, |
| 114 | /* parallel */ false, |
| 115 | &mut child_ids, |
| 116 | &mut writes, |
| 117 | &mut shell, |
| 118 | &mut network, |
| 119 | &mut secrets, |
| 120 | &mut worktree, |
| 121 | ); |
| 122 | |
| 123 | // Spec-level permissions also elevate. |
| 124 | merge_permissions( |
| 125 | &spec.permissions, |
| 126 | &mut writes, |
| 127 | &mut shell, |
| 128 | &mut network, |
| 129 | &mut secrets, |
| 130 | ); |
| 131 | |
| 132 | // The structured-plan lowerer stores its validated risk enum on |
| 133 | // `description`, while authored Workflow specs use that field for ordinary |
| 134 | // prose. Only consume recognized enum values here: treating free-form |
| 135 | // descriptions as unknown risk would falsely report writes, shell, and |
| 136 | // network in the approval receipt. Unknown planner risk remains fail-closed |
| 137 | // in `assess_plan_risk_string` and is rejected before structured lowering. |
| 138 | if let Some(risk) = embedded_plan_risk_hint(spec.description.as_deref()) { |
| 139 | apply_plan_risk_hint(Some(risk), &mut writes, &mut shell, &mut network); |
| 140 | } |
| 141 | |
| 142 | let effective_tokens = options |
| 143 | .token_budget |
| 144 | .or(spec.budget.max_tokens) |
| 145 | .filter(|n| *n > 0); |
| 146 | let high_budget = effective_tokens.is_some_and(|n| n > options.high_budget_threshold); |
| 147 | |
| 148 | let broader_authority = |
| 149 | (!options.parent_allows_write && writes) || (!options.parent_allows_network && network); |
| 150 | |
| 151 | let mut reasons = Vec::new(); |
| 152 | if writes { |
| 153 | reasons.push("writes".to_string()); |
| 154 | } |
| 155 | if shell { |
| 156 | reasons.push("shell".to_string()); |
| 157 | } |
| 158 | if network { |
| 159 | reasons.push("network".to_string()); |
| 160 | } |
| 161 | if secrets { |
| 162 | reasons.push("secrets".to_string()); |
| 163 | } |
| 164 | if worktree { |
| 165 | reasons.push("worktree".to_string()); |
| 166 | } |
| 167 | if high_budget { |
| 168 | reasons.push("high_budget".to_string()); |
| 169 | } |
| 170 | if broader_authority { |
| 171 | reasons.push("broader_authority".to_string()); |
| 172 | } |
| 173 | |
| 174 | let elevated = !reasons.is_empty(); |
| 175 | let child_count = child_ids.len(); |
| 176 | let child_summary = if child_ids.is_empty() { |
| 177 | "0 children".to_string() |
| 178 | } else if child_ids.len() <= 4 { |
| 179 | format!( |
| 180 | "{} child{}: {}", |
| 181 | child_ids.len(), |
| 182 | if child_ids.len() == 1 { "" } else { "ren" }, |
| 183 | child_ids.join(", ") |
| 184 | ) |
| 185 | } else { |
| 186 | format!( |
| 187 | "{} children: {}, {}… (+{})", |
| 188 | child_ids.len(), |
| 189 | child_ids[0], |
| 190 | child_ids[1], |
| 191 | child_ids.len() - 2 |
| 192 | ) |
| 193 | }; |
| 194 | |
| 195 | let budget_label = format_budget_label(effective_tokens, &spec.budget, high_budget); |
| 196 | |
| 197 | WorkflowPlanElevation { |
| 198 | elevated, |
| 199 | goal: spec.goal.clone(), |
| 200 | child_count, |
| 201 | child_summary, |
| 202 | writes, |
| 203 | shell, |
| 204 | network, |
| 205 | secrets, |
| 206 | worktree, |
| 207 | high_budget, |
| 208 | broader_authority, |
| 209 | budget_label, |
| 210 | reasons, |
| 211 | } |
| 212 | } |
| 213 | |
| 214 | /// Lightweight assessment from a planner `risk` string alone (before IR lower). |
| 215 | #[must_use] |
| 216 | pub fn assess_plan_risk_string(risk: Option<&str>) -> PlanRiskHint { |
| 217 | match risk.map(str::trim).filter(|s| !s.is_empty()) { |
| 218 | None | Some("read_only") | Some("readonly") | Some("low") | Some("safe") => { |
| 219 | PlanRiskHint::ReadOnly |
| 220 | } |
| 221 | Some("writes") | Some("write") | Some("read_write") | Some("readwrite") |
| 222 | | Some("medium") => PlanRiskHint::Writes, |
| 223 | Some("shell") => PlanRiskHint::Shell, |
| 224 | Some("network") => PlanRiskHint::Network, |
| 225 | Some("elevated") | Some("high") => PlanRiskHint::Elevated, |
| 226 | Some(_) => PlanRiskHint::Elevated, |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | /// Coarse risk classification from the structured plan `risk` field. |
| 231 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 232 | pub enum PlanRiskHint { |
| 233 | ReadOnly, |
| 234 | Writes, |
| 235 | Shell, |
| 236 | Network, |
| 237 | Elevated, |
| 238 | } |
| 239 | |
| 240 | impl PlanRiskHint { |
| 241 | #[must_use] |
| 242 | pub fn elevates(self) -> bool { |
| 243 | !matches!(self, Self::ReadOnly) |
| 244 | } |
| 245 | } |
| 246 | |
| 247 | fn apply_plan_risk_hint( |
| 248 | risk: Option<&str>, |
| 249 | writes: &mut bool, |
| 250 | shell: &mut bool, |
| 251 | network: &mut bool, |
| 252 | ) { |
| 253 | match assess_plan_risk_string(risk) { |
| 254 | PlanRiskHint::ReadOnly => {} |
| 255 | PlanRiskHint::Writes => *writes = true, |
| 256 | PlanRiskHint::Shell => { |
| 257 | *shell = true; |
| 258 | *writes = true; |
| 259 | } |
| 260 | PlanRiskHint::Network => { |
| 261 | *network = true; |
| 262 | } |
| 263 | PlanRiskHint::Elevated => { |
| 264 | *writes = true; |
| 265 | *shell = true; |
| 266 | *network = true; |
| 267 | } |
| 268 | } |
| 269 | } |
| 270 | |
| 271 | fn embedded_plan_risk_hint(description: Option<&str>) -> Option<&str> { |
| 272 | let value = description |
| 273 | .map(str::trim) |
| 274 | .filter(|value| !value.is_empty())?; |
| 275 | matches!( |
| 276 | value, |
| 277 | "read_only" |
| 278 | | "readonly" |
| 279 | | "low" |
| 280 | | "safe" |
| 281 | | "writes" |
| 282 | | "write" |
| 283 | | "read_write" |
| 284 | | "readwrite" |
| 285 | | "medium" |
| 286 | | "shell" |
| 287 | | "network" |
| 288 | | "elevated" |
| 289 | | "high" |
| 290 | ) |
| 291 | .then_some(value) |
| 292 | } |
| 293 | |
| 294 | fn format_budget_label( |
| 295 | effective_tokens: Option<u64>, |
| 296 | budget: &crate::BudgetSpec, |
| 297 | high_budget: bool, |
| 298 | ) -> String { |
| 299 | let mut parts = Vec::new(); |
| 300 | if let Some(tokens) = effective_tokens { |
| 301 | parts.push(format!("{tokens} tokens")); |
| 302 | } |
| 303 | if let Some(steps) = budget.max_steps { |
| 304 | parts.push(format!("max_steps={steps}")); |
| 305 | } |
| 306 | if let Some(timeout) = budget.timeout_secs { |
| 307 | parts.push(format!("timeout={timeout}s")); |
| 308 | } |
| 309 | if let Some(parallel) = budget.max_parallel { |
| 310 | parts.push(format!("max_parallel={parallel}")); |
| 311 | } |
| 312 | if parts.is_empty() { |
| 313 | "default".to_string() |
| 314 | } else if high_budget { |
| 315 | format!("{} (high)", parts.join(", ")) |
| 316 | } else { |
| 317 | parts.join(", ") |
| 318 | } |
| 319 | } |
| 320 | |
| 321 | #[allow(clippy::too_many_arguments)] |
| 322 | fn walk_nodes( |
| 323 | nodes: &[WorkflowNode], |
| 324 | parallel: bool, |
| 325 | child_ids: &mut Vec<String>, |
| 326 | writes: &mut bool, |
| 327 | shell: &mut bool, |
| 328 | network: &mut bool, |
| 329 | secrets: &mut bool, |
| 330 | worktree: &mut bool, |
| 331 | ) { |
| 332 | for node in nodes { |
| 333 | match node { |
| 334 | WorkflowNode::Leaf(leaf) => { |
| 335 | inspect_leaf( |
| 336 | leaf, parallel, child_ids, writes, shell, network, secrets, worktree, |
| 337 | ); |
| 338 | } |
| 339 | WorkflowNode::BranchSet(branch) => { |
| 340 | merge_permissions(&branch.permissions, writes, shell, network, secrets); |
| 341 | walk_nodes( |
| 342 | &branch.children, |
| 343 | branch.parallel || parallel, |
| 344 | child_ids, |
| 345 | writes, |
| 346 | shell, |
| 347 | network, |
| 348 | secrets, |
| 349 | worktree, |
| 350 | ); |
| 351 | } |
| 352 | WorkflowNode::Sequence(seq) => { |
| 353 | walk_nodes( |
| 354 | &seq.children, |
| 355 | parallel, |
| 356 | child_ids, |
| 357 | writes, |
| 358 | shell, |
| 359 | network, |
| 360 | secrets, |
| 361 | worktree, |
| 362 | ); |
| 363 | } |
| 364 | WorkflowNode::LoopUntil(loop_spec) => { |
| 365 | walk_nodes( |
| 366 | &loop_spec.children, |
| 367 | parallel, |
| 368 | child_ids, |
| 369 | writes, |
| 370 | shell, |
| 371 | network, |
| 372 | secrets, |
| 373 | worktree, |
| 374 | ); |
| 375 | } |
| 376 | WorkflowNode::Cond(cond) => { |
| 377 | walk_nodes( |
| 378 | &cond.then_nodes, |
| 379 | parallel, |
| 380 | child_ids, |
| 381 | writes, |
| 382 | shell, |
| 383 | network, |
| 384 | secrets, |
| 385 | worktree, |
| 386 | ); |
| 387 | walk_nodes( |
| 388 | &cond.else_nodes, |
| 389 | parallel, |
| 390 | child_ids, |
| 391 | writes, |
| 392 | shell, |
| 393 | network, |
| 394 | secrets, |
| 395 | worktree, |
| 396 | ); |
| 397 | } |
| 398 | WorkflowNode::Expand(expand) => { |
| 399 | if let Some(template) = expand.template.as_deref() { |
| 400 | walk_nodes( |
| 401 | std::slice::from_ref(template), |
| 402 | parallel, |
| 403 | child_ids, |
| 404 | writes, |
| 405 | shell, |
| 406 | network, |
| 407 | secrets, |
| 408 | worktree, |
| 409 | ); |
| 410 | } |
| 411 | } |
| 412 | WorkflowNode::Reduce(_) | WorkflowNode::TeacherReview(_) => { |
| 413 | // Control/reduce nodes do not spawn write-capable leaves themselves. |
| 414 | } |
| 415 | } |
| 416 | } |
| 417 | } |
| 418 | |
| 419 | #[allow(clippy::too_many_arguments)] |
| 420 | fn inspect_leaf( |
| 421 | leaf: &LeafSpec, |
| 422 | parallel: bool, |
| 423 | child_ids: &mut Vec<String>, |
| 424 | writes: &mut bool, |
| 425 | shell: &mut bool, |
| 426 | network: &mut bool, |
| 427 | secrets: &mut bool, |
| 428 | worktree: &mut bool, |
| 429 | ) { |
| 430 | child_ids.push(leaf.id.clone()); |
| 431 | if leaf_is_write_capable(leaf) { |
| 432 | *writes = true; |
| 433 | } |
| 434 | merge_permissions(&leaf.permissions, writes, shell, network, secrets); |
| 435 | if leaf_wants_worktree(leaf, parallel) || matches!(leaf.isolation, IsolationMode::Worktree) { |
| 436 | *worktree = true; |
| 437 | } |
| 438 | // Explicit read_write mode with shell tools already handled; implementer |
| 439 | // without a tool denylist can run shell. |
| 440 | if leaf.mode == TaskMode::ReadWrite |
| 441 | && leaf.permissions.allowed_tools.is_empty() |
| 442 | && matches!( |
| 443 | leaf.agent_type, |
| 444 | crate::AgentType::Implementer | crate::AgentType::General |
| 445 | ) |
| 446 | { |
| 447 | // Write-capable implementers/general agents may run shell beyond |
| 448 | // read-only — flag shell as elevated for the approval card. |
| 449 | *shell = true; |
| 450 | } |
| 451 | } |
| 452 | |
| 453 | fn merge_permissions( |
| 454 | permissions: &PermissionSpec, |
| 455 | writes: &mut bool, |
| 456 | shell: &mut bool, |
| 457 | network: &mut bool, |
| 458 | secrets: &mut bool, |
| 459 | ) { |
| 460 | if permissions.allow_write { |
| 461 | *writes = true; |
| 462 | } |
| 463 | if permissions.allow_network { |
| 464 | *network = true; |
| 465 | } |
| 466 | for tool in &permissions.allowed_tools { |
| 467 | let name = tool.trim(); |
| 468 | if is_write_tool(name) { |
| 469 | *writes = true; |
| 470 | } |
| 471 | if is_shell_tool(name) { |
| 472 | *shell = true; |
| 473 | } |
| 474 | if is_network_tool(name) { |
| 475 | *network = true; |
| 476 | } |
| 477 | if is_secret_tool(name) { |
| 478 | *secrets = true; |
| 479 | } |
| 480 | } |
| 481 | } |
| 482 | |
| 483 | /// True for a tool that can modify files. |
| 484 | /// |
| 485 | /// This is the one list. It previously existed twice — here and as half of |
| 486 | /// the TUI's `is_write_or_shell_tool` — and the two drifted: `Edit`, the |
| 487 | /// model-visible canonical name for the write tool, was in the TUI copy and |
| 488 | /// missing here, so a branch or sequence whose `allowed_tools` was `["Edit"]` |
| 489 | /// produced an approval card reporting `writes: false` for a spec that could |
| 490 | /// in fact write. |
| 491 | pub fn is_write_tool(tool: &str) -> bool { |
| 492 | matches!( |
| 493 | tool.trim(), |
| 494 | "Edit" | "write_file" | "edit_file" | "apply_patch" | "checklist_write" | "todo_write" |
| 495 | ) |
| 496 | } |
| 497 | |
| 498 | /// True for a tool that can run a shell command. |
| 499 | pub fn is_shell_tool(tool: &str) -> bool { |
| 500 | matches!( |
| 501 | tool.trim(), |
| 502 | "exec_shell" |
| 503 | | "exec_shell_wait" |
| 504 | | "exec_shell_interact" |
| 505 | | "exec_wait" |
| 506 | | "exec_interact" |
| 507 | | "task_shell_start" |
| 508 | | "task_shell_wait" |
| 509 | ) |
| 510 | } |
| 511 | |
| 512 | fn is_network_tool(tool: &str) -> bool { |
| 513 | matches!( |
| 514 | tool, |
| 515 | "web_search" | "web_run" | "fetch_url" | "wait_for_dev_server" |
| 516 | ) || tool.starts_with("mcp_") |
| 517 | } |
| 518 | |
| 519 | fn is_secret_tool(tool: &str) -> bool { |
| 520 | let lower = tool.to_ascii_lowercase(); |
| 521 | lower.contains("secret") |
| 522 | || lower.contains("credential") |
| 523 | || lower.contains("password") |
| 524 | || lower == "read_env" |
| 525 | || lower == "env" |
| 526 | } |
| 527 | |
| 528 | #[cfg(test)] |
| 529 | mod tests { |
| 530 | use super::*; |
| 531 | use crate::{ |
| 532 | AgentType, BranchSpec, BudgetSpec, LeafSpec, ModelPolicy, PermissionSpec, PromotionPolicy, |
| 533 | SequenceSpec, TaskMode, |
| 534 | }; |
| 535 | |
| 536 | fn leaf(id: &str, mode: TaskMode) -> LeafSpec { |
| 537 | LeafSpec { |
| 538 | id: id.to_string(), |
| 539 | prompt: format!("do {id}"), |
| 540 | agent_type: if mode == TaskMode::ReadWrite { |
| 541 | AgentType::Implementer |
| 542 | } else { |
| 543 | AgentType::Explore |
| 544 | }, |
| 545 | profile: None, |
| 546 | role: None, |
| 547 | mode, |
| 548 | isolation: IsolationMode::Auto, |
| 549 | file_scope: Vec::new(), |
| 550 | depends_on_results: Vec::new(), |
| 551 | budget: BudgetSpec::default(), |
| 552 | permissions: PermissionSpec::default(), |
| 553 | model_policy: ModelPolicy::default(), |
| 554 | } |
| 555 | } |
| 556 | |
| 557 | #[test] |
| 558 | fn edit_is_recognized_as_a_write_tool() { |
| 559 | // #4730: `Edit` is the model-visible canonical write-tool name. It |
| 560 | // lived only in the TUI's copy of this list, so the risk assessor |
| 561 | // didn't know it was a write. |
| 562 | assert!(is_write_tool("Edit")); |
| 563 | assert!(is_write_tool(" Edit ")); |
| 564 | for tool in [ |
| 565 | "write_file", |
| 566 | "edit_file", |
| 567 | "apply_patch", |
| 568 | "checklist_write", |
| 569 | "todo_write", |
| 570 | ] { |
| 571 | assert!(is_write_tool(tool), "{tool} must count as a write"); |
| 572 | } |
| 573 | assert!(!is_write_tool("read_file")); |
| 574 | assert!(!is_write_tool("Editor")); |
| 575 | } |
| 576 | |
| 577 | #[test] |
| 578 | fn branch_allowing_edit_reports_writes_in_its_risk_summary() { |
| 579 | // The tool-allowlist path is what produces branch/sequence-level |
| 580 | // permission summaries; a spec that can write must not present an |
| 581 | // approval card saying it cannot. |
| 582 | let spec = spec_with( |
| 583 | vec![WorkflowNode::BranchSet(BranchSpec { |
| 584 | id: "edits".to_string(), |
| 585 | description: None, |
| 586 | parallel: false, |
| 587 | budget: BudgetSpec::default(), |
| 588 | permissions: PermissionSpec { |
| 589 | allowed_tools: vec!["Edit".to_string()], |
| 590 | ..PermissionSpec::default() |
| 591 | }, |
| 592 | model_policy: ModelPolicy::default(), |
| 593 | children: vec![WorkflowNode::Leaf(leaf("child", TaskMode::ReadOnly))], |
| 594 | })], |
| 595 | None, |
| 596 | ); |
| 597 | |
| 598 | let elevation = assess_workflow_elevation(&spec, ElevationOptions::default()); |
| 599 | assert!( |
| 600 | elevation.writes, |
| 601 | "branch allowing Edit must report writes: {elevation:?}" |
| 602 | ); |
| 603 | } |
| 604 | |
| 605 | fn spec_with(nodes: Vec<WorkflowNode>, risk: Option<&str>) -> WorkflowSpec { |
| 606 | WorkflowSpec { |
| 607 | id: Some("test".to_string()), |
| 608 | goal: "ship feature".to_string(), |
| 609 | description: risk.map(str::to_string), |
| 610 | budget: BudgetSpec::default(), |
| 611 | permissions: PermissionSpec::default(), |
| 612 | model_policy: ModelPolicy::default(), |
| 613 | promotion_policy: PromotionPolicy::default(), |
| 614 | gates: Vec::new(), |
| 615 | nodes, |
| 616 | } |
| 617 | } |
| 618 | |
| 619 | #[test] |
| 620 | fn read_only_plan_is_not_elevated() { |
| 621 | let spec = spec_with( |
| 622 | vec![WorkflowNode::Leaf(leaf("scan", TaskMode::ReadOnly))], |
| 623 | Some("read_only"), |
| 624 | ); |
| 625 | let elevation = assess_workflow_elevation(&spec, ElevationOptions::default()); |
| 626 | assert!(!elevation.elevated, "{elevation:?}"); |
| 627 | assert!(elevation.is_read_only_envelope()); |
| 628 | assert_eq!(elevation.goal, "ship feature"); |
| 629 | assert!(elevation.child_summary.contains("scan")); |
| 630 | assert!(!elevation.writes); |
| 631 | assert!(!elevation.shell); |
| 632 | assert!(!elevation.network); |
| 633 | let fields = elevation.card_fields(); |
| 634 | assert_eq!(fields.len(), 6); |
| 635 | assert!( |
| 636 | fields |
| 637 | .iter() |
| 638 | .any(|(k, v)| *k == "Goal" && v == "ship feature") |
| 639 | ); |
| 640 | assert!(fields.iter().any(|(k, v)| *k == "Writes" && v == "no")); |
| 641 | } |
| 642 | |
| 643 | #[test] |
| 644 | fn free_form_description_is_not_treated_as_plan_risk() { |
| 645 | let spec = spec_with( |
| 646 | vec![WorkflowNode::Leaf(leaf("scan", TaskMode::ReadOnly))], |
| 647 | Some( |
| 648 | "Read-only release acceptance fixture; no step edits files or accesses the network.", |
| 649 | ), |
| 650 | ); |
| 651 | |
| 652 | let elevation = assess_workflow_elevation(&spec, ElevationOptions::default()); |
| 653 | assert!(elevation.is_read_only_envelope(), "{elevation:?}"); |
| 654 | assert!(!elevation.writes, "{elevation:?}"); |
| 655 | assert!(!elevation.shell, "{elevation:?}"); |
| 656 | assert!(!elevation.network, "{elevation:?}"); |
| 657 | assert!(elevation.reasons.is_empty(), "{elevation:?}"); |
| 658 | } |
| 659 | |
| 660 | #[test] |
| 661 | fn read_only_implementer_role_is_not_write_capable_or_elevated() { |
| 662 | let mut implementer = leaf("verify-only", TaskMode::ReadOnly); |
| 663 | implementer.agent_type = AgentType::Implementer; |
| 664 | implementer.role = Some("implementer".to_string()); |
| 665 | let spec = spec_with( |
| 666 | vec![WorkflowNode::BranchSet(BranchSpec { |
| 667 | id: "parallel-read-only".to_string(), |
| 668 | description: None, |
| 669 | parallel: true, |
| 670 | budget: BudgetSpec::default(), |
| 671 | permissions: PermissionSpec::default(), |
| 672 | model_policy: ModelPolicy::default(), |
| 673 | children: vec![WorkflowNode::Leaf(implementer)], |
| 674 | })], |
| 675 | Some("read_only"), |
| 676 | ); |
| 677 | |
| 678 | let elevation = assess_workflow_elevation(&spec, ElevationOptions::default()); |
| 679 | assert!(elevation.is_read_only_envelope(), "{elevation:?}"); |
| 680 | assert!(!elevation.elevated, "{elevation:?}"); |
| 681 | assert!(!elevation.writes, "{elevation:?}"); |
| 682 | assert!(!elevation.shell, "{elevation:?}"); |
| 683 | assert!(!elevation.worktree, "{elevation:?}"); |
| 684 | } |
| 685 | |
| 686 | #[test] |
| 687 | fn write_plan_elevates_and_flags_shell_for_implementer() { |
| 688 | let spec = spec_with( |
| 689 | vec![WorkflowNode::Leaf(leaf("impl", TaskMode::ReadWrite))], |
| 690 | Some("writes"), |
| 691 | ); |
| 692 | let elevation = assess_workflow_elevation(&spec, ElevationOptions::default()); |
| 693 | assert!(elevation.elevated); |
| 694 | assert!(elevation.writes); |
| 695 | assert!(elevation.shell); |
| 696 | assert!(elevation.reasons.iter().any(|r| r == "writes")); |
| 697 | } |
| 698 | |
| 699 | #[test] |
| 700 | fn network_and_secrets_tools_elevate() { |
| 701 | let mut network_leaf = leaf("fetch", TaskMode::ReadOnly); |
| 702 | network_leaf.permissions.allow_network = true; |
| 703 | network_leaf.permissions.allowed_tools = vec!["fetch_url".to_string()]; |
| 704 | |
| 705 | let mut secret_leaf = leaf("creds", TaskMode::ReadOnly); |
| 706 | secret_leaf.permissions.allowed_tools = vec!["read_secret".to_string()]; |
| 707 | |
| 708 | let spec = spec_with( |
| 709 | vec![WorkflowNode::Sequence(SequenceSpec { |
| 710 | id: "seq".to_string(), |
| 711 | children: vec![ |
| 712 | WorkflowNode::Leaf(network_leaf), |
| 713 | WorkflowNode::Leaf(secret_leaf), |
| 714 | ], |
| 715 | })], |
| 716 | None, |
| 717 | ); |
| 718 | let elevation = assess_workflow_elevation(&spec, ElevationOptions::default()); |
| 719 | assert!(elevation.elevated); |
| 720 | assert!(elevation.network); |
| 721 | assert!(elevation.secrets); |
| 722 | assert!(elevation.reasons.iter().any(|r| r == "network")); |
| 723 | assert!(elevation.reasons.iter().any(|r| r == "secrets")); |
| 724 | } |
| 725 | |
| 726 | #[test] |
| 727 | fn parallel_write_children_flag_worktree() { |
| 728 | let left = leaf("left", TaskMode::ReadWrite); |
| 729 | let right = leaf("right", TaskMode::ReadWrite); |
| 730 | let spec = spec_with( |
| 731 | vec![WorkflowNode::BranchSet(BranchSpec { |
| 732 | id: "parallel".to_string(), |
| 733 | description: None, |
| 734 | parallel: true, |
| 735 | budget: BudgetSpec::default(), |
| 736 | permissions: PermissionSpec::default(), |
| 737 | model_policy: ModelPolicy::default(), |
| 738 | children: vec![WorkflowNode::Leaf(left), WorkflowNode::Leaf(right)], |
| 739 | })], |
| 740 | Some("writes"), |
| 741 | ); |
| 742 | let elevation = assess_workflow_elevation(&spec, ElevationOptions::default()); |
| 743 | assert!(elevation.worktree, "{elevation:?}"); |
| 744 | assert!(elevation.writes); |
| 745 | } |
| 746 | |
| 747 | #[test] |
| 748 | fn high_budget_elevates() { |
| 749 | let mut spec = spec_with( |
| 750 | vec![WorkflowNode::Leaf(leaf("scan", TaskMode::ReadOnly))], |
| 751 | Some("read_only"), |
| 752 | ); |
| 753 | spec.budget.max_tokens = Some(250_000); |
| 754 | let elevation = assess_workflow_elevation(&spec, ElevationOptions::default()); |
| 755 | assert!(elevation.high_budget); |
| 756 | assert!(elevation.elevated); |
| 757 | assert!(elevation.budget_label.contains("high")); |
| 758 | } |
| 759 | |
| 760 | #[test] |
| 761 | fn broader_authority_when_parent_is_read_only() { |
| 762 | let spec = spec_with( |
| 763 | vec![WorkflowNode::Leaf(leaf("impl", TaskMode::ReadWrite))], |
| 764 | Some("writes"), |
| 765 | ); |
| 766 | let elevation = assess_workflow_elevation( |
| 767 | &spec, |
| 768 | ElevationOptions { |
| 769 | parent_allows_write: false, |
| 770 | parent_allows_network: false, |
| 771 | ..ElevationOptions::default() |
| 772 | }, |
| 773 | ); |
| 774 | assert!(elevation.broader_authority); |
| 775 | assert!(elevation.reasons.iter().any(|r| r == "broader_authority")); |
| 776 | } |
| 777 | |
| 778 | #[test] |
| 779 | fn plan_risk_string_classifies_elevated_variants() { |
| 780 | assert_eq!( |
| 781 | assess_plan_risk_string(Some("read_only")), |
| 782 | PlanRiskHint::ReadOnly |
| 783 | ); |
| 784 | assert_eq!( |
| 785 | assess_plan_risk_string(Some("writes")), |
| 786 | PlanRiskHint::Writes |
| 787 | ); |
| 788 | assert_eq!(assess_plan_risk_string(Some("shell")), PlanRiskHint::Shell); |
| 789 | assert_eq!( |
| 790 | assess_plan_risk_string(Some("network")), |
| 791 | PlanRiskHint::Network |
| 792 | ); |
| 793 | assert_eq!( |
| 794 | assess_plan_risk_string(Some("elevated")), |
| 795 | PlanRiskHint::Elevated |
| 796 | ); |
| 797 | assert_eq!( |
| 798 | assess_plan_risk_string(Some("unknown-risk")), |
| 799 | PlanRiskHint::Elevated, |
| 800 | "unknown planner risk must remain fail-closed" |
| 801 | ); |
| 802 | assert!(assess_plan_risk_string(Some("elevated")).elevates()); |
| 803 | assert!(!assess_plan_risk_string(Some("read_only")).elevates()); |
| 804 | } |
| 805 | |
| 806 | #[test] |
| 807 | fn card_fields_always_include_required_labels() { |
| 808 | let spec = spec_with( |
| 809 | vec![WorkflowNode::Leaf(leaf("a", TaskMode::ReadOnly))], |
| 810 | None, |
| 811 | ); |
| 812 | let fields = assess_workflow_elevation(&spec, ElevationOptions::default()).card_fields(); |
| 813 | let labels: Vec<_> = fields.iter().map(|(k, _)| *k).collect(); |
| 814 | assert_eq!( |
| 815 | labels, |
| 816 | vec!["Goal", "Children", "Writes", "Shell", "Network", "Budget"] |
| 817 | ); |
| 818 | } |
| 819 | } |
| 820 |