| 1 | //! Runtime for an **exact named Fleet** (`schema = "exact"`). |
| 2 | //! |
| 3 | //! The saved Fleet is the Fleet that runs. At Workflow start its definition is |
| 4 | //! read from the standard `FleetSearchRoot` locations, every worker route is |
| 5 | //! **preflighted and frozen**, the attached Reasoning Router service is |
| 6 | //! resolved, and the whole thing is captured into an immutable |
| 7 | //! [`FleetSnapshot`] projected onto the roster/profile machinery the in-process |
| 8 | //! spawn path already uses. |
| 9 | //! |
| 10 | //! Five invariants govern everything below. |
| 11 | //! |
| 12 | //! 1. **Routes freeze first, and are checked while freezing.** Provider |
| 13 | //! identity, canonical wire model, endpoint, local credential readiness, and |
| 14 | //! reasoning capability are all resolved before the Workflow starts — and |
| 15 | //! certainly before any Router is asked anything. Nothing downstream may |
| 16 | //! move them: not a task option, not the Router. |
| 17 | //! 2. **Admission comes before cost.** A task is resolved against the roster, |
| 18 | //! checked against gates, and given a concurrency slot *before* the Router |
| 19 | //! is called. A rejected or capacity-blocked task spends no Router tokens |
| 20 | //! and discloses nothing to a Router's provider. |
| 21 | //! 3. **Auto is a reasoning decision, and the attached Router makes it.** |
| 22 | //! `reasoning = "auto"` always goes to the Fleet's Reasoning Router — no |
| 23 | //! provider-native-adaptive bypass, no legacy model routing, no local |
| 24 | //! keyword heuristic. A manual tier calls no Router at all. |
| 25 | //! 4. **Runtime owns authority.** After exact member selection, Runtime maps |
| 26 | //! the semantic role onto its closed role policy and intersects that policy |
| 27 | //! with the live parent. Fleet identity never grants or withholds project |
| 28 | //! trust, tools, writes, network reach, shell, or delegation. |
| 29 | //! 5. **Receipts are truthful and content-free.** The tier a selector picked, |
| 30 | //! the control a provider actually receives, and what a Router cost are |
| 31 | //! recorded separately; task text never is. |
| 32 | |
| 33 | use std::sync::Arc; |
| 34 | |
| 35 | use async_trait::async_trait; |
| 36 | #[cfg(test)] |
| 37 | use codewhale_workflow::ShellCeiling; |
| 38 | use codewhale_workflow::{ |
| 39 | CapturedReasoningRouter, CredentialReadiness, EffectiveReasoning, EndpointIdentity, |
| 40 | FleetDocument, FleetRouterRef, FleetSearchRoot, FleetSnapshot, FleetSnapshotMember, |
| 41 | FleetTaskReceipt, NamedFleetError, PermissionCeiling, PreflightError, PreflightedRoute, |
| 42 | ProviderReasoningControl, QualifiedFleetId, ReasoningCapability, ReasoningRouterProfile, |
| 43 | ReasoningTier, ResolvedReasoning, RoutePreflight, RouterAvailability, RouterCallInput, |
| 44 | RouterCallPlan, RouterIdentity, RoutingDisclosure, bounded_routing_payload, |
| 45 | captured_legacy_inline_router, parse_router_decision, resolve_exact_member_reasoning, |
| 46 | router_call_plan, router_system_prompt, router_user_message, |
| 47 | }; |
| 48 | |
| 49 | use super::role::{ChildAuthority, public_role_label}; |
| 50 | #[cfg(test)] |
| 51 | use super::role::{ |
| 52 | NETWORK_DENIAL_SENTINEL, NETWORK_TOOL_DENYLIST, RAW_SHELL_SENTINEL, is_posture_denial, |
| 53 | session_shell_ceiling, |
| 54 | }; |
| 55 | use crate::config::{Config, ProviderKind}; |
| 56 | use crate::llm_client::LlmClient; |
| 57 | use crate::reasoning_preference::ReasoningEffort; |
| 58 | use codewhale_models::Role; |
| 59 | |
| 60 | /// Where exact Fleet definitions and Reasoning Router profiles are looked up, |
| 61 | /// labelled so an identity can be qualified (`workspace/glm-pair`) instead of |
| 62 | /// silently shadowed. |
| 63 | fn personal_fleet_root() -> anyhow::Result<std::path::PathBuf> { |
| 64 | codewhale_config::codewhale_home() |
| 65 | } |
| 66 | |
| 67 | pub(crate) fn personal_fleet_definitions_dir() -> anyhow::Result<std::path::PathBuf> { |
| 68 | Ok(personal_fleet_root()?.join("fleets")) |
| 69 | } |
| 70 | |
| 71 | /// The workspace has two origins. `workspace` is `<workspace>/.codewhale`, the |
| 72 | /// directory the Fleet store saves workspace-scoped Fleets to, so a Fleet |
| 73 | /// saved from the Fleet UI is found by name. `workspace_root` is the workspace |
| 74 | /// directory itself, which keeps checked-in `fleets/<name>.toml` rosters |
| 75 | /// loading as they always have. |
| 76 | #[must_use] |
| 77 | pub(crate) fn fleet_search_roots(workspace: &std::path::Path) -> Vec<FleetSearchRoot> { |
| 78 | let mut roots = Vec::new(); |
| 79 | if let Ok(home) = personal_fleet_root() { |
| 80 | roots.push(FleetSearchRoot::new("codewhale_home", home)); |
| 81 | } |
| 82 | roots.push(FleetSearchRoot::new( |
| 83 | "workspace", |
| 84 | workspace.join(".codewhale"), |
| 85 | )); |
| 86 | roots.push(FleetSearchRoot::new( |
| 87 | "workspace_root", |
| 88 | workspace.to_path_buf(), |
| 89 | )); |
| 90 | roots |
| 91 | } |
| 92 | |
| 93 | /// Load a Fleet document by (optionally qualified) name from the standard |
| 94 | /// roots. Ambiguity between origins is surfaced, never resolved by shadowing. |
| 95 | /// |
| 96 | /// Saved v2 Fleets (`schema = "fleet"`, from `.codewhale/fleets/` or |
| 97 | /// `$CODEWHALE_HOME/fleets/`) are looked up first and frozen into an exact |
| 98 | /// snapshot here — see [`freeze_saved_fleet`]. A miss falls through to the |
| 99 | /// workflow crate's legacy/exact loader. A bare name that exists both as a v2 |
| 100 | /// Fleet and as a legacy/exact file is ambiguous: neither shadows the other, |
| 101 | /// and the error names every path. v2 Fleets qualify as `user/<name>` and |
| 102 | /// `folder/<name>` (the store's own scope labels); a search-root origin |
| 103 | /// (`codewhale_home/`, `workspace/`, `workspace_root/`) reads that root's file |
| 104 | /// in whichever form it is. |
| 105 | /// |
| 106 | /// `config` is the session config the caller preflights with: inheriting |
| 107 | /// members resolve against it at this point, immediately before the same |
| 108 | /// config preflights the frozen routes, so a receipt names the route that ran. |
| 109 | /// |
| 110 | /// Synchronous file loading: async callers must run this on the blocking pool. |
| 111 | pub(crate) fn load_fleet_document( |
| 112 | name: &str, |
| 113 | workspace: &std::path::Path, |
| 114 | config: Option<&Config>, |
| 115 | ) -> Result<(FleetDocument, QualifiedFleetId), NamedFleetError> { |
| 116 | use super::store::{self, FleetScope}; |
| 117 | |
| 118 | let roots = fleet_search_roots(workspace); |
| 119 | // Validates the bare name before any path below is built from it. |
| 120 | let (origin, bare) = codewhale_workflow::split_qualified_fleet_name(name)?; |
| 121 | let store_error = |error: store::FleetStoreError| match error { |
| 122 | store::FleetStoreError::NotFound(what) => NamedFleetError::NotFound(what), |
| 123 | store::FleetStoreError::Io { path, message } => NamedFleetError::Io { path, message }, |
| 124 | store::FleetStoreError::Parse { path, message } => NamedFleetError::Parse { path, message }, |
| 125 | other => NamedFleetError::Parse { |
| 126 | path: bare.to_string(), |
| 127 | message: other.to_string(), |
| 128 | }, |
| 129 | }; |
| 130 | let v2_scope = match origin.map(str::to_ascii_lowercase).as_deref() { |
| 131 | None => None, |
| 132 | Some("user" | "personal") => Some(FleetScope::Personal), |
| 133 | Some("folder") => Some(FleetScope::Workspace), |
| 134 | // Any other origin names a legacy/exact search root. A saved v2 |
| 135 | // Fleet can live there too (the personal `fleets/` directory is |
| 136 | // shared), so the qualified file is read in whichever form it is. |
| 137 | Some(origin) => { |
| 138 | let saved = roots |
| 139 | .iter() |
| 140 | .find(|root| root.origin.eq_ignore_ascii_case(origin)) |
| 141 | .map(|root| { |
| 142 | root.root |
| 143 | .join(store::FLEET_DIR) |
| 144 | .join(format!("{bare}.toml")) |
| 145 | }) |
| 146 | .filter(|path| store::declares_v2_schema(path)); |
| 147 | let Some(path) = saved else { |
| 148 | return FleetDocument::load_by_name(name, &roots); |
| 149 | }; |
| 150 | let (fleet, scope) = store::load_fleet_at(&path).map_err(store_error)?; |
| 151 | return freeze_saved_fleet(&fleet, scope, &path, config); |
| 152 | } |
| 153 | }; |
| 154 | |
| 155 | if let Some(scope) = v2_scope { |
| 156 | let (fleet, path) = |
| 157 | store::load_fleet_in_scope(bare, scope, workspace).map_err(store_error)?; |
| 158 | return freeze_saved_fleet(&fleet, scope, &path, config); |
| 159 | } |
| 160 | |
| 161 | // Legacy/exact files under the same bare name, in any root. A v2 file in |
| 162 | // the shared personal directory is the store's, not a second Fleet. |
| 163 | let file_name = format!("{bare}.toml"); |
| 164 | let other_forms: Vec<String> = roots |
| 165 | .iter() |
| 166 | .filter_map(|root| { |
| 167 | let path = root.root.join(store::FLEET_DIR).join(&file_name); |
| 168 | let schema = store::read_declared_schema(&path)?; |
| 169 | (schema.as_deref() != Some(store::FLEET_SCHEMA_KIND)) |
| 170 | .then(|| format!("{}/{bare} ({})", root.origin, path.display())) |
| 171 | }) |
| 172 | .collect(); |
| 173 | |
| 174 | let v2_candidates = store::v2_fleet_candidates(bare, workspace); |
| 175 | let v2_labels = || { |
| 176 | v2_candidates |
| 177 | .iter() |
| 178 | .map(|(scope, path)| format!("{}/{bare} ({})", scope.label(), path.display())) |
| 179 | }; |
| 180 | if v2_candidates.len() > 1 || (!v2_candidates.is_empty() && !other_forms.is_empty()) { |
| 181 | return Err(NamedFleetError::AmbiguousFleet { |
| 182 | name: bare.to_string(), |
| 183 | origins: v2_labels().chain(other_forms).collect(), |
| 184 | }); |
| 185 | } |
| 186 | match store::load_fleet(bare, workspace) { |
| 187 | Ok((fleet, scope, path)) => freeze_saved_fleet(&fleet, scope, &path, config), |
| 188 | Err(store::FleetStoreError::NotFound(_)) => FleetDocument::load_by_name(name, &roots), |
| 189 | Err(error) => Err(store_error(error)), |
| 190 | } |
| 191 | } |
| 192 | |
| 193 | /// Freeze a saved v2 Fleet into an exact snapshot document. |
| 194 | /// |
| 195 | /// Every executable member leaves here with one concrete provider/model and |
| 196 | /// one concrete reasoning request: an explicit member pin wins, then the |
| 197 | /// Fleet's operator route, then the live session route from `config`. The |
| 198 | /// result is rendered as an exact document and parsed by the workflow crate's |
| 199 | /// own exact parser, so it passes the same validation as a hand-written exact |
| 200 | /// file, and the snapshot hash covers what was frozen. Editing the v2 file |
| 201 | /// afterwards changes only the next Workflow. |
| 202 | /// |
| 203 | /// Member `instructions` and `requires` are refused rather than dropped: the |
| 204 | /// exact snapshot has no field for either, and a Workflow that silently ran a |
| 205 | /// member without its instructions or capability requirement would not be the |
| 206 | /// saved Fleet. |
| 207 | fn freeze_saved_fleet( |
| 208 | fleet: &super::store::FleetFile, |
| 209 | scope: super::store::FleetScope, |
| 210 | path: &std::path::Path, |
| 211 | config: Option<&Config>, |
| 212 | ) -> Result<(FleetDocument, QualifiedFleetId), NamedFleetError> { |
| 213 | #[derive(serde::Serialize)] |
| 214 | struct FrozenFleet { |
| 215 | schema: &'static str, |
| 216 | schema_revision: u32, |
| 217 | name: String, |
| 218 | #[serde(skip_serializing_if = "Option::is_none")] |
| 219 | description: Option<String>, |
| 220 | members: Vec<FrozenMember>, |
| 221 | } |
| 222 | #[derive(serde::Serialize)] |
| 223 | struct FrozenMember { |
| 224 | id: String, |
| 225 | role: String, |
| 226 | provider: String, |
| 227 | model: String, |
| 228 | reasoning: String, |
| 229 | } |
| 230 | |
| 231 | let slug = fleet.file_slug(); |
| 232 | let fail = |message: String| NamedFleetError::Parse { |
| 233 | path: path.display().to_string(), |
| 234 | message, |
| 235 | }; |
| 236 | let operator = fleet.operator.as_ref(); |
| 237 | // A saved Fleet stores reasoning in the session vocabulary (`xhigh`, |
| 238 | // `ultra`, `minimal`, ... — what an imported agent profile carries); the |
| 239 | // exact schema names tiers. Map through the same effort-to-tier table the |
| 240 | // preflight uses, keep an explicit `auto` as a Router request, and treat a |
| 241 | // blank value as absent (inherit), as the selected-Fleet path does. |
| 242 | let frozen_reasoning = |raw: Option<&str>| -> Result<Option<String>, String> { |
| 243 | let Some(value) = raw.map(str::trim).filter(|value| !value.is_empty()) else { |
| 244 | return Ok(None); |
| 245 | }; |
| 246 | let effort = |
| 247 | ReasoningEffort::parse_strict(value).map_err(|error| format!("reasoning: {error}"))?; |
| 248 | Ok(Some( |
| 249 | tier_of(effort) |
| 250 | .map_or("auto", ReasoningTier::as_str) |
| 251 | .to_string(), |
| 252 | )) |
| 253 | }; |
| 254 | let session_route = config |
| 255 | .map(|config| { |
| 256 | config |
| 257 | .active_provider_identity() |
| 258 | .map(|identity| (identity.key.to_string(), config.default_model())) |
| 259 | }) |
| 260 | .transpose() |
| 261 | .map_err(fail)?; |
| 262 | let session_reasoning = || { |
| 263 | let effort = config |
| 264 | .and_then(Config::reasoning_effort) |
| 265 | .map(ReasoningEffort::from_setting) |
| 266 | .unwrap_or_default(); |
| 267 | // A session-level `auto` is per-turn adaptivity, not a Router |
| 268 | // request; a frozen member takes the concrete default tier instead. |
| 269 | tier_of(effort) |
| 270 | .unwrap_or(ReasoningTier::Max) |
| 271 | .as_str() |
| 272 | .to_string() |
| 273 | }; |
| 274 | |
| 275 | let mut unsupported = Vec::new(); |
| 276 | let mut members = Vec::new(); |
| 277 | for member in fleet.members.iter().filter(|member| !member.shortlist) { |
| 278 | let id = member.id.trim().to_string(); |
| 279 | if member |
| 280 | .instructions |
| 281 | .as_deref() |
| 282 | .is_some_and(|text| !text.trim().is_empty()) |
| 283 | { |
| 284 | unsupported.push(format!("`{id}` has instructions")); |
| 285 | } |
| 286 | if !member.requires.is_empty() { |
| 287 | unsupported.push(format!("`{id}` has requires")); |
| 288 | } |
| 289 | let (provider, model) = match (&member.provider, &member.model, operator, &session_route) { |
| 290 | (Some(provider), Some(model), _, _) => (provider.clone(), model.clone()), |
| 291 | (None, None, Some(operator), _) => (operator.provider.clone(), operator.model.clone()), |
| 292 | (None, None, None, Some((provider, model))) => (provider.clone(), model.clone()), |
| 293 | (None, None, None, None) => { |
| 294 | return Err(fail(format!( |
| 295 | "member `{id}` inherits the session route, but no session config is \ |
| 296 | available to resolve it" |
| 297 | ))); |
| 298 | } |
| 299 | _ => { |
| 300 | return Err(fail(format!( |
| 301 | "member `{id}` has a partial provider/model pin" |
| 302 | ))); |
| 303 | } |
| 304 | }; |
| 305 | let reasoning = match frozen_reasoning(member.reasoning.as_deref()) |
| 306 | .map_err(|error| fail(format!("member `{id}` {error}")))? |
| 307 | { |
| 308 | Some(tier) => tier, |
| 309 | None => frozen_reasoning(operator.and_then(|operator| operator.reasoning.as_deref())) |
| 310 | .map_err(|error| fail(format!("operator {error}")))? |
| 311 | .unwrap_or_else(session_reasoning), |
| 312 | }; |
| 313 | members.push(FrozenMember { |
| 314 | role: member.role_label().to_string(), |
| 315 | id, |
| 316 | provider, |
| 317 | model, |
| 318 | reasoning, |
| 319 | }); |
| 320 | } |
| 321 | if !unsupported.is_empty() { |
| 322 | return Err(fail(format!( |
| 323 | "saved Fleet `{}` cannot run as a Workflow Fleet yet: {}. Workflow snapshots freeze \ |
| 324 | each member's route and reasoning only; remove those fields or run the members \ |
| 325 | with `agent`.", |
| 326 | fleet.name, |
| 327 | unsupported.join(", ") |
| 328 | ))); |
| 329 | } |
| 330 | |
| 331 | let frozen = FrozenFleet { |
| 332 | schema: codewhale_workflow::EXACT_FLEET_SCHEMA_KIND, |
| 333 | schema_revision: codewhale_workflow::EXACT_FLEET_SCHEMA_REVISION, |
| 334 | name: slug.clone(), |
| 335 | description: fleet.description.clone(), |
| 336 | members, |
| 337 | }; |
| 338 | let text = toml::to_string(&frozen) |
| 339 | .map_err(|error| fail(format!("failed to freeze saved Fleet: {error}")))?; |
| 340 | let document = FleetDocument::from_frozen_saved_fleet(&text, path).map_err(|error| { |
| 341 | fail(format!( |
| 342 | "saved Fleet `{}` cannot run as a Workflow Fleet: {error}", |
| 343 | fleet.name |
| 344 | )) |
| 345 | })?; |
| 346 | Ok(( |
| 347 | document, |
| 348 | QualifiedFleetId { |
| 349 | name: slug, |
| 350 | origin: scope.label().to_string(), |
| 351 | }, |
| 352 | )) |
| 353 | } |
| 354 | |
| 355 | // ── Preflight: freeze the route, and check it while freezing ───────────────── |
| 356 | |
| 357 | /// Derive a route's real reasoning capability from the request shaping the |
| 358 | /// client actually performs, rather than from a hand-maintained claims table. |
| 359 | /// |
| 360 | /// The probe builds the request body this exact route would receive for every |
| 361 | /// tier and compares them. Two tiers that produce a byte-identical body are not |
| 362 | /// two provider-effective tiers, whatever the selector calls them — this is why |
| 363 | /// Z.AI's GLM routes come back as |
| 364 | /// [`ProviderReasoningControl::EnabledDisabled`] and why nothing here can claim |
| 365 | /// provider-native adaptive for a route whose body does not say so. |
| 366 | #[must_use] |
| 367 | pub(crate) fn reasoning_capability_for_route( |
| 368 | provider: ProviderKind, |
| 369 | base_url: &str, |
| 370 | wire_model: &str, |
| 371 | ) -> ReasoningCapability { |
| 372 | let body_for = |effort: ReasoningEffort| -> String { |
| 373 | let mut body = serde_json::json!({}); |
| 374 | let value = effort.api_value_for_route(provider, base_url, wire_model); |
| 375 | crate::client::apply_reasoning_effort(&mut body, value, provider); |
| 376 | // `reasoning_split` is a transport concern the client sets for every |
| 377 | // tier; it carries no reasoning depth, so it must not make tiers look |
| 378 | // distinct or make a no-control route look controllable. |
| 379 | if let Some(object) = body.as_object_mut() { |
| 380 | object.remove("reasoning_split"); |
| 381 | } |
| 382 | body.to_string() |
| 383 | }; |
| 384 | |
| 385 | let off = body_for(ReasoningEffort::Off); |
| 386 | let above_off: Vec<String> = [ |
| 387 | ReasoningEffort::Low, |
| 388 | ReasoningEffort::Medium, |
| 389 | ReasoningEffort::High, |
| 390 | ReasoningEffort::Max, |
| 391 | ] |
| 392 | .into_iter() |
| 393 | .map(body_for) |
| 394 | .collect(); |
| 395 | |
| 396 | let empty = "{}"; |
| 397 | let all_empty = off == empty && above_off.iter().all(|body| body == empty); |
| 398 | |
| 399 | let mut distinct = above_off.clone(); |
| 400 | distinct.sort(); |
| 401 | distinct.dedup(); |
| 402 | |
| 403 | let control = if all_empty { |
| 404 | ProviderReasoningControl::None |
| 405 | } else if distinct.len() == 1 && distinct[0] == off && off.contains("adaptive") { |
| 406 | // Every tier — including off — produces the same adaptive body: the |
| 407 | // provider genuinely chooses its own depth. Source-backed, not assumed. |
| 408 | ProviderReasoningControl::NativeAdaptive |
| 409 | } else if distinct.len() > 1 { |
| 410 | ProviderReasoningControl::Tiers |
| 411 | } else { |
| 412 | ProviderReasoningControl::EnabledDisabled |
| 413 | }; |
| 414 | |
| 415 | // What each requested tier actually becomes on the wire, straight from the |
| 416 | // route normalizer that shapes the real request. |
| 417 | // |
| 418 | // This subsumes a min/max floor-and-ceiling and expresses what one cannot: |
| 419 | // most non-Codex routes coerce `low` and `medium` to `high` while leaving |
| 420 | // `off` alone (first-party DeepSeek routes are the documented exception — |
| 421 | // their wire carries a real `low`), and an always-thinking route raises |
| 422 | // `off` instead. Reporting a `low` a route silently sends as `high` is |
| 423 | // the invisible substitution receipts exist to prevent, so the map — not |
| 424 | // a clamp — is the authority. |
| 425 | let wire_tiers = [ |
| 426 | ReasoningEffort::Off, |
| 427 | ReasoningEffort::Low, |
| 428 | ReasoningEffort::Medium, |
| 429 | ReasoningEffort::High, |
| 430 | ReasoningEffort::Max, |
| 431 | ] |
| 432 | .map(|effort| { |
| 433 | tier_of(effort.normalize_for_route(provider, base_url, wire_model)) |
| 434 | .unwrap_or(ReasoningTier::Off) |
| 435 | }); |
| 436 | |
| 437 | ReasoningCapability { |
| 438 | control, |
| 439 | min_tier: None, |
| 440 | max_tier: None, |
| 441 | wire_tiers: None, |
| 442 | } |
| 443 | .with_wire_tiers(wire_tiers) |
| 444 | } |
| 445 | |
| 446 | fn tier_of(effort: ReasoningEffort) -> Option<ReasoningTier> { |
| 447 | match effort { |
| 448 | ReasoningEffort::Off => Some(ReasoningTier::Off), |
| 449 | ReasoningEffort::Minimal => Some(ReasoningTier::Low), |
| 450 | ReasoningEffort::Low => Some(ReasoningTier::Low), |
| 451 | ReasoningEffort::Medium => Some(ReasoningTier::Medium), |
| 452 | ReasoningEffort::High => Some(ReasoningTier::High), |
| 453 | ReasoningEffort::XHigh => Some(ReasoningTier::Max), |
| 454 | ReasoningEffort::Ultra => Some(ReasoningTier::Max), |
| 455 | ReasoningEffort::Max => Some(ReasoningTier::Max), |
| 456 | ReasoningEffort::Auto => None, |
| 457 | } |
| 458 | } |
| 459 | |
| 460 | /// The **provider-facing** reasoning value for one tier on one exact route. |
| 461 | /// |
| 462 | /// A tier label (`off`, `max`) is a selector concept; what a request may carry |
| 463 | /// is a provider concept, and the two are not the same string. OpenAI Codex |
| 464 | /// routes spell the top tier `xhigh` and cannot express `off` at all, so |
| 465 | /// placing a bare tier label on a Codex request either sends a value the |
| 466 | /// provider does not accept or silently sends nothing and takes the provider |
| 467 | /// default while the receipt claims the tier. Reading the value back out of the |
| 468 | /// same route normalizer the client uses is what keeps the request and the |
| 469 | /// receipt describing each other. |
| 470 | #[must_use] |
| 471 | pub(crate) fn route_reasoning_setting( |
| 472 | provider: ProviderKind, |
| 473 | base_url: &str, |
| 474 | wire_model: &str, |
| 475 | tier: ReasoningTier, |
| 476 | ) -> String { |
| 477 | effort_of(tier) |
| 478 | .as_setting_for_route(provider, base_url, wire_model) |
| 479 | .to_string() |
| 480 | } |
| 481 | |
| 482 | fn effort_of(tier: ReasoningTier) -> ReasoningEffort { |
| 483 | match tier { |
| 484 | ReasoningTier::Off => ReasoningEffort::Off, |
| 485 | ReasoningTier::Low => ReasoningEffort::Low, |
| 486 | ReasoningTier::Medium => ReasoningEffort::Medium, |
| 487 | ReasoningTier::High => ReasoningEffort::High, |
| 488 | ReasoningTier::Max => ReasoningEffort::Max, |
| 489 | } |
| 490 | } |
| 491 | |
| 492 | /// Preflight one exact route: resolve the provider, canonicalize the model, |
| 493 | /// identify the endpoint, decide credential readiness **from local config**, |
| 494 | /// and derive the reasoning capability. |
| 495 | /// |
| 496 | /// No provider is contacted. Everything here is a configuration lookup, which |
| 497 | /// is what makes it safe to run before the operator's gates have fired. |
| 498 | pub(crate) fn preflight_route( |
| 499 | member_id: &str, |
| 500 | provider: &str, |
| 501 | model: &str, |
| 502 | config: &Config, |
| 503 | ) -> Result<PreflightedRoute, PreflightError> { |
| 504 | let identity = config |
| 505 | .resolve_provider_identity(provider.trim()) |
| 506 | .map_err(|detail| PreflightError::ProviderUnresolved { |
| 507 | member: member_id.to_string(), |
| 508 | provider: provider.to_string(), |
| 509 | detail, |
| 510 | })?; |
| 511 | |
| 512 | // The canonical wire model, resolved once. The receipt and the child spawn |
| 513 | // both read this value, so they cannot disagree about what actually ran. |
| 514 | let wire_model = crate::config::requested_model_for_provider(identity.provider, model.trim()) |
| 515 | .ok_or_else(|| PreflightError::ModelUnresolved { |
| 516 | member: member_id.to_string(), |
| 517 | provider: identity.key.to_string(), |
| 518 | model: model.to_string(), |
| 519 | detail: "not a known model for this provider".to_string(), |
| 520 | })?; |
| 521 | crate::config::validate_route(identity.provider, &wire_model).map_err(|detail| { |
| 522 | PreflightError::ModelUnresolved { |
| 523 | member: member_id.to_string(), |
| 524 | provider: identity.key.to_string(), |
| 525 | model: wire_model.clone(), |
| 526 | detail, |
| 527 | } |
| 528 | })?; |
| 529 | |
| 530 | let mut scoped = config.clone(); |
| 531 | scoped |
| 532 | .scope_to_provider_identity(&identity) |
| 533 | .map_err(|detail| PreflightError::ProviderUnresolved { |
| 534 | member: member_id.to_string(), |
| 535 | provider: provider.to_string(), |
| 536 | detail, |
| 537 | })?; |
| 538 | let base_url = scoped.active_route_base_url(); |
| 539 | |
| 540 | // Locally decided. A concrete loopback/self-hosted route is keyless by |
| 541 | // design, and that is a valid, first-class state — not a downgrade and |
| 542 | // not a missing credential. Ollama Cloud is hosted and falls through to |
| 543 | // the ordinary credential checks. |
| 544 | let credential = |
| 545 | if crate::config::provider_route_is_keyless_self_hosted(identity.provider, &base_url) { |
| 546 | CredentialReadiness::KeylessLocal |
| 547 | } else if crate::config::has_api_key_for(&scoped, &identity) { |
| 548 | CredentialReadiness::Configured |
| 549 | } else { |
| 550 | // The discriminant only. `Missing { detail }` names the provider table |
| 551 | // key, which for a custom route is the customer's own string. |
| 552 | codewhale_telemetry::session_counters() |
| 553 | .bump_error(codewhale_telemetry::ErrorCounter::AuthPreflightFailed); |
| 554 | CredentialReadiness::Missing { |
| 555 | detail: format!("no credential configured for `{}`", identity.key), |
| 556 | } |
| 557 | }; |
| 558 | |
| 559 | Ok(PreflightedRoute { |
| 560 | member_id: member_id.to_string(), |
| 561 | provider_id: identity.key.to_string(), |
| 562 | provider_config_id: identity |
| 563 | .migrated_legacy_ollama_cloud_route |
| 564 | .then(|| provider.trim().to_string()), |
| 565 | provider_kind: if identity.provider == ProviderKind::OllamaCloud { |
| 566 | identity.provider.as_str().to_string() |
| 567 | } else { |
| 568 | format!("{:?}", identity.provider).to_ascii_lowercase() |
| 569 | }, |
| 570 | declared_model: model.trim().to_string(), |
| 571 | wire_model: wire_model.clone(), |
| 572 | endpoint: EndpointIdentity::from_base_url(&base_url), |
| 573 | credential, |
| 574 | capability: reasoning_capability_for_route(identity.provider, &base_url, &wire_model), |
| 575 | }) |
| 576 | } |
| 577 | |
| 578 | /// Build the client one worker route would actually run on, and throw it away. |
| 579 | /// |
| 580 | /// Preflight resolves a route from *configuration*; this proves the same route |
| 581 | /// can be turned into a working client — the step that fails on a malformed |
| 582 | /// base URL, an unusable auth mode, or a transport CodeWhale cannot construct. |
| 583 | /// Doing it at Workflow start, for every member, is what stops a Fleet from |
| 584 | /// paying for a Router decision and only then discovering that the worker it |
| 585 | /// decided for could never have been launched. |
| 586 | /// |
| 587 | /// The client is deliberately not retained: the spawn path builds the child's |
| 588 | /// own client from the member's roster profile, and keeping a second one here |
| 589 | /// would create two objects that could drift apart. |
| 590 | fn validate_route_client(route: &PreflightedRoute, config: &Config) -> Result<(), String> { |
| 591 | let mut scoped = config.clone(); |
| 592 | let identity = config.resolve_provider_identity(route.provider_config_id())?; |
| 593 | scoped.scope_to_provider_identity(&identity)?; |
| 594 | crate::client::CodewhaleClient::new(&scoped) |
| 595 | .map(|_| ()) |
| 596 | .map_err(|error| { |
| 597 | format!( |
| 598 | "member `{}` is pinned to provider `{}` (model `{}`), whose client could not be \ |
| 599 | built on this machine: {error}", |
| 600 | route.member_id, route.provider_id, route.wire_model |
| 601 | ) |
| 602 | }) |
| 603 | } |
| 604 | |
| 605 | // ── The Reasoning Router, as a service ────────────────────────────────────── |
| 606 | |
| 607 | /// The seam a Reasoning Router call goes through. Implemented live against the |
| 608 | /// provider client, and by a fixture in tests so the whole reasoning path is |
| 609 | /// exercised without a network. |
| 610 | #[async_trait] |
| 611 | pub(crate) trait FleetRouterCaller: Send + Sync + std::fmt::Debug { |
| 612 | /// Return the router's raw text response for one worker task. |
| 613 | async fn decide(&self, input: &RouterCallInput) -> Result<String, String>; |
| 614 | |
| 615 | /// The Router service's exact identity, for the receipt. |
| 616 | fn identity(&self) -> RouterIdentity; |
| 617 | } |
| 618 | |
| 619 | /// A Reasoning Router bound to its own exact preflighted route. |
| 620 | #[derive(Clone)] |
| 621 | pub(crate) struct LiveFleetRouter { |
| 622 | client: crate::client::CodewhaleClient, |
| 623 | captured: CapturedReasoningRouter, |
| 624 | route: PreflightedRoute, |
| 625 | /// The Router route's provider kind and base URL, kept so the call's |
| 626 | /// reasoning value can be shaped by the *actual* configured route rather |
| 627 | /// than by a generic tier label. Never serialized — the base URL can carry |
| 628 | /// a credential and receipts are durable. |
| 629 | provider: ProviderKind, |
| 630 | base_url: String, |
| 631 | /// What the Router call is actually made at, plus the four-sided disclosure |
| 632 | /// for the receipt. Configured by the operator (`off` or `low`), normalized |
| 633 | /// only against what the Router's own route can express. |
| 634 | call: RouterCallPlan, |
| 635 | } |
| 636 | |
| 637 | impl std::fmt::Debug for LiveFleetRouter { |
| 638 | fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 639 | f.debug_struct("LiveFleetRouter") |
| 640 | .field("router", &self.captured.qualified()) |
| 641 | .field("provider", &self.route.provider_id) |
| 642 | .field("model", &self.route.wire_model) |
| 643 | .field("call_reasoning", &self.call.tier) |
| 644 | .field("client", &"<redacted>") |
| 645 | .finish() |
| 646 | } |
| 647 | } |
| 648 | |
| 649 | impl LiveFleetRouter { |
| 650 | /// Resolve the Router service's exact configured route and build its client. |
| 651 | /// |
| 652 | /// A Router that cannot be resolved is an error here — at Workflow start, |
| 653 | /// before any worker is dispatched — not a silent downgrade to legacy |
| 654 | /// routing. Readiness is decided from local configuration; no live probe. |
| 655 | pub(crate) fn bind( |
| 656 | captured: &CapturedReasoningRouter, |
| 657 | config: &Config, |
| 658 | ) -> Result<Self, RouterBindError> { |
| 659 | let route = preflight_route( |
| 660 | &captured.id, |
| 661 | &captured.route.provider, |
| 662 | &captured.route.model, |
| 663 | config, |
| 664 | ) |
| 665 | .map_err(|error| RouterBindError { |
| 666 | reason: error.to_string(), |
| 667 | })?; |
| 668 | route.require_ready().map_err(|error| RouterBindError { |
| 669 | reason: error.to_string(), |
| 670 | })?; |
| 671 | |
| 672 | let identity = config |
| 673 | .resolve_provider_identity(route.provider_config_id()) |
| 674 | .map_err(|detail| RouterBindError { |
| 675 | reason: format!( |
| 676 | "reasoning router provider `{}` did not resolve: {detail}", |
| 677 | route.provider_id |
| 678 | ), |
| 679 | })?; |
| 680 | let mut scoped = config.clone(); |
| 681 | scoped |
| 682 | .scope_to_provider_identity(&identity) |
| 683 | .map_err(|reason| RouterBindError { reason })?; |
| 684 | let base_url = scoped.active_route_base_url(); |
| 685 | let client = |
| 686 | crate::client::CodewhaleClient::new(&scoped).map_err(|error| RouterBindError { |
| 687 | reason: format!( |
| 688 | "reasoning router provider `{}` client could not be built: {error}", |
| 689 | route.provider_id |
| 690 | ), |
| 691 | })?; |
| 692 | |
| 693 | let call = router_call_plan(captured.requested_call_reasoning, &route.capability); |
| 694 | |
| 695 | Ok(Self { |
| 696 | client, |
| 697 | captured: captured.clone(), |
| 698 | route, |
| 699 | provider: identity.provider, |
| 700 | base_url, |
| 701 | call, |
| 702 | }) |
| 703 | } |
| 704 | |
| 705 | /// The preflighted Router route, for cross-provider disclosure. |
| 706 | #[must_use] |
| 707 | pub(crate) fn route(&self) -> &PreflightedRoute { |
| 708 | &self.route |
| 709 | } |
| 710 | } |
| 711 | |
| 712 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 713 | pub(crate) struct RouterBindError { |
| 714 | pub(crate) reason: String, |
| 715 | } |
| 716 | |
| 717 | #[async_trait] |
| 718 | impl FleetRouterCaller for LiveFleetRouter { |
| 719 | fn identity(&self) -> RouterIdentity { |
| 720 | RouterIdentity::from_captured( |
| 721 | &self.captured, |
| 722 | Some(&self.route), |
| 723 | Some(self.call.disclosure.clone()), |
| 724 | ) |
| 725 | } |
| 726 | |
| 727 | async fn decide(&self, input: &RouterCallInput) -> Result<String, String> { |
| 728 | use codewhale_models::{ContentBlock, Message, MessageRequest, SystemPrompt}; |
| 729 | |
| 730 | // The bounded, redacted summary is transmitted exactly once, in the |
| 731 | // user turn. The system prompt carries the contract and the frozen |
| 732 | // route, and no task content at all — sending it twice would double |
| 733 | // what leaves for this provider while the receipt counted one copy. |
| 734 | let request = MessageRequest { |
| 735 | model: self.route.wire_model.clone(), |
| 736 | messages: vec![Message { |
| 737 | role: Role::User, |
| 738 | content: vec![ContentBlock::Text { |
| 739 | text: router_user_message(input), |
| 740 | cache_control: None, |
| 741 | }], |
| 742 | }], |
| 743 | max_tokens: self |
| 744 | .client |
| 745 | .effective_max_output_tokens(&self.route.wire_model), |
| 746 | system: Some(SystemPrompt::Text(router_system_prompt(input))), |
| 747 | // A router receives no tools. Ever. |
| 748 | tools: None, |
| 749 | tool_choice: None, |
| 750 | metadata: None, |
| 751 | thinking: None, |
| 752 | // The operator-configured call tier remains authoritative. The |
| 753 | // normal route allowance above leaves room for its hidden |
| 754 | // reasoning before the small JSON answer is emitted. |
| 755 | reasoning_effort: Some(route_reasoning_setting( |
| 756 | self.provider, |
| 757 | &self.base_url, |
| 758 | &self.route.wire_model, |
| 759 | self.call.tier, |
| 760 | )), |
| 761 | stream: Some(false), |
| 762 | temperature: None, |
| 763 | top_p: None, |
| 764 | }; |
| 765 | |
| 766 | let response = self |
| 767 | .client |
| 768 | .create_message(request) |
| 769 | .await |
| 770 | .map_err(|error| error.to_string())?; |
| 771 | if codewhale_models::is_incomplete_stop_reason(response.stop_reason.as_deref()) { |
| 772 | return Err(format!( |
| 773 | "reasoning router response incomplete: provider stop reason `{}`", |
| 774 | codewhale_models::stop_reason_detail(response.stop_reason.as_deref()) |
| 775 | )); |
| 776 | } |
| 777 | let text = response |
| 778 | .content |
| 779 | .into_iter() |
| 780 | .filter_map(|block| match block { |
| 781 | ContentBlock::Text { text, .. } => Some(text), |
| 782 | _ => None, |
| 783 | }) |
| 784 | .collect::<Vec<_>>() |
| 785 | .join(""); |
| 786 | if text.trim().is_empty() { |
| 787 | return Err("reasoning router returned an empty response".to_string()); |
| 788 | } |
| 789 | Ok(text) |
| 790 | } |
| 791 | } |
| 792 | |
| 793 | // ── The Workflow ─────────────────────────────────────────────────────────── |
| 794 | |
| 795 | /// An exact Fleet, frozen at Workflow start. |
| 796 | /// |
| 797 | /// The snapshot and the preflight are immutable for the life of the run: |
| 798 | /// editing `fleets/<name>.toml` afterwards changes only the next Workflow. |
| 799 | /// Durable runs and in-process spawns bind the same frozen member. The |
| 800 | /// in-process path projects only that member onto its existing profile binder; |
| 801 | /// it does not read or replace the currently selected Fleet. |
| 802 | #[derive(Clone)] |
| 803 | pub(crate) struct ExactFleetWorkflow { |
| 804 | snapshot: Arc<FleetSnapshot>, |
| 805 | preflight: Arc<RoutePreflight>, |
| 806 | router: Option<Arc<dyn FleetRouterCaller>>, |
| 807 | router_unavailable: Option<String>, |
| 808 | } |
| 809 | |
| 810 | impl std::fmt::Debug for ExactFleetWorkflow { |
| 811 | fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 812 | f.debug_struct("ExactFleetWorkflow") |
| 813 | .field("fleet", &self.snapshot.fleet().qualified()) |
| 814 | .field("members", &self.snapshot.members().len()) |
| 815 | .field("router", &self.router.is_some()) |
| 816 | .finish() |
| 817 | } |
| 818 | } |
| 819 | |
| 820 | /// Whether `tool` is removed by a deny entry (exact, case-insensitive, or a |
| 821 | /// trailing-`*` prefix glob). |
| 822 | fn tool_denied_by(tool: &str, denied: &[String]) -> bool { |
| 823 | denied.iter().any(|entry| match entry.strip_suffix('*') { |
| 824 | Some(prefix) => tool |
| 825 | .to_ascii_lowercase() |
| 826 | .starts_with(&prefix.to_ascii_lowercase()), |
| 827 | None => entry.eq_ignore_ascii_case(tool), |
| 828 | }) |
| 829 | } |
| 830 | |
| 831 | /// Refuse a task whose explicitly requested tools do not all survive the |
| 832 | /// role ceiling and deny list (SHA-6734). The error names the requested |
| 833 | /// tools, what the role allows, and what was dropped, so the caller can fix |
| 834 | /// the request instead of launching a child that flounders without tools. |
| 835 | fn refuse_dropped_requested_tools( |
| 836 | member_id: &str, |
| 837 | member_role: &str, |
| 838 | requested: &[String], |
| 839 | ceiling: &ChildAuthority, |
| 840 | authority: &ChildAuthority, |
| 841 | ) -> Result<(), String> { |
| 842 | let mut requested: Vec<String> = requested.to_vec(); |
| 843 | requested.sort(); |
| 844 | requested.dedup(); |
| 845 | if requested.is_empty() { |
| 846 | // An explicit empty list is a deliberate tool-free child. |
| 847 | return Ok(()); |
| 848 | } |
| 849 | let survives = |tool: &String| { |
| 850 | authority |
| 851 | .allowed_tools |
| 852 | .as_ref() |
| 853 | .is_none_or(|allowed| allowed.contains(tool)) |
| 854 | && !tool_denied_by(tool, &authority.disallowed_tools) |
| 855 | }; |
| 856 | let dropped: Vec<&String> = requested.iter().filter(|tool| !survives(tool)).collect(); |
| 857 | if dropped.is_empty() { |
| 858 | return Ok(()); |
| 859 | } |
| 860 | let kept = requested.len() - dropped.len(); |
| 861 | let role_allows = match ceiling.allowed_tools.as_ref() { |
| 862 | Some(allowed) if allowed.is_empty() => "no tools".to_string(), |
| 863 | Some(allowed) => format!("[{}]", allowed.join(", ")), |
| 864 | None if ceiling.disallowed_tools.is_empty() => "all inherited tools".to_string(), |
| 865 | None => format!( |
| 866 | "all inherited tools except [{}]", |
| 867 | ceiling.disallowed_tools.join(", ") |
| 868 | ), |
| 869 | }; |
| 870 | let join = |tools: &[&String]| { |
| 871 | tools |
| 872 | .iter() |
| 873 | .map(|tool| tool.as_str()) |
| 874 | .collect::<Vec<_>>() |
| 875 | .join(", ") |
| 876 | }; |
| 877 | let requested_refs: Vec<&String> = requested.iter().collect(); |
| 878 | let headline = if kept == 0 { |
| 879 | format!("Agent '{member_id}' (role {member_role}) would start with no tools") |
| 880 | } else { |
| 881 | format!("Agent '{member_id}' (role {member_role}) would lose requested tools") |
| 882 | }; |
| 883 | Err(format!( |
| 884 | "{headline}: requested [{}], role allows {role_allows}, dropped [{}]. Request only tools \ |
| 885 | the role allows, or pick a member whose role carries them.", |
| 886 | join(&requested_refs), |
| 887 | join(&dropped), |
| 888 | )) |
| 889 | } |
| 890 | |
| 891 | /// One member, resolved and admitted — but **not yet routed**. |
| 892 | /// |
| 893 | /// This is the value the caller holds between admission and the Router call. |
| 894 | /// Producing it costs nothing: no provider is contacted, so a task that is |
| 895 | /// about to be rejected by a gate or blocked on capacity can be resolved |
| 896 | /// safely. |
| 897 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 898 | pub(crate) struct ExactMemberBinding { |
| 899 | /// Canonical member id from the frozen snapshot. |
| 900 | pub(crate) member_id: String, |
| 901 | /// Semantic role — what gates, handoffs, and records use. |
| 902 | pub(crate) member_role: String, |
| 903 | /// The preflighted, frozen route. |
| 904 | pub(crate) route: PreflightedRoute, |
| 905 | /// Whether this member's reasoning comes from the Router. |
| 906 | pub(crate) requires_router: bool, |
| 907 | /// The clamped authority the child will actually run under. |
| 908 | pub(crate) authority: ChildAuthority, |
| 909 | /// The live session posture this binding was clamped against, kept so the |
| 910 | /// launch half can **recompute** the authority instead of trusting the copy |
| 911 | /// it was handed. A binding travels across an await point (gates, a |
| 912 | /// concurrency slot, a router call); recomputing is what makes a stale or |
| 913 | /// tampered authority detectable rather than merely improbable. |
| 914 | pub(crate) session: PermissionCeiling, |
| 915 | /// Source layer captured with the snapshot, never refreshed at spawn. |
| 916 | profile_origin: super::roster::ProfileOrigin, |
| 917 | task_allowed_tools: Option<Vec<String>>, |
| 918 | task_disallowed_tools: Vec<String>, |
| 919 | task_worktree_write: bool, |
| 920 | } |
| 921 | |
| 922 | impl ExactMemberBinding { |
| 923 | /// Task options may narrow the Runtime/parent envelope, never replace the |
| 924 | /// frozen identity or grant authority. Preserve the narrowing for the |
| 925 | /// independent launch-time recomputation and its durable fingerprint. |
| 926 | pub(crate) fn narrow_for_task( |
| 927 | &mut self, |
| 928 | write_authority: Option<&str>, |
| 929 | allowed_tools: Option<&[String]>, |
| 930 | disallowed_tools: &[String], |
| 931 | max_depth: Option<u32>, |
| 932 | ) -> Result<(), String> { |
| 933 | match write_authority { |
| 934 | Some("read_only") => self.session.write = false, |
| 935 | Some("workspace_write" | "worktree_write") if !self.authority.ceiling.write => { |
| 936 | return Err(format!( |
| 937 | "member `{}` is read-only under its Runtime/parent ceiling; a task cannot request write authority via `write_authority`", |
| 938 | self.member_id, |
| 939 | )); |
| 940 | } |
| 941 | Some("workspace_write") if self.task_worktree_write => { |
| 942 | return Err("a task cannot remove its worktree isolation".to_string()); |
| 943 | } |
| 944 | Some("worktree_write") => self.task_worktree_write = true, |
| 945 | Some("workspace_write") | None => {} |
| 946 | Some(other) => return Err(format!("invalid task `write_authority` value `{other}`")), |
| 947 | } |
| 948 | if let Some(depth) = max_depth { |
| 949 | self.session.delegation_depth = self.session.delegation_depth.min(depth); |
| 950 | } |
| 951 | if let Some(tools) = allowed_tools { |
| 952 | let mut tools = tools.to_vec(); |
| 953 | if let Some(previous) = self.task_allowed_tools.as_ref() { |
| 954 | tools.retain(|tool| previous.contains(tool)); |
| 955 | } |
| 956 | tools.sort(); |
| 957 | tools.dedup(); |
| 958 | self.task_allowed_tools = Some(tools); |
| 959 | } |
| 960 | self.task_disallowed_tools |
| 961 | .extend_from_slice(disallowed_tools); |
| 962 | self.task_disallowed_tools.sort(); |
| 963 | self.task_disallowed_tools.dedup(); |
| 964 | let authority = self.recompute_authority(&self.member_role); |
| 965 | // SHA-6734: narrowing an explicit request against the role ceiling |
| 966 | // used to be silent and could start a child with no tools at all. |
| 967 | // Every tool the caller asked for by name must survive, or the spawn |
| 968 | // is refused with what was asked, allowed, and dropped. |
| 969 | if let Some(requested) = allowed_tools { |
| 970 | let ceiling = ChildAuthority::from_runtime_role(&self.member_role, self.session); |
| 971 | refuse_dropped_requested_tools( |
| 972 | &self.member_id, |
| 973 | &self.member_role, |
| 974 | requested, |
| 975 | &ceiling, |
| 976 | &authority, |
| 977 | )?; |
| 978 | } |
| 979 | self.authority = authority; |
| 980 | Ok(()) |
| 981 | } |
| 982 | |
| 983 | fn recompute_authority(&self, role: &str) -> ChildAuthority { |
| 984 | let mut authority = ChildAuthority::from_runtime_role(role, self.session); |
| 985 | if let Some(tools) = self.task_allowed_tools.as_ref() { |
| 986 | let mut tools = tools.clone(); |
| 987 | if let Some(ceiling) = authority.allowed_tools.as_ref() { |
| 988 | tools.retain(|tool| ceiling.contains(tool)); |
| 989 | } |
| 990 | authority.allowed_tools = Some(tools); |
| 991 | } |
| 992 | if !self.task_disallowed_tools.is_empty() { |
| 993 | authority |
| 994 | .disallowed_tools |
| 995 | .extend(self.task_disallowed_tools.iter().cloned()); |
| 996 | authority.disallowed_tools.sort(); |
| 997 | authority.disallowed_tools.dedup(); |
| 998 | } |
| 999 | if authority.ceiling.write && self.task_worktree_write { |
| 1000 | authority.write_authority = "worktree_write"; |
| 1001 | } |
| 1002 | authority |
| 1003 | } |
| 1004 | |
| 1005 | /// Project one preflighted member into the existing profile binder. The |
| 1006 | /// saved provider configuration key stays paired with the canonical wire |
| 1007 | /// model (including compatibility-migrated provider identities). |
| 1008 | pub(crate) fn spawn_profile(&self) -> super::profile::AgentProfile { |
| 1009 | super::profile::AgentProfile { |
| 1010 | native_preset: None, |
| 1011 | id: self.member_id.clone(), |
| 1012 | display_name: None, |
| 1013 | description: None, |
| 1014 | requires: Vec::new(), |
| 1015 | profile: codewhale_config::FleetProfile { |
| 1016 | slot: codewhale_config::FleetSlot::from_name(&self.member_role), |
| 1017 | role: codewhale_config::FleetRole { |
| 1018 | name: self.member_role.clone(), |
| 1019 | ..Default::default() |
| 1020 | }, |
| 1021 | provider: Some(self.route.provider_config_id().to_string()), |
| 1022 | model: Some(self.route.wire_model.clone()), |
| 1023 | ..Default::default() |
| 1024 | }, |
| 1025 | // Snapshot identity is recorded on the Workflow receipt; profile |
| 1026 | // application performs no source-file lookup. |
| 1027 | source: std::path::PathBuf::new(), |
| 1028 | origin: self.profile_origin, |
| 1029 | plugin_authority: None, |
| 1030 | } |
| 1031 | } |
| 1032 | } |
| 1033 | |
| 1034 | /// What a launched exact member resolves to, after routing. |
| 1035 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 1036 | pub(crate) struct ExactMemberLaunch { |
| 1037 | /// Canonical member id; also the roster profile id the spawn resolves. |
| 1038 | pub(crate) member_id: String, |
| 1039 | /// Semantic role, preserved for gates/handoffs/records. |
| 1040 | pub(crate) member_role: String, |
| 1041 | /// Frozen provider id. |
| 1042 | pub(crate) provider: String, |
| 1043 | /// Canonical wire model — the same string the receipt records. |
| 1044 | pub(crate) model: String, |
| 1045 | /// Concrete reasoning setting label for the spawn request. |
| 1046 | pub(crate) thinking: String, |
| 1047 | /// The full requested → effective story, for the receipt. |
| 1048 | pub(crate) reasoning: ResolvedReasoning, |
| 1049 | /// The clamped authority the child runs under. |
| 1050 | pub(crate) authority: ChildAuthority, |
| 1051 | /// The durable, visible receipt for this launch. |
| 1052 | pub(crate) receipt: FleetTaskReceipt, |
| 1053 | } |
| 1054 | |
| 1055 | impl ExactFleetWorkflow { |
| 1056 | /// Capture a Workflow from a parsed exact Fleet document. |
| 1057 | /// |
| 1058 | /// Everything that can fail locally fails here, before any worker is |
| 1059 | /// dispatched: an unresolvable provider, an unknown model, a missing |
| 1060 | /// credential, an unresolvable Reasoning Router profile, or an `auto` |
| 1061 | /// member with no usable Router. |
| 1062 | pub(crate) fn capture( |
| 1063 | document: &FleetDocument, |
| 1064 | id: QualifiedFleetId, |
| 1065 | captured_at: impl Into<String>, |
| 1066 | config: Option<&Config>, |
| 1067 | search_roots: &[FleetSearchRoot], |
| 1068 | ) -> Result<Self, String> { |
| 1069 | let exact = document |
| 1070 | .exact() |
| 1071 | .ok_or_else(|| "this Fleet is not an exact Fleet".to_string())?; |
| 1072 | |
| 1073 | // Resolve the attached Reasoning Router *reference* into the one |
| 1074 | // captured service both forms normalize onto. |
| 1075 | let captured_router = match exact.router_ref() { |
| 1076 | None => None, |
| 1077 | Some(FleetRouterRef::LegacyInline(_)) => captured_legacy_inline_router(exact), |
| 1078 | Some(FleetRouterRef::Profile { name }) => { |
| 1079 | let (profile, router_id) = |
| 1080 | ReasoningRouterProfile::load_by_name(&name, search_roots).map_err(|error| { |
| 1081 | format!( |
| 1082 | "exact Fleet `{}` references reasoning router `{name}`, which could \ |
| 1083 | not be loaded: {error}", |
| 1084 | id.qualified() |
| 1085 | ) |
| 1086 | })?; |
| 1087 | Some(CapturedReasoningRouter::from_profile( |
| 1088 | &profile, |
| 1089 | router_id.origin, |
| 1090 | )) |
| 1091 | } |
| 1092 | }; |
| 1093 | |
| 1094 | // Capture, then immediately verify the hash the receipt will vouch for. |
| 1095 | // `capture` computes it, so this can only fail if the value took a |
| 1096 | // detour through `Deserialize` — but that is exactly the case a receipt |
| 1097 | // must not certify, and checking here means no later caller has to |
| 1098 | // remember to. |
| 1099 | let snapshot = FleetSnapshot::capture(id, document, captured_at, captured_router.clone()) |
| 1100 | .and_then(FleetSnapshot::into_verified) |
| 1101 | .map_err(|error| error.to_string())?; |
| 1102 | |
| 1103 | // Preflight every worker route before anything else can happen. |
| 1104 | let (preflight, router) = Self::preflight_and_bind(&snapshot, captured_router, config)?; |
| 1105 | |
| 1106 | let router_unavailable = match (snapshot.router(), &router) { |
| 1107 | (Some(_), None) => { |
| 1108 | Some("the Fleet's reasoning router could not be bound on this machine".to_string()) |
| 1109 | } |
| 1110 | _ => None, |
| 1111 | }; |
| 1112 | |
| 1113 | let workflow = Self { |
| 1114 | snapshot: Arc::new(snapshot), |
| 1115 | preflight: Arc::new(preflight), |
| 1116 | router, |
| 1117 | router_unavailable, |
| 1118 | }; |
| 1119 | workflow.reject_unusable_auto_members()?; |
| 1120 | Ok(workflow) |
| 1121 | } |
| 1122 | |
| 1123 | /// Preflight every worker route and bind the Router, or fail the start. |
| 1124 | fn preflight_and_bind( |
| 1125 | snapshot: &FleetSnapshot, |
| 1126 | captured_router: Option<CapturedReasoningRouter>, |
| 1127 | config: Option<&Config>, |
| 1128 | ) -> Result<(RoutePreflight, Option<Arc<dyn FleetRouterCaller>>), String> { |
| 1129 | let Some(config) = config else { |
| 1130 | return Err(format!( |
| 1131 | "exact Fleet `{}` cannot start: no session config is available to preflight its \ |
| 1132 | members' providers and models. An exact Fleet fails closed here rather than \ |
| 1133 | dispatching a worker onto a route it never verified.", |
| 1134 | snapshot.fleet().qualified() |
| 1135 | )); |
| 1136 | }; |
| 1137 | |
| 1138 | let mut workers = Vec::with_capacity(snapshot.members().len()); |
| 1139 | for member in snapshot.members() { |
| 1140 | let route = preflight_route( |
| 1141 | &member.id, |
| 1142 | &member.route.provider, |
| 1143 | &member.route.model, |
| 1144 | config, |
| 1145 | ) |
| 1146 | .map_err(|error| { |
| 1147 | format!( |
| 1148 | "exact Fleet `{}` cannot start: {error}", |
| 1149 | snapshot.fleet().qualified() |
| 1150 | ) |
| 1151 | })?; |
| 1152 | route.require_ready().map_err(|error| { |
| 1153 | format!( |
| 1154 | "exact Fleet `{}` cannot start: {error}", |
| 1155 | snapshot.fleet().qualified() |
| 1156 | ) |
| 1157 | })?; |
| 1158 | workers.push(route); |
| 1159 | } |
| 1160 | |
| 1161 | // Every worker client is constructed and validated **before** the |
| 1162 | // Router is bound, let alone called. A member whose client cannot be |
| 1163 | // built is a start-time failure; discovering it after a Router decision |
| 1164 | // means the operator paid for a routing request for a task that could |
| 1165 | // never have run. |
| 1166 | for route in &workers { |
| 1167 | validate_route_client(route, config).map_err(|error| { |
| 1168 | format!( |
| 1169 | "exact Fleet `{}` cannot start: {error}", |
| 1170 | snapshot.fleet().qualified() |
| 1171 | ) |
| 1172 | })?; |
| 1173 | } |
| 1174 | |
| 1175 | let mut router: Option<Arc<dyn FleetRouterCaller>> = None; |
| 1176 | let mut router_route = None; |
| 1177 | if let Some(captured) = &captured_router { |
| 1178 | match LiveFleetRouter::bind(captured, config) { |
| 1179 | Ok(live) => { |
| 1180 | router_route = Some(live.route().clone()); |
| 1181 | router = Some(Arc::new(live)); |
| 1182 | } |
| 1183 | Err(error) => { |
| 1184 | // Recorded rather than raised: a Fleet with no `auto` |
| 1185 | // member does not need its router to be usable, and |
| 1186 | // failing the whole Workflow for an unused service would |
| 1187 | // be the wrong trade. |
| 1188 | if snapshot.has_auto_member() { |
| 1189 | return Err(format!( |
| 1190 | "exact Fleet `{}` cannot start: member(s) {} request reasoning \ |
| 1191 | `auto` but the Fleet's reasoning router is unusable ({}). Fix the \ |
| 1192 | router profile or pin an explicit reasoning tier — exact Fleets \ |
| 1193 | never fall back to legacy model routing or a local heuristic.", |
| 1194 | snapshot.fleet().qualified(), |
| 1195 | snapshot.auto_member_ids().join(", "), |
| 1196 | error.reason, |
| 1197 | )); |
| 1198 | } |
| 1199 | } |
| 1200 | } |
| 1201 | } |
| 1202 | |
| 1203 | Ok((RoutePreflight::new(workers, router_route), router)) |
| 1204 | } |
| 1205 | |
| 1206 | /// Fail at Workflow start — not at task launch — when a member requests |
| 1207 | /// `auto` and the Fleet has no Router it can actually call. |
| 1208 | fn reject_unusable_auto_members(&self) -> Result<(), String> { |
| 1209 | if !self.snapshot.has_auto_member() || self.router.is_some() { |
| 1210 | return Ok(()); |
| 1211 | } |
| 1212 | let reason = self |
| 1213 | .router_unavailable |
| 1214 | .clone() |
| 1215 | .unwrap_or_else(|| "this Fleet references no reasoning router".to_string()); |
| 1216 | Err(format!( |
| 1217 | "exact Fleet `{}` cannot start: member(s) {} request reasoning `auto` but the Fleet's \ |
| 1218 | reasoning router is unusable ({reason}). Attach a working reasoning router or pin an \ |
| 1219 | explicit reasoning tier — exact Fleets never fall back to legacy model routing or a \ |
| 1220 | local heuristic.", |
| 1221 | self.snapshot.fleet().qualified(), |
| 1222 | self.snapshot.auto_member_ids().join(", "), |
| 1223 | )) |
| 1224 | } |
| 1225 | |
| 1226 | #[must_use] |
| 1227 | pub(crate) fn snapshot(&self) -> &Arc<FleetSnapshot> { |
| 1228 | &self.snapshot |
| 1229 | } |
| 1230 | |
| 1231 | /// Human-readable roster listing for "unknown member" errors. |
| 1232 | #[must_use] |
| 1233 | pub(crate) fn member_names(&self) -> String { |
| 1234 | self.snapshot |
| 1235 | .members() |
| 1236 | .iter() |
| 1237 | .map(|member| { |
| 1238 | if member.role == member.id { |
| 1239 | member.id.clone() |
| 1240 | } else { |
| 1241 | format!("{} (role {})", member.id, member.role) |
| 1242 | } |
| 1243 | }) |
| 1244 | .collect::<Vec<_>>() |
| 1245 | .join(", ") |
| 1246 | } |
| 1247 | |
| 1248 | /// Resolve a task's `role`/`profile` to one admitted member, **without |
| 1249 | /// contacting any provider**. |
| 1250 | /// |
| 1251 | /// This is deliberately the cheap half of a launch. It runs before gate |
| 1252 | /// evaluation and before a concurrency slot is taken, so a task that is |
| 1253 | /// about to be rejected or queued costs nothing and discloses nothing. |
| 1254 | /// |
| 1255 | /// A task that names both a `profile` and a `role` which resolve to |
| 1256 | /// different members is **rejected**, not silently resolved by precedence: |
| 1257 | /// the two fields would then disagree about who ran, and the receipt could |
| 1258 | /// only record one of them. |
| 1259 | pub(crate) fn bind_member( |
| 1260 | &self, |
| 1261 | profile: Option<&str>, |
| 1262 | role: Option<&str>, |
| 1263 | session: PermissionCeiling, |
| 1264 | ) -> Result<ExactMemberBinding, String> { |
| 1265 | let fleet = self.snapshot.fleet().qualified(); |
| 1266 | let profile = profile.map(str::trim).filter(|key| !key.is_empty()); |
| 1267 | let role = role.map(str::trim).filter(|key| !key.is_empty()); |
| 1268 | |
| 1269 | let member = match (profile, role) { |
| 1270 | (None, None) => { |
| 1271 | return Err(format!( |
| 1272 | "Fleet `{fleet}` is an exact Fleet: every task must name a member via `role` \ |
| 1273 | or `profile`. Members: {}", |
| 1274 | self.member_names() |
| 1275 | )); |
| 1276 | } |
| 1277 | (Some(profile), None) => self.lookup(profile)?, |
| 1278 | (None, Some(role)) => self.lookup(role)?, |
| 1279 | (Some(profile), Some(role)) => { |
| 1280 | let by_profile = self.lookup(profile)?; |
| 1281 | let by_role = self.lookup(role)?; |
| 1282 | if by_profile.id != by_role.id { |
| 1283 | return Err(format!( |
| 1284 | "Fleet `{fleet}`: task names profile `{profile}` (member `{}`) and role \ |
| 1285 | `{role}` (member `{}`), which are different members. A task must name \ |
| 1286 | one member; the two fields cannot disagree about who ran.", |
| 1287 | by_profile.id, by_role.id |
| 1288 | )); |
| 1289 | } |
| 1290 | by_profile |
| 1291 | } |
| 1292 | }; |
| 1293 | |
| 1294 | let route = self.preflight.worker(&member.id).ok_or_else(|| { |
| 1295 | format!( |
| 1296 | "Fleet `{fleet}`: member `{}` has no preflighted route", |
| 1297 | member.id |
| 1298 | ) |
| 1299 | })?; |
| 1300 | |
| 1301 | Ok(ExactMemberBinding { |
| 1302 | member_id: member.id.clone(), |
| 1303 | member_role: public_role_label(&member.role), |
| 1304 | route: route.clone(), |
| 1305 | requires_router: member.requested_reasoning.is_auto(), |
| 1306 | authority: ChildAuthority::from_runtime_role(&member.role, session), |
| 1307 | session, |
| 1308 | profile_origin: match self.snapshot.fleet().origin.as_str() { |
| 1309 | "workspace" => super::roster::ProfileOrigin::Workspace, |
| 1310 | "codewhale_home" => super::roster::ProfileOrigin::Personal, |
| 1311 | _ => super::roster::ProfileOrigin::Config, |
| 1312 | }, |
| 1313 | task_allowed_tools: None, |
| 1314 | task_disallowed_tools: Vec::new(), |
| 1315 | task_worktree_write: false, |
| 1316 | }) |
| 1317 | } |
| 1318 | |
| 1319 | fn lookup(&self, key: &str) -> Result<&FleetSnapshotMember, String> { |
| 1320 | self.snapshot.member_by_id_or_role(key).ok_or_else(|| { |
| 1321 | format!( |
| 1322 | "unknown exact Fleet member `{key}` in `{}`. Members: {}", |
| 1323 | self.snapshot.fleet().qualified(), |
| 1324 | self.member_names() |
| 1325 | ) |
| 1326 | }) |
| 1327 | } |
| 1328 | |
| 1329 | /// Finish an **already admitted** binding: decide only how hard the already |
| 1330 | /// frozen model thinks, then build the receipt. |
| 1331 | /// |
| 1332 | /// This is the half that can cost money. Calling it means the task has |
| 1333 | /// already passed its gates and holds a concurrency slot. |
| 1334 | pub(crate) async fn route_admitted_task( |
| 1335 | &self, |
| 1336 | binding: &ExactMemberBinding, |
| 1337 | task_summary: &str, |
| 1338 | ) -> Result<ExactMemberLaunch, String> { |
| 1339 | // The receipt built at the end of this function stamps |
| 1340 | // `snapshot.content_hash()` as evidence that this launch matched a saved |
| 1341 | // definition. Verify the hash actually describes the snapshot *before* |
| 1342 | // spending a router call or emitting that claim — an unverified hash is |
| 1343 | // not weaker evidence, it is a false receipt. |
| 1344 | self.snapshot |
| 1345 | .verify_content_hash() |
| 1346 | .map_err(|error| error.to_string())?; |
| 1347 | |
| 1348 | let member = self.snapshot.member(&binding.member_id).ok_or_else(|| { |
| 1349 | format!( |
| 1350 | "Fleet `{}`: member `{}` vanished between admission and launch", |
| 1351 | self.snapshot.fleet().qualified(), |
| 1352 | binding.member_id |
| 1353 | ) |
| 1354 | })?; |
| 1355 | |
| 1356 | // Recompute authority from Runtime's role policy and the live-parent |
| 1357 | // posture this binding was admitted against, and require it to be |
| 1358 | // *identical* to the one the binding carries. The snapshot supplies |
| 1359 | // identity only; legacy internal `FleetProfilePermissions` input is never |
| 1360 | // consulted. |
| 1361 | // |
| 1362 | // A binding crosses gates, a concurrency wait, and (for `auto` members) |
| 1363 | // a router call before it gets here, so "the authority I was handed" and |
| 1364 | // "the authority this member actually has" are two different claims. The |
| 1365 | // launch below is the value the spawn path consumes, so it must be the |
| 1366 | // recomputed one; the equality check is what turns a divergence into a |
| 1367 | // refused launch instead of a silently widened child. |
| 1368 | let authority = binding.recompute_authority(&member.role); |
| 1369 | if authority != binding.authority { |
| 1370 | return Err(format!( |
| 1371 | "Fleet `{}`: member `{}` resolved a different permission envelope at launch than \ |
| 1372 | at admission, so the launch is refused. admitted={} launched={}", |
| 1373 | self.snapshot.fleet().qualified(), |
| 1374 | binding.member_id, |
| 1375 | binding.authority.fingerprint(), |
| 1376 | authority.fingerprint(), |
| 1377 | )); |
| 1378 | } |
| 1379 | |
| 1380 | // The route is already frozen and preflighted. Nothing below may move |
| 1381 | // it — not a task option, not the Router. |
| 1382 | let frozen = binding.route.frozen(); |
| 1383 | let capability = binding.route.capability; |
| 1384 | |
| 1385 | let availability = self.router_availability(); |
| 1386 | let mut router_identity = None; |
| 1387 | let mut routing_summary: Option<RoutingDisclosure> = None; |
| 1388 | let decision = if binding.requires_router { |
| 1389 | let router = self.router.as_ref().ok_or_else(|| { |
| 1390 | format!( |
| 1391 | "member `{}` requests reasoning `auto` but Fleet `{}` has no usable reasoning \ |
| 1392 | router", |
| 1393 | binding.member_id, |
| 1394 | self.snapshot.fleet().qualified() |
| 1395 | ) |
| 1396 | })?; |
| 1397 | let cross_provider = self.preflight.crosses_providers(&binding.member_id); |
| 1398 | let payload = bounded_routing_payload(task_summary).with_cross_provider(cross_provider); |
| 1399 | // What actually leaves for the router's provider, recorded so the |
| 1400 | // receipt discloses it — counts and hash only, never the text. |
| 1401 | routing_summary = Some(payload.disclosure().clone()); |
| 1402 | router_identity = Some(router.identity()); |
| 1403 | let input = RouterCallInput { |
| 1404 | fleet: self.snapshot.fleet().qualified(), |
| 1405 | member_id: binding.member_id.clone(), |
| 1406 | frozen: frozen.clone(), |
| 1407 | payload, |
| 1408 | }; |
| 1409 | let raw = router.decide(&input).await.map_err(|error| { |
| 1410 | format!( |
| 1411 | "reasoning router call failed for member `{}`: {error}", |
| 1412 | binding.member_id |
| 1413 | ) |
| 1414 | })?; |
| 1415 | Some(parse_router_decision(&raw).map_err(|error| { |
| 1416 | format!( |
| 1417 | "reasoning router returned an unusable decision for member `{}`: {error}", |
| 1418 | binding.member_id |
| 1419 | ) |
| 1420 | })?) |
| 1421 | } else { |
| 1422 | None |
| 1423 | }; |
| 1424 | |
| 1425 | let reasoning = resolve_exact_member_reasoning( |
| 1426 | &binding.member_id, |
| 1427 | &frozen, |
| 1428 | member.requested_reasoning, |
| 1429 | &capability, |
| 1430 | &availability, |
| 1431 | decision.as_ref(), |
| 1432 | router_identity.as_ref(), |
| 1433 | ) |
| 1434 | .map_err(|error| error.to_string())?; |
| 1435 | |
| 1436 | // Every exact launch carries a concrete tier. `auto` is resolved by the |
| 1437 | // router above and the literal sentinel never leaves this function. |
| 1438 | // |
| 1439 | // `NativeAdaptive` is no longer reachable here: removing the bypass |
| 1440 | // (so `auto` always asks the router) also removed the one path that |
| 1441 | // produced it. It used to be launched as `off`, which mislabelled the |
| 1442 | // request — a route choosing its own depth is not a route with thinking |
| 1443 | // disabled. Rather than re-introduce that lie, this fails loudly if the |
| 1444 | // variant ever comes back. |
| 1445 | let thinking = match reasoning.effective() { |
| 1446 | EffectiveReasoning::Tier(tier) => effort_of(tier).as_setting().to_string(), |
| 1447 | EffectiveReasoning::NativeAdaptive => { |
| 1448 | return Err(format!( |
| 1449 | "member `{}` resolved to provider-native adaptive reasoning, which an exact \ |
| 1450 | Fleet launch cannot place on a request. Pin an explicit reasoning tier.", |
| 1451 | binding.member_id |
| 1452 | )); |
| 1453 | } |
| 1454 | }; |
| 1455 | |
| 1456 | // The durable receipt. Built here, at the one place that knows every |
| 1457 | // side of the decision, so no consumer has to re-derive it. |
| 1458 | let receipt = FleetTaskReceipt::new( |
| 1459 | self.snapshot.fleet().qualified(), |
| 1460 | self.snapshot.schema_kind(), |
| 1461 | self.snapshot.schema_revision(), |
| 1462 | self.snapshot.content_hash(), |
| 1463 | binding.member_id.clone(), |
| 1464 | binding.member_role.clone(), |
| 1465 | &binding.route, |
| 1466 | &reasoning, |
| 1467 | routing_summary, |
| 1468 | binding.authority.ceiling.network_tool, |
| 1469 | ) |
| 1470 | // The fingerprint of the envelope this launch installs, carried on the |
| 1471 | // durable receipt so the spawn boundary has something to check against |
| 1472 | // rather than a sentinel it can only assume. |
| 1473 | .with_authority_fingerprint(authority.fingerprint()) |
| 1474 | // Semantic role and runtime posture stay two separate facts all the way |
| 1475 | // onto the durable receipt: `member_role` is what the operator named |
| 1476 | // and what gates key on, `posture_role` is the Runtime baseline role. |
| 1477 | // The fingerprint above records the effective parent-narrowed surface. |
| 1478 | .with_posture_role(binding.authority.posture_role); |
| 1479 | |
| 1480 | Ok(ExactMemberLaunch { |
| 1481 | member_id: binding.member_id.clone(), |
| 1482 | member_role: binding.member_role.clone(), |
| 1483 | provider: frozen.provider, |
| 1484 | model: frozen.model, |
| 1485 | thinking, |
| 1486 | reasoning, |
| 1487 | authority, |
| 1488 | receipt, |
| 1489 | }) |
| 1490 | } |
| 1491 | |
| 1492 | fn router_availability(&self) -> RouterAvailability { |
| 1493 | match (&self.router, &self.router_unavailable) { |
| 1494 | (Some(_), _) => RouterAvailability::Ready, |
| 1495 | (None, Some(reason)) => RouterAvailability::Unavailable { |
| 1496 | reason: reason.clone(), |
| 1497 | }, |
| 1498 | (None, None) => RouterAvailability::Absent, |
| 1499 | } |
| 1500 | } |
| 1501 | } |
| 1502 | |
| 1503 | // ── Test seams ────────────────────────────────────────────────────────────── |
| 1504 | |
| 1505 | /// A Router that answers with a fixed fixture string, recording what it saw. |
| 1506 | /// |
| 1507 | /// Test-only: it is how the exact-Fleet reasoning path is exercised end to end |
| 1508 | /// without a provider call, and how "the router was never called" is asserted. |
| 1509 | #[cfg(test)] |
| 1510 | #[derive(Debug)] |
| 1511 | pub(crate) struct StaticFleetRouter { |
| 1512 | response: String, |
| 1513 | identity: RouterIdentity, |
| 1514 | pub(crate) seen: std::sync::Mutex<Vec<RouterCallInput>>, |
| 1515 | } |
| 1516 | |
| 1517 | #[cfg(test)] |
| 1518 | impl StaticFleetRouter { |
| 1519 | pub(crate) fn new(response: impl Into<String>) -> Arc<Self> { |
| 1520 | Arc::new(Self { |
| 1521 | response: response.into(), |
| 1522 | identity: RouterIdentity { |
| 1523 | id: "luna-low".to_string(), |
| 1524 | origin: "workspace".to_string(), |
| 1525 | service_kind: codewhale_workflow::REASONING_ROUTER_SERVICE_KIND.to_string(), |
| 1526 | legacy_inline: false, |
| 1527 | provider: "openai".to_string(), |
| 1528 | model: "gpt-5.6-luna".to_string(), |
| 1529 | endpoint: Some(EndpointIdentity::from_base_url("https://api.openai.com/v1")), |
| 1530 | call: Some( |
| 1531 | router_call_plan( |
| 1532 | codewhale_workflow::RouterCallReasoning::Low, |
| 1533 | &ReasoningCapability::tiered(), |
| 1534 | ) |
| 1535 | .disclosure, |
| 1536 | ), |
| 1537 | }, |
| 1538 | seen: std::sync::Mutex::new(Vec::new()), |
| 1539 | }) |
| 1540 | } |
| 1541 | |
| 1542 | /// How many router calls were made. Zero is the assertion that matters for |
| 1543 | /// manual reasoning and for rejected/blocked tasks. |
| 1544 | pub(crate) fn call_count(&self) -> usize { |
| 1545 | self.seen.lock().expect("router log").len() |
| 1546 | } |
| 1547 | } |
| 1548 | |
| 1549 | #[cfg(test)] |
| 1550 | #[async_trait] |
| 1551 | impl FleetRouterCaller for StaticFleetRouter { |
| 1552 | fn identity(&self) -> RouterIdentity { |
| 1553 | self.identity.clone() |
| 1554 | } |
| 1555 | |
| 1556 | async fn decide(&self, input: &RouterCallInput) -> Result<String, String> { |
| 1557 | self.seen.lock().expect("router log").push(input.clone()); |
| 1558 | Ok(self.response.clone()) |
| 1559 | } |
| 1560 | } |
| 1561 | |
| 1562 | #[cfg(test)] |
| 1563 | impl ExactFleetWorkflow { |
| 1564 | /// Build a Workflow with an injected Router and a supplied capability, |
| 1565 | /// skipping provider binding so the reasoning path runs with no network and |
| 1566 | /// no configured provider. |
| 1567 | /// Takes the concrete fixture type rather than `Option<Arc<dyn ...>>`: |
| 1568 | /// `Option` does not coerce its payload, so the unsizing is done once here |
| 1569 | /// instead of at every call site. |
| 1570 | pub(crate) fn for_tests( |
| 1571 | document: &FleetDocument, |
| 1572 | id: QualifiedFleetId, |
| 1573 | router: Option<Arc<StaticFleetRouter>>, |
| 1574 | ) -> Self { |
| 1575 | Self::for_tests_with_capability(document, id, router, ReasoningCapability::tiered()) |
| 1576 | } |
| 1577 | |
| 1578 | pub(crate) fn for_tests_with_capability( |
| 1579 | document: &FleetDocument, |
| 1580 | id: QualifiedFleetId, |
| 1581 | router: Option<Arc<StaticFleetRouter>>, |
| 1582 | capability: ReasoningCapability, |
| 1583 | ) -> Self { |
| 1584 | let exact = document.exact().expect("exact Fleet"); |
| 1585 | let captured = captured_legacy_inline_router(exact).or_else(|| { |
| 1586 | exact.reasoning_router.as_ref().map(|name| { |
| 1587 | CapturedReasoningRouter::from_profile( |
| 1588 | &ReasoningRouterProfile::parse(&format!( |
| 1589 | "name = \"{name}\"\nschema = \"reasoning_router\"\nprovider = \ |
| 1590 | \"openai\"\nmodel = \"gpt-5.6-luna\"\ncall_reasoning = \"low\"\n" |
| 1591 | )) |
| 1592 | .expect("router profile"), |
| 1593 | "workspace", |
| 1594 | ) |
| 1595 | }) |
| 1596 | }); |
| 1597 | let snapshot = |
| 1598 | FleetSnapshot::capture(id, document, "2026-07-26T00:00:00Z", captured.clone()) |
| 1599 | .expect("valid roster"); |
| 1600 | |
| 1601 | let workers = snapshot |
| 1602 | .members() |
| 1603 | .iter() |
| 1604 | .map(|member| { |
| 1605 | test_route( |
| 1606 | &member.id, |
| 1607 | &member.route.provider, |
| 1608 | &member.route.model, |
| 1609 | capability, |
| 1610 | ) |
| 1611 | }) |
| 1612 | .collect::<Vec<_>>(); |
| 1613 | let router_route = captured.as_ref().map(|captured| { |
| 1614 | test_route( |
| 1615 | "router", |
| 1616 | &captured.route.provider, |
| 1617 | &captured.route.model, |
| 1618 | capability, |
| 1619 | ) |
| 1620 | }); |
| 1621 | let preflight = RoutePreflight::new(workers, router_route); |
| 1622 | |
| 1623 | Self { |
| 1624 | snapshot: Arc::new(snapshot), |
| 1625 | preflight: Arc::new(preflight), |
| 1626 | router: router.map(|router| { |
| 1627 | let router: Arc<dyn FleetRouterCaller> = router; |
| 1628 | router |
| 1629 | }), |
| 1630 | router_unavailable: None, |
| 1631 | } |
| 1632 | } |
| 1633 | |
| 1634 | /// A Workflow whose Router failed to bind locally — the shape |
| 1635 | /// [`Self::capture`] produces when a Router's provider has no credentials |
| 1636 | /// configured on this machine. No network is involved either way. |
| 1637 | pub(crate) fn for_tests_with_unavailable_router( |
| 1638 | document: &FleetDocument, |
| 1639 | id: QualifiedFleetId, |
| 1640 | reason: &str, |
| 1641 | ) -> Result<Self, String> { |
| 1642 | let mut workflow = Self::for_tests(document, id, None); |
| 1643 | workflow.router_unavailable = Some(reason.to_string()); |
| 1644 | workflow.reject_unusable_auto_members()?; |
| 1645 | Ok(workflow) |
| 1646 | } |
| 1647 | } |
| 1648 | |
| 1649 | #[cfg(test)] |
| 1650 | fn test_route( |
| 1651 | member: &str, |
| 1652 | provider: &str, |
| 1653 | model: &str, |
| 1654 | capability: ReasoningCapability, |
| 1655 | ) -> PreflightedRoute { |
| 1656 | PreflightedRoute { |
| 1657 | member_id: member.to_string(), |
| 1658 | provider_id: provider.to_string(), |
| 1659 | provider_config_id: None, |
| 1660 | provider_kind: provider.to_string(), |
| 1661 | declared_model: model.to_string(), |
| 1662 | wire_model: model.to_string(), |
| 1663 | endpoint: EndpointIdentity::from_base_url("https://api.example.test/v1"), |
| 1664 | credential: CredentialReadiness::Configured, |
| 1665 | capability, |
| 1666 | } |
| 1667 | } |
| 1668 | |
| 1669 | #[cfg(test)] |
| 1670 | mod shell_ceiling_tests { |
| 1671 | use super::*; |
| 1672 | |
| 1673 | fn ceiling(write: bool, shell: ShellCeiling) -> PermissionCeiling { |
| 1674 | PermissionCeiling { |
| 1675 | write, |
| 1676 | network_tool: false, |
| 1677 | shell, |
| 1678 | delegation_depth: 0, |
| 1679 | tools: true, |
| 1680 | } |
| 1681 | } |
| 1682 | |
| 1683 | fn session() -> PermissionCeiling { |
| 1684 | ceiling(true, ShellCeiling::Full) |
| 1685 | } |
| 1686 | |
| 1687 | fn denies_raw_shell(authority: &ChildAuthority) -> bool { |
| 1688 | authority |
| 1689 | .disallowed_tools |
| 1690 | .iter() |
| 1691 | .any(|rule| rule == RAW_SHELL_SENTINEL) |
| 1692 | } |
| 1693 | |
| 1694 | /// The `analyst` preset grants no shell. The envelope reads its shell bit |
| 1695 | /// back off the deny list, so the denial has to actually be installed — |
| 1696 | /// otherwise a shell-less ceiling reaches dispatch claiming full shell |
| 1697 | /// authority and can start a verification process. |
| 1698 | #[test] |
| 1699 | fn a_shell_less_ceiling_installs_the_raw_shell_denial() { |
| 1700 | for shell in [ShellCeiling::None, ShellCeiling::ReadOnly] { |
| 1701 | let authority = ChildAuthority::clamp(ceiling(false, shell), session()); |
| 1702 | assert!( |
| 1703 | denies_raw_shell(&authority), |
| 1704 | "{shell:?} must deny raw shell" |
| 1705 | ); |
| 1706 | } |
| 1707 | } |
| 1708 | |
| 1709 | /// The gap this repair closed: a write-capable member inside a session with |
| 1710 | /// no shell authority clamps to `write = true, shell = none`. Keying the |
| 1711 | /// denial on `write` alone left that combination with no denial installed — |
| 1712 | /// and therefore with an envelope that claimed shell authority the ceiling |
| 1713 | /// had refused. |
| 1714 | #[test] |
| 1715 | fn a_write_capable_member_clamped_to_no_shell_still_loses_raw_shell() { |
| 1716 | let authority = ChildAuthority::clamp( |
| 1717 | ceiling(true, ShellCeiling::Full), |
| 1718 | ceiling(true, ShellCeiling::None), |
| 1719 | ); |
| 1720 | |
| 1721 | assert_eq!(authority.ceiling.shell, ShellCeiling::None); |
| 1722 | assert!(authority.ceiling.write, "the write half is unchanged"); |
| 1723 | assert!(denies_raw_shell(&authority)); |
| 1724 | } |
| 1725 | |
| 1726 | /// Prior behavior preserved: a `verifier`/`tester` ceiling |
| 1727 | /// (`write = false, shell = "full"`) still loses raw shell, and a fully |
| 1728 | /// write-capable member still keeps it. |
| 1729 | #[test] |
| 1730 | fn the_existing_verifier_and_full_ceilings_are_unchanged() { |
| 1731 | let verifier = ChildAuthority::clamp(ceiling(false, ShellCeiling::Full), session()); |
| 1732 | assert!(denies_raw_shell(&verifier)); |
| 1733 | assert_eq!(verifier.posture_role, "test"); |
| 1734 | |
| 1735 | let full = ChildAuthority::clamp(ceiling(true, ShellCeiling::Full), session()); |
| 1736 | assert!(!denies_raw_shell(&full)); |
| 1737 | assert_eq!(full.posture_role, "implement"); |
| 1738 | } |
| 1739 | |
| 1740 | #[test] |
| 1741 | fn bounded_inspection_role_keeps_only_classifier_bounded_bash() { |
| 1742 | for role in ["scout", "reviewer", "planner"] { |
| 1743 | let authority = ChildAuthority::from_runtime_role(role, session()); |
| 1744 | assert!( |
| 1745 | !authority |
| 1746 | .disallowed_tools |
| 1747 | .iter() |
| 1748 | .any(|name| name.eq_ignore_ascii_case("Bash")), |
| 1749 | "{role} keeps canonical Bash for per-input classification" |
| 1750 | ); |
| 1751 | for denied in [ |
| 1752 | "exec_shell", |
| 1753 | "task_shell_start", |
| 1754 | "task_shell_wait", |
| 1755 | "terminal/*", |
| 1756 | "write_file", |
| 1757 | "apply_patch", |
| 1758 | ] { |
| 1759 | assert!( |
| 1760 | authority.disallowed_tools.iter().any(|name| name == denied), |
| 1761 | "{role} must still deny {denied}: {:?}", |
| 1762 | authority.disallowed_tools |
| 1763 | ); |
| 1764 | } |
| 1765 | } |
| 1766 | |
| 1767 | for role in ["consultant", "verifier"] { |
| 1768 | let authority = ChildAuthority::from_runtime_role(role, session()); |
| 1769 | assert!( |
| 1770 | authority |
| 1771 | .disallowed_tools |
| 1772 | .iter() |
| 1773 | .any(|name| name.eq_ignore_ascii_case("Bash")), |
| 1774 | "{role} must not gain the read-only inspection exception" |
| 1775 | ); |
| 1776 | } |
| 1777 | |
| 1778 | let parent_shell_off = |
| 1779 | ChildAuthority::from_runtime_role("scout", ceiling(true, ShellCeiling::None)); |
| 1780 | assert!( |
| 1781 | parent_shell_off |
| 1782 | .disallowed_tools |
| 1783 | .iter() |
| 1784 | .any(|name| name.eq_ignore_ascii_case("Bash")), |
| 1785 | "a named Scout may not turn a parent shell-off ceiling into ReadOnly" |
| 1786 | ); |
| 1787 | let planner_parent_shell_off = |
| 1788 | ChildAuthority::from_runtime_role("planner", ceiling(true, ShellCeiling::None)); |
| 1789 | assert!( |
| 1790 | planner_parent_shell_off |
| 1791 | .disallowed_tools |
| 1792 | .iter() |
| 1793 | .any(|name| name.eq_ignore_ascii_case("Bash")), |
| 1794 | "a named planner may not turn a parent shell-off ceiling into ReadOnly" |
| 1795 | ); |
| 1796 | assert_eq!(planner_parent_shell_off.posture_role, "planner"); |
| 1797 | assert_eq!( |
| 1798 | session_shell_ceiling(crate::worker_profile::ShellPolicy::Full, false), |
| 1799 | ShellCeiling::None |
| 1800 | ); |
| 1801 | } |
| 1802 | |
| 1803 | /// #5426 acceptance 2, made mechanical: delegation moves work, never |
| 1804 | /// authority. A read-only scout's own runtime posture is the "session" |
| 1805 | /// its children clamp against, so a Runtime `builder` dispatched from a |
| 1806 | /// read-only parent lands read-only — raw shell gone and mutating tools |
| 1807 | /// denied — while the Runtime posture remains separately identified and delegation stays |
| 1808 | /// available (the depth budget is the parent's, not zero). The escape |
| 1809 | /// hatch is work capacity, never a wider envelope. |
| 1810 | #[test] |
| 1811 | fn a_read_only_parents_delegation_never_widens_authority() { |
| 1812 | // The scout's live runtime posture, expressed as the session ceiling |
| 1813 | // a child clamps against: no writes, read-only shell, network kept, |
| 1814 | // one level of delegation budget left. |
| 1815 | let scout_runtime = PermissionCeiling { |
| 1816 | write: false, |
| 1817 | network_tool: true, |
| 1818 | shell: ShellCeiling::ReadOnly, |
| 1819 | delegation_depth: 1, |
| 1820 | tools: true, |
| 1821 | }; |
| 1822 | let authority = ChildAuthority::from_runtime_role("builder", scout_runtime); |
| 1823 | |
| 1824 | // Authority does not widen through delegation: the child is read-only. |
| 1825 | assert!(!authority.ceiling.write); |
| 1826 | assert_eq!(authority.ceiling.shell, ShellCeiling::ReadOnly); |
| 1827 | assert_eq!(authority.write_authority, "read_only"); |
| 1828 | assert_eq!(authority.posture_role, "implement"); |
| 1829 | assert!(denies_raw_shell(&authority)); |
| 1830 | for mutating in ["write_file", "apply_patch"] { |
| 1831 | assert!( |
| 1832 | authority |
| 1833 | .disallowed_tools |
| 1834 | .iter() |
| 1835 | .any(|rule| rule == mutating), |
| 1836 | "{mutating} must stay denied for a scout-delegated builder: {:?}", |
| 1837 | authority.disallowed_tools |
| 1838 | ); |
| 1839 | } |
| 1840 | |
| 1841 | // The escape hatch itself stays open: delegation is still possible |
| 1842 | // (the parent's budget is intact). But it is useless for shell: |
| 1843 | // canonical Bash is denied to a delegated child (it is not a bounded |
| 1844 | // inspection role), so a scout can never obtain bash by spawning — |
| 1845 | // the scout's own bounded read-only Bash from #5428 is the only shell |
| 1846 | // path a read-only parent has. |
| 1847 | assert_eq!(authority.max_depth, 1); |
| 1848 | assert!( |
| 1849 | authority |
| 1850 | .disallowed_tools |
| 1851 | .iter() |
| 1852 | .any(|name| name.eq_ignore_ascii_case("Bash")), |
| 1853 | "a scout-delegated child must not gain canonical Bash: {:?}", |
| 1854 | authority.disallowed_tools |
| 1855 | ); |
| 1856 | } |
| 1857 | |
| 1858 | /// The deny list feeds the fingerprint, so a ceiling that now denies more |
| 1859 | /// must fingerprint differently from one that does not. Two postures that |
| 1860 | /// install different surfaces may never share a fingerprint. |
| 1861 | #[test] |
| 1862 | fn the_shell_denial_is_visible_in_the_fingerprint() { |
| 1863 | let no_shell = ChildAuthority::clamp(ceiling(false, ShellCeiling::None), session()); |
| 1864 | let full = ChildAuthority::clamp(ceiling(true, ShellCeiling::Full), session()); |
| 1865 | |
| 1866 | assert_ne!(no_shell.fingerprint(), full.fingerprint()); |
| 1867 | assert!(no_shell.fingerprint().contains("shell=none")); |
| 1868 | } |
| 1869 | |
| 1870 | /// Every rule the shell clamp installs is a *posture* denial, so a |
| 1871 | /// grandchild spawned with `inherit_disallowed_tools: false` cannot drop it. |
| 1872 | #[test] |
| 1873 | fn the_installed_shell_denials_are_posture_denials() { |
| 1874 | let authority = ChildAuthority::clamp(ceiling(false, ShellCeiling::None), session()); |
| 1875 | for rule in &authority.disallowed_tools { |
| 1876 | assert!(is_posture_denial(rule), "{rule} must be a posture denial"); |
| 1877 | } |
| 1878 | } |
| 1879 | } |
| 1880 | |
| 1881 | #[cfg(test)] |
| 1882 | mod tests { |
| 1883 | use super::*; |
| 1884 | use codewhale_workflow::{ |
| 1885 | EffectiveReasoningSource, ProviderEffectiveReasoning, RequestedReasoning, |
| 1886 | }; |
| 1887 | |
| 1888 | /// A Fleet that references a saved, reusable Reasoning Router service. |
| 1889 | const GLM_FLEET: &str = r#" |
| 1890 | name = "glm-pair" |
| 1891 | schema = "exact" |
| 1892 | reasoning_router = "luna-low" |
| 1893 | |
| 1894 | [[members]] |
| 1895 | id = "implementer" |
| 1896 | role = "builder" |
| 1897 | provider = "zai" |
| 1898 | model = "glm-5" |
| 1899 | reasoning = "auto" |
| 1900 | permissions = "read_write" |
| 1901 | |
| 1902 | [[members]] |
| 1903 | id = "auditor" |
| 1904 | role = "reviewer" |
| 1905 | provider = "zai" |
| 1906 | model = "glm-5" |
| 1907 | reasoning = "high" |
| 1908 | permissions = "read_only" |
| 1909 | "#; |
| 1910 | |
| 1911 | fn id() -> QualifiedFleetId { |
| 1912 | QualifiedFleetId { |
| 1913 | name: "glm-pair".to_string(), |
| 1914 | origin: "workspace".to_string(), |
| 1915 | } |
| 1916 | } |
| 1917 | |
| 1918 | fn full_session() -> PermissionCeiling { |
| 1919 | PermissionCeiling { |
| 1920 | write: true, |
| 1921 | network_tool: true, |
| 1922 | shell: ShellCeiling::Full, |
| 1923 | delegation_depth: codewhale_config::DEFAULT_SPAWN_DEPTH, |
| 1924 | tools: true, |
| 1925 | } |
| 1926 | } |
| 1927 | |
| 1928 | /// Takes the concrete fixture type: `Option` does not coerce its payload, |
| 1929 | /// so the unsizing to `Arc<dyn FleetRouterCaller>` is spelled out here once |
| 1930 | /// rather than at every call site. |
| 1931 | fn workflow_with(router: Option<Arc<StaticFleetRouter>>, text: &str) -> ExactFleetWorkflow { |
| 1932 | let document = FleetDocument::parse(text).expect("parse"); |
| 1933 | ExactFleetWorkflow::for_tests(&document, id(), router) |
| 1934 | } |
| 1935 | |
| 1936 | #[tokio::test] |
| 1937 | async fn an_auto_member_takes_a_reasoning_only_router_decision_on_a_frozen_route() { |
| 1938 | let router = StaticFleetRouter::new(r#"{"reasoning":"max"}"#); |
| 1939 | let workflow = workflow_with(Some(router.clone()), GLM_FLEET); |
| 1940 | |
| 1941 | let binding = workflow |
| 1942 | .bind_member(None, Some("builder"), full_session()) |
| 1943 | .expect("role resolves"); |
| 1944 | assert_eq!( |
| 1945 | router.call_count(), |
| 1946 | 0, |
| 1947 | "binding a member must not cost a router call" |
| 1948 | ); |
| 1949 | |
| 1950 | let launch = workflow |
| 1951 | .route_admitted_task(&binding, "refactor three crates") |
| 1952 | .await |
| 1953 | .expect("auto resolves through the router"); |
| 1954 | |
| 1955 | // The route did not move. |
| 1956 | assert_eq!(launch.provider, "zai"); |
| 1957 | assert_eq!(launch.model, "glm-5"); |
| 1958 | assert_eq!(launch.thinking, "max"); |
| 1959 | assert_eq!(launch.member_id, "implementer"); |
| 1960 | assert_eq!(launch.member_role, "implement"); |
| 1961 | assert_eq!(launch.reasoning.requested(), RequestedReasoning::Auto); |
| 1962 | assert_eq!( |
| 1963 | launch.reasoning.source(), |
| 1964 | EffectiveReasoningSource::FleetRouter |
| 1965 | ); |
| 1966 | |
| 1967 | // The router saw the frozen route as context, never as a question, and |
| 1968 | // received the bounded payload rather than the raw task. |
| 1969 | let seen = router.seen.lock().expect("log"); |
| 1970 | assert_eq!(seen.len(), 1); |
| 1971 | assert_eq!(seen[0].frozen.model, "glm-5"); |
| 1972 | assert_eq!(seen[0].member_id, "implementer"); |
| 1973 | assert_eq!(seen[0].payload.text(), "refactor three crates"); |
| 1974 | } |
| 1975 | |
| 1976 | /// The semantic role must survive onto the launch and the receipt: a gate |
| 1977 | /// or handoff keyed on `builder` has to still see `builder` even though the |
| 1978 | /// roster resolves the distinct profile id `implementer`. |
| 1979 | #[tokio::test] |
| 1980 | async fn the_semantic_role_survives_while_the_id_addresses_the_roster() { |
| 1981 | let workflow = workflow_with( |
| 1982 | Some(StaticFleetRouter::new(r#"{"reasoning":"low"}"#)), |
| 1983 | GLM_FLEET, |
| 1984 | ); |
| 1985 | |
| 1986 | let binding = workflow |
| 1987 | .bind_member(None, Some("reviewer"), full_session()) |
| 1988 | .expect("role lookup"); |
| 1989 | assert_eq!(binding.member_id, "auditor"); |
| 1990 | assert_eq!(binding.member_role, "reviewer"); |
| 1991 | |
| 1992 | let launch = workflow |
| 1993 | .route_admitted_task(&binding, "read the diff") |
| 1994 | .await |
| 1995 | .expect("launch"); |
| 1996 | assert_eq!(launch.receipt.member_id, "auditor"); |
| 1997 | assert_eq!( |
| 1998 | launch.receipt.member_role, "reviewer", |
| 1999 | "the receipt records the semantic role, not the profile id" |
| 2000 | ); |
| 2001 | |
| 2002 | // The snapshot is addressed by id; the role is the semantic label. |
| 2003 | let member = workflow |
| 2004 | .snapshot() |
| 2005 | .member("auditor") |
| 2006 | .expect("snapshot entry"); |
| 2007 | assert_eq!(member.role, "reviewer"); |
| 2008 | } |
| 2009 | |
| 2010 | /// A task that names a profile and a role belonging to different members is |
| 2011 | /// rejected — the two fields cannot disagree about who ran. |
| 2012 | #[test] |
| 2013 | fn a_conflicting_task_role_and_profile_is_rejected() { |
| 2014 | let workflow = workflow_with(None, GLM_FLEET); |
| 2015 | |
| 2016 | let err = workflow |
| 2017 | .bind_member(Some("implementer"), Some("reviewer"), full_session()) |
| 2018 | .expect_err("conflicting identity"); |
| 2019 | assert!(err.contains("different members"), "{err}"); |
| 2020 | assert!(err.contains("implementer"), "{err}"); |
| 2021 | assert!(err.contains("auditor"), "{err}"); |
| 2022 | |
| 2023 | // Agreeing fields are fine: id plus that member's own role. |
| 2024 | let binding = workflow |
| 2025 | .bind_member(Some("implementer"), Some("builder"), full_session()) |
| 2026 | .expect("agreeing identity"); |
| 2027 | assert_eq!(binding.member_id, "implementer"); |
| 2028 | } |
| 2029 | |
| 2030 | /// Manual reasoning uses no Router at all — not a call whose answer is |
| 2031 | /// discarded, but zero calls. |
| 2032 | #[tokio::test] |
| 2033 | async fn an_explicit_tier_member_never_calls_the_router() { |
| 2034 | let router = StaticFleetRouter::new(r#"{"reasoning":"off"}"#); |
| 2035 | let workflow = workflow_with(Some(router.clone()), GLM_FLEET); |
| 2036 | |
| 2037 | let binding = workflow |
| 2038 | .bind_member(Some("auditor"), None, full_session()) |
| 2039 | .expect("bind"); |
| 2040 | assert!(!binding.requires_router); |
| 2041 | |
| 2042 | let launch = workflow |
| 2043 | .route_admitted_task(&binding, "read the diff") |
| 2044 | .await |
| 2045 | .expect("explicit tier"); |
| 2046 | |
| 2047 | assert_eq!(launch.thinking, "high"); |
| 2048 | assert_eq!( |
| 2049 | launch.reasoning.source(), |
| 2050 | EffectiveReasoningSource::MemberExplicit |
| 2051 | ); |
| 2052 | assert_eq!( |
| 2053 | router.call_count(), |
| 2054 | 0, |
| 2055 | "an explicit tier must not spend a router call" |
| 2056 | ); |
| 2057 | assert!(launch.receipt.router.is_none()); |
| 2058 | assert!(launch.receipt.routing_summary.is_none()); |
| 2059 | assert!(!launch.receipt.cross_provider_inference); |
| 2060 | } |
| 2061 | |
| 2062 | /// A task that never reaches admission must never reach the Router. This |
| 2063 | /// is the shape of a gate rejection or a capacity block: the caller binds, |
| 2064 | /// decides not to proceed, and no provider was contacted. |
| 2065 | #[test] |
| 2066 | fn a_task_that_is_never_admitted_costs_no_router_call() { |
| 2067 | let router = StaticFleetRouter::new(r#"{"reasoning":"max"}"#); |
| 2068 | let workflow = workflow_with(Some(router.clone()), GLM_FLEET); |
| 2069 | |
| 2070 | // Unknown member: rejected during binding, before any cost. |
| 2071 | assert!( |
| 2072 | workflow |
| 2073 | .bind_member(None, Some("wizard"), full_session()) |
| 2074 | .is_err() |
| 2075 | ); |
| 2076 | // Conflicting identity: likewise. |
| 2077 | assert!( |
| 2078 | workflow |
| 2079 | .bind_member(Some("implementer"), Some("reviewer"), full_session()) |
| 2080 | .is_err() |
| 2081 | ); |
| 2082 | // A valid binding that the caller then abandons (gate reject / no slot). |
| 2083 | let _binding = workflow |
| 2084 | .bind_member(None, Some("builder"), full_session()) |
| 2085 | .expect("valid binding"); |
| 2086 | |
| 2087 | assert_eq!( |
| 2088 | router.call_count(), |
| 2089 | 0, |
| 2090 | "no router call may happen before a task is admitted" |
| 2091 | ); |
| 2092 | } |
| 2093 | |
| 2094 | #[tokio::test] |
| 2095 | async fn a_router_that_tries_to_move_the_route_fails_the_launch() { |
| 2096 | let workflow = workflow_with( |
| 2097 | Some(StaticFleetRouter::new( |
| 2098 | r#"{"reasoning":"max","model":"glm-4"}"#, |
| 2099 | )), |
| 2100 | GLM_FLEET, |
| 2101 | ); |
| 2102 | let binding = workflow |
| 2103 | .bind_member(None, Some("builder"), full_session()) |
| 2104 | .expect("bind"); |
| 2105 | |
| 2106 | let err = workflow |
| 2107 | .route_admitted_task(&binding, "anything") |
| 2108 | .await |
| 2109 | .expect_err("a route mutation must fail the launch"); |
| 2110 | assert!(err.contains("frozen"), "{err}"); |
| 2111 | } |
| 2112 | |
| 2113 | #[tokio::test] |
| 2114 | async fn a_duplicate_reasoning_key_fails_the_launch() { |
| 2115 | let workflow = workflow_with( |
| 2116 | Some(StaticFleetRouter::new( |
| 2117 | r#"{"reasoning":"off","reasoning":"max"}"#, |
| 2118 | )), |
| 2119 | GLM_FLEET, |
| 2120 | ); |
| 2121 | let binding = workflow |
| 2122 | .bind_member(None, Some("builder"), full_session()) |
| 2123 | .expect("bind"); |
| 2124 | |
| 2125 | let err = workflow |
| 2126 | .route_admitted_task(&binding, "anything") |
| 2127 | .await |
| 2128 | .expect_err("duplicate key"); |
| 2129 | assert!(err.contains("more than once"), "{err}"); |
| 2130 | } |
| 2131 | |
| 2132 | #[test] |
| 2133 | fn a_missing_router_fails_before_any_worker_is_dispatched() { |
| 2134 | let router_less = GLM_FLEET.replace("reasoning_router = \"luna-low\"\n", ""); |
| 2135 | let document = FleetDocument::parse(&router_less).expect("parse"); |
| 2136 | let workflow = ExactFleetWorkflow::for_tests(&document, id(), None); |
| 2137 | |
| 2138 | let err = workflow |
| 2139 | .reject_unusable_auto_members() |
| 2140 | .expect_err("auto without a router must not start"); |
| 2141 | assert!(err.contains("implementer"), "{err}"); |
| 2142 | assert!(err.contains("reasoning router"), "{err}"); |
| 2143 | assert!( |
| 2144 | err.contains("never fall back"), |
| 2145 | "the error must rule out legacy fallback: {err}" |
| 2146 | ); |
| 2147 | } |
| 2148 | |
| 2149 | #[test] |
| 2150 | fn a_fleet_with_no_auto_member_starts_without_a_router() { |
| 2151 | let text = r#" |
| 2152 | name = "pinned" |
| 2153 | schema = "exact" |
| 2154 | |
| 2155 | [[members]] |
| 2156 | id = "auditor" |
| 2157 | provider = "zai" |
| 2158 | model = "glm-5" |
| 2159 | reasoning = "high" |
| 2160 | permissions = "read_only" |
| 2161 | "#; |
| 2162 | let document = FleetDocument::parse(text).expect("parse"); |
| 2163 | let workflow = ExactFleetWorkflow::for_tests( |
| 2164 | &document, |
| 2165 | QualifiedFleetId { |
| 2166 | name: "pinned".to_string(), |
| 2167 | origin: "workspace".to_string(), |
| 2168 | }, |
| 2169 | None, |
| 2170 | ); |
| 2171 | workflow |
| 2172 | .reject_unusable_auto_members() |
| 2173 | .expect("no auto member means no router requirement"); |
| 2174 | assert_eq!(workflow.snapshot().members().len(), 1); |
| 2175 | } |
| 2176 | |
| 2177 | /// A Router whose credentials are locally absent fails the Workflow before |
| 2178 | /// any worker is dispatched — decided from local config, never from a live |
| 2179 | /// probe of the provider. |
| 2180 | #[test] |
| 2181 | fn a_locally_unusable_router_fails_before_any_worker_is_dispatched() { |
| 2182 | let document = FleetDocument::parse(GLM_FLEET).expect("parse"); |
| 2183 | let err = ExactFleetWorkflow::for_tests_with_unavailable_router( |
| 2184 | &document, |
| 2185 | id(), |
| 2186 | "no credential configured for `openai`", |
| 2187 | ) |
| 2188 | .expect_err("an unusable router must not start an auto Fleet"); |
| 2189 | |
| 2190 | assert!(err.contains("cannot start"), "{err}"); |
| 2191 | assert!(err.contains("implementer"), "{err}"); |
| 2192 | assert!(err.contains("no credential configured"), "{err}"); |
| 2193 | assert!(err.contains("never fall back"), "{err}"); |
| 2194 | } |
| 2195 | |
| 2196 | #[test] |
| 2197 | fn the_frozen_route_pins_each_members_exact_provider_and_model() { |
| 2198 | let workflow = workflow_with(None, GLM_FLEET); |
| 2199 | let route = workflow |
| 2200 | .preflight |
| 2201 | .worker("implementer") |
| 2202 | .expect("preflighted worker"); |
| 2203 | |
| 2204 | assert_eq!(route.provider_id, "zai"); |
| 2205 | assert_eq!(route.wire_model, "glm-5"); |
| 2206 | let member = workflow |
| 2207 | .snapshot() |
| 2208 | .member("implementer") |
| 2209 | .expect("snapshot entry"); |
| 2210 | assert!( |
| 2211 | member.requested_reasoning.is_auto(), |
| 2212 | "reasoning is decided per task, not baked into the frozen route" |
| 2213 | ); |
| 2214 | } |
| 2215 | |
| 2216 | /// Binding carries route and Runtime role, but no Fleet-owned authority. |
| 2217 | #[test] |
| 2218 | fn bound_members_use_runtime_roles_and_neutral_compatibility_fields() { |
| 2219 | let workflow = workflow_with(None, GLM_FLEET); |
| 2220 | for (id, expected_posture, expected_write) in [ |
| 2221 | ("auditor", "reviewer", "read_only"), |
| 2222 | ("implementer", "implement", "workspace_write"), |
| 2223 | ] { |
| 2224 | let binding = workflow |
| 2225 | .bind_member(Some(id), None, full_session()) |
| 2226 | .expect("bind"); |
| 2227 | assert_eq!( |
| 2228 | binding.authority.posture_role, expected_posture, |
| 2229 | "{id} must resolve through Runtime's closed role policy" |
| 2230 | ); |
| 2231 | assert_eq!( |
| 2232 | binding.authority.write_authority, expected_write, |
| 2233 | "{id} authority comes from the role posture, never a Fleet permissions block" |
| 2234 | ); |
| 2235 | } |
| 2236 | } |
| 2237 | |
| 2238 | /// #5575: a free-form member role keeps its identity and **fails closed**. |
| 2239 | /// |
| 2240 | /// This test previously asserted the opposite — `posture_role == "custom"` |
| 2241 | /// and `write_authority == "workspace_write"` — which is exactly the defect: |
| 2242 | /// `audit-lead` is a name nobody declared, and the exact driver answered it |
| 2243 | /// with the widest posture there is while the durable driver answered the |
| 2244 | /// same class of name with `general`. Identity is still preserved verbatim |
| 2245 | /// (`member_role`), but an undeclared label now buys the narrowest useful |
| 2246 | /// posture, not write authority. |
| 2247 | #[test] |
| 2248 | fn a_free_form_member_role_fails_closed_without_losing_identity() { |
| 2249 | const AUDIT_FLEET: &str = r#" |
| 2250 | name = "audit" |
| 2251 | schema = "exact" |
| 2252 | |
| 2253 | [[members]] |
| 2254 | id = "auditor-one" |
| 2255 | role = "audit-lead" |
| 2256 | provider = "zai" |
| 2257 | model = "glm-5" |
| 2258 | permissions = "read_only" |
| 2259 | "#; |
| 2260 | let workflow = workflow_with(None, AUDIT_FLEET); |
| 2261 | let binding = workflow |
| 2262 | .bind_member(Some("auditor-one"), None, full_session()) |
| 2263 | .expect("bind"); |
| 2264 | |
| 2265 | assert_eq!(binding.member_role, "audit-lead"); |
| 2266 | assert_eq!(binding.authority.posture_role, "explore"); |
| 2267 | assert_eq!( |
| 2268 | binding.authority.write_authority, "read_only", |
| 2269 | "an undeclared role name must never grant write authority; an \ |
| 2270 | operator who wants the parent's posture spells the role `custom`" |
| 2271 | ); |
| 2272 | |
| 2273 | // The escape hatch is a declared role, not a typo. |
| 2274 | assert_eq!( |
| 2275 | ChildAuthority::from_runtime_role("custom", full_session()).write_authority, |
| 2276 | "workspace_write" |
| 2277 | ); |
| 2278 | } |
| 2279 | |
| 2280 | // ── Permission ceilings, as the child actually experiences them ───────── |
| 2281 | |
| 2282 | /// `tools = false` means zero model tools — an empty allowlist, which the |
| 2283 | /// child registry treats as "nothing is visible and nothing is callable". |
| 2284 | #[test] |
| 2285 | fn spawn_refuses_empty_effective_toolset() { |
| 2286 | let authority = ChildAuthority::clamp(PermissionCeiling::ROUTER, full_session()); |
| 2287 | let err = refuse_dropped_requested_tools( |
| 2288 | "router", |
| 2289 | "advisor", |
| 2290 | &["read_file".to_string(), "grep_files".to_string()], |
| 2291 | &authority, |
| 2292 | &authority, |
| 2293 | ) |
| 2294 | .expect_err("a child that would start with no tools is refused"); |
| 2295 | assert!( |
| 2296 | err.contains("Agent 'router' (role advisor) would start with no tools"), |
| 2297 | "{err}" |
| 2298 | ); |
| 2299 | assert!(err.contains("requested [grep_files, read_file]"), "{err}"); |
| 2300 | assert!(err.contains("role allows no tools"), "{err}"); |
| 2301 | // An explicit empty request stays a deliberate tool-free child. |
| 2302 | refuse_dropped_requested_tools("router", "advisor", &[], &authority, &authority) |
| 2303 | .expect("explicit empty toolset is allowed"); |
| 2304 | } |
| 2305 | |
| 2306 | #[test] |
| 2307 | fn spawn_error_names_dropped_tools() { |
| 2308 | let workflow = workflow_with(None, GLM_FLEET); |
| 2309 | let mut binding = workflow |
| 2310 | .bind_member(None, Some("reviewer"), full_session()) |
| 2311 | .expect("role lookup"); |
| 2312 | let err = binding |
| 2313 | .narrow_for_task( |
| 2314 | None, |
| 2315 | Some(&["read_file".to_string(), "exec_shell".to_string()]), |
| 2316 | &[], |
| 2317 | None, |
| 2318 | ) |
| 2319 | .expect_err("a requested tool the role denies is refused"); |
| 2320 | assert!(err.contains("would lose requested tools"), "{err}"); |
| 2321 | assert!(err.contains("dropped [exec_shell]"), "{err}"); |
| 2322 | assert!(err.contains("requested [exec_shell, read_file]"), "{err}"); |
| 2323 | assert!( |
| 2324 | err.contains("role allows all inherited tools except ["), |
| 2325 | "{err}" |
| 2326 | ); |
| 2327 | |
| 2328 | // Tools the role allows narrow cleanly. |
| 2329 | binding |
| 2330 | .narrow_for_task(None, Some(&["read_file".to_string()]), &[], None) |
| 2331 | .expect("an allowed narrowing succeeds"); |
| 2332 | assert_eq!( |
| 2333 | binding.authority.allowed_tools.as_deref(), |
| 2334 | Some(&["read_file".to_string()] as &[String]) |
| 2335 | ); |
| 2336 | } |
| 2337 | |
| 2338 | #[test] |
| 2339 | fn tools_false_yields_an_empty_tool_surface() { |
| 2340 | let authority = ChildAuthority::clamp(PermissionCeiling::ROUTER, full_session()); |
| 2341 | |
| 2342 | assert!(!authority.ceiling.tools); |
| 2343 | assert_eq!( |
| 2344 | authority.allowed_tools.as_deref(), |
| 2345 | Some(&[] as &[String]), |
| 2346 | "tools = false must be an empty allowlist, not an absent one" |
| 2347 | ); |
| 2348 | assert_eq!(authority.write_authority, "read_only"); |
| 2349 | assert_eq!(authority.max_depth, 0); |
| 2350 | } |
| 2351 | |
| 2352 | /// `network_tool = false` removes every model-visible network, browser, |
| 2353 | /// and remote-MCP surface except the `Web` family's two read-only actions |
| 2354 | /// — even when `tools = true`. The family *name* must survive the deny |
| 2355 | /// list so the child registry's action seam can grant exactly |
| 2356 | /// `search`/`fetch`; every other browsing spelling is denied. |
| 2357 | #[test] |
| 2358 | fn network_disabled_denies_every_network_surface_even_with_tools_enabled() { |
| 2359 | let member = PermissionCeiling::preset("read_write").expect("preset"); |
| 2360 | assert!(member.tools); |
| 2361 | assert!(!member.network_tool); |
| 2362 | |
| 2363 | let authority = ChildAuthority::clamp(member, full_session()); |
| 2364 | |
| 2365 | assert!( |
| 2366 | authority.allowed_tools.is_none(), |
| 2367 | "a tool-using member keeps full inheritance, narrowed by the deny list" |
| 2368 | ); |
| 2369 | for expected in [ |
| 2370 | "web.run", |
| 2371 | "web_run", |
| 2372 | "web_search", |
| 2373 | "fetch_url", |
| 2374 | "wait_for_dev_server", |
| 2375 | "github", |
| 2376 | "mcp*", |
| 2377 | ] { |
| 2378 | assert!( |
| 2379 | authority |
| 2380 | .disallowed_tools |
| 2381 | .iter() |
| 2382 | .any(|name| name == expected), |
| 2383 | "{expected} must be denied: {:?}", |
| 2384 | authority.disallowed_tools |
| 2385 | ); |
| 2386 | } |
| 2387 | // The canonical family name is what the read-only web surface |
| 2388 | // dispatches under; only its reaching spellings are denied. |
| 2389 | assert!( |
| 2390 | !authority.disallowed_tools.iter().any(|name| name == "Web"), |
| 2391 | "the Web family name must survive so search/fetch stay reachable: {:?}", |
| 2392 | authority.disallowed_tools |
| 2393 | ); |
| 2394 | |
| 2395 | // A member that IS allowed a network tool gets no such deny list. |
| 2396 | let networked = ChildAuthority::clamp( |
| 2397 | PermissionCeiling::preset("full").expect("preset"), |
| 2398 | full_session(), |
| 2399 | ); |
| 2400 | assert!(networked.ceiling.network_tool); |
| 2401 | assert!(networked.disallowed_tools.is_empty()); |
| 2402 | } |
| 2403 | |
| 2404 | /// The browsing capability is registered under several names, and `web.run` |
| 2405 | /// is the one a deny list stopping at the `Web` family name leaves behind. |
| 2406 | /// A network-denied member that can still call `web.run` is not |
| 2407 | /// network-denied, so every spelling *except* the family name itself — |
| 2408 | /// which the action seam bounds to `search`/`fetch` — stays on the list. |
| 2409 | #[test] |
| 2410 | fn network_disabled_denies_the_canonical_web_run_surface_and_its_aliases() { |
| 2411 | let authority = ChildAuthority::clamp( |
| 2412 | PermissionCeiling::preset("read_write").expect("preset"), |
| 2413 | full_session(), |
| 2414 | ); |
| 2415 | |
| 2416 | let denied = |name: &str| { |
| 2417 | let lowered = name.to_ascii_lowercase(); |
| 2418 | authority.disallowed_tools.iter().any(|rule| { |
| 2419 | let rule = rule.to_ascii_lowercase(); |
| 2420 | rule.strip_suffix('*') |
| 2421 | .map_or(rule == lowered, |prefix| lowered.starts_with(prefix)) |
| 2422 | }) |
| 2423 | }; |
| 2424 | |
| 2425 | for name in [ |
| 2426 | "web.run", |
| 2427 | "web_run", |
| 2428 | "web_search", |
| 2429 | "web.fetch", |
| 2430 | "web_fetch", |
| 2431 | "fetch_url", |
| 2432 | "wait_for_dev_server", |
| 2433 | "browse", |
| 2434 | "browser", |
| 2435 | ] { |
| 2436 | assert!( |
| 2437 | denied(name), |
| 2438 | "{name} must be denied: {:?}", |
| 2439 | authority.disallowed_tools |
| 2440 | ); |
| 2441 | } |
| 2442 | // The family name itself is what the read-only search/fetch surface |
| 2443 | // dispatches under; the action seam and the URL-input guard bound it. |
| 2444 | assert!( |
| 2445 | !denied("Web"), |
| 2446 | "the Web family name must survive a network denial: {:?}", |
| 2447 | authority.disallowed_tools |
| 2448 | ); |
| 2449 | // The globs must not reach past the browsing family. |
| 2450 | for name in ["read_file", "run_tests", "Git", "grep_files"] { |
| 2451 | assert!(!denied(name), "{name} is not a network surface"); |
| 2452 | } |
| 2453 | } |
| 2454 | |
| 2455 | /// `rlm` reaches the network without ever naming a network tool: `open` |
| 2456 | /// fetches a `url` by calling `FetchUrlTool` in-process, and `eval` runs |
| 2457 | /// Python that owns a socket API. Denying `fetch_url` sees neither call, so |
| 2458 | /// both actions carry their own deny-list entries. |
| 2459 | #[test] |
| 2460 | fn network_disabled_denies_the_in_process_rlm_reach() { |
| 2461 | let authority = ChildAuthority::clamp( |
| 2462 | PermissionCeiling::preset("read_write").expect("preset"), |
| 2463 | full_session(), |
| 2464 | ); |
| 2465 | |
| 2466 | let denied = |name: &str| { |
| 2467 | let lowered = name.to_ascii_lowercase(); |
| 2468 | authority.disallowed_tools.iter().any(|rule| { |
| 2469 | let rule = rule.to_ascii_lowercase(); |
| 2470 | rule.strip_suffix('*') |
| 2471 | .map_or(rule == lowered, |prefix| lowered.starts_with(prefix)) |
| 2472 | }) |
| 2473 | }; |
| 2474 | |
| 2475 | for reaching in ["rlm_open", "rlm_eval"] { |
| 2476 | assert!( |
| 2477 | denied(reaching), |
| 2478 | "{reaching} reaches the network in-process and must be denied: {:?}", |
| 2479 | authority.disallowed_tools |
| 2480 | ); |
| 2481 | } |
| 2482 | // The fail-closed narrowing is deliberate but *bounded*: the bounded |
| 2483 | // local metadata actions survive, and so does the family itself, so the |
| 2484 | // per-action seam has something left to permit. |
| 2485 | for kept in ["rlm", "rlm_session_objects", "rlm_configure", "rlm_close"] { |
| 2486 | assert!( |
| 2487 | !denied(kept), |
| 2488 | "{kept} is bounded local metadata and must survive a network denial" |
| 2489 | ); |
| 2490 | } |
| 2491 | } |
| 2492 | |
| 2493 | /// The deny-list sentinel has to actually be on the deny list, or every |
| 2494 | /// posture check derived from it silently reads "network allowed". |
| 2495 | #[test] |
| 2496 | fn the_network_denial_sentinel_is_installed_by_a_network_denial() { |
| 2497 | assert!( |
| 2498 | NETWORK_TOOL_DENYLIST.contains(&NETWORK_DENIAL_SENTINEL), |
| 2499 | "{NETWORK_DENIAL_SENTINEL} must be an explicit entry, not a glob match" |
| 2500 | ); |
| 2501 | let authority = ChildAuthority::clamp( |
| 2502 | PermissionCeiling::preset("read_write").expect("preset"), |
| 2503 | full_session(), |
| 2504 | ); |
| 2505 | assert!( |
| 2506 | authority |
| 2507 | .disallowed_tools |
| 2508 | .iter() |
| 2509 | .any(|rule| rule == NETWORK_DENIAL_SENTINEL), |
| 2510 | "a network denial must install the sentinel verbatim: {:?}", |
| 2511 | authority.disallowed_tools |
| 2512 | ); |
| 2513 | // …and a network-*capable* member must not, or the sentinel would read |
| 2514 | // as denied for everyone. |
| 2515 | let networked = ChildAuthority::clamp( |
| 2516 | PermissionCeiling::preset("full").expect("preset"), |
| 2517 | full_session(), |
| 2518 | ); |
| 2519 | assert!( |
| 2520 | !networked |
| 2521 | .disallowed_tools |
| 2522 | .iter() |
| 2523 | .any(|rule| rule == NETWORK_DENIAL_SENTINEL) |
| 2524 | ); |
| 2525 | } |
| 2526 | |
| 2527 | /// Every network-denied preset — read_only/read-only inspection included — leaves the |
| 2528 | /// `Web` family name reachable and seals each of its reaching spellings. |
| 2529 | /// This is the deny-list half of the read-only web-search contract; the |
| 2530 | /// registry-side half (exactly `search`/`fetch`, with URL-addressed calls |
| 2531 | /// refused) is asserted in `subagent/tests.rs`. |
| 2532 | #[test] |
| 2533 | fn every_network_denial_leaves_web_search_reachable_by_family_name() { |
| 2534 | for preset in ["analyst", "read_only", "verifier", "read_write"] { |
| 2535 | let authority = ChildAuthority::clamp( |
| 2536 | PermissionCeiling::preset(preset).expect("preset"), |
| 2537 | full_session(), |
| 2538 | ); |
| 2539 | assert!( |
| 2540 | !authority.ceiling.network_tool, |
| 2541 | "{preset} is network-denied" |
| 2542 | ); |
| 2543 | assert!( |
| 2544 | !authority.disallowed_tools.iter().any(|rule| rule == "Web"), |
| 2545 | "{preset} must keep the Web family name: {:?}", |
| 2546 | authority.disallowed_tools |
| 2547 | ); |
| 2548 | for sealed in [ |
| 2549 | "web_*", |
| 2550 | "web.*", |
| 2551 | "web.run", |
| 2552 | "web_run", |
| 2553 | "web_search", |
| 2554 | "web.fetch", |
| 2555 | "web_fetch", |
| 2556 | "fetch_url", |
| 2557 | "wait_for_dev_server", |
| 2558 | "github", |
| 2559 | "mcp*", |
| 2560 | ] { |
| 2561 | assert!( |
| 2562 | authority.disallowed_tools.iter().any(|rule| rule == sealed), |
| 2563 | "{preset} must deny {sealed}: {:?}", |
| 2564 | authority.disallowed_tools |
| 2565 | ); |
| 2566 | } |
| 2567 | } |
| 2568 | } |
| 2569 | |
| 2570 | /// A member saved as `write = false` must not receive a mutating surface — |
| 2571 | /// including the raw shell a `verifier`-shaped ceiling keeps for running |
| 2572 | /// checks. `rm -rf` mutates a workspace exactly as well as `write_file`, |
| 2573 | /// and a receipt that says `write=false` while the child holds `exec_shell` |
| 2574 | /// is not true. |
| 2575 | #[test] |
| 2576 | fn a_read_only_member_gets_a_truthful_non_mutating_tool_contract() { |
| 2577 | let verifier = PermissionCeiling::preset("verifier").expect("preset"); |
| 2578 | assert!(!verifier.write); |
| 2579 | assert_eq!(verifier.shell, ShellCeiling::Full); |
| 2580 | |
| 2581 | let authority = ChildAuthority::clamp(verifier, full_session()); |
| 2582 | assert_eq!(authority.write_authority, "read_only"); |
| 2583 | |
| 2584 | let denied = |name: &str| { |
| 2585 | authority.disallowed_tools.iter().any(|rule| { |
| 2586 | rule == name || rule.strip_suffix('*').is_some_and(|p| name.starts_with(p)) |
| 2587 | }) |
| 2588 | }; |
| 2589 | |
| 2590 | // `rlm_eval` belongs on this list for the same reason `exec_shell` does: |
| 2591 | // the Python it runs writes files. A tool is a mutation primitive |
| 2592 | // because of what it can do, not because of what it is called. |
| 2593 | for mutating in [ |
| 2594 | "write_file", |
| 2595 | "edit_file", |
| 2596 | "apply_patch", |
| 2597 | "fim_edit", |
| 2598 | "rlm_eval", |
| 2599 | ] { |
| 2600 | assert!( |
| 2601 | denied(mutating), |
| 2602 | "{mutating} must be denied for a read-only member: {:?}", |
| 2603 | authority.disallowed_tools |
| 2604 | ); |
| 2605 | } |
| 2606 | for raw_shell in [ |
| 2607 | "Bash", |
| 2608 | "exec_shell", |
| 2609 | "exec_shell_interact", |
| 2610 | "task_shell_start", |
| 2611 | "terminal/run", |
| 2612 | ] { |
| 2613 | assert!( |
| 2614 | denied(raw_shell), |
| 2615 | "{raw_shell} is a general mutation primitive: {:?}", |
| 2616 | authority.disallowed_tools |
| 2617 | ); |
| 2618 | } |
| 2619 | // What the member is *for* survives: the bounded verification surface. |
| 2620 | // (`rlm_open` is absent from this list only because the `verifier` |
| 2621 | // preset is also network-denied; the write contract alone keeps it — |
| 2622 | // see `a_write_denial_alone_keeps_local_rlm_loading`.) |
| 2623 | for kept in [ |
| 2624 | "Run", |
| 2625 | "run_tests", |
| 2626 | "run_verifiers", |
| 2627 | "read_file", |
| 2628 | "grep_files", |
| 2629 | "rlm", |
| 2630 | ] { |
| 2631 | assert!(!denied(kept), "{kept} must stay available to a verifier"); |
| 2632 | } |
| 2633 | |
| 2634 | // A write-capable member is untouched by this contract. |
| 2635 | let builder = ChildAuthority::clamp( |
| 2636 | PermissionCeiling::preset("read_write").expect("preset"), |
| 2637 | full_session(), |
| 2638 | ); |
| 2639 | assert!(builder.ceiling.write); |
| 2640 | for kept in ["write_file", "apply_patch", "exec_shell"] { |
| 2641 | assert!( |
| 2642 | !builder.disallowed_tools.iter().any(|rule| rule == kept), |
| 2643 | "{kept} must stay available to a write-capable member" |
| 2644 | ); |
| 2645 | } |
| 2646 | } |
| 2647 | |
| 2648 | /// The two denials are separate contracts and must not bleed into each |
| 2649 | /// other. A member that may not *write* can still load a large local file |
| 2650 | /// into an RLM kernel and read it — that is analysis, not mutation. Only |
| 2651 | /// `eval` goes, because only `eval` runs code. |
| 2652 | #[test] |
| 2653 | fn a_write_denial_alone_keeps_local_rlm_loading() { |
| 2654 | let member = PermissionCeiling { |
| 2655 | write: false, |
| 2656 | network_tool: true, |
| 2657 | shell: ShellCeiling::ReadOnly, |
| 2658 | delegation_depth: 0, |
| 2659 | tools: true, |
| 2660 | }; |
| 2661 | let authority = ChildAuthority::clamp(member, full_session()); |
| 2662 | assert!(!authority.ceiling.write); |
| 2663 | assert!(authority.ceiling.network_tool); |
| 2664 | |
| 2665 | let denied = |name: &str| authority.disallowed_tools.iter().any(|rule| rule == name); |
| 2666 | |
| 2667 | assert!(denied("rlm_eval"), "eval runs code, so it mutates"); |
| 2668 | for kept in ["rlm", "rlm_open", "rlm_session_objects", "rlm_close"] { |
| 2669 | assert!( |
| 2670 | !denied(kept), |
| 2671 | "{kept} loads and inspects; it does not mutate: {:?}", |
| 2672 | authority.disallowed_tools |
| 2673 | ); |
| 2674 | } |
| 2675 | } |
| 2676 | |
| 2677 | /// The parent posture always wins. A saved `full` member inside a |
| 2678 | /// read-only, no-network, no-shell session runs at the session's ceiling. |
| 2679 | #[test] |
| 2680 | fn the_parent_ceiling_wins_over_a_wider_saved_member() { |
| 2681 | let session = PermissionCeiling { |
| 2682 | write: false, |
| 2683 | network_tool: false, |
| 2684 | shell: ShellCeiling::ReadOnly, |
| 2685 | delegation_depth: 0, |
| 2686 | tools: true, |
| 2687 | }; |
| 2688 | let member = PermissionCeiling::preset("full").expect("preset"); |
| 2689 | assert!(member.write && member.network_tool); |
| 2690 | |
| 2691 | let authority = ChildAuthority::clamp(member, session); |
| 2692 | |
| 2693 | assert!(!authority.ceiling.write, "a Fleet may not grant write"); |
| 2694 | assert!( |
| 2695 | !authority.ceiling.network_tool, |
| 2696 | "a Fleet may not grant a network tool" |
| 2697 | ); |
| 2698 | assert_eq!(authority.ceiling.shell, ShellCeiling::ReadOnly); |
| 2699 | assert_eq!(authority.ceiling.delegation_depth, 0); |
| 2700 | assert_eq!(authority.write_authority, "read_only"); |
| 2701 | assert_eq!(authority.max_depth, 0); |
| 2702 | assert_eq!(authority.posture_role, "explore"); |
| 2703 | assert!(!authority.disallowed_tools.is_empty()); |
| 2704 | } |
| 2705 | |
| 2706 | /// A read-only session cannot be widened by a session that *is* permissive |
| 2707 | /// either — clamping is symmetric, and takes the narrower side each way. |
| 2708 | #[test] |
| 2709 | fn clamping_takes_the_narrower_side_of_every_field() { |
| 2710 | let narrow_member = PermissionCeiling { |
| 2711 | write: false, |
| 2712 | network_tool: false, |
| 2713 | shell: ShellCeiling::None, |
| 2714 | delegation_depth: 0, |
| 2715 | tools: true, |
| 2716 | }; |
| 2717 | let authority = ChildAuthority::clamp(narrow_member, full_session()); |
| 2718 | |
| 2719 | assert!(!authority.ceiling.write); |
| 2720 | assert_eq!(authority.ceiling.shell, ShellCeiling::None); |
| 2721 | assert_eq!(authority.ceiling.delegation_depth, 0); |
| 2722 | } |
| 2723 | |
| 2724 | // ── Preflight ────────────────────────────────────────────────────────── |
| 2725 | |
| 2726 | /// Z.AI GLM routes express only thinking enabled/disabled, so `high` and |
| 2727 | /// `max` must not be reported as two distinct provider-effective tiers. |
| 2728 | #[test] |
| 2729 | fn glm_routes_report_an_enabled_disabled_provider_control() { |
| 2730 | let capability = reasoning_capability_for_route( |
| 2731 | ProviderKind::Zai, |
| 2732 | crate::config::DEFAULT_ZAI_BASE_URL, |
| 2733 | crate::config::ZAI_GLM_5_2_MODEL, |
| 2734 | ); |
| 2735 | |
| 2736 | assert_eq!( |
| 2737 | capability.control, |
| 2738 | ProviderReasoningControl::EnabledDisabled, |
| 2739 | "Z.AI's request shaping emits only thinking enabled/disabled" |
| 2740 | ); |
| 2741 | assert!(!capability.supports_native_adaptive()); |
| 2742 | assert_eq!( |
| 2743 | capability.provider_effective(ReasoningTier::High), |
| 2744 | ProviderEffectiveReasoning::Enabled |
| 2745 | ); |
| 2746 | assert_eq!( |
| 2747 | capability.provider_effective(ReasoningTier::Off), |
| 2748 | ProviderEffectiveReasoning::Disabled |
| 2749 | ); |
| 2750 | } |
| 2751 | |
| 2752 | /// DeepSeek varies `reasoning_effort` per tier, so its tiers are real. |
| 2753 | #[test] |
| 2754 | fn a_route_that_varies_its_wire_value_reports_distinct_tiers() { |
| 2755 | let capability = reasoning_capability_for_route( |
| 2756 | ProviderKind::Deepseek, |
| 2757 | crate::config::DEFAULT_DEEPSEEK_BASE_URL, |
| 2758 | "deepseek-v4-pro", |
| 2759 | ); |
| 2760 | assert_eq!(capability.control, ProviderReasoningControl::Tiers); |
| 2761 | } |
| 2762 | |
| 2763 | /// First-party DeepSeek routes document `reasoning_effort` low/high/max |
| 2764 | /// on the wire (no medium), so `low` is a real tier there. The capability |
| 2765 | /// must report the tier the route *sends*, not the tier the selector |
| 2766 | /// named: low reaches the wire as low, medium rounds up to high because |
| 2767 | /// the dialect has no such value (#52). |
| 2768 | #[test] |
| 2769 | fn a_deepseek_route_reports_low_as_low_and_medium_as_high() { |
| 2770 | let capability = reasoning_capability_for_route( |
| 2771 | ProviderKind::Deepseek, |
| 2772 | crate::config::DEFAULT_DEEPSEEK_BASE_URL, |
| 2773 | "deepseek-v4-pro", |
| 2774 | ); |
| 2775 | |
| 2776 | // Exactly what the request shaping does, read back off the capability. |
| 2777 | for (requested, expected) in [ |
| 2778 | (ReasoningTier::Low, ReasoningTier::Low), |
| 2779 | (ReasoningTier::Medium, ReasoningTier::High), |
| 2780 | (ReasoningTier::High, ReasoningTier::High), |
| 2781 | (ReasoningTier::Max, ReasoningTier::Max), |
| 2782 | (ReasoningTier::Off, ReasoningTier::Off), |
| 2783 | ] { |
| 2784 | assert_eq!( |
| 2785 | capability.wire_tier(requested), |
| 2786 | expected, |
| 2787 | "requested {requested:?} must be reported as what the wire carries" |
| 2788 | ); |
| 2789 | let (effective, normalized) = capability.normalize(requested); |
| 2790 | assert_eq!(effective, expected); |
| 2791 | assert_eq!(normalized, requested != expected); |
| 2792 | } |
| 2793 | |
| 2794 | // And the resolver carries that all the way onto the receipt. |
| 2795 | let resolved = codewhale_workflow::resolve_exact_member_reasoning( |
| 2796 | "implementer", |
| 2797 | &codewhale_workflow::FrozenRoute { |
| 2798 | provider: "deepseek".to_string(), |
| 2799 | model: "deepseek-v4-pro".to_string(), |
| 2800 | }, |
| 2801 | RequestedReasoning::Low, |
| 2802 | &capability, |
| 2803 | &RouterAvailability::Absent, |
| 2804 | None, |
| 2805 | None, |
| 2806 | ) |
| 2807 | .expect("resolve"); |
| 2808 | assert_eq!(resolved.requested(), RequestedReasoning::Low); |
| 2809 | assert_eq!( |
| 2810 | resolved.effective(), |
| 2811 | codewhale_workflow::EffectiveReasoning::Tier(ReasoningTier::Low) |
| 2812 | ); |
| 2813 | assert!(!resolved.capability_normalized()); |
| 2814 | } |
| 2815 | |
| 2816 | /// Routes whose dialect has no low tier still collapse low onto high, and |
| 2817 | /// the capability must say so instead of reporting a `low` the wire never |
| 2818 | /// carried. CodeWhale's normalizer keeps the historic low/medium → high |
| 2819 | /// coercion for these DeepSeek-compatible hosted routes because their own |
| 2820 | /// wire contracts are not verified. |
| 2821 | #[test] |
| 2822 | fn a_route_that_collapses_low_onto_high_says_so_instead_of_reporting_low() { |
| 2823 | let capability = reasoning_capability_for_route( |
| 2824 | ProviderKind::Siliconflow, |
| 2825 | crate::config::DEFAULT_SILICONFLOW_BASE_URL, |
| 2826 | "deepseek-ai/DeepSeek-V4-Pro", |
| 2827 | ); |
| 2828 | |
| 2829 | for (requested, expected) in [ |
| 2830 | (ReasoningTier::Low, ReasoningTier::High), |
| 2831 | (ReasoningTier::Medium, ReasoningTier::High), |
| 2832 | (ReasoningTier::High, ReasoningTier::High), |
| 2833 | (ReasoningTier::Max, ReasoningTier::Max), |
| 2834 | (ReasoningTier::Off, ReasoningTier::Off), |
| 2835 | ] { |
| 2836 | assert_eq!( |
| 2837 | capability.wire_tier(requested), |
| 2838 | expected, |
| 2839 | "requested {requested:?} must be reported as what the wire carries" |
| 2840 | ); |
| 2841 | let (effective, normalized) = capability.normalize(requested); |
| 2842 | assert_eq!(effective, expected); |
| 2843 | assert_eq!(normalized, requested != expected); |
| 2844 | } |
| 2845 | } |
| 2846 | |
| 2847 | /// Preflight resolves the provider, canonicalizes the model, identifies the |
| 2848 | /// endpoint, and decides credential readiness — all from local config. |
| 2849 | #[test] |
| 2850 | fn preflight_freezes_provider_model_endpoint_and_local_readiness() { |
| 2851 | let _env_lock = crate::test_support::lock_test_env(); |
| 2852 | let _key = crate::test_support::EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 2853 | let config = Config { |
| 2854 | provider: Some("zai".to_string()), |
| 2855 | ..Default::default() |
| 2856 | }; |
| 2857 | |
| 2858 | let route = preflight_route( |
| 2859 | "implementer", |
| 2860 | "zai", |
| 2861 | crate::config::ZAI_GLM_5_2_MODEL, |
| 2862 | &config, |
| 2863 | ) |
| 2864 | .expect("preflight"); |
| 2865 | |
| 2866 | assert_eq!(route.member_id, "implementer"); |
| 2867 | assert_eq!(route.provider_kind, "zai"); |
| 2868 | assert_eq!(route.wire_model, crate::config::ZAI_GLM_5_2_MODEL); |
| 2869 | assert!(!route.endpoint.host.is_empty()); |
| 2870 | assert!(!route.endpoint.host.contains('/')); |
| 2871 | assert_eq!(route.credential, CredentialReadiness::Configured); |
| 2872 | route.require_ready().expect("ready"); |
| 2873 | |
| 2874 | // The receipt and the child spawn read the same canonical wire model. |
| 2875 | assert_eq!(route.frozen().model, route.wire_model); |
| 2876 | } |
| 2877 | |
| 2878 | /// A keyless local provider is valid, and is decided without a probe. |
| 2879 | #[test] |
| 2880 | fn a_keyless_local_provider_preflights_as_ready() { |
| 2881 | let _env_lock = crate::test_support::lock_test_env(); |
| 2882 | let config = Config { |
| 2883 | provider: Some("ollama".to_string()), |
| 2884 | ..Default::default() |
| 2885 | }; |
| 2886 | |
| 2887 | let Ok(route) = preflight_route("worker", "ollama", "qwen3", &config) else { |
| 2888 | // A model id this build does not know is a different failure than |
| 2889 | // the one under test; skip rather than assert on the catalog. |
| 2890 | return; |
| 2891 | }; |
| 2892 | assert_eq!(route.credential, CredentialReadiness::KeylessLocal); |
| 2893 | assert!(route.credential.is_ready()); |
| 2894 | route.require_ready().expect("keyless local is valid"); |
| 2895 | assert!(route.endpoint.local, "a local runtime is marked local"); |
| 2896 | } |
| 2897 | |
| 2898 | #[test] |
| 2899 | fn ollama_cloud_and_custom_remote_preflight_require_route_scoped_credentials() { |
| 2900 | let _env_lock = crate::test_support::lock_test_env(); |
| 2901 | let temp = tempfile::tempdir().expect("isolated credential home"); |
| 2902 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", temp.path()); |
| 2903 | let _backend = crate::test_support::EnvVarGuard::set("CODEWHALE_SECRET_BACKEND", "file"); |
| 2904 | let _ollama_cloud_key = crate::test_support::EnvVarGuard::remove("OLLAMA_CLOUD_API_KEY"); |
| 2905 | let _ollama_key = crate::test_support::EnvVarGuard::remove("OLLAMA_API_KEY"); |
| 2906 | let _cli_source = crate::test_support::EnvVarGuard::remove("DEEPSEEK_API_KEY_SOURCE"); |
| 2907 | let _cli_key = crate::test_support::EnvVarGuard::remove("CODEWHALE_CLI_API_KEY"); |
| 2908 | codewhale_secrets::Secrets::auto_detect() |
| 2909 | .set("ollama", "legacy-cloud-key") |
| 2910 | .expect("seed released Ollama Cloud slot"); |
| 2911 | |
| 2912 | let cloud = Config { |
| 2913 | provider: Some("deepseek".to_string()), |
| 2914 | providers: Some(crate::config::ProvidersConfig { |
| 2915 | ollama: crate::config::ProviderConfig { |
| 2916 | base_url: Some(codewhale_config::provider::OLLAMA_CLOUD_BASE_URL.to_string()), |
| 2917 | ..Default::default() |
| 2918 | }, |
| 2919 | ..Default::default() |
| 2920 | }), |
| 2921 | ..Default::default() |
| 2922 | }; |
| 2923 | let cloud_route = preflight_route( |
| 2924 | "cloud-worker", |
| 2925 | "ollama", |
| 2926 | crate::config::DEFAULT_OLLAMA_MODEL, |
| 2927 | &cloud, |
| 2928 | ) |
| 2929 | .expect("official Cloud route"); |
| 2930 | assert_eq!(cloud_route.provider_id, "ollama-cloud"); |
| 2931 | assert_eq!(cloud_route.provider_config_id.as_deref(), Some("ollama")); |
| 2932 | assert_eq!(cloud_route.provider_kind, "ollama-cloud"); |
| 2933 | assert_eq!(cloud_route.credential, CredentialReadiness::Configured); |
| 2934 | assert!(!cloud_route.endpoint.local); |
| 2935 | cloud_route.require_ready().expect("Cloud env key is ready"); |
| 2936 | |
| 2937 | let custom_remote = Config { |
| 2938 | provider: Some("ollama".to_string()), |
| 2939 | providers: Some(crate::config::ProvidersConfig { |
| 2940 | ollama: crate::config::ProviderConfig { |
| 2941 | base_url: Some("https://ollama-gateway.example.test/v1".to_string()), |
| 2942 | ..Default::default() |
| 2943 | }, |
| 2944 | ..Default::default() |
| 2945 | }), |
| 2946 | ..Default::default() |
| 2947 | }; |
| 2948 | let custom_route = preflight_route( |
| 2949 | "custom-worker", |
| 2950 | "ollama", |
| 2951 | crate::config::DEFAULT_OLLAMA_MODEL, |
| 2952 | &custom_remote, |
| 2953 | ) |
| 2954 | .expect("custom route still resolves structurally"); |
| 2955 | assert!(matches!( |
| 2956 | custom_route.credential, |
| 2957 | CredentialReadiness::Missing { .. } |
| 2958 | )); |
| 2959 | assert!(!custom_route.endpoint.local); |
| 2960 | assert!(custom_route.require_ready().is_err()); |
| 2961 | } |
| 2962 | |
| 2963 | #[tokio::test] |
| 2964 | async fn legacy_ollama_cloud_fleet_start_builds_clients_from_the_frozen_source_route() { |
| 2965 | let _env_lock = crate::test_support::lock_test_env(); |
| 2966 | let temp = tempfile::tempdir().expect("isolated credential home"); |
| 2967 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", temp.path()); |
| 2968 | let _backend = crate::test_support::EnvVarGuard::set("CODEWHALE_SECRET_BACKEND", "file"); |
| 2969 | let _cloud_env = crate::test_support::EnvVarGuard::remove("OLLAMA_CLOUD_API_KEY"); |
| 2970 | let _official_env = crate::test_support::EnvVarGuard::remove("OLLAMA_API_KEY"); |
| 2971 | codewhale_secrets::Secrets::auto_detect() |
| 2972 | .set("ollama", "legacy-cloud-fleet-key") |
| 2973 | .expect("seed released Ollama Cloud slot"); |
| 2974 | |
| 2975 | let config = Config { |
| 2976 | provider: Some("deepseek".to_string()), |
| 2977 | providers: Some(crate::config::ProvidersConfig { |
| 2978 | ollama: crate::config::ProviderConfig { |
| 2979 | base_url: Some(codewhale_config::provider::OLLAMA_CLOUD_BASE_URL.to_string()), |
| 2980 | model: Some(crate::config::DEFAULT_OLLAMA_CLOUD_MODEL.to_string()), |
| 2981 | ..Default::default() |
| 2982 | }, |
| 2983 | ..Default::default() |
| 2984 | }), |
| 2985 | ..Default::default() |
| 2986 | }; |
| 2987 | let document = FleetDocument::parse(&format!( |
| 2988 | r#" |
| 2989 | name = "glm-pair" |
| 2990 | schema = "exact" |
| 2991 | |
| 2992 | [[members]] |
| 2993 | id = "cloud-worker" |
| 2994 | role = "builder" |
| 2995 | provider = "ollama" |
| 2996 | model = "{}" |
| 2997 | reasoning = "medium" |
| 2998 | permissions = "read_only" |
| 2999 | "#, |
| 3000 | crate::config::DEFAULT_OLLAMA_CLOUD_MODEL |
| 3001 | )) |
| 3002 | .expect("legacy Cloud Fleet parses"); |
| 3003 | |
| 3004 | // `capture` is the real Workflow-start path: it preflights readiness, |
| 3005 | // constructs every worker client, and freezes the snapshot. |
| 3006 | let workflow = ExactFleetWorkflow::capture( |
| 3007 | &document, |
| 3008 | id(), |
| 3009 | "2026-08-14T00:00:00Z", |
| 3010 | Some(&config), |
| 3011 | &[], |
| 3012 | ) |
| 3013 | .expect("legacy Cloud Fleet starts"); |
| 3014 | let route = workflow |
| 3015 | .preflight |
| 3016 | .worker("cloud-worker") |
| 3017 | .expect("preflighted worker"); |
| 3018 | assert_eq!(route.provider_id, "ollama-cloud"); |
| 3019 | assert_eq!(route.provider_config_id.as_deref(), Some("ollama")); |
| 3020 | |
| 3021 | let binding = workflow |
| 3022 | .bind_member(Some("cloud-worker"), None, full_session()) |
| 3023 | .expect("worker binds"); |
| 3024 | let launch = workflow |
| 3025 | .route_admitted_task(&binding, "verify the frozen Cloud route") |
| 3026 | .await |
| 3027 | .expect("manual-tier launch needs no provider call"); |
| 3028 | assert_eq!(launch.provider, "ollama-cloud"); |
| 3029 | assert_eq!(launch.receipt.provider, "ollama-cloud"); |
| 3030 | |
| 3031 | let router_profile = ReasoningRouterProfile::parse(&format!( |
| 3032 | r#" |
| 3033 | name = "legacy-cloud-router" |
| 3034 | schema = "reasoning_router" |
| 3035 | provider = "ollama" |
| 3036 | model = "{}" |
| 3037 | call_reasoning = "low" |
| 3038 | "#, |
| 3039 | crate::config::DEFAULT_OLLAMA_CLOUD_MODEL |
| 3040 | )) |
| 3041 | .expect("legacy Cloud router profile parses"); |
| 3042 | let captured = |
| 3043 | CapturedReasoningRouter::from_profile(&router_profile, "workspace".to_string()); |
| 3044 | let live = LiveFleetRouter::bind(&captured, &config) |
| 3045 | .expect("legacy Cloud Router binds its source table and secret"); |
| 3046 | assert_eq!(live.route.provider_id, "ollama-cloud"); |
| 3047 | assert_eq!(live.route.provider_config_id.as_deref(), Some("ollama")); |
| 3048 | assert_eq!(live.client.api_provider(), ProviderKind::OllamaCloud); |
| 3049 | assert_eq!( |
| 3050 | live.client.base_url(), |
| 3051 | codewhale_config::provider::OLLAMA_CLOUD_BASE_URL |
| 3052 | ); |
| 3053 | } |
| 3054 | |
| 3055 | /// A tier label is a selector concept; what a request may carry is a |
| 3056 | /// provider concept. The value placed on a call must come from the route |
| 3057 | /// normalizer the client actually uses, or a Codex-routed Router is called |
| 3058 | /// at the provider default while its receipt claims a tier. |
| 3059 | #[test] |
| 3060 | fn a_call_reasoning_value_is_shaped_by_the_configured_route_not_a_tier_label() { |
| 3061 | // A tiered non-Codex route spells the tiers the ordinary way, after |
| 3062 | // the same route normalization the client performs (first-party |
| 3063 | // DeepSeek keeps a real `low`; medium still rounds up to high). |
| 3064 | for (tier, expected) in [ |
| 3065 | (ReasoningTier::Off, "off"), |
| 3066 | (ReasoningTier::High, "high"), |
| 3067 | (ReasoningTier::Max, "max"), |
| 3068 | ] { |
| 3069 | assert_eq!( |
| 3070 | route_reasoning_setting( |
| 3071 | ProviderKind::Deepseek, |
| 3072 | crate::config::DEFAULT_DEEPSEEK_BASE_URL, |
| 3073 | "deepseek-v4-pro", |
| 3074 | tier, |
| 3075 | ), |
| 3076 | expected, |
| 3077 | "{tier:?} on a deepseek route" |
| 3078 | ); |
| 3079 | } |
| 3080 | |
| 3081 | // Codex is the case a bare tier label gets wrong in both directions: |
| 3082 | // it has no `off`, and its ladder now spells three separate top rungs |
| 3083 | // (`xhigh`, `max`, `ultra`) that the roster publishes per model. |
| 3084 | let codex = |tier| { |
| 3085 | route_reasoning_setting( |
| 3086 | ProviderKind::OpenaiCodex, |
| 3087 | "https://chatgpt.com/backend-api/codex", |
| 3088 | "gpt-5.6-codex", |
| 3089 | tier, |
| 3090 | ) |
| 3091 | }; |
| 3092 | assert_eq!(codex(ReasoningTier::Max), "max"); |
| 3093 | assert_eq!(codex(ReasoningTier::Low), "low"); |
| 3094 | assert_ne!( |
| 3095 | codex(ReasoningTier::Off), |
| 3096 | "off", |
| 3097 | "an always-thinking route cannot be asked for `off`; sending the label \ |
| 3098 | would take the provider default while the receipt claimed a tier" |
| 3099 | ); |
| 3100 | } |
| 3101 | |
| 3102 | /// An unresolvable provider fails preflight rather than reaching a launch. |
| 3103 | #[test] |
| 3104 | fn an_unresolvable_provider_fails_preflight() { |
| 3105 | let config = Config::default(); |
| 3106 | let err = preflight_route("implementer", "not-a-provider", "whatever", &config) |
| 3107 | .expect_err("unresolvable provider"); |
| 3108 | assert!(matches!(err, PreflightError::ProviderUnresolved { .. })); |
| 3109 | } |
| 3110 | |
| 3111 | // ── Receipts ─────────────────────────────────────────────────────────── |
| 3112 | |
| 3113 | /// The receipt is the durable artifact. It must carry every side of the |
| 3114 | /// decision — including which service chose the tier and what that call was |
| 3115 | /// configured to cost — and must store no task text, path, or key. |
| 3116 | #[tokio::test] |
| 3117 | async fn a_launch_receipt_names_the_service_route_and_call_cost_without_content() { |
| 3118 | let workflow = workflow_with( |
| 3119 | Some(StaticFleetRouter::new(r#"{"reasoning":"max"}"#)), |
| 3120 | GLM_FLEET, |
| 3121 | ); |
| 3122 | let binding = workflow |
| 3123 | .bind_member(None, Some("builder"), full_session()) |
| 3124 | .expect("bind"); |
| 3125 | |
| 3126 | let launch = workflow |
| 3127 | .route_admitted_task(&binding, "refactor /Users/hunter/app with ZAI_API_KEY=zzz") |
| 3128 | .await |
| 3129 | .expect("launch"); |
| 3130 | let receipt = &launch.receipt; |
| 3131 | |
| 3132 | assert_eq!(receipt.fleet, "workspace/glm-pair"); |
| 3133 | assert_eq!(receipt.schema_kind, "exact"); |
| 3134 | assert_eq!(receipt.member_id, "implementer"); |
| 3135 | assert_eq!(receipt.member_role, "implement"); |
| 3136 | assert_eq!(receipt.provider, "zai"); |
| 3137 | assert_eq!(receipt.model, "glm-5"); |
| 3138 | assert_eq!(receipt.requested_reasoning, "auto"); |
| 3139 | assert_eq!(receipt.effective_reasoning, "max"); |
| 3140 | assert_eq!(receipt.selection_source, "fleet_router"); |
| 3141 | assert!(!receipt.content_hash.is_empty()); |
| 3142 | |
| 3143 | // The service is labelled as a service, with its exact route and the |
| 3144 | // configured requested → provider-effective call reasoning. |
| 3145 | let router = receipt.router.as_ref().expect("router identity"); |
| 3146 | assert_eq!(router.service_kind, "reasoning_router"); |
| 3147 | assert_eq!(router.qualified(), "workspace/luna-low"); |
| 3148 | assert_eq!(router.provider, "openai"); |
| 3149 | assert_eq!(router.model, "gpt-5.6-luna"); |
| 3150 | let call = router.call.as_ref().expect("call disclosure"); |
| 3151 | assert_eq!(call.requested, "low"); |
| 3152 | assert_eq!(call.effective, "low"); |
| 3153 | assert_eq!(call.provider_effective, "low"); |
| 3154 | |
| 3155 | // Cross-provider inference happened (zai worker, openai router) and is |
| 3156 | // disclosed rather than implied away. |
| 3157 | assert!(receipt.cross_provider_inference); |
| 3158 | assert!( |
| 3159 | receipt.transport.contains("different provider"), |
| 3160 | "{}", |
| 3161 | receipt.transport |
| 3162 | ); |
| 3163 | |
| 3164 | // Disclosure without content. |
| 3165 | let disclosure = receipt.routing_summary.as_ref().expect("disclosure"); |
| 3166 | assert!(disclosure.transmitted_bytes > 0); |
| 3167 | assert!(disclosure.content_hash.starts_with("sha256:")); |
| 3168 | assert!(disclosure.redacted); |
| 3169 | |
| 3170 | let json = serde_json::to_string(receipt).expect("serialize"); |
| 3171 | for forbidden in ["/Users/", "/home/", ".toml", "api_key", "zzz", "refactor"] { |
| 3172 | assert!(!json.contains(forbidden), "{forbidden} in {json}"); |
| 3173 | } |
| 3174 | |
| 3175 | // The visible line names every side and echoes no content. |
| 3176 | let line = receipt.line(); |
| 3177 | for expected in [ |
| 3178 | "requested=auto", |
| 3179 | "effective=max", |
| 3180 | "source=fleet_router", |
| 3181 | "reasoning_router:workspace/luna-low", |
| 3182 | "router_call_requested=low", |
| 3183 | ] { |
| 3184 | assert!(line.contains(expected), "{expected} missing from {line}"); |
| 3185 | } |
| 3186 | assert!(!line.contains("refactor"), "{line}"); |
| 3187 | } |
| 3188 | |
| 3189 | /// A member's semantic role and its Runtime posture are separate |
| 3190 | /// facts and the receipt keeps both. An operator who named a member |
| 3191 | /// `auditor` must see `auditor` on the receipt, while the surface actually |
| 3192 | /// selected (`explore`, the fail-closed posture an undeclared role gets |
| 3193 | /// since #5575) is disclosed rather than substituted for the name. |
| 3194 | #[tokio::test] |
| 3195 | async fn a_receipt_records_the_posture_without_renaming_the_members_role() { |
| 3196 | const AUDIT_FLEET: &str = r#" |
| 3197 | name = "glm-pair" |
| 3198 | schema = "exact" |
| 3199 | |
| 3200 | [[members]] |
| 3201 | id = "auditor" |
| 3202 | role = "auditor" |
| 3203 | provider = "zai" |
| 3204 | model = "glm-5" |
| 3205 | reasoning = "high" |
| 3206 | permissions = "read_only" |
| 3207 | "#; |
| 3208 | let workflow = workflow_with(None, AUDIT_FLEET); |
| 3209 | let binding = workflow |
| 3210 | .bind_member(None, Some("auditor"), full_session()) |
| 3211 | .expect("bind"); |
| 3212 | |
| 3213 | // Enforcement uses the posture; it is not the operator's role name. |
| 3214 | assert_eq!(binding.member_role, "auditor"); |
| 3215 | assert_eq!(binding.authority.posture_role, "explore"); |
| 3216 | |
| 3217 | let launch = workflow |
| 3218 | .route_admitted_task(&binding, "review the queue") |
| 3219 | .await |
| 3220 | .expect("launch"); |
| 3221 | let receipt = &launch.receipt; |
| 3222 | |
| 3223 | assert_eq!(receipt.member_role, "auditor"); |
| 3224 | assert_eq!(receipt.posture_role.as_deref(), Some("explore")); |
| 3225 | let line = receipt.line(); |
| 3226 | assert!(line.contains("(role auditor)"), "{line}"); |
| 3227 | assert!(line.contains("posture=explore"), "{line}"); |
| 3228 | } |
| 3229 | |
| 3230 | // ── Search roots: where a workspace Fleet lives ──────────────────────── |
| 3231 | |
| 3232 | /// The Fleet store saves workspace Fleets under `<workspace>/.codewhale`, |
| 3233 | /// so that is the primary `workspace` origin; the workspace root stays a |
| 3234 | /// second origin for checked-in `fleets/<name>.toml` rosters. |
| 3235 | #[test] |
| 3236 | fn workspace_fleets_load_from_dot_codewhale_and_the_legacy_root() { |
| 3237 | let _lock = crate::test_support::lock_test_env(); |
| 3238 | let home = tempfile::tempdir().expect("home"); |
| 3239 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3240 | let ws = tempfile::tempdir().expect("workspace"); |
| 3241 | |
| 3242 | let saved = ws.path().join(".codewhale").join("fleets"); |
| 3243 | std::fs::create_dir_all(&saved).expect("saved fleets dir"); |
| 3244 | std::fs::write(saved.join("glm-pair.toml"), GLM_FLEET).expect("write saved"); |
| 3245 | let (document, id) = |
| 3246 | load_fleet_document("glm-pair", ws.path(), None).expect("saved fleet loads"); |
| 3247 | assert_eq!(document.name(), "glm-pair"); |
| 3248 | assert_eq!(id.origin, "workspace"); |
| 3249 | |
| 3250 | let checked_in = ws.path().join("fleets"); |
| 3251 | std::fs::create_dir_all(&checked_in).expect("checked-in fleets dir"); |
| 3252 | std::fs::write( |
| 3253 | checked_in.join("stopship.toml"), |
| 3254 | "name = \"stopship\"\n\n[roles]\nscout = \"scout\"\n", |
| 3255 | ) |
| 3256 | .expect("write checked-in"); |
| 3257 | let (document, id) = |
| 3258 | load_fleet_document("stopship", ws.path(), None).expect("checked-in fleet still loads"); |
| 3259 | assert_eq!(document.name(), "stopship"); |
| 3260 | assert_eq!(id.origin, "workspace_root"); |
| 3261 | |
| 3262 | // An exact Fleet in both workspace origins is ambiguous, and each |
| 3263 | // origin can be named explicitly. |
| 3264 | std::fs::write(checked_in.join("glm-pair.toml"), GLM_FLEET).expect("write twin"); |
| 3265 | assert!(matches!( |
| 3266 | load_fleet_document("glm-pair", ws.path(), None), |
| 3267 | Err(NamedFleetError::AmbiguousFleet { .. }) |
| 3268 | )); |
| 3269 | let (_, id) = |
| 3270 | load_fleet_document("workspace_root/glm-pair", ws.path(), None).expect("qualified"); |
| 3271 | assert_eq!(id.origin, "workspace_root"); |
| 3272 | } |
| 3273 | |
| 3274 | #[test] |
| 3275 | fn fleet_names_that_leave_the_fleets_directory_are_refused() { |
| 3276 | let _lock = crate::test_support::lock_test_env(); |
| 3277 | let home = tempfile::tempdir().expect("home"); |
| 3278 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3279 | let outer = tempfile::tempdir().expect("outer"); |
| 3280 | let ws = outer.path().join("ws"); |
| 3281 | std::fs::create_dir_all(ws.join("fleets")).expect("fleets dir"); |
| 3282 | std::fs::write(outer.path().join("outside.toml"), GLM_FLEET).expect("outside"); |
| 3283 | let absolute = outer.path().join("outside"); |
| 3284 | |
| 3285 | for name in [ |
| 3286 | absolute.to_string_lossy().to_string(), |
| 3287 | "workspace_root/../../outside".to_string(), |
| 3288 | "codewhale_home/../outside".to_string(), |
| 3289 | "user/../outside".to_string(), |
| 3290 | ] { |
| 3291 | let err = load_fleet_document(&name, &ws, None).expect_err(&name); |
| 3292 | assert!( |
| 3293 | matches!(err, NamedFleetError::InvalidName), |
| 3294 | "{name}: expected InvalidName, got {err:?}" |
| 3295 | ); |
| 3296 | let message = err.to_string(); |
| 3297 | assert!( |
| 3298 | !message.contains(outer.path().to_string_lossy().as_ref()), |
| 3299 | "{message}" |
| 3300 | ); |
| 3301 | } |
| 3302 | } |
| 3303 | |
| 3304 | /// A Fleet saved through the store at workspace scope is found by the |
| 3305 | /// Workflow loader instead of being reported missing. Today the store's |
| 3306 | /// `schema = "fleet"` revision-2 document is not a schema the Workflow |
| 3307 | /// loader parses, so the load names that exact file and its schema; if a |
| 3308 | /// v2 bridge lands, the same call succeeds from the `workspace` origin. |
| 3309 | #[test] |
| 3310 | fn a_store_saved_workspace_fleet_is_found_by_load_fleet_document() { |
| 3311 | use crate::fleet::store::{FleetFile, FleetScope, save_fleet}; |
| 3312 | |
| 3313 | let _lock = crate::test_support::lock_test_env(); |
| 3314 | let home = tempfile::tempdir().expect("home"); |
| 3315 | let _home = crate::test_support::EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3316 | let ws = tempfile::tempdir().expect("workspace"); |
| 3317 | |
| 3318 | let fleet = FleetFile::new("Folder Pair".to_string(), None).expect("fleet"); |
| 3319 | let path = save_fleet(&fleet, FleetScope::Workspace, ws.path()).expect("save"); |
| 3320 | |
| 3321 | match load_fleet_document(&fleet.file_slug(), ws.path(), None) { |
| 3322 | // A v2 bridge may label the store scope `folder` rather than the |
| 3323 | // `workspace` search-root origin; either names this workspace. |
| 3324 | Ok((_, id)) => assert!( |
| 3325 | matches!(id.origin.as_str(), "workspace" | "folder"), |
| 3326 | "{}", |
| 3327 | id.origin |
| 3328 | ), |
| 3329 | Err(err) => { |
| 3330 | assert!( |
| 3331 | !matches!(err, NamedFleetError::NotFound(_)), |
| 3332 | "the saved Fleet must be found, got {err}" |
| 3333 | ); |
| 3334 | let message = err.to_string(); |
| 3335 | assert!(message.contains(&path.display().to_string()), "{message}"); |
| 3336 | } |
| 3337 | } |
| 3338 | } |
| 3339 | } |
| 3340 | |
| 3341 | /// `workflow(fleet:)` resolving saved v2 Fleets (store-first lookup, freeze |
| 3342 | /// into an exact snapshot, ambiguity against legacy/exact files). |
| 3343 | #[cfg(test)] |
| 3344 | mod saved_fleet_tests { |
| 3345 | use super::*; |
| 3346 | use crate::fleet::store::{FleetFile, FleetMember, FleetOperator, FleetScope, save_fleet}; |
| 3347 | use crate::test_support::{EnvVarGuard, lock_test_env}; |
| 3348 | |
| 3349 | fn member(id: &str, pin: Option<(&str, &str)>) -> FleetMember { |
| 3350 | FleetMember { |
| 3351 | id: id.to_string(), |
| 3352 | display_name: None, |
| 3353 | shortlist: false, |
| 3354 | role: String::new(), |
| 3355 | model: pin.map(|(_, model)| model.to_string()), |
| 3356 | provider: pin.map(|(provider, _)| provider.to_string()), |
| 3357 | reasoning: None, |
| 3358 | instructions: None, |
| 3359 | requires: Vec::new(), |
| 3360 | } |
| 3361 | } |
| 3362 | |
| 3363 | fn fleet(name: &str, members: Vec<FleetMember>) -> FleetFile { |
| 3364 | let mut fleet = FleetFile::new(name.to_string(), None).expect("fleet"); |
| 3365 | fleet.members = members; |
| 3366 | fleet |
| 3367 | } |
| 3368 | |
| 3369 | fn zai_session() -> Config { |
| 3370 | Config { |
| 3371 | provider: Some("zai".to_string()), |
| 3372 | reasoning_effort: Some("high".to_string()), |
| 3373 | ..Default::default() |
| 3374 | } |
| 3375 | } |
| 3376 | |
| 3377 | fn session_ceiling() -> PermissionCeiling { |
| 3378 | PermissionCeiling { |
| 3379 | write: true, |
| 3380 | network_tool: true, |
| 3381 | shell: codewhale_workflow::ShellCeiling::Full, |
| 3382 | delegation_depth: codewhale_config::DEFAULT_SPAWN_DEPTH, |
| 3383 | tools: true, |
| 3384 | } |
| 3385 | } |
| 3386 | |
| 3387 | #[test] |
| 3388 | fn a_personal_saved_fleet_loads_as_a_frozen_exact_document() { |
| 3389 | let _lock = lock_test_env(); |
| 3390 | let home = tempfile::tempdir().expect("home"); |
| 3391 | let _home = EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3392 | let ws = tempfile::tempdir().expect("workspace"); |
| 3393 | let config = zai_session(); |
| 3394 | let saved = fleet( |
| 3395 | "My fleet", |
| 3396 | vec![ |
| 3397 | member("builder", Some(("zai", crate::config::ZAI_GLM_5_2_MODEL))), |
| 3398 | member("reviewer", None), |
| 3399 | ], |
| 3400 | ); |
| 3401 | let path = save_fleet(&saved, FleetScope::Personal, ws.path()).expect("save"); |
| 3402 | |
| 3403 | let (document, id) = |
| 3404 | load_fleet_document("My fleet", ws.path(), Some(&config)).expect("v2 loads"); |
| 3405 | |
| 3406 | assert_eq!(id.origin, "user"); |
| 3407 | assert_eq!(id.name, "my-fleet"); |
| 3408 | assert_eq!(document.source_path(), Some(path.as_path())); |
| 3409 | let exact = document.exact().expect("frozen into the exact schema"); |
| 3410 | let builder = exact.member("builder").expect("builder"); |
| 3411 | assert_eq!( |
| 3412 | (builder.provider.as_str(), builder.model.as_str()), |
| 3413 | ("zai", crate::config::ZAI_GLM_5_2_MODEL) |
| 3414 | ); |
| 3415 | // No pin and no operator: the member inherits the live session route |
| 3416 | // and tier, resolved now rather than left open. |
| 3417 | let reviewer = exact.member("reviewer").expect("reviewer"); |
| 3418 | assert_eq!( |
| 3419 | reviewer.provider, |
| 3420 | config.active_provider_identity().unwrap().key.as_str() |
| 3421 | ); |
| 3422 | assert_eq!(reviewer.model, config.default_model()); |
| 3423 | assert_eq!(reviewer.reasoning.as_str(), "high"); |
| 3424 | |
| 3425 | // The slug also resolves, and so does the qualified store scope. |
| 3426 | load_fleet_document("my-fleet", ws.path(), Some(&config)).expect("slug loads"); |
| 3427 | load_fleet_document("user/My fleet", ws.path(), Some(&config)).expect("user/ loads"); |
| 3428 | } |
| 3429 | |
| 3430 | #[test] |
| 3431 | fn a_workspace_saved_fleet_loads_and_members_follow_the_operator_route() { |
| 3432 | let _lock = lock_test_env(); |
| 3433 | let home = tempfile::tempdir().expect("home"); |
| 3434 | let _home = EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3435 | let ws = tempfile::tempdir().expect("workspace"); |
| 3436 | let mut saved = fleet("reviewers", vec![member("auditor", None)]); |
| 3437 | saved.operator = Some(FleetOperator { |
| 3438 | provider: "zai".to_string(), |
| 3439 | model: crate::config::ZAI_GLM_5_2_MODEL.to_string(), |
| 3440 | reasoning: Some("low".to_string()), |
| 3441 | }); |
| 3442 | let path = save_fleet(&saved, FleetScope::Workspace, ws.path()).expect("save"); |
| 3443 | assert!(path.starts_with(ws.path().join(".codewhale").join("fleets"))); |
| 3444 | |
| 3445 | // No session config is needed: nothing inherits the session route. |
| 3446 | let (document, id) = load_fleet_document("reviewers", ws.path(), None).expect("loads"); |
| 3447 | assert_eq!(id.origin, "folder"); |
| 3448 | let auditor = document.exact().unwrap().member("auditor").unwrap(); |
| 3449 | assert_eq!(auditor.model, crate::config::ZAI_GLM_5_2_MODEL); |
| 3450 | assert_eq!(auditor.reasoning.as_str(), "low"); |
| 3451 | } |
| 3452 | |
| 3453 | #[test] |
| 3454 | fn a_saved_fleet_colliding_with_an_exact_file_is_ambiguous_and_names_both_paths() { |
| 3455 | let _lock = lock_test_env(); |
| 3456 | let home = tempfile::tempdir().expect("home"); |
| 3457 | let _home = EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3458 | let ws = tempfile::tempdir().expect("workspace"); |
| 3459 | let saved_path = save_fleet( |
| 3460 | &fleet("glm-pair", vec![member("builder", None)]), |
| 3461 | FleetScope::Personal, |
| 3462 | ws.path(), |
| 3463 | ) |
| 3464 | .expect("save"); |
| 3465 | let exact_dir = ws.path().join("fleets"); |
| 3466 | std::fs::create_dir_all(&exact_dir).unwrap(); |
| 3467 | let exact_path = exact_dir.join("glm-pair.toml"); |
| 3468 | std::fs::write( |
| 3469 | &exact_path, |
| 3470 | "name = \"glm-pair\"\nschema = \"exact\"\n\n[[members]]\nid = \"builder\"\nprovider = \"zai\"\nmodel = \"glm-5\"\n", |
| 3471 | ) |
| 3472 | .unwrap(); |
| 3473 | |
| 3474 | let error = load_fleet_document("glm-pair", ws.path(), Some(&zai_session())) |
| 3475 | .expect_err("a v2 and an exact Fleet of one name must not shadow each other"); |
| 3476 | let message = error.to_string(); |
| 3477 | assert!( |
| 3478 | matches!(error, NamedFleetError::AmbiguousFleet { .. }), |
| 3479 | "{message}" |
| 3480 | ); |
| 3481 | assert!( |
| 3482 | message.contains(&saved_path.display().to_string()), |
| 3483 | "{message}" |
| 3484 | ); |
| 3485 | assert!( |
| 3486 | message.contains(&exact_path.display().to_string()), |
| 3487 | "{message}" |
| 3488 | ); |
| 3489 | |
| 3490 | // Qualifying either side resolves it. |
| 3491 | let (document, _) = |
| 3492 | load_fleet_document("user/glm-pair", ws.path(), Some(&zai_session())).expect("v2"); |
| 3493 | assert_eq!(document.source_path(), Some(saved_path.as_path())); |
| 3494 | } |
| 3495 | |
| 3496 | /// Saved Fleets carry session-vocabulary reasoning (an imported agent |
| 3497 | /// profile stores `ultra`, `xhigh`, `minimal`); freezing maps it onto an |
| 3498 | /// exact tier instead of failing the exact parser, a blank value inherits, |
| 3499 | /// and an unknown value is refused with the member named. |
| 3500 | #[test] |
| 3501 | fn saved_fleet_reasoning_in_session_vocabulary_freezes_to_exact_tiers() { |
| 3502 | let _lock = lock_test_env(); |
| 3503 | let home = tempfile::tempdir().expect("home"); |
| 3504 | let _home = EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3505 | let ws = tempfile::tempdir().expect("workspace"); |
| 3506 | let pin = Some(("zai", crate::config::ZAI_GLM_5_2_MODEL)); |
| 3507 | let mut ultra = member("ultra", pin); |
| 3508 | ultra.reasoning = Some("ultra".to_string()); |
| 3509 | let mut minimal = member("minimal", pin); |
| 3510 | minimal.reasoning = Some("minimal".to_string()); |
| 3511 | let mut blank = member("blank", pin); |
| 3512 | blank.reasoning = Some(" ".to_string()); |
| 3513 | let mut saved = fleet("tiers", vec![ultra, minimal, blank]); |
| 3514 | saved.operator = Some(FleetOperator { |
| 3515 | provider: "zai".to_string(), |
| 3516 | model: crate::config::ZAI_GLM_5_2_MODEL.to_string(), |
| 3517 | reasoning: Some("xhigh".to_string()), |
| 3518 | }); |
| 3519 | save_fleet(&saved, FleetScope::Workspace, ws.path()).expect("save"); |
| 3520 | |
| 3521 | let (document, _) = load_fleet_document("tiers", ws.path(), None).expect("freezes"); |
| 3522 | let exact = document.exact().expect("exact"); |
| 3523 | let tier = |id: &str| exact.member(id).expect(id).reasoning.as_str(); |
| 3524 | assert_eq!(tier("ultra"), "max"); |
| 3525 | assert_eq!(tier("minimal"), "low"); |
| 3526 | // Blank inherits the operator's `xhigh`, which is the `max` tier. |
| 3527 | assert_eq!(tier("blank"), "max"); |
| 3528 | |
| 3529 | let mut bad = member("bad", pin); |
| 3530 | bad.reasoning = Some("turbo".to_string()); |
| 3531 | save_fleet( |
| 3532 | &fleet("bad-tier", vec![bad]), |
| 3533 | FleetScope::Workspace, |
| 3534 | ws.path(), |
| 3535 | ) |
| 3536 | .expect("save"); |
| 3537 | let error = load_fleet_document("bad-tier", ws.path(), None).expect_err("refused"); |
| 3538 | assert!( |
| 3539 | error.to_string().contains("member `bad` reasoning"), |
| 3540 | "{error}" |
| 3541 | ); |
| 3542 | } |
| 3543 | |
| 3544 | #[test] |
| 3545 | fn member_instructions_are_refused_rather_than_silently_dropped() { |
| 3546 | let _lock = lock_test_env(); |
| 3547 | let home = tempfile::tempdir().expect("home"); |
| 3548 | let _home = EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3549 | let ws = tempfile::tempdir().expect("workspace"); |
| 3550 | let mut coach = member("coach", Some(("zai", crate::config::ZAI_GLM_5_2_MODEL))); |
| 3551 | coach.instructions = Some("Always cite sources.".to_string()); |
| 3552 | save_fleet( |
| 3553 | &fleet("coached", vec![coach]), |
| 3554 | FleetScope::Workspace, |
| 3555 | ws.path(), |
| 3556 | ) |
| 3557 | .expect("save"); |
| 3558 | |
| 3559 | let error = load_fleet_document("coached", ws.path(), None).expect_err("refused"); |
| 3560 | assert!( |
| 3561 | error.to_string().contains("`coach` has instructions"), |
| 3562 | "{error}" |
| 3563 | ); |
| 3564 | } |
| 3565 | |
| 3566 | /// The frozen snapshot is what runs: editing the saved file after capture |
| 3567 | /// moves nothing, and the inherited member's preflighted route is the same |
| 3568 | /// session route the snapshot names. |
| 3569 | #[test] |
| 3570 | fn frozen_routes_survive_a_mid_run_edit_and_inherit_matches_preflight() { |
| 3571 | let _lock = lock_test_env(); |
| 3572 | let home = tempfile::tempdir().expect("home"); |
| 3573 | let _home = EnvVarGuard::set("CODEWHALE_HOME", home.path()); |
| 3574 | let _key = EnvVarGuard::set("ZAI_API_KEY", "zai-key"); |
| 3575 | let ws = tempfile::tempdir().expect("workspace"); |
| 3576 | let config = zai_session(); |
| 3577 | let mut saved = fleet( |
| 3578 | "release", |
| 3579 | vec![ |
| 3580 | member("builder", Some(("zai", crate::config::ZAI_GLM_5_2_MODEL))), |
| 3581 | member("reviewer", None), |
| 3582 | ], |
| 3583 | ); |
| 3584 | save_fleet(&saved, FleetScope::Workspace, ws.path()).expect("save"); |
| 3585 | |
| 3586 | let (document, id) = |
| 3587 | load_fleet_document("release", ws.path(), Some(&config)).expect("loads"); |
| 3588 | let roots = fleet_search_roots(ws.path()); |
| 3589 | let workflow = ExactFleetWorkflow::capture( |
| 3590 | &document, |
| 3591 | id, |
| 3592 | "2026-09-22T00:00:00Z", |
| 3593 | Some(&config), |
| 3594 | &roots, |
| 3595 | ) |
| 3596 | .expect("capture"); |
| 3597 | |
| 3598 | // Edit the saved Fleet mid-run. |
| 3599 | saved.members[0].model = Some("glm-4.6".to_string()); |
| 3600 | save_fleet(&saved, FleetScope::Workspace, ws.path()).expect("re-save"); |
| 3601 | |
| 3602 | let builder = workflow |
| 3603 | .bind_member(Some("builder"), None, session_ceiling()) |
| 3604 | .expect("bind builder"); |
| 3605 | assert_eq!(builder.route.wire_model, crate::config::ZAI_GLM_5_2_MODEL); |
| 3606 | |
| 3607 | let reviewer = workflow |
| 3608 | .bind_member(Some("reviewer"), None, session_ceiling()) |
| 3609 | .expect("bind reviewer"); |
| 3610 | let frozen = workflow |
| 3611 | .snapshot() |
| 3612 | .members() |
| 3613 | .iter() |
| 3614 | .find(|member| member.id == "reviewer") |
| 3615 | .expect("reviewer in snapshot"); |
| 3616 | assert_eq!(frozen.route.model, config.default_model()); |
| 3617 | assert_eq!(reviewer.route.frozen().model, reviewer.route.wire_model); |
| 3618 | assert_eq!( |
| 3619 | reviewer.route.wire_model, |
| 3620 | crate::config::requested_model_for_provider( |
| 3621 | config.active_provider_identity().unwrap().provider, |
| 3622 | &config.default_model() |
| 3623 | ) |
| 3624 | .expect("session model is a known route") |
| 3625 | ); |
| 3626 | } |
| 3627 | } |
| 3628 |