| 1 | //! Fleet roster — the persistent, inspectable party of named agent roles. |
| 2 | //! |
| 3 | //! The roster merges four layers into one config-backed lineup shared by |
| 4 | //! model-spawned sub-agents and fleet dispatch (#fleet-roster cutover |
| 5 | //! (v0.8.67)): |
| 6 | //! |
| 7 | //! - built-in members (the default party, always available), |
| 8 | //! - `[fleet.profiles]` entries from config.toml, |
| 9 | //! - personal `$CODEWHALE_HOME/agents/*.toml` profile files, |
| 10 | //! - workspace `.codewhale/agents/*.toml` profile files. |
| 11 | //! |
| 12 | //! Precedence is Workspace > Personal > Config > BuiltIn, merged by id. Loading never |
| 13 | //! fails the session: an unreadable workspace profile dir degrades to the |
| 14 | //! built-in + config layers with a log line. |
| 15 | //! |
| 16 | //! Two guardrails (#5098): |
| 17 | //! |
| 18 | //! - Shadowing is recorded, not silent: when a higher layer displaces a |
| 19 | //! lower-precedence file for the same id, the roster keeps a |
| 20 | //! [`ShadowedProfile`] receipt (logged at load, badged in the roster view) |
| 21 | //! so an edit in the losing layer is visibly ignored rather than dropped. |
| 22 | //! - Project-scope profiles (`.codewhale/agents/*.toml`) join the roster only |
| 23 | //! when project-level config is trusted for the launch; `--no-project-config` |
| 24 | //! opts the whole layer out, same as `.codewhale/config.toml` (#485). |
| 25 | |
| 26 | #![allow(dead_code)] |
| 27 | |
| 28 | use std::collections::HashMap; |
| 29 | use std::path::{Path, PathBuf}; |
| 30 | |
| 31 | use serde::{Deserialize, Serialize}; |
| 32 | |
| 33 | use codewhale_config::{ |
| 34 | FleetConfigToml, FleetDelegationHints, FleetLoadout, FleetProfile, FleetProfilePermissions, |
| 35 | FleetRole, FleetSlot, |
| 36 | }; |
| 37 | |
| 38 | use super::profile::{ |
| 39 | AgentProfile, load_agent_profiles_from_dir_tolerant, load_workspace_agent_profiles_tolerant, |
| 40 | personal_agent_profile_dir, |
| 41 | }; |
| 42 | |
| 43 | /// Which layer a roster member came from. Higher layers override lower ones |
| 44 | /// by id (Workspace > Personal > Config > BuiltIn). |
| 45 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)] |
| 46 | #[serde(rename_all = "snake_case")] |
| 47 | pub enum ProfileOrigin { |
| 48 | BuiltIn, |
| 49 | Config, |
| 50 | Personal, |
| 51 | Workspace, |
| 52 | } |
| 53 | |
| 54 | impl std::fmt::Display for ProfileOrigin { |
| 55 | fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 56 | f.write_str(match self { |
| 57 | Self::BuiltIn => "built-in", |
| 58 | Self::Config => "config", |
| 59 | Self::Personal => "personal", |
| 60 | Self::Workspace => "project", |
| 61 | }) |
| 62 | } |
| 63 | } |
| 64 | |
| 65 | /// The merged fleet roster. Think RPG saved party / K8s runconfig: a stable, |
| 66 | /// named lineup of agent roles the session can inspect and dispatch against. |
| 67 | #[derive(Debug, Clone)] |
| 68 | pub struct FleetRoster { |
| 69 | members: Vec<AgentProfile>, |
| 70 | /// Lower-precedence profiles displaced by a higher layer for the same id |
| 71 | /// (#5098). Shadowing is normal precedence, but it must be VISIBLE: a |
| 72 | /// personal edit that loses to a stale project copy otherwise changes |
| 73 | /// nothing anywhere with no signal why. |
| 74 | shadowed: Vec<ShadowedProfile>, |
| 75 | } |
| 76 | |
| 77 | /// A lower-precedence profile displaced by a higher layer for the same id. |
| 78 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 79 | pub struct ShadowedProfile { |
| 80 | pub id: String, |
| 81 | pub shadowed_origin: ProfileOrigin, |
| 82 | pub shadowed_source: PathBuf, |
| 83 | pub winner_origin: ProfileOrigin, |
| 84 | pub winner_source: PathBuf, |
| 85 | } |
| 86 | |
| 87 | /// Process-launch decision: whether project-scope agent profiles |
| 88 | /// (`.codewhale/agents/*.toml`) may join the dispatch roster (#5098). Set |
| 89 | /// once from `--no-project-config` at launch so every roster re-read (spawn |
| 90 | /// refresh, dispatch, views) honors the same trust decision other |
| 91 | /// project-level config already has (#485). Defaults to enabled, matching |
| 92 | /// project config itself. |
| 93 | static PROJECT_AGENT_PROFILES_ENABLED: std::sync::atomic::AtomicBool = |
| 94 | std::sync::atomic::AtomicBool::new(true); |
| 95 | |
| 96 | /// Record the launch-time trust decision for project-scope agent profiles. |
| 97 | pub fn set_project_agent_profiles_enabled(enabled: bool) { |
| 98 | PROJECT_AGENT_PROFILES_ENABLED.store(enabled, std::sync::atomic::Ordering::Relaxed); |
| 99 | } |
| 100 | |
| 101 | /// Whether project-scope agent profiles join the roster in this process. |
| 102 | #[must_use] |
| 103 | pub fn project_agent_profiles_enabled() -> bool { |
| 104 | PROJECT_AGENT_PROFILES_ENABLED.load(std::sync::atomic::Ordering::Relaxed) |
| 105 | } |
| 106 | |
| 107 | impl FleetRoster { |
| 108 | /// Roster containing only the built-in party. Used as the runtime default |
| 109 | /// before config/workspace layers are wired in. |
| 110 | #[must_use] |
| 111 | pub fn built_ins_only() -> Self { |
| 112 | Self { |
| 113 | members: Self::built_in_members(), |
| 114 | shadowed: Vec::new(), |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | /// A roster built from an explicit member list. |
| 119 | /// |
| 120 | /// Used for run-scoped rosters that are not a merge of the config layers — |
| 121 | /// notably an exact named Fleet, whose members are frozen at Workflow |
| 122 | /// start and must not pick up built-in or workspace profiles by name. |
| 123 | #[must_use] |
| 124 | pub fn from_members(members: Vec<AgentProfile>) -> Self { |
| 125 | Self { |
| 126 | members, |
| 127 | shadowed: Vec::new(), |
| 128 | } |
| 129 | } |
| 130 | |
| 131 | /// Load and merge the full roster for a workspace. |
| 132 | /// |
| 133 | /// Config members come from `[fleet.profiles]` (id = map key). Personal |
| 134 | /// members come from `$CODEWHALE_HOME/agents/*.toml`, and workspace members |
| 135 | /// come from `.codewhale/agents/*.toml`. A load failure is logged and |
| 136 | /// skipped so one broken profile layer cannot take down the session. |
| 137 | #[must_use] |
| 138 | pub fn load(fleet_config: &FleetConfigToml, workspace: &Path) -> Self { |
| 139 | let personal_dir = personal_agent_profile_dir().ok(); |
| 140 | Self::load_with_personal_dir( |
| 141 | fleet_config, |
| 142 | workspace, |
| 143 | personal_dir.as_deref(), |
| 144 | project_agent_profiles_enabled(), |
| 145 | ) |
| 146 | } |
| 147 | |
| 148 | fn load_with_personal_dir( |
| 149 | fleet_config: &FleetConfigToml, |
| 150 | workspace: &Path, |
| 151 | personal_dir: Option<&Path>, |
| 152 | include_workspace_profiles: bool, |
| 153 | ) -> Self { |
| 154 | let mut built_ins = Self::built_in_members(); |
| 155 | let mut extras: Vec<AgentProfile> = Vec::new(); |
| 156 | let mut shadowed: Vec<ShadowedProfile> = Vec::new(); |
| 157 | |
| 158 | for (id, profile) in &fleet_config.profiles { |
| 159 | let mut profile = profile.clone(); |
| 160 | profile.role.name = super::profile::canonical_public_role_name(&profile.role.name); |
| 161 | profile.slot = FleetSlot::from_name(&profile.role.name); |
| 162 | let member = AgentProfile { |
| 163 | id: id.clone(), |
| 164 | display_name: None, |
| 165 | description: profile.role.description.clone(), |
| 166 | profile, |
| 167 | source: PathBuf::from("config.toml"), |
| 168 | origin: ProfileOrigin::Config, |
| 169 | }; |
| 170 | record_shadow( |
| 171 | merge_member(&mut built_ins, &mut extras, member), |
| 172 | &mut shadowed, |
| 173 | ); |
| 174 | } |
| 175 | |
| 176 | if let Some(personal_dir) = personal_dir { |
| 177 | match load_agent_profiles_from_dir_tolerant(personal_dir, ProfileOrigin::Personal) { |
| 178 | Ok((profiles, issues)) => { |
| 179 | for issue in issues { |
| 180 | tracing::warn!( |
| 181 | "fleet roster: skipping invalid personal agent profile: {issue}" |
| 182 | ); |
| 183 | } |
| 184 | for member in profiles { |
| 185 | record_shadow( |
| 186 | merge_member(&mut built_ins, &mut extras, member), |
| 187 | &mut shadowed, |
| 188 | ); |
| 189 | } |
| 190 | } |
| 191 | Err(err) => { |
| 192 | tracing::warn!("fleet roster: skipping personal agent profiles: {err:#}"); |
| 193 | } |
| 194 | } |
| 195 | } |
| 196 | |
| 197 | // #5098: project-scope profiles join the dispatch roster only when the |
| 198 | // launch trusted project-level config (`--no-project-config` opts the |
| 199 | // whole layer out, same as `.codewhale/config.toml`). |
| 200 | if include_workspace_profiles { |
| 201 | match load_workspace_agent_profiles_tolerant(workspace) { |
| 202 | Ok((profiles, issues)) => { |
| 203 | for issue in issues { |
| 204 | tracing::warn!( |
| 205 | workspace = %workspace.display(), |
| 206 | "fleet roster: skipping invalid workspace agent profile: {issue}" |
| 207 | ); |
| 208 | } |
| 209 | for member in profiles { |
| 210 | record_shadow( |
| 211 | merge_member(&mut built_ins, &mut extras, member), |
| 212 | &mut shadowed, |
| 213 | ); |
| 214 | } |
| 215 | } |
| 216 | Err(err) => { |
| 217 | tracing::warn!( |
| 218 | workspace = %workspace.display(), |
| 219 | "fleet roster: skipping workspace agent profiles: {err:#}" |
| 220 | ); |
| 221 | } |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | for shadow in &shadowed { |
| 226 | // Overriding a built-in is the intended customization path — |
| 227 | // keep it quiet. A file layer (config/personal) losing to another |
| 228 | // file layer is the #5098 footgun: the edit changes nothing |
| 229 | // anywhere and must be visible. |
| 230 | if shadow.shadowed_origin == ProfileOrigin::BuiltIn { |
| 231 | tracing::debug!( |
| 232 | "fleet roster: '{}' {} copy at {} overrides the built-in default", |
| 233 | shadow.id, |
| 234 | shadow.winner_origin, |
| 235 | shadow.winner_source.display() |
| 236 | ); |
| 237 | } else { |
| 238 | tracing::warn!( |
| 239 | "fleet roster: '{}' {} copy at {} shadows the {} copy at {} (ignored)", |
| 240 | shadow.id, |
| 241 | shadow.winner_origin, |
| 242 | shadow.winner_source.display(), |
| 243 | shadow.shadowed_origin, |
| 244 | shadow.shadowed_source.display() |
| 245 | ); |
| 246 | } |
| 247 | } |
| 248 | |
| 249 | // Built-ins keep their canonical slot order (overrides included); |
| 250 | // config/workspace-only extras follow alphabetically. |
| 251 | extras.sort_by_key(|a| a.id.to_lowercase()); |
| 252 | let mut members = built_ins; |
| 253 | members.extend(extras); |
| 254 | Self { members, shadowed } |
| 255 | } |
| 256 | |
| 257 | /// The default party. Built-ins carry no permission grants (permissions |
| 258 | /// stay at the [`FleetProfilePermissions::default`] floor); behavior comes |
| 259 | /// from the role posture / system prompts plus the role `instructions` |
| 260 | /// below, which encode the coordination hierarchy: the **operator** (the |
| 261 | /// session's `/model` selection) directs the work and assigns managers |
| 262 | /// to workflows; a **manager** is the middle manager of one workflow. |
| 263 | #[must_use] |
| 264 | pub fn built_in_members() -> Vec<AgentProfile> { |
| 265 | [ |
| 266 | ( |
| 267 | "manager", |
| 268 | FleetSlot::Manager, |
| 269 | FleetLoadout::Inherit, |
| 270 | "Middle manager for one workflow: decomposes it into bounded tasks, dispatches workers, integrates results, and reports to the operator.", |
| 271 | Some( |
| 272 | "You lead exactly one workflow. Decompose it into bounded tasks, dispatch them to the right roles, keep work-in-progress small, integrate the results, and report a concise receipt (what was done, evidence, gaps) upward. Do not take on work outside your workflow.", |
| 273 | ), |
| 274 | ), |
| 275 | ( |
| 276 | "operator", |
| 277 | FleetSlot::Operator, |
| 278 | FleetLoadout::Inherit, |
| 279 | "The helm of the session — the session's /model selection. Assigns managers to Workflows, routes work between them, arbitrates conflicts, and reviews what comes back.", |
| 280 | Some( |
| 281 | "You direct the overall work, not individual Workflow steps. Assign a manager per Workflow, route work and context between them, arbitrate conflicts and priorities, review the receipts that come back, and decide what runs next. Delegate execution; keep judgment.", |
| 282 | ), |
| 283 | ), |
| 284 | ( |
| 285 | "scout", |
| 286 | FleetSlot::Scout, |
| 287 | FleetLoadout::Inherit, |
| 288 | "Read-only reconnaissance: find files, map code, gather evidence.", |
| 289 | None, |
| 290 | ), |
| 291 | ( |
| 292 | "builder", |
| 293 | FleetSlot::Implementer, |
| 294 | FleetLoadout::Inherit, |
| 295 | "Writes code: implements bounded tasks with write and shell access.", |
| 296 | None, |
| 297 | ), |
| 298 | ( |
| 299 | "reviewer", |
| 300 | FleetSlot::Reviewer, |
| 301 | FleetLoadout::Inherit, |
| 302 | "Adversarial code review: assumes the change is broken and tries to prove it — regressions, missing tests, unhandled cases. Read-only.", |
| 303 | Some( |
| 304 | "Be adversarial: assume the change is wrong until the evidence proves otherwise. Actively try to refute the claims made about the work — hunt regressions, missing tests, unhandled edge cases, and quiet behavior changes. Report severity-scored findings with file:line evidence; if nothing survives your attack, say so plainly. Never patch.", |
| 305 | ), |
| 306 | ), |
| 307 | ( |
| 308 | "verifier", |
| 309 | FleetSlot::Verifier, |
| 310 | FleetLoadout::Inherit, |
| 311 | "Runs builds and tests to verify claims; reports evidence, does not patch.", |
| 312 | None, |
| 313 | ), |
| 314 | ( |
| 315 | "consultant", |
| 316 | FleetSlot::Custom("consultant".to_string()), |
| 317 | FleetLoadout::Inherit, |
| 318 | "Short-lived, high-reasoning, read-only counsel for difficult decisions and overlooked risks.", |
| 319 | Some( |
| 320 | "Give the operator a direct second opinion grounded in what you can read. Surface the decisive tradeoff, overlooked failure mode, and your recommendation. Advise only: do not edit files or run commands.", |
| 321 | ), |
| 322 | ), |
| 323 | ( |
| 324 | "synthesizer", |
| 325 | FleetSlot::Summarizer, |
| 326 | FleetLoadout::Inherit, |
| 327 | "Read-only synthesis: merge findings into one coherent report.", |
| 328 | None, |
| 329 | ), |
| 330 | ( |
| 331 | "general", |
| 332 | FleetSlot::General, |
| 333 | FleetLoadout::Inherit, |
| 334 | "General-purpose worker with full capabilities.", |
| 335 | None, |
| 336 | ), |
| 337 | ] |
| 338 | .into_iter() |
| 339 | .map(|(id, slot, loadout, description, instructions)| AgentProfile { |
| 340 | id: id.to_string(), |
| 341 | display_name: None, |
| 342 | description: Some(description.to_string()), |
| 343 | profile: FleetProfile { |
| 344 | slot, |
| 345 | role: FleetRole { |
| 346 | name: id.to_string(), |
| 347 | description: Some(description.to_string()), |
| 348 | instructions: instructions.map(str::to_string), |
| 349 | }, |
| 350 | loadout, |
| 351 | model: None, |
| 352 | provider: None, |
| 353 | reasoning_effort: (id == "consultant").then(|| "high".to_string()), |
| 354 | permissions: FleetProfilePermissions::default(), |
| 355 | delegation: FleetDelegationHints::default(), |
| 356 | }, |
| 357 | source: PathBuf::from("built-in"), |
| 358 | origin: ProfileOrigin::BuiltIn, |
| 359 | }) |
| 360 | .collect() |
| 361 | } |
| 362 | |
| 363 | /// Look up a member by id (trimmed, case-insensitive). |
| 364 | #[must_use] |
| 365 | pub fn get(&self, id: &str) -> Option<&AgentProfile> { |
| 366 | let id = id.trim(); |
| 367 | self.members |
| 368 | .iter() |
| 369 | .find(|member| member.id.trim().eq_ignore_ascii_case(id)) |
| 370 | } |
| 371 | |
| 372 | /// All members in stable order: built-in canonical order first (an |
| 373 | /// overridden built-in keeps its slot but shows its overriding origin), |
| 374 | /// then extra config/workspace-only members alphabetically. |
| 375 | #[must_use] |
| 376 | pub fn members(&self) -> &[AgentProfile] { |
| 377 | &self.members |
| 378 | } |
| 379 | |
| 380 | /// Per-member explicit model pins, keyed by lowercased member id. |
| 381 | /// Feeds the sub-agent `role_models` lookup; explicit `[subagents]` |
| 382 | /// overrides are merged on top by the engine and win. |
| 383 | /// |
| 384 | /// Members that ALSO pin a provider are deliberately excluded. This map is |
| 385 | /// provider-less by construction: the sub-agent role/type lookup |
| 386 | /// (`configured_model_for_role_or_type`) applies whatever it finds against |
| 387 | /// the *session* provider's client, so exporting a provider-pinned model |
| 388 | /// here strips the only thing that made the id routable. A profile pinning |
| 389 | /// `provider = "deepseek"` + `model = "deepseek-v4-flash"` then leaks that |
| 390 | /// bare id onto an unrelated session route — and on a pass-through |
| 391 | /// provider (Alibaba Model Studio, whose Token Plan actually serves |
| 392 | /// `deepseek-v4-flash-0731`) nothing downstream rejects it, so the child |
| 393 | /// dies on the provider's own denial instead of inheriting the parent's |
| 394 | /// working model. Provider-pinned profiles keep their full route through |
| 395 | /// the profile spawn path (`child_provider_binding`), which builds a client |
| 396 | /// for the pinned provider and carries the model with it. |
| 397 | #[must_use] |
| 398 | pub fn model_overrides(&self) -> HashMap<String, String> { |
| 399 | self.members |
| 400 | .iter() |
| 401 | .filter_map(|member| { |
| 402 | if member |
| 403 | .profile |
| 404 | .provider |
| 405 | .as_deref() |
| 406 | .is_some_and(|provider| !provider.trim().is_empty()) |
| 407 | { |
| 408 | return None; |
| 409 | } |
| 410 | let model = member.profile.model.as_deref()?.trim(); |
| 411 | (!model.is_empty()).then(|| (member.id.to_lowercase(), model.to_string())) |
| 412 | }) |
| 413 | .collect() |
| 414 | } |
| 415 | /// Lower-precedence profiles displaced by higher layers (#5098). Empty |
| 416 | /// for `built_ins_only` / `from_members` rosters. |
| 417 | #[must_use] |
| 418 | pub fn shadowed(&self) -> &[ShadowedProfile] { |
| 419 | &self.shadowed |
| 420 | } |
| 421 | |
| 422 | /// Shadow records for one member id (trimmed, case-insensitive). |
| 423 | pub fn shadowed_for<'a>(&'a self, id: &'a str) -> impl Iterator<Item = &'a ShadowedProfile> { |
| 424 | let id = id.trim().to_lowercase(); |
| 425 | self.shadowed |
| 426 | .iter() |
| 427 | .filter(move |shadow| shadow.id.trim().eq_ignore_ascii_case(&id)) |
| 428 | } |
| 429 | } |
| 430 | |
| 431 | /// Fold a displaced layer (if any) into the shadow log. |
| 432 | fn record_shadow(displaced: Option<ShadowedProfile>, shadowed: &mut Vec<ShadowedProfile>) { |
| 433 | if let Some(shadow) = displaced { |
| 434 | shadowed.push(shadow); |
| 435 | } |
| 436 | } |
| 437 | |
| 438 | /// Overlay `member` onto the roster layers: replace an existing member with |
| 439 | /// the same id (case-insensitive) in place, otherwise collect it as an extra. |
| 440 | /// Returns a shadow record when a lower-precedence layer was displaced so the |
| 441 | /// load can log it and the roster can surface it (#5098). |
| 442 | fn merge_member( |
| 443 | built_ins: &mut [AgentProfile], |
| 444 | extras: &mut Vec<AgentProfile>, |
| 445 | member: AgentProfile, |
| 446 | ) -> Option<ShadowedProfile> { |
| 447 | let matches = |
| 448 | |existing: &AgentProfile| existing.id.trim().eq_ignore_ascii_case(member.id.trim()); |
| 449 | let slot = built_ins |
| 450 | .iter_mut() |
| 451 | .find(|existing| matches(existing)) |
| 452 | .or_else(|| extras.iter_mut().find(|existing| matches(existing))); |
| 453 | match slot { |
| 454 | Some(existing) => { |
| 455 | let shadow = ShadowedProfile { |
| 456 | id: existing.id.clone(), |
| 457 | shadowed_origin: existing.origin, |
| 458 | shadowed_source: existing.source.clone(), |
| 459 | winner_origin: member.origin, |
| 460 | winner_source: member.source.clone(), |
| 461 | }; |
| 462 | *existing = member; |
| 463 | Some(shadow) |
| 464 | } |
| 465 | None => { |
| 466 | extras.push(member); |
| 467 | None |
| 468 | } |
| 469 | } |
| 470 | } |
| 471 | |
| 472 | #[cfg(test)] |
| 473 | mod tests { |
| 474 | use super::*; |
| 475 | use std::collections::BTreeMap; |
| 476 | use tempfile::TempDir; |
| 477 | |
| 478 | fn config_with_profiles(profiles: BTreeMap<String, FleetProfile>) -> FleetConfigToml { |
| 479 | FleetConfigToml { |
| 480 | profiles, |
| 481 | ..FleetConfigToml::default() |
| 482 | } |
| 483 | } |
| 484 | |
| 485 | fn config_profile(role: &str, model: Option<&str>) -> FleetProfile { |
| 486 | FleetProfile { |
| 487 | slot: FleetSlot::from_name(role), |
| 488 | role: FleetRole { |
| 489 | name: role.to_string(), |
| 490 | description: Some(format!("{role} from config")), |
| 491 | instructions: None, |
| 492 | }, |
| 493 | loadout: FleetLoadout::Inherit, |
| 494 | model: model.map(str::to_string), |
| 495 | provider: None, |
| 496 | reasoning_effort: None, |
| 497 | permissions: FleetProfilePermissions::default(), |
| 498 | delegation: FleetDelegationHints::default(), |
| 499 | } |
| 500 | } |
| 501 | |
| 502 | fn write_workspace_profile(workspace: &Path, filename: &str, contents: &str) { |
| 503 | let dir = workspace.join(super::super::profile::WORKSPACE_AGENT_PROFILE_DIR); |
| 504 | std::fs::create_dir_all(&dir).unwrap(); |
| 505 | std::fs::write(dir.join(filename), contents).unwrap(); |
| 506 | } |
| 507 | |
| 508 | #[test] |
| 509 | fn built_in_party_is_complete_with_floor_permissions() { |
| 510 | let members = FleetRoster::built_in_members(); |
| 511 | let ids: Vec<&str> = members.iter().map(|m| m.id.as_str()).collect(); |
| 512 | assert_eq!( |
| 513 | ids, |
| 514 | [ |
| 515 | "manager", |
| 516 | "operator", |
| 517 | "scout", |
| 518 | "builder", |
| 519 | "reviewer", |
| 520 | "verifier", |
| 521 | "consultant", |
| 522 | "synthesizer", |
| 523 | "general" |
| 524 | ] |
| 525 | ); |
| 526 | for member in &members { |
| 527 | assert_eq!(member.origin, ProfileOrigin::BuiltIn, "{}", member.id); |
| 528 | assert_eq!( |
| 529 | member.profile.permissions, |
| 530 | FleetProfilePermissions::default(), |
| 531 | "built-in {} must stay at the permission floor", |
| 532 | member.id |
| 533 | ); |
| 534 | assert_eq!( |
| 535 | member.profile.delegation, |
| 536 | FleetDelegationHints::default(), |
| 537 | "{}", |
| 538 | member.id |
| 539 | ); |
| 540 | assert!(member.profile.model.is_none(), "{}", member.id); |
| 541 | assert_eq!( |
| 542 | member.profile.reasoning_effort.as_deref(), |
| 543 | (member.id == "consultant").then_some("high"), |
| 544 | "built-in {} reasoning", |
| 545 | member.id |
| 546 | ); |
| 547 | // The coordination hierarchy (operator/manager) and the |
| 548 | // adversarial reviewer carry role doctrine; the remaining |
| 549 | // built-ins get behavior from posture / system prompts alone. |
| 550 | let carries_doctrine = matches!( |
| 551 | member.id.as_str(), |
| 552 | "manager" | "operator" | "reviewer" | "consultant" |
| 553 | ); |
| 554 | assert_eq!( |
| 555 | member.profile.role.instructions.is_some(), |
| 556 | carries_doctrine, |
| 557 | "built-in {} instructions presence", |
| 558 | member.id |
| 559 | ); |
| 560 | assert!(member.description.is_some(), "{}", member.id); |
| 561 | } |
| 562 | assert_eq!(members[0].profile.slot, FleetSlot::Manager); |
| 563 | assert_eq!(members[1].profile.slot, FleetSlot::Operator); |
| 564 | assert_eq!(members[2].profile.loadout, FleetLoadout::Inherit); |
| 565 | assert_eq!(members[6].profile.slot.as_str(), "consultant"); |
| 566 | assert_eq!(members[7].profile.slot, FleetSlot::Summarizer); |
| 567 | assert_eq!(members[7].profile.loadout, FleetLoadout::Inherit); |
| 568 | } |
| 569 | |
| 570 | #[test] |
| 571 | fn config_member_overrides_built_in_and_extras_sort_alphabetically() { |
| 572 | let _env_lock = crate::test_support::lock_test_env(); |
| 573 | let home = TempDir::new().unwrap(); |
| 574 | let _codewhale_home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 575 | let tmp = TempDir::new().unwrap(); |
| 576 | let config = config_with_profiles(BTreeMap::from([ |
| 577 | ( |
| 578 | "reviewer".to_string(), |
| 579 | config_profile("reviewer", Some("deepseek-v4-pro")), |
| 580 | ), |
| 581 | ("zeta".to_string(), config_profile("scout", None)), |
| 582 | ("alpha".to_string(), config_profile("builder", None)), |
| 583 | ])); |
| 584 | |
| 585 | // Isolate from ambient personal agent profiles on developer machines. |
| 586 | let roster = FleetRoster::load_with_personal_dir(&config, tmp.path(), None, true); |
| 587 | |
| 588 | let ids: Vec<&str> = roster.members().iter().map(|m| m.id.as_str()).collect(); |
| 589 | assert_eq!( |
| 590 | ids, |
| 591 | [ |
| 592 | "manager", |
| 593 | "operator", |
| 594 | "scout", |
| 595 | "builder", |
| 596 | "reviewer", |
| 597 | "verifier", |
| 598 | "consultant", |
| 599 | "synthesizer", |
| 600 | "general", |
| 601 | "alpha", |
| 602 | "zeta" |
| 603 | ], |
| 604 | "overridden built-in keeps its slot; extras follow alphabetically" |
| 605 | ); |
| 606 | let reviewer = roster.get("reviewer").unwrap(); |
| 607 | assert_eq!(reviewer.origin, ProfileOrigin::Config); |
| 608 | assert_eq!(reviewer.profile.model.as_deref(), Some("deepseek-v4-pro")); |
| 609 | assert_eq!(reviewer.source, PathBuf::from("config.toml")); |
| 610 | } |
| 611 | |
| 612 | #[test] |
| 613 | fn workspace_member_wins_over_config_and_built_in() { |
| 614 | let tmp = TempDir::new().unwrap(); |
| 615 | write_workspace_profile( |
| 616 | tmp.path(), |
| 617 | "reviewer.toml", |
| 618 | "id = \"reviewer\"\nrole_hint = \"reviewer\"\nmodel = \"glm-5.2\"\n", |
| 619 | ); |
| 620 | let config = config_with_profiles(BTreeMap::from([( |
| 621 | "reviewer".to_string(), |
| 622 | config_profile("reviewer", Some("deepseek-v4-pro")), |
| 623 | )])); |
| 624 | |
| 625 | let roster = FleetRoster::load(&config, tmp.path()); |
| 626 | |
| 627 | let reviewer = roster.get("reviewer").unwrap(); |
| 628 | assert_eq!(reviewer.origin, ProfileOrigin::Workspace); |
| 629 | assert_eq!(reviewer.profile.model.as_deref(), Some("glm-5.2")); |
| 630 | // Precedence must not duplicate the member. |
| 631 | assert_eq!( |
| 632 | roster |
| 633 | .members() |
| 634 | .iter() |
| 635 | .filter(|m| m.id == "reviewer") |
| 636 | .count(), |
| 637 | 1 |
| 638 | ); |
| 639 | } |
| 640 | |
| 641 | #[test] |
| 642 | fn personal_member_applies_across_projects_but_project_still_wins() { |
| 643 | let tmp = TempDir::new().unwrap(); |
| 644 | let personal_dir = tmp.path().join("personal-agents"); |
| 645 | std::fs::create_dir_all(&personal_dir).unwrap(); |
| 646 | std::fs::write( |
| 647 | personal_dir.join("reviewer.toml"), |
| 648 | "id = \"reviewer\"\nrole_hint = \"reviewer\"\nmodel = \"deepseek-v4-flash\"\n", |
| 649 | ) |
| 650 | .unwrap(); |
| 651 | let workspace = tmp.path().join("workspace"); |
| 652 | std::fs::create_dir_all(&workspace).unwrap(); |
| 653 | |
| 654 | let personal = FleetRoster::load_with_personal_dir( |
| 655 | &FleetConfigToml::default(), |
| 656 | &workspace, |
| 657 | Some(&personal_dir), |
| 658 | true, |
| 659 | ); |
| 660 | let reviewer = personal.get("reviewer").unwrap(); |
| 661 | assert_eq!(reviewer.origin, ProfileOrigin::Personal); |
| 662 | assert_eq!(reviewer.profile.model.as_deref(), Some("deepseek-v4-flash")); |
| 663 | |
| 664 | write_workspace_profile( |
| 665 | &workspace, |
| 666 | "reviewer.toml", |
| 667 | "id = \"reviewer\"\nrole_hint = \"reviewer\"\nmodel = \"glm-5.2\"\n", |
| 668 | ); |
| 669 | let project = FleetRoster::load_with_personal_dir( |
| 670 | &FleetConfigToml::default(), |
| 671 | &workspace, |
| 672 | Some(&personal_dir), |
| 673 | true, |
| 674 | ); |
| 675 | let reviewer = project.get("reviewer").unwrap(); |
| 676 | assert_eq!(reviewer.origin, ProfileOrigin::Workspace); |
| 677 | assert_eq!(reviewer.profile.model.as_deref(), Some("glm-5.2")); |
| 678 | } |
| 679 | |
| 680 | #[test] |
| 681 | fn personal_setup_target_round_trips_through_the_runtime_roster() { |
| 682 | let _env_lock = crate::test_support::lock_test_env(); |
| 683 | let home = TempDir::new().unwrap(); |
| 684 | let _codewhale_home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 685 | let workspace = TempDir::new().unwrap(); |
| 686 | let personal_dir = super::super::profile::agent_profile_dir_for_scope( |
| 687 | super::super::profile::FleetProfileScope::Personal, |
| 688 | workspace.path(), |
| 689 | ) |
| 690 | .expect("personal profile directory"); |
| 691 | assert_eq!(personal_dir, home.path().join("agents")); |
| 692 | |
| 693 | let target = personal_dir.join("reviewer.toml"); |
| 694 | let mut transaction = codewhale_config::persistence::SetupTransaction::new(); |
| 695 | transaction.stage( |
| 696 | target.clone(), |
| 697 | b"id = \"reviewer\"\nrole_hint = \"reviewer\"\nprovider = \"deepseek\"\nmodel = \"deepseek-v4-flash\"\n" |
| 698 | .to_vec(), |
| 699 | ); |
| 700 | transaction.commit().expect("atomic personal save"); |
| 701 | assert!(target.is_file(), "save must land under CODEWHALE_HOME"); |
| 702 | |
| 703 | let roster = FleetRoster::load(&FleetConfigToml::default(), workspace.path()); |
| 704 | let reviewer = roster |
| 705 | .get("reviewer") |
| 706 | .expect("saved personal profile must be loaded"); |
| 707 | assert_eq!(reviewer.origin, ProfileOrigin::Personal); |
| 708 | assert_eq!(reviewer.source, target); |
| 709 | assert_eq!(reviewer.profile.provider.as_deref(), Some("deepseek")); |
| 710 | assert_eq!(reviewer.profile.model.as_deref(), Some("deepseek-v4-flash")); |
| 711 | } |
| 712 | |
| 713 | #[test] |
| 714 | fn broken_workspace_dir_degrades_to_built_ins_and_config() { |
| 715 | let tmp = TempDir::new().unwrap(); |
| 716 | // A malformed provider token is still a load failure (#4093 / #3965): |
| 717 | // profile pins may name built-ins or simple custom ids like |
| 718 | // `lm-studio`, but whitespace/punctuation is rejected so a broken |
| 719 | // workspace dir still degrades to built-ins + config. |
| 720 | write_workspace_profile( |
| 721 | tmp.path(), |
| 722 | "broken.toml", |
| 723 | "provider = \"not a real provider\"\n", |
| 724 | ); |
| 725 | let config = config_with_profiles(BTreeMap::from([( |
| 726 | "extra".to_string(), |
| 727 | config_profile("scout", None), |
| 728 | )])); |
| 729 | |
| 730 | let roster = FleetRoster::load(&config, tmp.path()); |
| 731 | |
| 732 | assert!(roster.get("extra").is_some()); |
| 733 | assert_eq!( |
| 734 | roster.members().len(), |
| 735 | FleetRoster::built_in_members().len() + 1 |
| 736 | ); |
| 737 | } |
| 738 | |
| 739 | #[test] |
| 740 | fn invalid_legacy_profile_does_not_hide_valid_scout_neighbor() { |
| 741 | let _env_lock = crate::test_support::lock_test_env(); |
| 742 | let home = TempDir::new().unwrap(); |
| 743 | let _codewhale_home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 744 | let tmp = TempDir::new().unwrap(); |
| 745 | write_workspace_profile( |
| 746 | tmp.path(), |
| 747 | "reviewer.toml", |
| 748 | "id = \"reviewer\"\nmodel_class_hint = \"heavy\"\n", |
| 749 | ); |
| 750 | write_workspace_profile( |
| 751 | tmp.path(), |
| 752 | "scout.toml", |
| 753 | "id = \"scout\"\nrole_hint = \"scout\"\nprovider = \"deepseek\"\nmodel = \"deepseek-v4-flash\"\n", |
| 754 | ); |
| 755 | |
| 756 | // Isolate from ambient personal agent profiles on developer machines. |
| 757 | let roster = FleetRoster::load_with_personal_dir( |
| 758 | &FleetConfigToml::default(), |
| 759 | tmp.path(), |
| 760 | None, |
| 761 | true, |
| 762 | ); |
| 763 | |
| 764 | let scout = roster.get("scout").expect("valid scout remains visible"); |
| 765 | assert_eq!(scout.origin, ProfileOrigin::Workspace); |
| 766 | assert_eq!(scout.profile.provider.as_deref(), Some("deepseek")); |
| 767 | assert_eq!(scout.profile.model.as_deref(), Some("deepseek-v4-flash")); |
| 768 | assert_eq!( |
| 769 | roster.get("reviewer").unwrap().origin, |
| 770 | ProfileOrigin::BuiltIn, |
| 771 | "invalid legacy override must fall back to the safe built-in" |
| 772 | ); |
| 773 | } |
| 774 | |
| 775 | #[test] |
| 776 | fn model_overrides_use_lowercased_ids_and_only_explicit_models() { |
| 777 | let _env_lock = crate::test_support::lock_test_env(); |
| 778 | let home = TempDir::new().unwrap(); |
| 779 | let _codewhale_home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 780 | // Isolate personal `$CODEWHALE_HOME/agents` so ambient developer |
| 781 | // profiles cannot pin built-ins like manager during unit tests. |
| 782 | let tmp = TempDir::new().unwrap(); |
| 783 | let config = config_with_profiles(BTreeMap::from([ |
| 784 | ( |
| 785 | "Reviewer".to_string(), |
| 786 | config_profile("reviewer", Some("deepseek-v4-pro")), |
| 787 | ), |
| 788 | ("scout".to_string(), config_profile("scout", None)), |
| 789 | ])); |
| 790 | |
| 791 | let roster = FleetRoster::load(&config, tmp.path()); |
| 792 | let overrides = roster.model_overrides(); |
| 793 | |
| 794 | assert_eq!( |
| 795 | overrides, |
| 796 | HashMap::from([("reviewer".to_string(), "deepseek-v4-pro".to_string())]), |
| 797 | "only members with explicit models are pinned, keyed lowercased" |
| 798 | ); |
| 799 | } |
| 800 | |
| 801 | /// A profile that pins BOTH a provider and a model must not contribute to |
| 802 | /// the provider-less `role_models` map. That map is applied against the |
| 803 | /// session provider's client, so exporting `deepseek-v4-flash` from a |
| 804 | /// `provider = "deepseek"` scout profile sent a bare DeepSeek id onto an |
| 805 | /// Alibaba Model Studio session (a pass-through provider, so nothing |
| 806 | /// downstream rejected it) and the scout died on the provider's denial — |
| 807 | /// the Model Studio Token Plan roster serves `deepseek-v4-flash-0731`, not |
| 808 | /// `deepseek-v4-flash`. Provider-pinned profiles keep their model via the |
| 809 | /// profile spawn path, which builds a client for the pinned provider. |
| 810 | #[test] |
| 811 | fn model_overrides_skip_provider_pinned_profiles() { |
| 812 | let _env_lock = crate::test_support::lock_test_env(); |
| 813 | let home = TempDir::new().unwrap(); |
| 814 | let _codewhale_home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 815 | let tmp = TempDir::new().unwrap(); |
| 816 | |
| 817 | let mut pinned = config_profile("scout", Some("deepseek-v4-flash")); |
| 818 | pinned.provider = Some("deepseek".to_string()); |
| 819 | let mut blank_provider = config_profile("builder", Some("deepseek-v4-pro")); |
| 820 | blank_provider.provider = Some(" ".to_string()); |
| 821 | |
| 822 | let config = config_with_profiles(BTreeMap::from([ |
| 823 | ("scout".to_string(), pinned), |
| 824 | ("builder".to_string(), blank_provider), |
| 825 | ])); |
| 826 | |
| 827 | let roster = FleetRoster::load(&config, tmp.path()); |
| 828 | let overrides = roster.model_overrides(); |
| 829 | |
| 830 | assert!( |
| 831 | !overrides.contains_key("scout"), |
| 832 | "a provider-pinned profile must not leak its model into the \ |
| 833 | provider-less role_models map: {overrides:?}" |
| 834 | ); |
| 835 | assert_eq!( |
| 836 | overrides.get("builder").map(String::as_str), |
| 837 | Some("deepseek-v4-pro"), |
| 838 | "a blank provider pin is still provider-less: {overrides:?}" |
| 839 | ); |
| 840 | // The pin itself survives on the member for the profile spawn path. |
| 841 | let scout = roster.get("scout").expect("scout member"); |
| 842 | assert_eq!(scout.profile.provider.as_deref(), Some("deepseek")); |
| 843 | assert_eq!(scout.profile.model.as_deref(), Some("deepseek-v4-flash")); |
| 844 | } |
| 845 | |
| 846 | #[test] |
| 847 | fn get_is_trimmed_and_case_insensitive() { |
| 848 | let roster = FleetRoster::built_ins_only(); |
| 849 | assert!(roster.get(" Reviewer ").is_some()); |
| 850 | assert!(roster.get("SYNTHESIZER").is_some()); |
| 851 | assert!(roster.get("nonexistent").is_none()); |
| 852 | } |
| 853 | |
| 854 | #[test] |
| 855 | fn origin_labels_are_stable() { |
| 856 | assert_eq!(ProfileOrigin::BuiltIn.to_string(), "built-in"); |
| 857 | assert_eq!(ProfileOrigin::Config.to_string(), "config"); |
| 858 | assert_eq!(ProfileOrigin::Personal.to_string(), "personal"); |
| 859 | assert_eq!(ProfileOrigin::Workspace.to_string(), "project"); |
| 860 | } |
| 861 | } |
| 862 | |
| 863 | #[cfg(test)] |
| 864 | mod shadow_and_trust_tests { |
| 865 | use super::*; |
| 866 | use tempfile::TempDir; |
| 867 | |
| 868 | fn write_profile(dir: &Path, filename: &str, contents: &str) { |
| 869 | std::fs::create_dir_all(dir).unwrap(); |
| 870 | std::fs::write(dir.join(filename), contents).unwrap(); |
| 871 | } |
| 872 | |
| 873 | #[test] |
| 874 | fn workspace_shadow_of_personal_file_is_recorded_and_reported() { |
| 875 | // #5098: editing the personal builder.toml changed nothing because a |
| 876 | // project copy silently shadowed it. The roster must report that the |
| 877 | // shadowed personal file exists and is ignored. |
| 878 | let tmp = TempDir::new().unwrap(); |
| 879 | let personal_dir = tmp.path().join("personal"); |
| 880 | let workspace = tmp.path().join("workspace"); |
| 881 | std::fs::create_dir_all(&workspace).unwrap(); |
| 882 | write_profile( |
| 883 | &personal_dir, |
| 884 | "builder.toml", |
| 885 | "id = \"builder\"\nrole_hint = \"builder\"\nmodel = \"deepseek-v4-flash\"\n", |
| 886 | ); |
| 887 | write_profile( |
| 888 | &workspace.join(".codewhale").join("agents"), |
| 889 | "builder.toml", |
| 890 | "id = \"builder\"\nrole_hint = \"builder\"\nmodel = \"deepseek-v4-pro\"\n", |
| 891 | ); |
| 892 | |
| 893 | let roster = FleetRoster::load_with_personal_dir( |
| 894 | &FleetConfigToml::default(), |
| 895 | &workspace, |
| 896 | Some(&personal_dir), |
| 897 | true, |
| 898 | ); |
| 899 | |
| 900 | let builder = roster.get("builder").expect("builder member"); |
| 901 | assert_eq!(builder.origin, ProfileOrigin::Workspace); |
| 902 | let shadows: Vec<_> = roster.shadowed_for("builder").collect(); |
| 903 | // The chain is built-in → personal → workspace; both displacements |
| 904 | // are recorded, and the file-on-file one names the ignored personal |
| 905 | // copy explicitly. |
| 906 | assert_eq!(shadows.len(), 2, "full shadow chain: {shadows:?}"); |
| 907 | let shadow = shadows |
| 908 | .iter() |
| 909 | .find(|shadow| shadow.shadowed_origin == ProfileOrigin::Personal) |
| 910 | .expect("personal file shadow is recorded"); |
| 911 | assert!(shadow.shadowed_source.ends_with("builder.toml")); |
| 912 | assert_eq!(shadow.winner_origin, ProfileOrigin::Workspace); |
| 913 | assert!( |
| 914 | shadows |
| 915 | .iter() |
| 916 | .any(|shadow| shadow.shadowed_origin == ProfileOrigin::BuiltIn), |
| 917 | "the built-in displacement is recorded too: {shadows:?}" |
| 918 | ); |
| 919 | assert!( |
| 920 | roster.shadowed().iter().any(|s| s.id == "builder"), |
| 921 | "roster-level shadow log carries the record" |
| 922 | ); |
| 923 | } |
| 924 | |
| 925 | #[test] |
| 926 | fn project_scope_profiles_are_skipped_when_the_layer_is_not_trusted() { |
| 927 | // #5098: `load_workspace_agent_profiles_tolerant` applied no trust |
| 928 | // check — a cloned repo's .codewhale/agents/*.toml silently joined |
| 929 | // the dispatch roster. With project config disabled |
| 930 | // (`--no-project-config`), the whole layer stays out. |
| 931 | let tmp = TempDir::new().unwrap(); |
| 932 | let workspace = tmp.path().join("workspace"); |
| 933 | write_profile( |
| 934 | &workspace.join(".codewhale").join("agents"), |
| 935 | "builder.toml", |
| 936 | "id = \"builder\"\nrole_hint = \"builder\"\nmodel = \"gpt-5.6-luna\"\n", |
| 937 | ); |
| 938 | |
| 939 | let gated = FleetRoster::load_with_personal_dir( |
| 940 | &FleetConfigToml::default(), |
| 941 | &workspace, |
| 942 | None, |
| 943 | false, |
| 944 | ); |
| 945 | let builder = gated.get("builder").expect("built-in builder remains"); |
| 946 | assert_eq!( |
| 947 | builder.origin, |
| 948 | ProfileOrigin::BuiltIn, |
| 949 | "untrusted project profile must not join the roster" |
| 950 | ); |
| 951 | assert_ne!( |
| 952 | builder.profile.model.as_deref(), |
| 953 | Some("gpt-5.6-luna"), |
| 954 | "foreign project pin must not reach dispatch" |
| 955 | ); |
| 956 | |
| 957 | let trusted = FleetRoster::load_with_personal_dir( |
| 958 | &FleetConfigToml::default(), |
| 959 | &workspace, |
| 960 | None, |
| 961 | true, |
| 962 | ); |
| 963 | assert_eq!( |
| 964 | trusted.get("builder").expect("builder").origin, |
| 965 | ProfileOrigin::Workspace, |
| 966 | "trusted project profile wins as before" |
| 967 | ); |
| 968 | } |
| 969 | } |
| 970 |