| 1 | //! Fleet executor — runs a Fleet worker as a real `codewhale exec` subprocess. |
| 2 | //! |
| 3 | //! A Fleet worker IS a headless `codewhale exec` run. There is no separate |
| 4 | //! "Fleet worker" execution engine: the sub-agent runtime, full tool surface, |
| 5 | //! and recursion depth all come from the one `codewhale exec` runtime, so |
| 6 | //! Fleets and sub-agents are one substrate (not two moving targets). |
| 7 | //! |
| 8 | //! This module is the bridge: |
| 9 | //! - [`build_worker_exec_command`] turns a `FleetTaskSpec` + `FleetExecConfig` |
| 10 | //! into the `codewhale [route flags] exec --output-format stream-json …` |
| 11 | //! argv that a host adapter ([`super::host`]) launches locally or over SSH. |
| 12 | //! - [`map_exec_stream_line`] maps one stream-json line emitted by that worker |
| 13 | //! into a [`FleetWorkerEventPayload`] for the durable ledger, so the ledger |
| 14 | //! persists the worker's own event vocabulary instead of a simulated one. |
| 15 | //! - [`classify_worker_exit`] turns the process exit into a terminal event. |
| 16 | //! |
| 17 | //! The TUI/CLI/Runtime API observe the ledger's compact event stream — they |
| 18 | //! never render a child session, which is what keeps the orchestrator light at |
| 19 | //! high fanout. |
| 20 | |
| 21 | #![allow(dead_code)] |
| 22 | |
| 23 | use anyhow::Result; |
| 24 | use codewhale_config::FleetExecConfig; |
| 25 | use codewhale_protocol::fleet::{FleetHostSpec, FleetTaskSpec, FleetWorkerEventPayload}; |
| 26 | |
| 27 | use super::host::{FleetHostAdapter, FleetWorkerCommand}; |
| 28 | use super::profile::AgentProfile; |
| 29 | use super::task_spec::FleetWorkerFinalAnswer; |
| 30 | use super::worker_runtime::{ |
| 31 | fleet_task_prompt, fleet_task_prompt_with_profiles, fleet_worker_launch_reasoning_effort, |
| 32 | fleet_worker_launch_route, |
| 33 | }; |
| 34 | use crate::tools::spec::{ |
| 35 | ToolAuthorityEnvelope, ToolMutationAuthority, ToolShellAuthority, ToolVerificationAuthority, |
| 36 | }; |
| 37 | use crate::tools::subagent::{AgentWorkerSpec, FleetRole}; |
| 38 | |
| 39 | #[derive(Clone, Copy, Default)] |
| 40 | struct WorkerExecRoute<'a> { |
| 41 | model: Option<&'a str>, |
| 42 | provider: Option<&'a str>, |
| 43 | reasoning_effort: Option<&'a str>, |
| 44 | } |
| 45 | |
| 46 | #[derive(Clone, Copy, Default)] |
| 47 | struct WorkerExecLimits { |
| 48 | max_turns: Option<u32>, |
| 49 | max_tool_calls: Option<u32>, |
| 50 | } |
| 51 | |
| 52 | /// Resolve the executable used for Fleet worker subprocesses. |
| 53 | /// |
| 54 | /// Kept here so every long-lived surface (CLI and Runtime API) launches the |
| 55 | /// same configured worker binary instead of silently diverging. |
| 56 | pub fn configured_codewhale_binary() -> String { |
| 57 | std::env::var("CODEWHALE_FLEET_CODEWHALE_BINARY") |
| 58 | .ok() |
| 59 | .map(|value| value.trim().to_string()) |
| 60 | .filter(|value| !value.is_empty()) |
| 61 | .unwrap_or_else(|| "codewhale".to_string()) |
| 62 | } |
| 63 | |
| 64 | /// Build the `codewhale exec` argv that runs a fleet task headlessly. |
| 65 | /// |
| 66 | /// `--auto` is always passed: a headless worker has no human to approve tool |
| 67 | /// calls, so it runs with full (policy-gated) tool access. `--output-format |
| 68 | /// stream-json` makes the worker emit the NDJSON event stream this module |
| 69 | /// parses. A worker launched with the v0.9.1 machine-readable outer authority |
| 70 | /// cap is a truthful leaf (`max_spawn_depth = 0`): the nested-agent surface is |
| 71 | /// disabled until authority scopes can be intersected across |
| 72 | /// process/workspace boundaries. |
| 73 | /// |
| 74 | /// Secrets are NEVER placed on the argv: provider credentials are resolved by |
| 75 | /// the worker process from its own config/keyring exactly like an interactive |
| 76 | /// run. The host adapter additionally refuses secret-bearing env keys. The |
| 77 | /// `--provider` flag threaded by [`build_worker_exec_command_with_profiles`] is |
| 78 | /// a non-secret provider *identifier* only (#4093) — the worker still resolves |
| 79 | /// that provider's credentials from its own env/config, so this invariant |
| 80 | /// holds. |
| 81 | pub fn build_worker_exec_command( |
| 82 | codewhale_binary: &str, |
| 83 | task_spec: &FleetTaskSpec, |
| 84 | exec_config: &FleetExecConfig, |
| 85 | model: Option<&str>, |
| 86 | ) -> FleetWorkerCommand { |
| 87 | let max_turns = effective_task_max_turns(task_spec, exec_config); |
| 88 | let max_tool_calls = task_max_tool_calls(task_spec); |
| 89 | build_worker_exec_command_from_prompt( |
| 90 | codewhale_binary, |
| 91 | fleet_task_prompt(task_spec), |
| 92 | exec_config, |
| 93 | WorkerExecRoute { |
| 94 | model, |
| 95 | ..WorkerExecRoute::default() |
| 96 | }, |
| 97 | None, |
| 98 | WorkerExecLimits { |
| 99 | max_turns, |
| 100 | max_tool_calls, |
| 101 | }, |
| 102 | ) |
| 103 | } |
| 104 | |
| 105 | /// Build a worker command after resolving workspace Fleet profile input. |
| 106 | /// |
| 107 | /// The launched subprocess runs on the worker's RESOLVED route, not blindly on |
| 108 | /// the run-level session model (#4093 AC #4): the per-worker model+provider are |
| 109 | /// resolved from the task's agent profile via the same explicit-only path the |
| 110 | /// receipt uses ([`fleet_worker_launch_route`]). A worker whose profile pins |
| 111 | /// provider B thus launches on provider B's model even when the parent session |
| 112 | /// is on provider A. Workers with no profile-bound provider fall back to the |
| 113 | /// run-level model and emit no `--provider`, so the worker keeps its own |
| 114 | /// session default (today's behavior, unchanged). |
| 115 | pub fn build_worker_exec_command_with_profiles( |
| 116 | codewhale_binary: &str, |
| 117 | task_spec: &FleetTaskSpec, |
| 118 | exec_config: &FleetExecConfig, |
| 119 | model: Option<&str>, |
| 120 | agent_profiles: &[AgentProfile], |
| 121 | ) -> Result<FleetWorkerCommand> { |
| 122 | let (worker_model, worker_provider) = |
| 123 | fleet_worker_launch_route(task_spec, agent_profiles, model.unwrap_or_default()); |
| 124 | let worker_reasoning_effort = fleet_worker_launch_reasoning_effort(task_spec, agent_profiles); |
| 125 | let max_turns = effective_task_max_turns(task_spec, exec_config); |
| 126 | let max_tool_calls = task_max_tool_calls(task_spec); |
| 127 | Ok(build_worker_exec_command_from_prompt( |
| 128 | codewhale_binary, |
| 129 | fleet_task_prompt_with_profiles(task_spec, agent_profiles)?, |
| 130 | exec_config, |
| 131 | WorkerExecRoute { |
| 132 | model: Some(worker_model.as_str()), |
| 133 | provider: worker_provider.as_deref(), |
| 134 | reasoning_effort: worker_reasoning_effort.as_deref(), |
| 135 | }, |
| 136 | None, |
| 137 | WorkerExecLimits { |
| 138 | max_turns, |
| 139 | max_tool_calls, |
| 140 | }, |
| 141 | )) |
| 142 | } |
| 143 | |
| 144 | /// Build the exact Fleet subprocess command from the coordination-registered |
| 145 | /// worker spec. Unlike the compatibility helpers above, production dispatch |
| 146 | /// uses the projected objective and carries a machine-readable outer authority |
| 147 | /// envelope into the child process. |
| 148 | pub fn build_worker_exec_command_with_launch_spec( |
| 149 | codewhale_binary: &str, |
| 150 | task_spec: &FleetTaskSpec, |
| 151 | launch_spec: &AgentWorkerSpec, |
| 152 | exec_config: &FleetExecConfig, |
| 153 | model: Option<&str>, |
| 154 | agent_profiles: &[AgentProfile], |
| 155 | ) -> Result<FleetWorkerCommand> { |
| 156 | let (worker_model, worker_provider) = |
| 157 | fleet_worker_launch_route(task_spec, agent_profiles, model.unwrap_or_default()); |
| 158 | let worker_reasoning_effort = fleet_worker_launch_reasoning_effort(task_spec, agent_profiles); |
| 159 | let authority = authority_envelope_for_worker(launch_spec, task_spec)?; |
| 160 | // Production dispatch receives an already-hardened launch spec from the |
| 161 | // Fleet manager. Its positive max_steps value has therefore already been |
| 162 | // intersected with a positive FleetExecConfig.max_turns; zero remains the |
| 163 | // explicit unbounded sentinel. |
| 164 | let max_turns = (launch_spec.max_steps > 0).then_some(launch_spec.max_steps); |
| 165 | let max_tool_calls = task_max_tool_calls(task_spec); |
| 166 | Ok(build_worker_exec_command_from_prompt( |
| 167 | codewhale_binary, |
| 168 | launch_spec.objective.clone(), |
| 169 | exec_config, |
| 170 | WorkerExecRoute { |
| 171 | model: Some(worker_model.as_str()), |
| 172 | provider: worker_provider.as_deref(), |
| 173 | reasoning_effort: worker_reasoning_effort.as_deref(), |
| 174 | }, |
| 175 | Some(&authority), |
| 176 | WorkerExecLimits { |
| 177 | max_turns, |
| 178 | max_tool_calls, |
| 179 | }, |
| 180 | )) |
| 181 | } |
| 182 | |
| 183 | fn task_max_steps(task_spec: &FleetTaskSpec) -> Option<u32> { |
| 184 | task_spec |
| 185 | .budget |
| 186 | .as_ref() |
| 187 | .and_then(|budget| budget.max_steps) |
| 188 | .filter(|max_steps| *max_steps > 0) |
| 189 | } |
| 190 | |
| 191 | fn task_max_tool_calls(task_spec: &FleetTaskSpec) -> Option<u32> { |
| 192 | task_spec |
| 193 | .budget |
| 194 | .as_ref() |
| 195 | .and_then(|budget| budget.max_tool_calls) |
| 196 | .filter(|max_tool_calls| *max_tool_calls > 0) |
| 197 | } |
| 198 | |
| 199 | fn effective_task_max_turns( |
| 200 | task_spec: &FleetTaskSpec, |
| 201 | exec_config: &FleetExecConfig, |
| 202 | ) -> Option<u32> { |
| 203 | match (task_max_steps(task_spec), exec_config.max_turns) { |
| 204 | (Some(task_max_steps), fleet_max_turns) if fleet_max_turns > 0 => { |
| 205 | Some(task_max_steps.min(fleet_max_turns)) |
| 206 | } |
| 207 | (Some(task_max_steps), _) => Some(task_max_steps), |
| 208 | (None, fleet_max_turns) if fleet_max_turns > 0 => Some(fleet_max_turns), |
| 209 | (None, _) => None, |
| 210 | } |
| 211 | } |
| 212 | |
| 213 | pub(crate) fn authority_envelope_for_worker( |
| 214 | spec: &AgentWorkerSpec, |
| 215 | task_spec: &FleetTaskSpec, |
| 216 | ) -> Result<ToolAuthorityEnvelope> { |
| 217 | let (authority, writable_roots, writable_files, coordination_contracts) = |
| 218 | if spec.runtime_profile.permissions.write { |
| 219 | let manifest = spec.launch_manifest.as_ref().ok_or_else(|| { |
| 220 | anyhow::anyhow!( |
| 221 | "write-capable Fleet worker '{}' has no launch manifest", |
| 222 | spec.worker_id |
| 223 | ) |
| 224 | })?; |
| 225 | ( |
| 226 | ToolMutationAuthority::ScopedWrite, |
| 227 | super::worker_runtime::fleet_runtime_write_roots(task_spec)?, |
| 228 | manifest.writable_files.clone(), |
| 229 | manifest.coordination_contracts.clone(), |
| 230 | ) |
| 231 | } else { |
| 232 | ( |
| 233 | ToolMutationAuthority::ReadOnly, |
| 234 | Vec::new(), |
| 235 | Vec::new(), |
| 236 | Vec::new(), |
| 237 | ) |
| 238 | }; |
| 239 | let shell = if authority == ToolMutationAuthority::ReadOnly |
| 240 | && matches!( |
| 241 | &spec.agent_type, |
| 242 | FleetRole::Scout | FleetRole::Reviewer | FleetRole::Planner |
| 243 | ) |
| 244 | && spec.runtime_profile.shell.allows_shell() |
| 245 | { |
| 246 | // Scout, Reviewer, and Planner get one classifier-bounded |
| 247 | // foreground shell. Verifiers keep the dedicated Run surface, |
| 248 | // consultants remain shell-less, and writers retain the historical |
| 249 | // subprocess contract. |
| 250 | ToolShellAuthority::ReadOnly |
| 251 | } else { |
| 252 | ToolShellAuthority::None |
| 253 | }; |
| 254 | let verification = if authority == ToolMutationAuthority::ReadOnly |
| 255 | && matches!(&spec.agent_type, FleetRole::Verifier) |
| 256 | && spec.runtime_profile.shell.allows_shell() |
| 257 | { |
| 258 | ToolVerificationAuthority::Bounded |
| 259 | } else { |
| 260 | ToolVerificationAuthority::None |
| 261 | }; |
| 262 | ToolAuthorityEnvelope { |
| 263 | schema_version: 1, |
| 264 | owner: spec.worker_id.clone(), |
| 265 | authority, |
| 266 | network_access: Some(spec.runtime_profile.permissions.network), |
| 267 | shell, |
| 268 | verification, |
| 269 | writable_roots, |
| 270 | writable_files, |
| 271 | coordination_contracts, |
| 272 | } |
| 273 | .normalized() |
| 274 | .map_err(anyhow::Error::msg) |
| 275 | } |
| 276 | |
| 277 | fn build_worker_exec_command_from_prompt( |
| 278 | codewhale_binary: &str, |
| 279 | task_prompt: String, |
| 280 | exec_config: &FleetExecConfig, |
| 281 | route: WorkerExecRoute<'_>, |
| 282 | authority: Option<&ToolAuthorityEnvelope>, |
| 283 | limits: WorkerExecLimits, |
| 284 | ) -> FleetWorkerCommand { |
| 285 | let mut args: Vec<String> = Vec::new(); |
| 286 | |
| 287 | // The canonical `codewhale` dispatcher owns these route overrides as |
| 288 | // global flags and deliberately rejects them after `exec`. Keep them in |
| 289 | // front of the subcommand so Fleet commands work through the installed |
| 290 | // dispatcher as well as when a host points directly at `codewhale-tui`. |
| 291 | if let Some(model) = route.model.map(str::trim).filter(|m| !m.is_empty()) { |
| 292 | args.push("--model".to_string()); |
| 293 | args.push(model.to_string()); |
| 294 | } |
| 295 | |
| 296 | // Non-secret provider identifier only (#4093): the worker resolves the |
| 297 | // provider's credentials from its own env/config. Emitted ONLY when the |
| 298 | // worker's profile explicitly pins a provider, so profile-less workers keep |
| 299 | // their own session default exactly as before. |
| 300 | if let Some(provider) = route.provider.map(str::trim).filter(|p| !p.is_empty()) { |
| 301 | args.push("--provider".to_string()); |
| 302 | args.push(provider.to_string()); |
| 303 | } |
| 304 | |
| 305 | args.extend([ |
| 306 | "exec".to_string(), |
| 307 | "--auto".to_string(), |
| 308 | "--output-format".to_string(), |
| 309 | "stream-json".to_string(), |
| 310 | // R7: the worker shuts itself down when the manager dies (stdin EOF). |
| 311 | "--parent-death-watch".to_string(), |
| 312 | ]); |
| 313 | |
| 314 | // Non-secret thinking tier only (#4137). This is profile metadata and |
| 315 | // follows the same explicit-only policy as provider: omit it when the |
| 316 | // worker profile inherits the session/default reasoning setting. |
| 317 | if let Some(reasoning_effort) = route |
| 318 | .reasoning_effort |
| 319 | .map(str::trim) |
| 320 | .filter(|e| !e.is_empty()) |
| 321 | { |
| 322 | args.push("--reasoning-effort".to_string()); |
| 323 | args.push(reasoning_effort.to_string()); |
| 324 | } |
| 325 | |
| 326 | if !exec_config.allowed_tools.is_empty() { |
| 327 | args.push("--allowed-tools".to_string()); |
| 328 | args.push(exec_config.allowed_tools.join(",")); |
| 329 | } |
| 330 | if !exec_config.disallowed_tools.is_empty() { |
| 331 | args.push("--disallowed-tools".to_string()); |
| 332 | args.push(exec_config.disallowed_tools.join(",")); |
| 333 | } |
| 334 | if let Some(max_turns) = limits.max_turns { |
| 335 | args.push("--max-turns".to_string()); |
| 336 | args.push(max_turns.to_string()); |
| 337 | } |
| 338 | if let Some(max_tool_calls) = limits.max_tool_calls { |
| 339 | args.push("--max-tool-calls".to_string()); |
| 340 | args.push(max_tool_calls.to_string()); |
| 341 | } |
| 342 | if !exec_config.append_system_prompt.trim().is_empty() { |
| 343 | args.push("--append-system-prompt".to_string()); |
| 344 | args.push(exec_config.append_system_prompt.clone()); |
| 345 | } |
| 346 | |
| 347 | if let Some(authority) = authority { |
| 348 | args.push("--tool-authority-json".to_string()); |
| 349 | args.push( |
| 350 | serde_json::to_string(authority) |
| 351 | .expect("validated Fleet tool authority envelope must serialize"), |
| 352 | ); |
| 353 | } |
| 354 | |
| 355 | // The composed task prompt is the final positional argument. |
| 356 | args.push(task_prompt); |
| 357 | |
| 358 | FleetWorkerCommand::new(codewhale_binary.to_string(), args) |
| 359 | } |
| 360 | |
| 361 | /// Map one `codewhale exec` stream-json line into a fleet ledger event. |
| 362 | /// |
| 363 | /// Returns `None` for lines that don't correspond to a worker lifecycle |
| 364 | /// transition (e.g. `session_capture`, `metadata`). The exec event schema is |
| 365 | /// `{"type": "...", ...}` (see `ExecStreamEvent` in `main.rs`). |
| 366 | pub fn map_exec_stream_line(line: &str) -> Option<FleetWorkerEventPayload> { |
| 367 | let value: serde_json::Value = serde_json::from_str(line.trim()).ok()?; |
| 368 | map_exec_stream_value(&value) |
| 369 | } |
| 370 | |
| 371 | /// [`map_exec_stream_line`] on an already-parsed line, so the incremental |
| 372 | /// stream reader parses each frame exactly once. |
| 373 | fn map_exec_stream_value(value: &serde_json::Value) -> Option<FleetWorkerEventPayload> { |
| 374 | match value.get("type").and_then(serde_json::Value::as_str)? { |
| 375 | "tool_use" => { |
| 376 | let tool = value |
| 377 | .get("name") |
| 378 | .and_then(serde_json::Value::as_str) |
| 379 | .unwrap_or("tool") |
| 380 | .to_string(); |
| 381 | let call_id = value |
| 382 | .get("id") |
| 383 | .and_then(serde_json::Value::as_str) |
| 384 | .map(str::to_string); |
| 385 | Some(FleetWorkerEventPayload::RunningTool { tool, call_id }) |
| 386 | } |
| 387 | "workflow_event" => Some(FleetWorkerEventPayload::WorkflowEvent { |
| 388 | workflow_run_id: value.get("run_id")?.as_str()?.to_string(), |
| 389 | event: value.get("event")?.clone(), |
| 390 | }), |
| 391 | // Streaming model output / tool results mean the worker is alive and |
| 392 | // making progress; surface a coarse Running heartbeat. |
| 393 | "content" | "tool_result" => Some(FleetWorkerEventPayload::Running), |
| 394 | // Per-step usage receipts feed the run-level accumulator behind the |
| 395 | // fleet usage ceiling (R6, #5567); a malformed line still counts as |
| 396 | // liveness. |
| 397 | "turn_usage" => { |
| 398 | let tokens = |field: &str| { |
| 399 | value |
| 400 | .get(field) |
| 401 | .and_then(serde_json::Value::as_u64) |
| 402 | .unwrap_or(0) |
| 403 | }; |
| 404 | let input_tokens = tokens("input_tokens"); |
| 405 | let output_tokens = tokens("output_tokens"); |
| 406 | if input_tokens == 0 && output_tokens == 0 { |
| 407 | Some(FleetWorkerEventPayload::Running) |
| 408 | } else { |
| 409 | Some(FleetWorkerEventPayload::UsageReport { |
| 410 | input_tokens, |
| 411 | output_tokens, |
| 412 | }) |
| 413 | } |
| 414 | } |
| 415 | "done" => Some(FleetWorkerEventPayload::Completed { |
| 416 | exit_code: Some(0), |
| 417 | summary: None, |
| 418 | }), |
| 419 | "error" => { |
| 420 | let reason = value |
| 421 | .get("error") |
| 422 | .and_then(serde_json::Value::as_str) |
| 423 | .unwrap_or("worker reported an error") |
| 424 | .to_string(); |
| 425 | Some(FleetWorkerEventPayload::Failed { |
| 426 | reason, |
| 427 | recoverable: false, |
| 428 | }) |
| 429 | } |
| 430 | _ => None, |
| 431 | } |
| 432 | } |
| 433 | |
| 434 | #[derive(Debug)] |
| 435 | enum ParsedTerminalRoute { |
| 436 | NotTerminal, |
| 437 | Valid(FleetWorkerReportedRoute), |
| 438 | Invalid, |
| 439 | } |
| 440 | |
| 441 | /// The `meta` object of a terminal exec receipt, or `None` for every other |
| 442 | /// stream line. |
| 443 | fn exec_terminal_meta( |
| 444 | value: &serde_json::Value, |
| 445 | ) -> Option<&serde_json::Map<String, serde_json::Value>> { |
| 446 | if value.get("type").and_then(serde_json::Value::as_str) != Some("metadata") { |
| 447 | return None; |
| 448 | } |
| 449 | let meta = value.get("meta").and_then(serde_json::Value::as_object)?; |
| 450 | (meta.get("receipt_kind").and_then(serde_json::Value::as_str) == Some("terminal")) |
| 451 | .then_some(meta) |
| 452 | } |
| 453 | |
| 454 | /// The worker's visible final answer from a terminal exec receipt: the |
| 455 | /// emitter already bounded and redacted `visible_final_answer_excerpt`, and |
| 456 | /// `visible_final_answer_chars` is the real pre-truncation length. |
| 457 | fn parse_exec_terminal_final_answer(value: &serde_json::Value) -> Option<FleetWorkerFinalAnswer> { |
| 458 | let meta = exec_terminal_meta(value)?; |
| 459 | let excerpt = meta |
| 460 | .get("visible_final_answer_excerpt") |
| 461 | .and_then(serde_json::Value::as_str)? |
| 462 | .trim(); |
| 463 | if excerpt.is_empty() { |
| 464 | return None; |
| 465 | } |
| 466 | let chars = meta |
| 467 | .get("visible_final_answer_chars") |
| 468 | .and_then(serde_json::Value::as_u64) |
| 469 | .and_then(|chars| usize::try_from(chars).ok()) |
| 470 | .unwrap_or_else(|| excerpt.chars().count()); |
| 471 | Some(FleetWorkerFinalAnswer { |
| 472 | excerpt: crate::exec_stream_final_answer_excerpt(excerpt), |
| 473 | chars, |
| 474 | }) |
| 475 | } |
| 476 | |
| 477 | /// Parse one allowlisted, secret-free route identity from terminal exec |
| 478 | /// metadata. Once a line declares itself as a terminal receipt, malformed |
| 479 | /// route fields are distinct from ordinary non-terminal stream noise so a |
| 480 | /// prior valid record cannot survive contradictory evidence. |
| 481 | fn parse_exec_terminal_route(value: &serde_json::Value) -> ParsedTerminalRoute { |
| 482 | let Some(meta) = exec_terminal_meta(value) else { |
| 483 | return ParsedTerminalRoute::NotTerminal; |
| 484 | }; |
| 485 | |
| 486 | let route = (|| { |
| 487 | let provider = meta.get("provider")?.as_str()?.trim(); |
| 488 | let model = meta.get("model")?.as_str()?.trim(); |
| 489 | if provider.is_empty() || model.is_empty() { |
| 490 | return None; |
| 491 | } |
| 492 | let provider_kind = crate::config::ApiProvider::parse(provider)?; |
| 493 | let provider_exact_id = match meta.get("provider_id") { |
| 494 | None => None, |
| 495 | Some(value) => { |
| 496 | let id = value.as_str()?.trim(); |
| 497 | if id.is_empty() { |
| 498 | return None; |
| 499 | } |
| 500 | Some(id.to_string()) |
| 501 | } |
| 502 | }; |
| 503 | if provider_exact_id.is_some() && provider_kind != crate::config::ApiProvider::Custom { |
| 504 | return None; |
| 505 | } |
| 506 | Some(FleetWorkerReportedRoute { |
| 507 | provider: provider.to_string(), |
| 508 | provider_exact_id, |
| 509 | model: model.to_string(), |
| 510 | }) |
| 511 | })(); |
| 512 | |
| 513 | route.map_or(ParsedTerminalRoute::Invalid, ParsedTerminalRoute::Valid) |
| 514 | } |
| 515 | |
| 516 | #[cfg(test)] |
| 517 | fn map_exec_terminal_route(line: &str) -> Option<FleetWorkerReportedRoute> { |
| 518 | let value: serde_json::Value = serde_json::from_str(line).ok()?; |
| 519 | match parse_exec_terminal_route(&value) { |
| 520 | ParsedTerminalRoute::Valid(route) => Some(route), |
| 521 | ParsedTerminalRoute::NotTerminal | ParsedTerminalRoute::Invalid => None, |
| 522 | } |
| 523 | } |
| 524 | |
| 525 | /// Classify a worker process exit into a terminal fleet event. |
| 526 | /// |
| 527 | /// `stopped` means the operator stopped the worker (cancellation), which takes |
| 528 | /// precedence over the exit code. |
| 529 | pub fn classify_worker_exit(exit_code: Option<i32>, stopped: bool) -> FleetWorkerEventPayload { |
| 530 | if stopped { |
| 531 | return FleetWorkerEventPayload::Cancelled { cancelled_by: None }; |
| 532 | } |
| 533 | match exit_code { |
| 534 | Some(0) => FleetWorkerEventPayload::Completed { |
| 535 | exit_code: Some(0), |
| 536 | summary: None, |
| 537 | }, |
| 538 | Some(code) => FleetWorkerEventPayload::Failed { |
| 539 | reason: format!("worker exited with code {code}"), |
| 540 | recoverable: true, |
| 541 | }, |
| 542 | None => FleetWorkerEventPayload::Failed { |
| 543 | reason: "worker exited without a status code".to_string(), |
| 544 | recoverable: true, |
| 545 | }, |
| 546 | } |
| 547 | } |
| 548 | |
| 549 | /// Drives fleet workers as real `codewhale exec` subprocesses on the local |
| 550 | /// host, incrementally draining each worker's stream-json output into fleet |
| 551 | /// ledger events. |
| 552 | /// |
| 553 | /// The caller (the `codewhale fleet run` loop / `FleetManager`) owns the |
| 554 | /// ledger; the executor owns the OS process boundary and the incremental log |
| 555 | /// parse. Because the worker is a separate process, its heavy runtime/tool |
| 556 | /// construction never touches the orchestrator — the parent only ingests a |
| 557 | /// compact event stream, which is what keeps it light at high fanout. |
| 558 | pub struct FleetExecutor { |
| 559 | workspace: std::path::PathBuf, |
| 560 | sessions_dir: Option<std::path::PathBuf>, |
| 561 | adapter: super::host::LocalProcessFleetHostAdapter, |
| 562 | ssh_adapters: std::collections::BTreeMap<String, super::host::SshFleetHostAdapter>, |
| 563 | streams: std::collections::BTreeMap<String, WorkerStream>, |
| 564 | } |
| 565 | |
| 566 | /// Durable lease identity owned by one concrete host process. |
| 567 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 568 | pub struct FleetExecutorAttempt { |
| 569 | pub run_id: codewhale_protocol::fleet::FleetRunId, |
| 570 | pub task_id: String, |
| 571 | pub attempt: u32, |
| 572 | } |
| 573 | |
| 574 | struct WorkerStream { |
| 575 | log_path: std::path::PathBuf, |
| 576 | host: WorkerStreamHost, |
| 577 | attempt: Option<FleetExecutorAttempt>, |
| 578 | offset: u64, |
| 579 | // Keep incomplete stream frames as bytes. Decoding each read separately |
| 580 | // corrupts valid UTF-8 when a multibyte code point crosses a read boundary. |
| 581 | pending: Vec<u8>, |
| 582 | terminal: bool, |
| 583 | terminal_route: TerminalRouteEvidence, |
| 584 | /// When this worker process was started, for per-task wall-clock limits (R5). |
| 585 | started_at: std::time::Instant, |
| 586 | /// The worker's visible final answer from its terminal exec receipt. This |
| 587 | /// is the task's deliverable for report/summary work that produces no |
| 588 | /// file artifact; surfaced as `Completed.summary` and in the receipt note |
| 589 | /// so receipts stop reporting "no verifiable output" for a worker that |
| 590 | /// wrote a full report. Bounded by the emitter, so nothing accumulates |
| 591 | /// here. |
| 592 | final_answer: Option<FleetWorkerFinalAnswer>, |
| 593 | /// Saved exec session id reported by the worker's `session_capture` event. |
| 594 | /// Resolving it via `GET /v1/sessions/{id}` yields the full transcript |
| 595 | /// (the worker's final assistant reply). |
| 596 | saved_session_id: Option<String>, |
| 597 | /// Parent-owned capture identity in the Runtime's existing session store. |
| 598 | session_capture: Option<(String, std::path::PathBuf)>, |
| 599 | } |
| 600 | |
| 601 | impl WorkerStream { |
| 602 | /// Observe one raw stream-json frame: record route evidence, the final |
| 603 | /// answer, and the saved-session id, and map it to a ledger payload. The |
| 604 | /// frame is parsed exactly once. |
| 605 | fn observe_line(&mut self, line: &[u8]) -> Option<FleetWorkerEventPayload> { |
| 606 | let Ok(line) = std::str::from_utf8(line) else { |
| 607 | // stream-json is a UTF-8 contract. Never accept a lossy-decoded route |
| 608 | // receipt: replacement characters could turn corrupt provider/model |
| 609 | // bytes into apparently valid provenance. |
| 610 | self.terminal_route.observe(ParsedTerminalRoute::Invalid); |
| 611 | return None; |
| 612 | }; |
| 613 | let value: serde_json::Value = serde_json::from_str(line.trim()).ok()?; |
| 614 | self.terminal_route |
| 615 | .observe(parse_exec_terminal_route(&value)); |
| 616 | if exec_terminal_meta(&value).is_some() { |
| 617 | self.final_answer = parse_exec_terminal_final_answer(&value); |
| 618 | } |
| 619 | if value.get("type").and_then(serde_json::Value::as_str) == Some("session_capture") |
| 620 | && let Some(id) = value |
| 621 | .get("saved_session_id") |
| 622 | .and_then(serde_json::Value::as_str) |
| 623 | .map(str::trim) |
| 624 | .filter(|id| !id.is_empty()) |
| 625 | { |
| 626 | if self |
| 627 | .session_capture |
| 628 | .as_ref() |
| 629 | .is_some_and(|(expected, _)| expected == id) |
| 630 | { |
| 631 | self.saved_session_id = Some(id.to_string()); |
| 632 | } else { |
| 633 | // A worker log cannot select an unrelated local transcript. |
| 634 | self.session_capture = None; |
| 635 | self.saved_session_id = None; |
| 636 | } |
| 637 | } |
| 638 | map_exec_stream_value(&value) |
| 639 | } |
| 640 | } |
| 641 | |
| 642 | #[derive(Debug, Clone, Default)] |
| 643 | enum TerminalRouteEvidence { |
| 644 | #[default] |
| 645 | Missing, |
| 646 | Valid(FleetWorkerReportedRoute), |
| 647 | InvalidOrAmbiguous, |
| 648 | } |
| 649 | |
| 650 | impl TerminalRouteEvidence { |
| 651 | fn observe(&mut self, parsed: ParsedTerminalRoute) { |
| 652 | match parsed { |
| 653 | ParsedTerminalRoute::NotTerminal => {} |
| 654 | ParsedTerminalRoute::Invalid => *self = Self::InvalidOrAmbiguous, |
| 655 | ParsedTerminalRoute::Valid(route) => { |
| 656 | *self = if matches!(&*self, Self::Missing) { |
| 657 | Self::Valid(route) |
| 658 | } else { |
| 659 | // The stream contract emits exactly one terminal receipt. |
| 660 | // Any second record, even an identical one, is ambiguous |
| 661 | // provenance and must permanently fail closed. |
| 662 | Self::InvalidOrAmbiguous |
| 663 | }; |
| 664 | } |
| 665 | } |
| 666 | } |
| 667 | |
| 668 | fn reported_route(&self) -> Option<&FleetWorkerReportedRoute> { |
| 669 | match self { |
| 670 | Self::Valid(route) => Some(route), |
| 671 | Self::Missing | Self::InvalidOrAmbiguous => None, |
| 672 | } |
| 673 | } |
| 674 | } |
| 675 | |
| 676 | enum WorkerStreamHost { |
| 677 | Local, |
| 678 | Ssh(String), |
| 679 | } |
| 680 | |
| 681 | #[derive(Debug, Clone)] |
| 682 | pub struct FleetWorkerReportedRoute { |
| 683 | pub provider: String, |
| 684 | pub provider_exact_id: Option<String>, |
| 685 | pub model: String, |
| 686 | } |
| 687 | |
| 688 | #[derive(Debug, Clone)] |
| 689 | pub struct FleetWorkerTerminalEvent { |
| 690 | pub payload: FleetWorkerEventPayload, |
| 691 | pub exit_code: Option<i32>, |
| 692 | /// Non-terminal payloads discovered by the mandatory post-exit drain. |
| 693 | pub tail_payloads: Vec<FleetWorkerEventPayload>, |
| 694 | pub reported_route: Option<FleetWorkerReportedRoute>, |
| 695 | /// The worker's visible final answer from its terminal exec receipt, |
| 696 | /// whatever the outcome: a worker that fails after writing most of a |
| 697 | /// report keeps the text on its receipt. |
| 698 | pub final_answer: Option<FleetWorkerFinalAnswer>, |
| 699 | /// Saved exec session id reported by the worker's `session_capture` event, |
| 700 | /// when one was persisted on completion. |
| 701 | pub saved_session_id: Option<String>, |
| 702 | /// A real headless exec process must report its actual route. Callers use |
| 703 | /// this bit to distinguish a missing/invalid report (fail closed) from |
| 704 | /// pre-launch or simulated paths that only have declared route intent. |
| 705 | pub requires_reported_route: bool, |
| 706 | } |
| 707 | |
| 708 | impl FleetExecutor { |
| 709 | pub fn new(workspace: impl AsRef<std::path::Path>) -> Self { |
| 710 | let workspace = workspace.as_ref().to_path_buf(); |
| 711 | Self { |
| 712 | adapter: super::host::LocalProcessFleetHostAdapter::new(&workspace), |
| 713 | workspace, |
| 714 | sessions_dir: None, |
| 715 | ssh_adapters: std::collections::BTreeMap::new(), |
| 716 | streams: std::collections::BTreeMap::new(), |
| 717 | } |
| 718 | } |
| 719 | |
| 720 | /// Share the caller's SessionManager directory; never create a second store. |
| 721 | pub fn with_sessions_dir(mut self, sessions_dir: std::path::PathBuf) -> Self { |
| 722 | self.sessions_dir = Some(sessions_dir); |
| 723 | self |
| 724 | } |
| 725 | |
| 726 | /// Start a worker process and begin tracking its event stream. |
| 727 | pub fn start_worker( |
| 728 | &mut self, |
| 729 | worker_id: &str, |
| 730 | command: FleetWorkerCommand, |
| 731 | cwd: Option<std::path::PathBuf>, |
| 732 | ) -> super::host::FleetHostResult<super::host::FleetWorkerHandle> { |
| 733 | self.start_worker_on_host(worker_id, &FleetHostSpec::Local, command, cwd) |
| 734 | } |
| 735 | |
| 736 | /// Start a worker on the requested fleet host. |
| 737 | pub fn start_worker_on_host( |
| 738 | &mut self, |
| 739 | worker_id: &str, |
| 740 | host: &FleetHostSpec, |
| 741 | command: FleetWorkerCommand, |
| 742 | cwd: Option<std::path::PathBuf>, |
| 743 | ) -> super::host::FleetHostResult<super::host::FleetWorkerHandle> { |
| 744 | self.start_worker_on_host_inner(worker_id, host, command, cwd, None) |
| 745 | } |
| 746 | |
| 747 | /// Start the concrete process for one exact durable Fleet lease. |
| 748 | pub fn start_worker_attempt_on_host( |
| 749 | &mut self, |
| 750 | worker_id: &str, |
| 751 | host: &FleetHostSpec, |
| 752 | command: FleetWorkerCommand, |
| 753 | cwd: Option<std::path::PathBuf>, |
| 754 | attempt: FleetExecutorAttempt, |
| 755 | ) -> super::host::FleetHostResult<super::host::FleetWorkerHandle> { |
| 756 | self.start_worker_on_host_inner(worker_id, host, command, cwd, Some(attempt)) |
| 757 | } |
| 758 | |
| 759 | fn start_worker_on_host_inner( |
| 760 | &mut self, |
| 761 | worker_id: &str, |
| 762 | host: &FleetHostSpec, |
| 763 | command: FleetWorkerCommand, |
| 764 | cwd: Option<std::path::PathBuf>, |
| 765 | attempt: Option<FleetExecutorAttempt>, |
| 766 | ) -> super::host::FleetHostResult<super::host::FleetWorkerHandle> { |
| 767 | let mut request = super::host::FleetWorkerStartRequest::new(worker_id, command); |
| 768 | request.cwd = cwd; |
| 769 | let session_capture = if matches!(host, FleetHostSpec::Local) { |
| 770 | self.sessions_dir.as_ref().map(|dir| { |
| 771 | let id = uuid::Uuid::new_v4().to_string(); |
| 772 | for (key, value) in [ |
| 773 | ("CODEWHALE_FLEET_CAPTURE_ID", id.clone()), |
| 774 | ( |
| 775 | "CODEWHALE_FLEET_CAPTURE_DIR", |
| 776 | dir.to_string_lossy().into_owned(), |
| 777 | ), |
| 778 | ] { |
| 779 | request.env.insert(key.to_string(), value); |
| 780 | request.env_allowlist.insert(key.to_string()); |
| 781 | } |
| 782 | // Preserve an explicit local account/config root as well. |
| 783 | if let Some(home) = std::env::var_os("CODEWHALE_HOME") { |
| 784 | request |
| 785 | .env |
| 786 | .insert("CODEWHALE_HOME".into(), home.to_string_lossy().into_owned()); |
| 787 | request.env_allowlist.insert("CODEWHALE_HOME".into()); |
| 788 | } |
| 789 | (id, dir.clone()) |
| 790 | }) |
| 791 | } else { |
| 792 | // SSH transcripts live remotely and have no local retrieval link. |
| 793 | None |
| 794 | }; |
| 795 | let (handle, host) = match host { |
| 796 | FleetHostSpec::Local => { |
| 797 | let handle = self.adapter.start_worker(request)?; |
| 798 | (handle, WorkerStreamHost::Local) |
| 799 | } |
| 800 | FleetHostSpec::Ssh { .. } => { |
| 801 | let config = super::host::SshFleetHostConfig::from_host_spec(host)?; |
| 802 | let key = worker_id.to_string(); |
| 803 | let adapter = self.ssh_adapters.entry(key.clone()).or_insert( |
| 804 | super::host::SshFleetHostAdapter::new(&self.workspace, config)?, |
| 805 | ); |
| 806 | let handle = adapter.start_worker(request)?; |
| 807 | (handle, WorkerStreamHost::Ssh(key)) |
| 808 | } |
| 809 | FleetHostSpec::Docker { image, .. } => { |
| 810 | return Err(super::host::FleetHostError { |
| 811 | kind: super::host::FleetHostErrorKind::Configuration, |
| 812 | message: format!("docker Fleet workers are not wired yet (image {image})"), |
| 813 | }); |
| 814 | } |
| 815 | }; |
| 816 | self.streams.insert( |
| 817 | worker_id.to_string(), |
| 818 | WorkerStream { |
| 819 | log_path: handle.log_path.clone(), |
| 820 | host, |
| 821 | attempt, |
| 822 | offset: 0, |
| 823 | pending: Vec::new(), |
| 824 | terminal: false, |
| 825 | terminal_route: TerminalRouteEvidence::default(), |
| 826 | started_at: std::time::Instant::now(), |
| 827 | final_answer: None, |
| 828 | saved_session_id: None, |
| 829 | session_capture, |
| 830 | }, |
| 831 | ); |
| 832 | Ok(handle) |
| 833 | } |
| 834 | |
| 835 | pub fn is_tracking(&self, worker_id: &str) -> bool { |
| 836 | self.streams.contains_key(worker_id) |
| 837 | } |
| 838 | |
| 839 | pub fn worker_ids(&self) -> Vec<String> { |
| 840 | self.streams.keys().cloned().collect() |
| 841 | } |
| 842 | |
| 843 | pub fn tracked_attempt(&self, worker_id: &str) -> Option<FleetExecutorAttempt> { |
| 844 | self.streams |
| 845 | .get(worker_id) |
| 846 | .and_then(|stream| stream.attempt.clone()) |
| 847 | } |
| 848 | |
| 849 | /// Wall-clock time this worker process has been running (R5). `None` when |
| 850 | /// the worker is not tracked. |
| 851 | #[must_use] |
| 852 | pub fn worker_running_for(&self, worker_id: &str) -> Option<std::time::Duration> { |
| 853 | self.streams |
| 854 | .get(worker_id) |
| 855 | .map(|stream| stream.started_at.elapsed()) |
| 856 | } |
| 857 | |
| 858 | /// Stop a tracked worker at the host boundary. |
| 859 | /// |
| 860 | /// Operator controls run in a separate process from the foreground Fleet |
| 861 | /// manager, so they communicate cancellation through the durable ledger. |
| 862 | /// The manager calls this method after observing that terminal state; the |
| 863 | /// executor is the only owner that can reliably reach the live local/SSH |
| 864 | /// adapter handle. |
| 865 | pub fn stop_worker(&mut self, worker_id: &str) -> Result<()> { |
| 866 | let ssh_key = match self.streams.get(worker_id).map(|stream| &stream.host) { |
| 867 | Some(WorkerStreamHost::Local) => None, |
| 868 | Some(WorkerStreamHost::Ssh(key)) => Some(key.clone()), |
| 869 | None => return Ok(()), |
| 870 | }; |
| 871 | if let Some(key) = ssh_key { |
| 872 | let adapter = self.ssh_adapters.get_mut(&key).ok_or_else(|| { |
| 873 | anyhow::anyhow!("tracked SSH Fleet worker {worker_id} has no host adapter") |
| 874 | })?; |
| 875 | adapter.stop_worker(worker_id)?; |
| 876 | } else { |
| 877 | self.adapter.stop_worker(worker_id)?; |
| 878 | } |
| 879 | Ok(()) |
| 880 | } |
| 881 | |
| 882 | /// Stop tracking a terminal worker so the scheduler can reuse the same |
| 883 | /// logical worker id for the next queued task. |
| 884 | pub fn forget_worker(&mut self, worker_id: &str) { |
| 885 | let Some(stream) = self.streams.remove(worker_id) else { |
| 886 | return; |
| 887 | }; |
| 888 | match stream.host { |
| 889 | WorkerStreamHost::Local => { |
| 890 | let _ = self.adapter.cleanup_worker(worker_id); |
| 891 | } |
| 892 | WorkerStreamHost::Ssh(key) => { |
| 893 | if let Some(adapter) = self.ssh_adapters.get_mut(&key) { |
| 894 | let _ = adapter.cleanup_worker(worker_id); |
| 895 | } |
| 896 | self.ssh_adapters.remove(&key); |
| 897 | } |
| 898 | } |
| 899 | } |
| 900 | |
| 901 | /// Maximum bytes drained from a worker log per call, and the cap for the |
| 902 | /// buffered partial line. Event lines are small; the remainder stays for |
| 903 | /// the next drain. |
| 904 | const MAX_DRAIN_BYTES: u64 = 1024 * 1024; |
| 905 | |
| 906 | /// Read any newly-written stream-json lines for a worker and map them to |
| 907 | /// fleet ledger events. Safe to call repeatedly; only new bytes are parsed, |
| 908 | /// and a trailing partial line is buffered until its newline arrives. |
| 909 | pub fn drain_events(&mut self, worker_id: &str) -> Vec<FleetWorkerEventPayload> { |
| 910 | let Some(stream) = self.streams.get_mut(worker_id) else { |
| 911 | return Vec::new(); |
| 912 | }; |
| 913 | let mut events = Vec::new(); |
| 914 | let Ok(mut file) = std::fs::File::open(&stream.log_path) else { |
| 915 | return events; |
| 916 | }; |
| 917 | use std::io::{Read, Seek, SeekFrom}; |
| 918 | if file.seek(SeekFrom::Start(stream.offset)).is_err() { |
| 919 | return events; |
| 920 | } |
| 921 | let mut buf = Vec::new(); |
| 922 | if let Ok(read) = file.take(Self::MAX_DRAIN_BYTES).read_to_end(&mut buf) { |
| 923 | stream.offset += read as u64; |
| 924 | stream.pending.extend_from_slice(&buf); |
| 925 | while let Some(idx) = stream.pending.iter().position(|byte| *byte == b'\n') { |
| 926 | let line: Vec<u8> = stream.pending.drain(..=idx).collect(); |
| 927 | if let Some(event) = stream.observe_line(&line) { |
| 928 | events.push(event); |
| 929 | } |
| 930 | } |
| 931 | // Whatever remains has no newline; drop it rather than buffering |
| 932 | // a newline-free flood forever. |
| 933 | if stream.pending.len() as u64 > Self::MAX_DRAIN_BYTES { |
| 934 | tracing::debug!( |
| 935 | worker_id, |
| 936 | dropped_bytes = stream.pending.len(), |
| 937 | "fleet drain dropped a newline-free flood exceeding the per-call budget" |
| 938 | ); |
| 939 | stream.pending.clear(); |
| 940 | } |
| 941 | } |
| 942 | events |
| 943 | } |
| 944 | |
| 945 | /// Poll the worker process; once it exits, return the terminal event exactly |
| 946 | /// once. Returns `None` while the worker is still running or already |
| 947 | /// finalized. |
| 948 | pub fn poll_terminal(&mut self, worker_id: &str) -> Option<FleetWorkerEventPayload> { |
| 949 | self.poll_terminal_with_status(worker_id) |
| 950 | .map(|event| event.payload) |
| 951 | } |
| 952 | |
| 953 | /// Poll the worker process and include the raw exit code for receipt |
| 954 | /// verification. |
| 955 | pub fn poll_terminal_with_status( |
| 956 | &mut self, |
| 957 | worker_id: &str, |
| 958 | ) -> Option<FleetWorkerTerminalEvent> { |
| 959 | if self.streams.get(worker_id).is_none_or(|s| s.terminal) { |
| 960 | return None; |
| 961 | } |
| 962 | let status = match self.streams.get(worker_id).map(|s| &s.host)? { |
| 963 | WorkerStreamHost::Local => self.adapter.read_status(worker_id).ok()?, |
| 964 | WorkerStreamHost::Ssh(key) => self |
| 965 | .ssh_adapters |
| 966 | .get_mut(key) |
| 967 | .and_then(|adapter| adapter.read_status(worker_id).ok())?, |
| 968 | }; |
| 969 | let mut terminal = match status.state { |
| 970 | super::host::FleetHostWorkerState::Running |
| 971 | | super::host::FleetHostWorkerState::Draining |
| 972 | | super::host::FleetHostWorkerState::Unknown => return None, |
| 973 | super::host::FleetHostWorkerState::Stopped => { |
| 974 | classify_worker_exit(status.exit_code, true) |
| 975 | } |
| 976 | super::host::FleetHostWorkerState::Exited |
| 977 | | super::host::FleetHostWorkerState::Failed => { |
| 978 | classify_worker_exit(status.exit_code, false) |
| 979 | } |
| 980 | }; |
| 981 | // Once status is terminal the worker can no longer append. Drain one |
| 982 | // final time before snapshotting route evidence so metadata written |
| 983 | // between the scheduler's ordinary drain and this status poll cannot |
| 984 | // be lost when the worker is forgotten. |
| 985 | let mut tail_payloads = self.drain_events(worker_id); |
| 986 | let stream = self.streams.get_mut(worker_id)?; |
| 987 | let trailing_line = std::mem::take(&mut stream.pending); |
| 988 | if trailing_line.iter().any(|byte| !byte.is_ascii_whitespace()) |
| 989 | && let Some(payload) = stream.observe_line(&trailing_line) |
| 990 | { |
| 991 | tail_payloads.push(payload); |
| 992 | } |
| 993 | stream.terminal = true; |
| 994 | let final_answer = stream.final_answer.take(); |
| 995 | // Surface the visible final answer on a successful completion so |
| 996 | // report/summary tasks (no scorer, no file artifact) show their |
| 997 | // deliverable instead of "no verifiable output". `Failed` has no |
| 998 | // summary slot; the receipt keeps the text via `final_answer`. |
| 999 | if let (Some(answer), FleetWorkerEventPayload::Completed { summary, .. }) = |
| 1000 | (final_answer.as_ref(), &mut terminal) |
| 1001 | { |
| 1002 | *summary = Some(answer.excerpt.clone()); |
| 1003 | } |
| 1004 | Some(FleetWorkerTerminalEvent { |
| 1005 | payload: terminal, |
| 1006 | exit_code: status.exit_code, |
| 1007 | tail_payloads, |
| 1008 | reported_route: stream.terminal_route.reported_route().cloned(), |
| 1009 | final_answer, |
| 1010 | saved_session_id: stream.saved_session_id.take().filter(|id| { |
| 1011 | stream |
| 1012 | .session_capture |
| 1013 | .as_ref() |
| 1014 | .is_some_and(|(expected, dir)| { |
| 1015 | id == expected |
| 1016 | && crate::session_manager::SessionManager::new(dir.clone()) |
| 1017 | .and_then(|manager| manager.load_session(id)) |
| 1018 | .is_ok() |
| 1019 | }) |
| 1020 | }), |
| 1021 | requires_reported_route: true, |
| 1022 | }) |
| 1023 | } |
| 1024 | |
| 1025 | /// True once every started worker has reached a terminal state. |
| 1026 | pub fn all_terminal(&self) -> bool { |
| 1027 | !self.streams.is_empty() && self.streams.values().all(|s| s.terminal) |
| 1028 | } |
| 1029 | } |
| 1030 | |
| 1031 | #[cfg(test)] |
| 1032 | mod tests { |
| 1033 | use super::*; |
| 1034 | use codewhale_config::{ |
| 1035 | FleetDelegationHints, FleetLoadout, FleetProfile, FleetProfilePermissions, FleetRole, |
| 1036 | FleetSlot, |
| 1037 | }; |
| 1038 | use codewhale_protocol::fleet::{ |
| 1039 | FleetHostSpec, FleetTaskBudget, FleetTaskSpec, FleetTaskWorkerProfile, FleetWorkerSpec, |
| 1040 | FleetWorkspaceRequirements, |
| 1041 | }; |
| 1042 | use std::collections::BTreeMap; |
| 1043 | use tempfile::TempDir; |
| 1044 | |
| 1045 | fn task(instructions: &str) -> FleetTaskSpec { |
| 1046 | FleetTaskSpec { |
| 1047 | id: "t1".to_string(), |
| 1048 | name: "Smoke".to_string(), |
| 1049 | description: None, |
| 1050 | objective: Some("prove it runs".to_string()), |
| 1051 | instructions: instructions.to_string(), |
| 1052 | worker: Some(FleetTaskWorkerProfile { |
| 1053 | agent_profile: None, |
| 1054 | role: Some("reviewer".to_string()), |
| 1055 | loadout: None, |
| 1056 | model_class: None, |
| 1057 | model: None, |
| 1058 | tool_profile: Some("read-only".to_string()), |
| 1059 | tools: vec![], |
| 1060 | capabilities: vec![], |
| 1061 | }), |
| 1062 | workspace: None, |
| 1063 | input_files: vec![], |
| 1064 | context: vec![], |
| 1065 | budget: None, |
| 1066 | tags: vec![], |
| 1067 | expected_artifacts: vec![], |
| 1068 | scorer: None, |
| 1069 | retry_policy: None, |
| 1070 | alert_policy: None, |
| 1071 | timeout_seconds: None, |
| 1072 | metadata: BTreeMap::new(), |
| 1073 | } |
| 1074 | } |
| 1075 | |
| 1076 | fn agent_profile(id: &str, role: &str, instructions: &str) -> AgentProfile { |
| 1077 | AgentProfile { |
| 1078 | id: id.to_string(), |
| 1079 | display_name: Some(format!("{role} profile")), |
| 1080 | description: Some(format!("{role} description")), |
| 1081 | requires: Vec::new(), |
| 1082 | profile: FleetProfile { |
| 1083 | slot: FleetSlot::from_name(role), |
| 1084 | role: FleetRole { |
| 1085 | name: role.to_string(), |
| 1086 | description: None, |
| 1087 | instructions: Some(instructions.to_string()), |
| 1088 | }, |
| 1089 | loadout: FleetLoadout::Inherit, |
| 1090 | model: None, |
| 1091 | provider: None, |
| 1092 | reasoning_effort: None, |
| 1093 | permissions: FleetProfilePermissions::default(), |
| 1094 | delegation: FleetDelegationHints::default(), |
| 1095 | }, |
| 1096 | source: std::path::PathBuf::from(format!("{id}.toml")), |
| 1097 | origin: crate::fleet::roster::ProfileOrigin::Workspace, |
| 1098 | plugin_authority: None, |
| 1099 | } |
| 1100 | } |
| 1101 | |
| 1102 | fn launch_spec(task: &FleetTaskSpec, workspace: &std::path::Path) -> AgentWorkerSpec { |
| 1103 | let worker = FleetWorkerSpec { |
| 1104 | id: "worker-1".to_string(), |
| 1105 | name: "Worker 1".to_string(), |
| 1106 | host: FleetHostSpec::Local, |
| 1107 | trust_level: None, |
| 1108 | labels: BTreeMap::new(), |
| 1109 | capabilities: Vec::new(), |
| 1110 | max_concurrent_tasks: Some(1), |
| 1111 | }; |
| 1112 | crate::fleet::worker_runtime::fleet_task_to_worker_spec_with_profiles( |
| 1113 | "worker-1", |
| 1114 | "run-1", |
| 1115 | task, |
| 1116 | &worker, |
| 1117 | "auto", |
| 1118 | workspace, |
| 1119 | workspace, |
| 1120 | &[], |
| 1121 | None, |
| 1122 | ) |
| 1123 | .unwrap() |
| 1124 | } |
| 1125 | |
| 1126 | fn track_test_stream( |
| 1127 | executor: &mut FleetExecutor, |
| 1128 | worker_id: &str, |
| 1129 | log_path: std::path::PathBuf, |
| 1130 | ) { |
| 1131 | executor.streams.insert( |
| 1132 | worker_id.to_string(), |
| 1133 | WorkerStream { |
| 1134 | log_path, |
| 1135 | host: WorkerStreamHost::Local, |
| 1136 | attempt: None, |
| 1137 | offset: 0, |
| 1138 | pending: Vec::new(), |
| 1139 | terminal: false, |
| 1140 | terminal_route: TerminalRouteEvidence::default(), |
| 1141 | started_at: std::time::Instant::now(), |
| 1142 | final_answer: None, |
| 1143 | saved_session_id: None, |
| 1144 | session_capture: None, |
| 1145 | }, |
| 1146 | ); |
| 1147 | } |
| 1148 | |
| 1149 | fn append_test_stream(path: &std::path::Path, bytes: &[u8]) { |
| 1150 | use std::io::Write as _; |
| 1151 | |
| 1152 | std::fs::OpenOptions::new() |
| 1153 | .append(true) |
| 1154 | .open(path) |
| 1155 | .unwrap() |
| 1156 | .write_all(bytes) |
| 1157 | .unwrap(); |
| 1158 | } |
| 1159 | |
| 1160 | #[test] |
| 1161 | fn worker_command_is_a_headless_codewhale_exec_run() { |
| 1162 | let exec = FleetExecConfig::default(); |
| 1163 | let cmd = build_worker_exec_command("codewhale", &task("read the file"), &exec, None); |
| 1164 | assert_eq!(cmd.program, "codewhale"); |
| 1165 | assert_eq!(cmd.args[0], "exec"); |
| 1166 | assert!(cmd.args.contains(&"--auto".to_string())); |
| 1167 | // stream-json so the executor can ingest the worker's event stream. |
| 1168 | let joined = cmd.args.join(" "); |
| 1169 | assert!(joined.contains("--output-format stream-json")); |
| 1170 | // The task instructions ride in the positional prompt (last arg). |
| 1171 | assert!(cmd.args.last().unwrap().contains("read the file")); |
| 1172 | } |
| 1173 | |
| 1174 | #[test] |
| 1175 | fn worker_command_threads_exec_hardening_flags() { |
| 1176 | let exec = FleetExecConfig { |
| 1177 | allowed_tools: vec!["read_file".to_string(), "grep_files".to_string()], |
| 1178 | disallowed_tools: vec!["exec_shell".to_string()], |
| 1179 | max_turns: 40, |
| 1180 | append_system_prompt: "never push to main".to_string(), |
| 1181 | ..FleetExecConfig::default() |
| 1182 | }; |
| 1183 | let cmd = build_worker_exec_command("codewhale", &task("audit"), &exec, Some("glm-5.1")); |
| 1184 | let exec_idx = cmd |
| 1185 | .args |
| 1186 | .iter() |
| 1187 | .position(|arg| arg == "exec") |
| 1188 | .expect("worker command must contain exec"); |
| 1189 | let model_idx = cmd |
| 1190 | .args |
| 1191 | .iter() |
| 1192 | .position(|arg| arg == "--model") |
| 1193 | .expect("worker command must contain --model"); |
| 1194 | assert!( |
| 1195 | model_idx < exec_idx, |
| 1196 | "global --model must precede exec: {:?}", |
| 1197 | cmd.args |
| 1198 | ); |
| 1199 | let joined = cmd.args.join(" "); |
| 1200 | assert!(joined.contains("--model glm-5.1")); |
| 1201 | assert!(joined.contains("--allowed-tools read_file,grep_files")); |
| 1202 | assert!(joined.contains("--disallowed-tools exec_shell")); |
| 1203 | assert!(joined.contains("--max-turns 40")); |
| 1204 | assert!(cmd.args.iter().any(|a| a == "never push to main")); |
| 1205 | } |
| 1206 | |
| 1207 | #[test] |
| 1208 | fn worker_command_threads_positive_task_budgets_and_caps_steps() { |
| 1209 | let mut task = task("audit"); |
| 1210 | task.budget = Some(FleetTaskBudget { |
| 1211 | max_steps: Some(75), |
| 1212 | max_tool_calls: Some(11), |
| 1213 | ..FleetTaskBudget::default() |
| 1214 | }); |
| 1215 | let exec = FleetExecConfig { |
| 1216 | max_turns: 40, |
| 1217 | ..FleetExecConfig::default() |
| 1218 | }; |
| 1219 | |
| 1220 | let cmd = build_worker_exec_command("codewhale", &task, &exec, None); |
| 1221 | let max_turns_idx = cmd |
| 1222 | .args |
| 1223 | .iter() |
| 1224 | .position(|arg| arg == "--max-turns") |
| 1225 | .expect("positive max_steps must reach exec"); |
| 1226 | let max_tool_calls_idx = cmd |
| 1227 | .args |
| 1228 | .iter() |
| 1229 | .position(|arg| arg == "--max-tool-calls") |
| 1230 | .expect("positive max_tool_calls must reach exec"); |
| 1231 | |
| 1232 | assert_eq!(cmd.args[max_turns_idx + 1], "40"); |
| 1233 | assert_eq!(cmd.args[max_tool_calls_idx + 1], "11"); |
| 1234 | } |
| 1235 | |
| 1236 | #[test] |
| 1237 | fn production_worker_command_uses_hardened_launch_spec_steps() { |
| 1238 | let tmp = TempDir::new().unwrap(); |
| 1239 | let mut task = task("audit"); |
| 1240 | task.budget = Some(FleetTaskBudget { |
| 1241 | max_steps: Some(75), |
| 1242 | max_tool_calls: Some(11), |
| 1243 | ..FleetTaskBudget::default() |
| 1244 | }); |
| 1245 | let mut launch_spec = launch_spec(&task, tmp.path()); |
| 1246 | // The manager owns this hardening step. A different value here proves |
| 1247 | // production argv comes from the registered spec, not a second budget |
| 1248 | // projection from the task document. |
| 1249 | launch_spec.max_steps = 13; |
| 1250 | |
| 1251 | let cmd = build_worker_exec_command_with_launch_spec( |
| 1252 | "codewhale", |
| 1253 | &task, |
| 1254 | &launch_spec, |
| 1255 | &FleetExecConfig { |
| 1256 | max_turns: 40, |
| 1257 | ..FleetExecConfig::default() |
| 1258 | }, |
| 1259 | None, |
| 1260 | &[], |
| 1261 | ) |
| 1262 | .unwrap(); |
| 1263 | let max_turns_idx = cmd |
| 1264 | .args |
| 1265 | .iter() |
| 1266 | .position(|arg| arg == "--max-turns") |
| 1267 | .expect("hardened launch max_steps must reach exec"); |
| 1268 | let max_tool_calls_idx = cmd |
| 1269 | .args |
| 1270 | .iter() |
| 1271 | .position(|arg| arg == "--max-tool-calls") |
| 1272 | .expect("task max_tool_calls must reach exec"); |
| 1273 | |
| 1274 | assert_eq!(cmd.args[max_turns_idx + 1], "13"); |
| 1275 | assert_eq!(cmd.args[max_tool_calls_idx + 1], "11"); |
| 1276 | } |
| 1277 | |
| 1278 | #[test] |
| 1279 | fn worker_command_threads_agent_profile_prompt() { |
| 1280 | let mut task = task("audit"); |
| 1281 | task.worker.as_mut().unwrap().agent_profile = Some("reviewer".to_string()); |
| 1282 | let cmd = build_worker_exec_command_with_profiles( |
| 1283 | "codewhale", |
| 1284 | &task, |
| 1285 | &FleetExecConfig::default(), |
| 1286 | None, |
| 1287 | &[agent_profile( |
| 1288 | "reviewer", |
| 1289 | "reviewer", |
| 1290 | "Focus on defects, regressions, and missing tests.", |
| 1291 | )], |
| 1292 | ) |
| 1293 | .unwrap(); |
| 1294 | let prompt = cmd.args.last().unwrap(); |
| 1295 | |
| 1296 | assert!(prompt.contains("Fleet profile: reviewer")); |
| 1297 | assert!(prompt.contains("Focus on defects, regressions, and missing tests.")); |
| 1298 | } |
| 1299 | |
| 1300 | #[test] |
| 1301 | fn consultant_launch_spec_carries_read_only_authority_with_network_reads() { |
| 1302 | let tmp = TempDir::new().unwrap(); |
| 1303 | let mut task = task("advise on the release candidate"); |
| 1304 | task.worker.as_mut().unwrap().role = Some("consultant".to_string()); |
| 1305 | let launch_spec = launch_spec(&task, tmp.path()); |
| 1306 | assert_eq!( |
| 1307 | launch_spec.agent_type, |
| 1308 | crate::tools::subagent::FleetRole::Consultant |
| 1309 | ); |
| 1310 | // Counsel only: never writes the workspace, but may read the web. |
| 1311 | assert!(!launch_spec.runtime_profile.permissions.write); |
| 1312 | assert!(launch_spec.runtime_profile.permissions.network); |
| 1313 | |
| 1314 | let cmd = build_worker_exec_command_with_launch_spec( |
| 1315 | "codewhale", |
| 1316 | &task, |
| 1317 | &launch_spec, |
| 1318 | &FleetExecConfig::default(), |
| 1319 | None, |
| 1320 | &[], |
| 1321 | ) |
| 1322 | .unwrap(); |
| 1323 | |
| 1324 | assert_eq!(cmd.args.last(), Some(&launch_spec.objective)); |
| 1325 | let authority_index = cmd |
| 1326 | .args |
| 1327 | .iter() |
| 1328 | .position(|arg| arg == "--tool-authority-json") |
| 1329 | .expect("launch command must carry machine-readable authority"); |
| 1330 | let authority = ToolAuthorityEnvelope::from_json(&cmd.args[authority_index + 1]).unwrap(); |
| 1331 | assert_eq!(authority.owner, "worker-1"); |
| 1332 | assert_eq!(authority.authority, ToolMutationAuthority::ReadOnly); |
| 1333 | assert_eq!(authority.network_access, Some(true)); |
| 1334 | assert_eq!(authority.shell, ToolShellAuthority::None); |
| 1335 | assert_eq!(authority.verification, ToolVerificationAuthority::None); |
| 1336 | assert!(authority.writable_roots.is_empty()); |
| 1337 | assert!(authority.writable_files.is_empty()); |
| 1338 | assert!(authority.coordination_contracts.is_empty()); |
| 1339 | } |
| 1340 | |
| 1341 | #[test] |
| 1342 | fn scout_reviewer_and_planner_launches_carry_only_read_only_shell_authority() { |
| 1343 | let tmp = TempDir::new().unwrap(); |
| 1344 | for role in ["scout", "reviewer", "planner"] { |
| 1345 | let mut task = task("inspect repository and GitHub state"); |
| 1346 | task.worker.as_mut().unwrap().role = Some(role.to_string()); |
| 1347 | let launch_spec = launch_spec(&task, tmp.path()); |
| 1348 | let cmd = build_worker_exec_command_with_launch_spec( |
| 1349 | "codewhale", |
| 1350 | &task, |
| 1351 | &launch_spec, |
| 1352 | &FleetExecConfig::default(), |
| 1353 | None, |
| 1354 | &[], |
| 1355 | ) |
| 1356 | .unwrap(); |
| 1357 | let authority_index = cmd |
| 1358 | .args |
| 1359 | .iter() |
| 1360 | .position(|arg| arg == "--tool-authority-json") |
| 1361 | .expect("launch command must carry machine-readable authority"); |
| 1362 | let authority = |
| 1363 | ToolAuthorityEnvelope::from_json(&cmd.args[authority_index + 1]).unwrap(); |
| 1364 | |
| 1365 | assert_eq!(authority.authority, ToolMutationAuthority::ReadOnly); |
| 1366 | assert_eq!(authority.network_access, Some(true)); |
| 1367 | assert_eq!(authority.shell, ToolShellAuthority::ReadOnly); |
| 1368 | assert_eq!(authority.verification, ToolVerificationAuthority::None); |
| 1369 | } |
| 1370 | } |
| 1371 | |
| 1372 | #[test] |
| 1373 | fn verifier_launch_carries_only_bounded_verification_process_authority() { |
| 1374 | let tmp = TempDir::new().unwrap(); |
| 1375 | let mut task = task("run the focused release checks"); |
| 1376 | task.worker.as_mut().unwrap().role = Some("verifier".to_string()); |
| 1377 | let mut launch_spec = launch_spec(&task, tmp.path()); |
| 1378 | let authority = authority_envelope_for_worker(&launch_spec, &task).unwrap(); |
| 1379 | |
| 1380 | assert_eq!(authority.authority, ToolMutationAuthority::ReadOnly); |
| 1381 | assert_eq!(authority.shell, ToolShellAuthority::None); |
| 1382 | assert_eq!(authority.verification, ToolVerificationAuthority::Bounded); |
| 1383 | |
| 1384 | launch_spec.runtime_profile.shell = crate::worker_profile::ShellPolicy::None; |
| 1385 | let shellless = authority_envelope_for_worker(&launch_spec, &task).unwrap(); |
| 1386 | assert_eq!( |
| 1387 | shellless.verification, |
| 1388 | ToolVerificationAuthority::None, |
| 1389 | "the parent shell ceiling also removes bounded verification process authority" |
| 1390 | ); |
| 1391 | } |
| 1392 | |
| 1393 | #[test] |
| 1394 | fn launch_spec_command_preserves_exact_write_scope() { |
| 1395 | let tmp = TempDir::new().unwrap(); |
| 1396 | let mut task = task("edit the bounded source tree"); |
| 1397 | let worker = task.worker.as_mut().unwrap(); |
| 1398 | worker.role = Some("implementer".to_string()); |
| 1399 | worker.tool_profile = None; |
| 1400 | task.workspace = Some(FleetWorkspaceRequirements { |
| 1401 | writable_paths: vec![std::path::PathBuf::from("src")], |
| 1402 | ..FleetWorkspaceRequirements::default() |
| 1403 | }); |
| 1404 | let launch_spec = launch_spec(&task, tmp.path()); |
| 1405 | |
| 1406 | let cmd = build_worker_exec_command_with_launch_spec( |
| 1407 | "codewhale", |
| 1408 | &task, |
| 1409 | &launch_spec, |
| 1410 | &FleetExecConfig::default(), |
| 1411 | None, |
| 1412 | &[], |
| 1413 | ) |
| 1414 | .unwrap(); |
| 1415 | let authority_index = cmd |
| 1416 | .args |
| 1417 | .iter() |
| 1418 | .position(|arg| arg == "--tool-authority-json") |
| 1419 | .expect("launch command must carry machine-readable authority"); |
| 1420 | let authority = ToolAuthorityEnvelope::from_json(&cmd.args[authority_index + 1]).unwrap(); |
| 1421 | |
| 1422 | assert_eq!(authority.authority, ToolMutationAuthority::ScopedWrite); |
| 1423 | assert_eq!(authority.network_access, Some(true)); |
| 1424 | assert_eq!(authority.shell, ToolShellAuthority::None); |
| 1425 | assert_eq!(authority.verification, ToolVerificationAuthority::None); |
| 1426 | assert_eq!(authority.writable_roots, ["src"]); |
| 1427 | assert!(authority.writable_files.is_empty()); |
| 1428 | assert!(authority.coordination_contracts.is_empty()); |
| 1429 | assert_eq!(cmd.args.last(), Some(&launch_spec.objective)); |
| 1430 | } |
| 1431 | |
| 1432 | /// #4093 AC #4 at the LAUNCH boundary (not just the receipt): a worker whose |
| 1433 | /// profile pins a DIFFERENT provider+model than the parent session must |
| 1434 | /// actually launch on the profile's route and saved reasoning tier. The |
| 1435 | /// parent session is DeepSeek here (`--model deepseek-v4-pro`); the profile |
| 1436 | /// pins OpenRouter + glm-5.2 + max thinking. The emitted argv must carry |
| 1437 | /// OpenRouter's id, the profile's model, and the profile's thinking tier as |
| 1438 | /// paired flag/values — never the parent's model. This is the gap the |
| 1439 | /// save→load→resolve receipt tests never covered. |
| 1440 | #[test] |
| 1441 | fn worker_command_launches_profile_bound_provider_and_model_not_the_parent() { |
| 1442 | let mut task = task("audit"); |
| 1443 | task.worker.as_mut().unwrap().agent_profile = Some("cross".to_string()); |
| 1444 | |
| 1445 | let mut profile = agent_profile("cross", "scout", "Read first."); |
| 1446 | profile.profile.provider = Some("openrouter".to_string()); |
| 1447 | profile.profile.model = Some("glm-5.2".to_string()); |
| 1448 | profile.profile.reasoning_effort = Some("max".to_string()); |
| 1449 | |
| 1450 | let cmd = build_worker_exec_command_with_profiles( |
| 1451 | "codewhale", |
| 1452 | &task, |
| 1453 | &FleetExecConfig::default(), |
| 1454 | Some("deepseek-v4-pro"), // parent/session model on provider A. |
| 1455 | &[profile], |
| 1456 | ) |
| 1457 | .unwrap(); |
| 1458 | |
| 1459 | // Assert the flag/value PAIRS, so the provider and model are proven to |
| 1460 | // ride together rather than merely appearing somewhere on the argv. |
| 1461 | let provider_idx = cmd |
| 1462 | .args |
| 1463 | .iter() |
| 1464 | .position(|a| a == "--provider") |
| 1465 | .expect("--provider must be threaded for a provider-pinned worker"); |
| 1466 | let exec_idx = cmd |
| 1467 | .args |
| 1468 | .iter() |
| 1469 | .position(|a| a == "exec") |
| 1470 | .expect("worker command must contain exec"); |
| 1471 | assert_eq!( |
| 1472 | cmd.args.get(provider_idx + 1).map(String::as_str), |
| 1473 | Some("openrouter"), |
| 1474 | "{:?}", |
| 1475 | cmd.args |
| 1476 | ); |
| 1477 | assert!( |
| 1478 | provider_idx < exec_idx, |
| 1479 | "global --provider must precede exec: {:?}", |
| 1480 | cmd.args |
| 1481 | ); |
| 1482 | let model_idx = cmd |
| 1483 | .args |
| 1484 | .iter() |
| 1485 | .position(|a| a == "--model") |
| 1486 | .expect("--model must be present"); |
| 1487 | assert_eq!( |
| 1488 | cmd.args.get(model_idx + 1).map(String::as_str), |
| 1489 | Some("glm-5.2"), |
| 1490 | "{:?}", |
| 1491 | cmd.args |
| 1492 | ); |
| 1493 | assert!( |
| 1494 | model_idx < exec_idx, |
| 1495 | "global --model must precede exec: {:?}", |
| 1496 | cmd.args |
| 1497 | ); |
| 1498 | let reasoning_idx = cmd |
| 1499 | .args |
| 1500 | .iter() |
| 1501 | .position(|a| a == "--reasoning-effort") |
| 1502 | .expect("--reasoning-effort must be present for a thinking-pinned worker"); |
| 1503 | assert_eq!( |
| 1504 | cmd.args.get(reasoning_idx + 1).map(String::as_str), |
| 1505 | Some("max"), |
| 1506 | "{:?}", |
| 1507 | cmd.args |
| 1508 | ); |
| 1509 | assert!( |
| 1510 | reasoning_idx > exec_idx, |
| 1511 | "exec-only --reasoning-effort must follow exec: {:?}", |
| 1512 | cmd.args |
| 1513 | ); |
| 1514 | |
| 1515 | assert_eq!( |
| 1516 | &cmd.args[..exec_idx], |
| 1517 | ["--model", "glm-5.2", "--provider", "openrouter"], |
| 1518 | "route flags must form the complete global prefix: {:?}", |
| 1519 | cmd.args |
| 1520 | ); |
| 1521 | assert_eq!( |
| 1522 | &cmd.args[exec_idx..exec_idx + 5], |
| 1523 | [ |
| 1524 | "exec", |
| 1525 | "--auto", |
| 1526 | "--output-format", |
| 1527 | "stream-json", |
| 1528 | "--parent-death-watch" |
| 1529 | ], |
| 1530 | "exec flags must remain behind the subcommand: {:?}", |
| 1531 | cmd.args |
| 1532 | ); |
| 1533 | |
| 1534 | // The parent/session model must NOT leak onto the argv. |
| 1535 | assert!( |
| 1536 | !cmd.args.iter().any(|a| a == "deepseek-v4-pro"), |
| 1537 | "parent model leaked into a profile-pinned worker's argv: {:?}", |
| 1538 | cmd.args |
| 1539 | ); |
| 1540 | } |
| 1541 | |
| 1542 | #[test] |
| 1543 | fn worker_command_threads_custom_profile_provider_name() { |
| 1544 | let mut task = task("format"); |
| 1545 | task.worker.as_mut().unwrap().agent_profile = Some("local".to_string()); |
| 1546 | |
| 1547 | let mut profile = agent_profile("local", "formatter", "Keep edits tight."); |
| 1548 | profile.profile.provider = Some("lm-studio".to_string()); |
| 1549 | profile.profile.model = Some("qwen-2.5-7b".to_string()); |
| 1550 | |
| 1551 | let cmd = build_worker_exec_command_with_profiles( |
| 1552 | "codewhale", |
| 1553 | &task, |
| 1554 | &FleetExecConfig::default(), |
| 1555 | Some("deepseek-v4-pro"), |
| 1556 | &[profile], |
| 1557 | ) |
| 1558 | .unwrap(); |
| 1559 | |
| 1560 | let provider_idx = cmd |
| 1561 | .args |
| 1562 | .iter() |
| 1563 | .position(|a| a == "--provider") |
| 1564 | .expect("--provider must be threaded for a custom provider pin"); |
| 1565 | assert_eq!( |
| 1566 | cmd.args.get(provider_idx + 1).map(String::as_str), |
| 1567 | Some("lm-studio"), |
| 1568 | "{:?}", |
| 1569 | cmd.args |
| 1570 | ); |
| 1571 | let exec_idx = cmd |
| 1572 | .args |
| 1573 | .iter() |
| 1574 | .position(|a| a == "exec") |
| 1575 | .expect("worker command must contain exec"); |
| 1576 | assert!( |
| 1577 | provider_idx < exec_idx, |
| 1578 | "global --provider must precede exec: {:?}", |
| 1579 | cmd.args |
| 1580 | ); |
| 1581 | let model_idx = cmd |
| 1582 | .args |
| 1583 | .iter() |
| 1584 | .position(|a| a == "--model") |
| 1585 | .expect("--model must be present"); |
| 1586 | assert_eq!( |
| 1587 | cmd.args.get(model_idx + 1).map(String::as_str), |
| 1588 | Some("qwen-2.5-7b"), |
| 1589 | "{:?}", |
| 1590 | cmd.args |
| 1591 | ); |
| 1592 | assert!( |
| 1593 | model_idx < exec_idx, |
| 1594 | "global --model must precede exec: {:?}", |
| 1595 | cmd.args |
| 1596 | ); |
| 1597 | } |
| 1598 | |
| 1599 | /// A worker with no profile-bound provider preserves today's behavior: the |
| 1600 | /// run-level model on `--model`, and NO `--provider` (the worker keeps its |
| 1601 | /// own session default). Guards against regressing profile-less workers. |
| 1602 | #[test] |
| 1603 | fn worker_command_without_profile_provider_omits_provider_and_keeps_run_model() { |
| 1604 | let cmd = build_worker_exec_command_with_profiles( |
| 1605 | "codewhale", |
| 1606 | &task("read"), |
| 1607 | &FleetExecConfig::default(), |
| 1608 | Some("deepseek-v4-pro"), |
| 1609 | &[], |
| 1610 | ) |
| 1611 | .unwrap(); |
| 1612 | |
| 1613 | assert!( |
| 1614 | !cmd.args.iter().any(|a| a == "--provider"), |
| 1615 | "profile-less worker must not carry --provider: {:?}", |
| 1616 | cmd.args |
| 1617 | ); |
| 1618 | assert!( |
| 1619 | !cmd.args.iter().any(|a| a == "--reasoning-effort"), |
| 1620 | "profile-less worker must not carry --reasoning-effort: {:?}", |
| 1621 | cmd.args |
| 1622 | ); |
| 1623 | let model_idx = cmd |
| 1624 | .args |
| 1625 | .iter() |
| 1626 | .position(|a| a == "--model") |
| 1627 | .expect("--model must be present"); |
| 1628 | assert_eq!( |
| 1629 | cmd.args.get(model_idx + 1).map(String::as_str), |
| 1630 | Some("deepseek-v4-pro"), |
| 1631 | "{:?}", |
| 1632 | cmd.args |
| 1633 | ); |
| 1634 | let exec_idx = cmd |
| 1635 | .args |
| 1636 | .iter() |
| 1637 | .position(|a| a == "exec") |
| 1638 | .expect("worker command must contain exec"); |
| 1639 | assert!( |
| 1640 | model_idx < exec_idx, |
| 1641 | "global --model must precede exec: {:?}", |
| 1642 | cmd.args |
| 1643 | ); |
| 1644 | } |
| 1645 | |
| 1646 | #[test] |
| 1647 | fn zero_max_turns_is_not_passed() { |
| 1648 | // Zero task budgets and max_turns mean "no cap"; neither flag should |
| 1649 | // appear in the command. |
| 1650 | let exec = FleetExecConfig { |
| 1651 | max_turns: 0, |
| 1652 | ..Default::default() |
| 1653 | }; |
| 1654 | let mut task = task("x"); |
| 1655 | task.budget = Some(FleetTaskBudget { |
| 1656 | max_steps: Some(0), |
| 1657 | max_tool_calls: Some(0), |
| 1658 | ..FleetTaskBudget::default() |
| 1659 | }); |
| 1660 | let cmd = build_worker_exec_command("codewhale", &task, &exec, None); |
| 1661 | assert!(!cmd.args.join(" ").contains("--max-turns")); |
| 1662 | assert!(!cmd.args.join(" ").contains("--max-tool-calls")); |
| 1663 | } |
| 1664 | |
| 1665 | #[test] |
| 1666 | fn default_max_turns_does_not_add_a_hidden_worker_cap() { |
| 1667 | use clap::Parser; |
| 1668 | |
| 1669 | let exec = FleetExecConfig::default(); |
| 1670 | let workspace = TempDir::new().unwrap(); |
| 1671 | for requested_steps in [None, Some(0), Some(13)] { |
| 1672 | let mut task = task("x"); |
| 1673 | task.budget = requested_steps.map(|max_steps| FleetTaskBudget { |
| 1674 | max_steps: Some(max_steps), |
| 1675 | ..FleetTaskBudget::default() |
| 1676 | }); |
| 1677 | let spec = launch_spec(&task, workspace.path()); |
| 1678 | let cmd = build_worker_exec_command_with_launch_spec( |
| 1679 | "codewhale", |
| 1680 | &task, |
| 1681 | &spec, |
| 1682 | &exec, |
| 1683 | None, |
| 1684 | &[], |
| 1685 | ) |
| 1686 | .expect("production worker command"); |
| 1687 | let cli = crate::Cli::try_parse_from( |
| 1688 | std::iter::once("codewhale").chain(cmd.args.iter().map(String::as_str)), |
| 1689 | ) |
| 1690 | .expect("production worker args must parse"); |
| 1691 | let Some(crate::Commands::Exec(args)) = cli.command else { |
| 1692 | panic!("expected worker exec command"); |
| 1693 | }; |
| 1694 | let expected = requested_steps.filter(|steps| *steps > 0); |
| 1695 | assert_eq!(args.max_turns, expected); |
| 1696 | assert_eq!(args.max_tool_calls, None); |
| 1697 | // Follow omission beyond argv through the real CLI resolver. This |
| 1698 | // previously installed 200 despite correct unbounded launch args. |
| 1699 | let turn = crate::core::turn::TurnContext::new(crate::exec_max_steps(args.max_turns)); |
| 1700 | assert_eq!(turn.step_limit(), expected); |
| 1701 | assert_eq!(turn.stop_diagnostics.effective_max_steps, expected); |
| 1702 | } |
| 1703 | } |
| 1704 | |
| 1705 | #[test] |
| 1706 | fn stream_line_maps_tool_use_to_running_tool() { |
| 1707 | let line = r#"{"type":"tool_use","name":"read_file","id":"call-7","input":{}}"#; |
| 1708 | match map_exec_stream_line(line) { |
| 1709 | Some(FleetWorkerEventPayload::RunningTool { tool, call_id }) => { |
| 1710 | assert_eq!(tool, "read_file"); |
| 1711 | assert_eq!(call_id.as_deref(), Some("call-7")); |
| 1712 | } |
| 1713 | other => panic!("expected RunningTool, got {other:?}"), |
| 1714 | } |
| 1715 | } |
| 1716 | |
| 1717 | #[test] |
| 1718 | fn stream_line_maps_done_and_error() { |
| 1719 | assert!(matches!( |
| 1720 | map_exec_stream_line(r#"{"type":"done"}"#), |
| 1721 | Some(FleetWorkerEventPayload::Completed { .. }) |
| 1722 | )); |
| 1723 | match map_exec_stream_line(r#"{"type":"error","error":"boom"}"#) { |
| 1724 | Some(FleetWorkerEventPayload::Failed { reason, .. }) => assert_eq!(reason, "boom"), |
| 1725 | other => panic!("expected Failed, got {other:?}"), |
| 1726 | } |
| 1727 | } |
| 1728 | |
| 1729 | #[test] |
| 1730 | fn stream_line_maps_workflow_receipt_to_typed_event() { |
| 1731 | let line = |
| 1732 | r#"{"type":"workflow_event","run_id":"workflow_1","event":{"type":"task_completed"}}"#; |
| 1733 | match map_exec_stream_line(line) { |
| 1734 | Some(FleetWorkerEventPayload::WorkflowEvent { |
| 1735 | workflow_run_id, |
| 1736 | event, |
| 1737 | }) => { |
| 1738 | assert_eq!(workflow_run_id, "workflow_1"); |
| 1739 | assert_eq!(event["type"], "task_completed"); |
| 1740 | } |
| 1741 | other => panic!("expected typed workflow receipt, got {other:?}"), |
| 1742 | } |
| 1743 | } |
| 1744 | |
| 1745 | #[test] |
| 1746 | fn stream_line_ignores_noise_and_bad_json() { |
| 1747 | assert!(map_exec_stream_line(r#"{"type":"session_capture","content":"x"}"#).is_none()); |
| 1748 | assert!(map_exec_stream_line("not json").is_none()); |
| 1749 | assert!(map_exec_stream_line("").is_none()); |
| 1750 | } |
| 1751 | |
| 1752 | #[test] |
| 1753 | fn terminal_route_keeps_exact_literal_custom_distinct_from_idless_root_and_redacts() { |
| 1754 | let exact = map_exec_terminal_route( |
| 1755 | r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"custom","model":"literal-model","base_url":"https://must-not-cross.invalid/v1","api_key":"sk-must-not-cross"}}"#, |
| 1756 | ) |
| 1757 | .expect("literal custom terminal route"); |
| 1758 | assert_eq!(exact.provider, "custom"); |
| 1759 | assert_eq!(exact.provider_exact_id.as_deref(), Some("custom")); |
| 1760 | assert_eq!(exact.model, "literal-model"); |
| 1761 | |
| 1762 | let root = map_exec_terminal_route( |
| 1763 | r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","model":"root-model"}}"#, |
| 1764 | ) |
| 1765 | .expect("idless root custom terminal route"); |
| 1766 | assert_eq!(root.provider, "custom"); |
| 1767 | assert_eq!(root.provider_exact_id, None); |
| 1768 | assert_eq!(root.model, "root-model"); |
| 1769 | |
| 1770 | let named = map_exec_terminal_route( |
| 1771 | r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"lm-studio","model":"local-model"}}"#, |
| 1772 | ) |
| 1773 | .expect("named custom terminal route"); |
| 1774 | assert_eq!(named.provider, "custom"); |
| 1775 | assert_eq!(named.provider_exact_id.as_deref(), Some("lm-studio")); |
| 1776 | assert_eq!(named.model, "local-model"); |
| 1777 | |
| 1778 | let reported = format!("{exact:?}").to_ascii_lowercase(); |
| 1779 | for forbidden in ["base_url", "https://", "api_key", "sk-must-not-cross"] { |
| 1780 | assert!( |
| 1781 | !reported.contains(forbidden), |
| 1782 | "allowlisted terminal route leaked {forbidden:?}: {reported}" |
| 1783 | ); |
| 1784 | } |
| 1785 | |
| 1786 | for malformed in [ |
| 1787 | r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"","model":"root-model"}}"#, |
| 1788 | r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":" ","model":"root-model"}}"#, |
| 1789 | r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":7,"model":"root-model"}}"#, |
| 1790 | r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"deepseek","provider_id":"custom-x","model":"deepseek-v4-pro"}}"#, |
| 1791 | r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"unknown-kind","model":"unknown-model"}}"#, |
| 1792 | ] { |
| 1793 | assert!( |
| 1794 | map_exec_terminal_route(malformed).is_none(), |
| 1795 | "malformed present exact id must not collapse to idless root: {malformed}" |
| 1796 | ); |
| 1797 | } |
| 1798 | } |
| 1799 | |
| 1800 | #[test] |
| 1801 | fn terminal_route_evidence_requires_exactly_one_valid_envelope() { |
| 1802 | let route_x = r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"remote-x","model":"worker-model-x"}}"#; |
| 1803 | let route_y = r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"remote-y","model":"worker-model-y"}}"#; |
| 1804 | let malformed = r#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"","model":"worker-model-x"}}"#; |
| 1805 | let noise = r#"{"type":"content","delta":"progress"}"#; |
| 1806 | |
| 1807 | let observe = |lines: &[&str]| { |
| 1808 | let mut evidence = TerminalRouteEvidence::default(); |
| 1809 | for line in lines { |
| 1810 | let value: serde_json::Value = serde_json::from_str(line).unwrap(); |
| 1811 | evidence.observe(parse_exec_terminal_route(&value)); |
| 1812 | } |
| 1813 | evidence.reported_route().cloned() |
| 1814 | }; |
| 1815 | |
| 1816 | let only = observe(&[noise, route_x]).expect("one valid route"); |
| 1817 | assert_eq!(only.provider_exact_id.as_deref(), Some("remote-x")); |
| 1818 | assert!( |
| 1819 | observe(&[route_x, malformed]).is_none(), |
| 1820 | "valid then malformed must invalidate stale evidence" |
| 1821 | ); |
| 1822 | assert!( |
| 1823 | observe(&[malformed, route_x]).is_none(), |
| 1824 | "malformed then valid must remain invalid" |
| 1825 | ); |
| 1826 | assert!( |
| 1827 | observe(&[route_x, route_y]).is_none(), |
| 1828 | "conflicting valid routes must be ambiguous" |
| 1829 | ); |
| 1830 | assert!( |
| 1831 | observe(&[route_x, route_x]).is_none(), |
| 1832 | "even identical duplicates violate the exactly-one contract" |
| 1833 | ); |
| 1834 | } |
| 1835 | |
| 1836 | #[test] |
| 1837 | fn exit_classification() { |
| 1838 | assert!(matches!( |
| 1839 | classify_worker_exit(Some(0), false), |
| 1840 | FleetWorkerEventPayload::Completed { .. } |
| 1841 | )); |
| 1842 | assert!(matches!( |
| 1843 | classify_worker_exit(Some(1), false), |
| 1844 | FleetWorkerEventPayload::Failed { |
| 1845 | recoverable: true, |
| 1846 | .. |
| 1847 | } |
| 1848 | )); |
| 1849 | assert!(matches!( |
| 1850 | classify_worker_exit(Some(0), true), |
| 1851 | FleetWorkerEventPayload::Cancelled { .. } |
| 1852 | )); |
| 1853 | } |
| 1854 | |
| 1855 | /// End-to-end: run a REAL subprocess that emits stream-json (standing in for |
| 1856 | /// `codewhale exec`), and prove the executor drains its events and terminal |
| 1857 | /// exit through the real host adapter — no codewhale binary needed. This is |
| 1858 | /// the verifiable proof that a fleet worker is an out-of-process exec run. |
| 1859 | #[cfg(unix)] |
| 1860 | #[test] |
| 1861 | fn executor_runs_real_process_and_drains_stream_json_into_ledger_events() { |
| 1862 | let tmp = tempfile::TempDir::new().unwrap(); |
| 1863 | let mut exec = FleetExecutor::new(tmp.path()); |
| 1864 | let script = r#"printf '{"type":"tool_use","name":"read_file","id":"c1","input":{}}\n'; printf '{"type":"done"}\n'"#; |
| 1865 | let command = FleetWorkerCommand::new("sh", vec!["-c".to_string(), script.to_string()]); |
| 1866 | exec.start_worker("w1", command, None).unwrap(); |
| 1867 | |
| 1868 | let mut events = Vec::new(); |
| 1869 | let deadline = std::time::Instant::now() + std::time::Duration::from_secs(5); |
| 1870 | loop { |
| 1871 | events.extend(exec.drain_events("w1")); |
| 1872 | if let Some(term) = exec.poll_terminal("w1") { |
| 1873 | events.extend(exec.drain_events("w1")); // final flush after exit |
| 1874 | events.push(term); |
| 1875 | break; |
| 1876 | } |
| 1877 | assert!( |
| 1878 | std::time::Instant::now() < deadline, |
| 1879 | "worker did not terminate; events so far: {events:?}" |
| 1880 | ); |
| 1881 | std::thread::sleep(std::time::Duration::from_millis(20)); |
| 1882 | } |
| 1883 | |
| 1884 | assert!( |
| 1885 | events.iter().any(|e| matches!( |
| 1886 | e, |
| 1887 | FleetWorkerEventPayload::RunningTool { tool, .. } if tool == "read_file" |
| 1888 | )), |
| 1889 | "expected a RunningTool(read_file) event, got {events:?}" |
| 1890 | ); |
| 1891 | assert!( |
| 1892 | events |
| 1893 | .iter() |
| 1894 | .any(|e| matches!(e, FleetWorkerEventPayload::Completed { .. })), |
| 1895 | "expected a terminal Completed event, got {events:?}" |
| 1896 | ); |
| 1897 | assert!(exec.all_terminal()); |
| 1898 | } |
| 1899 | |
| 1900 | #[cfg(unix)] |
| 1901 | fn run_worker_to_terminal(script: &str, worker_id: &str) -> FleetWorkerTerminalEvent { |
| 1902 | let tmp = tempfile::TempDir::new().unwrap(); |
| 1903 | let mut exec = FleetExecutor::new(tmp.path()); |
| 1904 | let command = FleetWorkerCommand::new("sh", vec!["-c".to_string(), script.to_string()]); |
| 1905 | exec.start_worker(worker_id, command, None).unwrap(); |
| 1906 | |
| 1907 | let deadline = std::time::Instant::now() + std::time::Duration::from_secs(5); |
| 1908 | loop { |
| 1909 | exec.drain_events(worker_id); |
| 1910 | if let Some(term) = exec.poll_terminal_with_status(worker_id) { |
| 1911 | break term; |
| 1912 | } |
| 1913 | assert!( |
| 1914 | std::time::Instant::now() < deadline, |
| 1915 | "worker did not terminate in time" |
| 1916 | ); |
| 1917 | std::thread::sleep(std::time::Duration::from_millis(20)); |
| 1918 | } |
| 1919 | } |
| 1920 | |
| 1921 | #[cfg(unix)] |
| 1922 | #[test] |
| 1923 | fn completed_worker_surfaces_terminal_final_answer_as_summary() { |
| 1924 | // Report/summary tasks produce their deliverable as the final |
| 1925 | // assistant reply, not a file artifact. The exec side emits a bounded |
| 1926 | // excerpt plus the real length on its terminal receipt; the executor |
| 1927 | // reads that (never the streamed `content` deltas, which are the run |
| 1928 | // thinking out loud) and attaches it to `Completed.summary` and the |
| 1929 | // terminal event so a receipt can show the actual result. |
| 1930 | let script = r#"printf '%s\n' '{"type":"content","content":"let me look first"}' '{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"remote-x","model":"worker-model","visible_final_answer_chars":9000,"visible_final_answer_excerpt":"the report..."}}' '{"type":"done"}'"#; |
| 1931 | let terminal = run_worker_to_terminal(script, "w1"); |
| 1932 | |
| 1933 | match &terminal.payload { |
| 1934 | FleetWorkerEventPayload::Completed { summary, .. } => { |
| 1935 | assert_eq!(summary.as_deref(), Some("the report...")); |
| 1936 | } |
| 1937 | other => panic!("expected Completed, got {other:?}"), |
| 1938 | } |
| 1939 | assert_eq!( |
| 1940 | terminal.final_answer, |
| 1941 | Some(FleetWorkerFinalAnswer { |
| 1942 | excerpt: "the report...".to_string(), |
| 1943 | chars: 9000, |
| 1944 | }) |
| 1945 | ); |
| 1946 | } |
| 1947 | |
| 1948 | #[cfg(unix)] |
| 1949 | #[test] |
| 1950 | fn failed_worker_keeps_terminal_final_answer_on_terminal_event() { |
| 1951 | // A worker that fails after writing most of a report still reports |
| 1952 | // its visible answer on the terminal receipt; the executor keeps it |
| 1953 | // on the terminal event so the receipt can retain the text. |
| 1954 | let script = r#"printf '%s\n' '{"type":"error","error":"boom"}' '{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"remote-x","model":"worker-model","visible_final_answer_chars":12,"visible_final_answer_excerpt":"partial text"}}'; exit 1"#; |
| 1955 | let terminal = run_worker_to_terminal(script, "w-failed"); |
| 1956 | |
| 1957 | assert!( |
| 1958 | matches!(terminal.payload, FleetWorkerEventPayload::Failed { .. }), |
| 1959 | "{:?}", |
| 1960 | terminal.payload |
| 1961 | ); |
| 1962 | assert_eq!( |
| 1963 | terminal |
| 1964 | .final_answer |
| 1965 | .as_ref() |
| 1966 | .map(|answer| answer.excerpt.as_str()), |
| 1967 | Some("partial text") |
| 1968 | ); |
| 1969 | } |
| 1970 | |
| 1971 | #[cfg(unix)] |
| 1972 | #[test] |
| 1973 | fn unbound_worker_session_claim_is_not_advertised() { |
| 1974 | // The worker persists its full transcript as a saved session and |
| 1975 | // reports the recoverable id via `session_capture.saved_session_id`. |
| 1976 | // The executor must capture that id on the terminal event so a caller |
| 1977 | // can resolve the final assistant reply through `GET /v1/sessions/{id}`. |
| 1978 | let script = r#"printf '%s\n' '{"type":"session_capture","content":"<redacted:log-only>","saved_session_id":"session-abc"}' '{"type":"done"}'"#; |
| 1979 | let terminal = run_worker_to_terminal(script, "w-session"); |
| 1980 | |
| 1981 | assert!(terminal.saved_session_id.is_none()); |
| 1982 | assert!(terminal.final_answer.is_none()); |
| 1983 | } |
| 1984 | |
| 1985 | #[cfg(unix)] |
| 1986 | #[test] |
| 1987 | fn worker_session_capture_requires_the_parent_id_and_a_saved_local_transcript() { |
| 1988 | for (persist, forged) in [(true, false), (false, false), (true, true)] { |
| 1989 | let tmp = tempfile::TempDir::new().unwrap(); |
| 1990 | let sessions_dir = tmp.path().join("runtime-sessions"); |
| 1991 | let manager = |
| 1992 | crate::session_manager::SessionManager::new(sessions_dir.clone()).unwrap(); |
| 1993 | let mut executor = FleetExecutor::new(tmp.path()).with_sessions_dir(sessions_dir); |
| 1994 | let script = if forged { |
| 1995 | r#"printf '{"type":"session_capture","saved_session_id":"%s"}\n' "$CODEWHALE_FLEET_CAPTURE_ID"; printf '%s\n' '{"type":"session_capture","saved_session_id":"unrelated-session"}'"# |
| 1996 | } else { |
| 1997 | r#"printf '{"type":"session_capture","saved_session_id":"%s"}\n' "$CODEWHALE_FLEET_CAPTURE_ID""# |
| 1998 | }; |
| 1999 | executor |
| 2000 | .start_worker( |
| 2001 | "capture-worker", |
| 2002 | FleetWorkerCommand::new("sh", ["-c", script]), |
| 2003 | None, |
| 2004 | ) |
| 2005 | .unwrap(); |
| 2006 | let expected = executor.streams["capture-worker"] |
| 2007 | .session_capture |
| 2008 | .as_ref() |
| 2009 | .unwrap() |
| 2010 | .0 |
| 2011 | .clone(); |
| 2012 | if persist { |
| 2013 | let saved = crate::session_manager::create_saved_session_with_id_and_mode( |
| 2014 | expected.clone(), |
| 2015 | &[], |
| 2016 | "fixture-model", |
| 2017 | tmp.path(), |
| 2018 | 0, |
| 2019 | None, |
| 2020 | Some("exec"), |
| 2021 | ); |
| 2022 | manager.save_session(&saved).unwrap(); |
| 2023 | } |
| 2024 | let deadline = std::time::Instant::now() + std::time::Duration::from_secs(5); |
| 2025 | let terminal = loop { |
| 2026 | if let Some(terminal) = executor.poll_terminal_with_status("capture-worker") { |
| 2027 | break terminal; |
| 2028 | } |
| 2029 | assert!(std::time::Instant::now() < deadline); |
| 2030 | std::thread::sleep(std::time::Duration::from_millis(20)); |
| 2031 | }; |
| 2032 | assert_eq!( |
| 2033 | terminal.saved_session_id, |
| 2034 | (persist && !forged).then_some(expected) |
| 2035 | ); |
| 2036 | } |
| 2037 | } |
| 2038 | |
| 2039 | #[test] |
| 2040 | fn terminal_final_answer_is_bounded_and_redacted_at_the_worker_boundary() { |
| 2041 | let answer = parse_exec_terminal_final_answer(&serde_json::json!({ |
| 2042 | "type": "metadata", "meta": { "receipt_kind": "terminal", |
| 2043 | "visible_final_answer_excerpt": format!("sk-ant-must-not-leak-1234567890 {}", "x".repeat(8000)), |
| 2044 | "visible_final_answer_chars": 9000, |
| 2045 | } |
| 2046 | })).unwrap(); |
| 2047 | assert!(!answer.excerpt.contains("sk-ant-must-not-leak")); |
| 2048 | assert!( |
| 2049 | answer.excerpt.chars().count() <= crate::EXEC_STREAM_FINAL_ANSWER_EXCERPT_CHARS + 3 |
| 2050 | ); |
| 2051 | assert_eq!(answer.chars, 9000); |
| 2052 | } |
| 2053 | |
| 2054 | #[test] |
| 2055 | fn terminal_final_answer_ignores_empty_and_nonterminal_receipts() { |
| 2056 | let parse = |
| 2057 | |line: &str| parse_exec_terminal_final_answer(&serde_json::from_str(line).unwrap()); |
| 2058 | assert!(parse(r#"{"type":"content","content":"streamed"}"#).is_none()); |
| 2059 | assert!( |
| 2060 | parse(r#"{"type":"metadata","meta":{"receipt_kind":"turn","visible_final_answer_excerpt":"x"}}"#) |
| 2061 | .is_none() |
| 2062 | ); |
| 2063 | assert!( |
| 2064 | parse(r#"{"type":"metadata","meta":{"receipt_kind":"terminal","visible_final_answer_excerpt":" "}}"#) |
| 2065 | .is_none() |
| 2066 | ); |
| 2067 | // A receipt without the count falls back to the excerpt length. |
| 2068 | assert_eq!( |
| 2069 | parse( |
| 2070 | r#"{"type":"metadata","meta":{"receipt_kind":"terminal","visible_final_answer_excerpt":"héllo"}}"# |
| 2071 | ), |
| 2072 | Some(FleetWorkerFinalAnswer { |
| 2073 | excerpt: "héllo".to_string(), |
| 2074 | chars: 5, |
| 2075 | }) |
| 2076 | ); |
| 2077 | } |
| 2078 | |
| 2079 | #[cfg(unix)] |
| 2080 | #[test] |
| 2081 | fn terminal_poll_final_drains_route_metadata_and_tail_payloads() { |
| 2082 | let tmp = tempfile::TempDir::new().unwrap(); |
| 2083 | let mut exec = FleetExecutor::new(tmp.path()); |
| 2084 | let script = r#"printf '%s\n' '{"type":"content","delta":"tail progress"}'; printf '%s' '{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"remote-x","model":"worker-model-x"}}'"#; |
| 2085 | let command = FleetWorkerCommand::new("sh", vec!["-c".to_string(), script.to_string()]); |
| 2086 | exec.start_worker("tail-worker", command, None).unwrap(); |
| 2087 | |
| 2088 | // Deliberately do not call the ordinary event drain. Poll only after |
| 2089 | // exit, reproducing the scheduler gap where the previous poll saw EOF |
| 2090 | // just before the worker wrote its terminal tail. |
| 2091 | let deadline = std::time::Instant::now() + std::time::Duration::from_secs(5); |
| 2092 | let terminal = loop { |
| 2093 | if let Some(terminal) = exec.poll_terminal_with_status("tail-worker") { |
| 2094 | break terminal; |
| 2095 | } |
| 2096 | assert!(std::time::Instant::now() < deadline, "terminal worker"); |
| 2097 | std::thread::sleep(std::time::Duration::from_millis(10)); |
| 2098 | }; |
| 2099 | let route = terminal.reported_route.expect("final-drained route"); |
| 2100 | assert_eq!(route.provider, "custom"); |
| 2101 | assert_eq!(route.provider_exact_id.as_deref(), Some("remote-x")); |
| 2102 | assert_eq!(route.model, "worker-model-x"); |
| 2103 | assert!( |
| 2104 | terminal |
| 2105 | .tail_payloads |
| 2106 | .iter() |
| 2107 | .any(|payload| matches!(payload, FleetWorkerEventPayload::Running)) |
| 2108 | ); |
| 2109 | } |
| 2110 | |
| 2111 | #[cfg(unix)] |
| 2112 | #[test] |
| 2113 | fn terminal_poll_trailing_malformed_route_invalidates_prior_valid_route() { |
| 2114 | let tmp = tempfile::TempDir::new().unwrap(); |
| 2115 | let mut exec = FleetExecutor::new(tmp.path()); |
| 2116 | let script = r#"printf '%s\n' '{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"remote-x","model":"worker-model-x"}}'; printf '%s' '{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"","model":"worker-model-x"}}'"#; |
| 2117 | let command = FleetWorkerCommand::new("sh", vec!["-c".to_string(), script.to_string()]); |
| 2118 | exec.start_worker("ambiguous-tail-worker", command, None) |
| 2119 | .unwrap(); |
| 2120 | |
| 2121 | // Keep all output for the terminal drain, but wait for real worker |
| 2122 | // exit instead of assuming the shell finishes within one timer tick. |
| 2123 | let deadline = std::time::Instant::now() + std::time::Duration::from_secs(5); |
| 2124 | let terminal = loop { |
| 2125 | if let Some(terminal) = exec.poll_terminal_with_status("ambiguous-tail-worker") { |
| 2126 | break terminal; |
| 2127 | } |
| 2128 | assert!(std::time::Instant::now() < deadline, "terminal worker"); |
| 2129 | std::thread::sleep(std::time::Duration::from_millis(10)); |
| 2130 | }; |
| 2131 | assert!( |
| 2132 | terminal.reported_route.is_none(), |
| 2133 | "malformed trailing terminal evidence must invalidate the prior valid route" |
| 2134 | ); |
| 2135 | } |
| 2136 | |
| 2137 | /// Dogfood smoke (#3166): several concurrent exec-style workers with one |
| 2138 | /// injected failure. Proves the executor drives a small fleet to terminal |
| 2139 | /// outcomes and that a failing worker is classified distinctly from the |
| 2140 | /// passing ones — all without the codewhale binary. |
| 2141 | #[cfg(unix)] |
| 2142 | #[test] |
| 2143 | fn executor_drives_concurrent_workers_with_injected_failure() { |
| 2144 | let tmp = tempfile::TempDir::new().unwrap(); |
| 2145 | let mut exec = FleetExecutor::new(tmp.path()); |
| 2146 | |
| 2147 | // Three healthy workers emit a tool_use + done; one injected-failure |
| 2148 | // worker emits an error event and exits non-zero. |
| 2149 | let ok = r#"printf '{"type":"tool_use","name":"grep_files","id":"c","input":{}}\n{"type":"done"}\n'"#; |
| 2150 | let bad = r#"printf '{"type":"error","error":"injected failure"}\n'; exit 7"#; |
| 2151 | for id in ["w1", "w2", "w3"] { |
| 2152 | exec.start_worker( |
| 2153 | id, |
| 2154 | FleetWorkerCommand::new("sh", vec!["-c".to_string(), ok.to_string()]), |
| 2155 | None, |
| 2156 | ) |
| 2157 | .unwrap(); |
| 2158 | } |
| 2159 | exec.start_worker( |
| 2160 | "w-fail", |
| 2161 | FleetWorkerCommand::new("sh", vec!["-c".to_string(), bad.to_string()]), |
| 2162 | None, |
| 2163 | ) |
| 2164 | .unwrap(); |
| 2165 | |
| 2166 | let ids = ["w1", "w2", "w3", "w-fail"]; |
| 2167 | let mut terminals: std::collections::BTreeMap<&str, FleetWorkerEventPayload> = |
| 2168 | std::collections::BTreeMap::new(); |
| 2169 | let deadline = std::time::Instant::now() + std::time::Duration::from_secs(8); |
| 2170 | while terminals.len() < ids.len() { |
| 2171 | for id in ids { |
| 2172 | let _ = exec.drain_events(id); |
| 2173 | if let Some(term) = exec.poll_terminal(id) { |
| 2174 | terminals.insert(id, term); |
| 2175 | } |
| 2176 | } |
| 2177 | assert!( |
| 2178 | std::time::Instant::now() < deadline, |
| 2179 | "not all workers terminated: {terminals:?}" |
| 2180 | ); |
| 2181 | std::thread::sleep(std::time::Duration::from_millis(20)); |
| 2182 | } |
| 2183 | |
| 2184 | assert!(exec.all_terminal()); |
| 2185 | for id in ["w1", "w2", "w3"] { |
| 2186 | assert!( |
| 2187 | matches!(terminals[id], FleetWorkerEventPayload::Completed { .. }), |
| 2188 | "{id} should pass, got {:?}", |
| 2189 | terminals[id] |
| 2190 | ); |
| 2191 | } |
| 2192 | assert!( |
| 2193 | matches!(terminals["w-fail"], FleetWorkerEventPayload::Failed { .. }), |
| 2194 | "injected-failure worker should fail, got {:?}", |
| 2195 | terminals["w-fail"] |
| 2196 | ); |
| 2197 | } |
| 2198 | |
| 2199 | #[test] |
| 2200 | fn terminal_route_preserves_multibyte_identity_across_read_boundaries() { |
| 2201 | let tmp = tempfile::TempDir::new().unwrap(); |
| 2202 | let log_path = tmp.path().join("split-utf8.jsonl"); |
| 2203 | std::fs::write(&log_path, []).unwrap(); |
| 2204 | let mut executor = FleetExecutor::new(tmp.path()); |
| 2205 | track_test_stream(&mut executor, "split-utf8", log_path.clone()); |
| 2206 | |
| 2207 | let provider_id = "深海鲸-供应商"; |
| 2208 | let model = "深潜-模型"; |
| 2209 | let line = format!( |
| 2210 | "{{\"type\":\"metadata\",\"meta\":{{\"receipt_kind\":\"terminal\",\"provider\":\"custom\",\"provider_id\":\"{provider_id}\",\"model\":\"{model}\"}}}}\n" |
| 2211 | ); |
| 2212 | let bytes = line.as_bytes(); |
| 2213 | let provider_start = bytes |
| 2214 | .windows("鲸".len()) |
| 2215 | .position(|window| window == "鲸".as_bytes()) |
| 2216 | .unwrap(); |
| 2217 | let model_start = bytes |
| 2218 | .windows("潜".len()) |
| 2219 | .position(|window| window == "潜".as_bytes()) |
| 2220 | .unwrap(); |
| 2221 | let provider_split = provider_start + 1; |
| 2222 | let model_split = model_start + 2; |
| 2223 | |
| 2224 | append_test_stream(&log_path, &bytes[..provider_split]); |
| 2225 | assert!(executor.drain_events("split-utf8").is_empty()); |
| 2226 | append_test_stream(&log_path, &bytes[provider_split..model_split]); |
| 2227 | assert!(executor.drain_events("split-utf8").is_empty()); |
| 2228 | append_test_stream(&log_path, &bytes[model_split..]); |
| 2229 | assert!(executor.drain_events("split-utf8").is_empty()); |
| 2230 | |
| 2231 | let route = executor |
| 2232 | .streams |
| 2233 | .get("split-utf8") |
| 2234 | .and_then(|stream| stream.terminal_route.reported_route()) |
| 2235 | .expect("one exact terminal route"); |
| 2236 | assert_eq!(route.provider, "custom"); |
| 2237 | assert_eq!(route.provider_exact_id.as_deref(), Some(provider_id)); |
| 2238 | assert_eq!(route.model, model); |
| 2239 | } |
| 2240 | |
| 2241 | #[test] |
| 2242 | fn invalid_utf8_terminal_route_fails_closed_without_lossy_identity() { |
| 2243 | let tmp = tempfile::TempDir::new().unwrap(); |
| 2244 | let log_path = tmp.path().join("invalid-utf8.jsonl"); |
| 2245 | let mut line = br#"{"type":"metadata","meta":{"receipt_kind":"terminal","provider":"custom","provider_id":"remote-x","model":"worker-model"}}"#.to_vec(); |
| 2246 | let invalid_at = line |
| 2247 | .windows(b"remote-x".len()) |
| 2248 | .position(|window| window == b"remote-x") |
| 2249 | .unwrap() |
| 2250 | + 3; |
| 2251 | line[invalid_at] = 0xff; |
| 2252 | line.push(b'\n'); |
| 2253 | std::fs::write(&log_path, line).unwrap(); |
| 2254 | |
| 2255 | let mut executor = FleetExecutor::new(tmp.path()); |
| 2256 | track_test_stream(&mut executor, "invalid-utf8", log_path); |
| 2257 | assert!(executor.drain_events("invalid-utf8").is_empty()); |
| 2258 | assert!(matches!( |
| 2259 | executor |
| 2260 | .streams |
| 2261 | .get("invalid-utf8") |
| 2262 | .map(|stream| &stream.terminal_route), |
| 2263 | Some(TerminalRouteEvidence::InvalidOrAmbiguous) |
| 2264 | )); |
| 2265 | } |
| 2266 | |
| 2267 | #[test] |
| 2268 | fn invalid_utf8_nonterminal_line_cannot_synthesize_route_evidence() { |
| 2269 | let tmp = tempfile::TempDir::new().unwrap(); |
| 2270 | let log_path = tmp.path().join("invalid-nonterminal.jsonl"); |
| 2271 | let mut line = br#"{"type":"content","delta":"ordinary-output"}"#.to_vec(); |
| 2272 | let invalid_at = line |
| 2273 | .windows(b"ordinary-output".len()) |
| 2274 | .position(|window| window == b"ordinary-output") |
| 2275 | .unwrap() |
| 2276 | + 4; |
| 2277 | line[invalid_at] = 0xff; |
| 2278 | line.push(b'\n'); |
| 2279 | std::fs::write(&log_path, line).unwrap(); |
| 2280 | |
| 2281 | let mut executor = FleetExecutor::new(tmp.path()); |
| 2282 | track_test_stream(&mut executor, "invalid-nonterminal", log_path); |
| 2283 | assert!(executor.drain_events("invalid-nonterminal").is_empty()); |
| 2284 | assert!( |
| 2285 | executor |
| 2286 | .streams |
| 2287 | .get("invalid-nonterminal") |
| 2288 | .and_then(|stream| stream.terminal_route.reported_route()) |
| 2289 | .is_none() |
| 2290 | ); |
| 2291 | } |
| 2292 | } |
| 2293 |