| 1 | use std::collections::{HashMap, VecDeque}; |
| 2 | use std::net::SocketAddr; |
| 3 | use std::path::{Path, PathBuf}; |
| 4 | use std::process::{Child, Command, Stdio}; |
| 5 | use std::sync::Arc; |
| 6 | use std::time::{Duration, Instant}; |
| 7 | |
| 8 | use anyhow::{Context, Result, anyhow, bail}; |
| 9 | use axum::extract::{DefaultBodyLimit, Request, State}; |
| 10 | use axum::http::{HeaderValue, Method, StatusCode, header}; |
| 11 | use axum::middleware::{self, Next}; |
| 12 | use axum::response::{IntoResponse, Response}; |
| 13 | use axum::routing::{get, post}; |
| 14 | use axum::{Json, Router}; |
| 15 | use codewhale_agent::ModelRegistry; |
| 16 | use codewhale_config::{CliRuntimeOverrides, ConfigStore}; |
| 17 | use codewhale_core::Runtime; |
| 18 | use codewhale_hooks::{HookDispatcher, JsonlHookSink, StdoutHookSink, UnixSocketHookSink}; |
| 19 | use codewhale_mcp::McpManager; |
| 20 | use codewhale_protocol::{ |
| 21 | AppRequest, AppResponse, PromptRequest, PromptResponse, ThreadGoalClearParams, |
| 22 | ThreadGoalGetParams, ThreadGoalSetParams, ThreadRequest, ThreadResponse, UserInputAnswerEvent, |
| 23 | }; |
| 24 | use codewhale_state::StateStore; |
| 25 | use codewhale_tools::{ToolCall, ToolRegistry}; |
| 26 | use serde::de::DeserializeOwned; |
| 27 | use serde::{Deserialize, Serialize}; |
| 28 | use serde_json::{Value, json}; |
| 29 | use tokio::io::{AsyncBufRead, AsyncBufReadExt, AsyncWrite, AsyncWriteExt, BufReader}; |
| 30 | use tokio::sync::{Mutex, RwLock}; |
| 31 | use tower_http::cors::CorsLayer; |
| 32 | use uuid::Uuid; |
| 33 | |
| 34 | /// Answers submitted for a pending `request_user_input` clarification. |
| 35 | /// |
| 36 | /// The headless runtime emits [`codewhale_protocol::EventFrame::UserInputRequest`] |
| 37 | /// fire-and-return (it has no resume channel, mirroring headless approval). |
| 38 | /// Clients POST answers back via [`AppRequest::SubmitUserInput`]; we record |
| 39 | /// them here keyed by `request_id` so a driver can retrieve and feed them into |
| 40 | /// the next turn as structured context. True in-flight resume would require an |
| 41 | /// awaiter in `invoke_tool` and is left as a follow-up. |
| 42 | type PendingUserInputAnswers = Vec<UserInputAnswerEvent>; |
| 43 | |
| 44 | mod chat_completions; |
| 45 | |
| 46 | /// Legacy DeepSeek-era naming kept for external compatibility. |
| 47 | /// |
| 48 | /// CodeWhale began life as a DeepSeek CLI; existing health probes, SDK |
| 49 | /// harnesses, and on-disk layouts still key off these names. Every remaining |
| 50 | /// legacy reference in this crate routes through this shim so a future |
| 51 | /// coordinated migration touches exactly one place (repo policy: preserve |
| 52 | /// legacy migration care). |
| 53 | mod legacy_deepseek_compat { |
| 54 | use std::path::PathBuf; |
| 55 | |
| 56 | /// Service name advertised by the HTTP and stdio health probes. |
| 57 | pub(crate) const SERVICE_NAME: &str = "deepseek-app-server"; |
| 58 | |
| 59 | /// Fallback hook-event log location used when no config path is |
| 60 | /// provided (legacy `.deepseek/` dot-directory layout). |
| 61 | pub(crate) fn default_events_log_path() -> PathBuf { |
| 62 | PathBuf::from(".deepseek/events.jsonl") |
| 63 | } |
| 64 | } |
| 65 | |
| 66 | /// Upper bound on JSON request bodies accepted by the HTTP app-server. |
| 67 | const MAX_HTTP_BODY_BYTES: usize = 16 * 1024 * 1024; |
| 68 | const MAX_SSE_FRAME_BYTES: usize = 16 * 1024 * 1024; |
| 69 | |
| 70 | const DEFAULT_CORS_ORIGINS: &[&str] = &[ |
| 71 | "http://localhost", |
| 72 | "http://localhost:1420", |
| 73 | "http://localhost:3000", |
| 74 | "http://localhost:5173", |
| 75 | "http://127.0.0.1", |
| 76 | "http://127.0.0.1:1420", |
| 77 | "tauri://localhost", |
| 78 | ]; |
| 79 | |
| 80 | #[derive(Clone)] |
| 81 | pub struct AppServerOptions { |
| 82 | pub listen: SocketAddr, |
| 83 | pub config_path: Option<PathBuf>, |
| 84 | pub auth_token: Option<String>, |
| 85 | pub insecure_no_auth: bool, |
| 86 | pub cors_origins: Vec<String>, |
| 87 | } |
| 88 | |
| 89 | impl std::fmt::Debug for AppServerOptions { |
| 90 | fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 91 | f.debug_struct("AppServerOptions") |
| 92 | .field("listen", &self.listen) |
| 93 | .field("config_path", &self.config_path) |
| 94 | .field( |
| 95 | "auth_token", |
| 96 | &self.auth_token.as_ref().map(|_| "<redacted>"), |
| 97 | ) |
| 98 | .field("insecure_no_auth", &self.insecure_no_auth) |
| 99 | .field("cors_origins", &self.cors_origins) |
| 100 | .finish() |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | /// Cached stdio→runtime bridge handle. |
| 105 | /// |
| 106 | /// The outer [`AppState::stdio_bridge`] mutex guards only the cache slot; |
| 107 | /// this inner mutex serializes traffic on one bridge (single child process |
| 108 | /// plus per-thread seq bookkeeping requires ordered access). |
| 109 | type SharedRuntimeBridge = Arc<Mutex<RuntimeBridge>>; |
| 110 | |
| 111 | #[derive(Clone)] |
| 112 | struct AppState { |
| 113 | config_path: Option<PathBuf>, |
| 114 | config: Arc<RwLock<codewhale_config::ConfigToml>>, |
| 115 | /// Read/write split mirrors [`Runtime`]'s own receivers: `&self` |
| 116 | /// operations (tool calls, status, MCP startup) share a read guard and |
| 117 | /// run concurrently; `&mut self` turns (prompt/thread) and config pushes |
| 118 | /// take the write guard because the runtime genuinely requires |
| 119 | /// exclusivity there. |
| 120 | runtime: Arc<RwLock<Runtime>>, |
| 121 | registry: ModelRegistry, |
| 122 | auth_token: Option<String>, |
| 123 | stdio_bridge: Arc<Mutex<Option<SharedRuntimeBridge>>>, |
| 124 | stdio_thread_hints: Arc<Mutex<HashMap<String, RuntimeThreadHint>>>, |
| 125 | /// Answers submitted via `AppRequest::SubmitUserInput`, keyed by |
| 126 | /// `request_id`. A driver polls this to resolve clarification questions |
| 127 | /// raised by the model during a headless run. |
| 128 | pending_user_input: Arc<Mutex<std::collections::HashMap<String, PendingUserInputAnswers>>>, |
| 129 | /// Turns currently streaming over stdio, keyed by stdio thread id. |
| 130 | /// |
| 131 | /// Deliberately kept *outside* the bridge mutex: a streaming turn holds |
| 132 | /// that mutex for its entire duration, so anything reachable only through |
| 133 | /// it cannot be used to stop the turn. This holds its own copy of what an |
| 134 | /// interrupt needs, so a cancel never waits on the turn it is cancelling. |
| 135 | in_flight_turns: Arc<Mutex<HashMap<String, InFlightTurn>>>, |
| 136 | } |
| 137 | |
| 138 | /// Everything needed to interrupt a running turn without the bridge lock. |
| 139 | #[derive(Debug, Clone)] |
| 140 | struct InFlightTurn { |
| 141 | base_url: String, |
| 142 | auth_token: Option<String>, |
| 143 | /// Thread id as the *runtime* knows it, not the stdio-facing id. |
| 144 | runtime_thread_id: String, |
| 145 | turn_id: String, |
| 146 | } |
| 147 | |
| 148 | type TurnRegistry = Arc<Mutex<HashMap<String, InFlightTurn>>>; |
| 149 | |
| 150 | #[derive(Debug, Clone, Serialize, Deserialize)] |
| 151 | struct ToolCallRequest { |
| 152 | call: ToolCall, |
| 153 | #[serde(default)] |
| 154 | cwd: Option<PathBuf>, |
| 155 | } |
| 156 | |
| 157 | #[derive(Debug, Deserialize)] |
| 158 | struct JsonRpcRequest { |
| 159 | #[serde(default)] |
| 160 | jsonrpc: Option<String>, |
| 161 | #[serde(default)] |
| 162 | id: Option<Value>, |
| 163 | method: String, |
| 164 | #[serde(default)] |
| 165 | params: Value, |
| 166 | } |
| 167 | |
| 168 | #[derive(Debug)] |
| 169 | struct JsonRpcError { |
| 170 | code: i64, |
| 171 | message: String, |
| 172 | data: Option<Value>, |
| 173 | } |
| 174 | |
| 175 | #[derive(Debug)] |
| 176 | struct StdioDispatchResult { |
| 177 | result: Value, |
| 178 | should_exit: bool, |
| 179 | } |
| 180 | |
| 181 | #[derive(Debug)] |
| 182 | struct RuntimeBridge { |
| 183 | base_url: String, |
| 184 | client: reqwest::Client, |
| 185 | auth_token: Option<String>, |
| 186 | child: Option<Child>, |
| 187 | thread_map: HashMap<String, String>, |
| 188 | last_seq_by_thread: HashMap<String, u64>, |
| 189 | } |
| 190 | |
| 191 | #[derive(Debug, Clone, Default)] |
| 192 | struct RuntimeThreadHint { |
| 193 | model: Option<String>, |
| 194 | workspace: Option<PathBuf>, |
| 195 | } |
| 196 | |
| 197 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 198 | enum TurnTerminalStatus { |
| 199 | Completed, |
| 200 | Failed, |
| 201 | Interrupted, |
| 202 | Canceled, |
| 203 | } |
| 204 | |
| 205 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 206 | enum AppTransport { |
| 207 | Http, |
| 208 | Stdio, |
| 209 | } |
| 210 | |
| 211 | #[derive(Debug, Deserialize)] |
| 212 | struct ConfigGetParams { |
| 213 | key: String, |
| 214 | } |
| 215 | |
| 216 | #[derive(Debug, Deserialize)] |
| 217 | struct ConfigSetParams { |
| 218 | key: String, |
| 219 | value: String, |
| 220 | } |
| 221 | |
| 222 | #[derive(Debug, Deserialize)] |
| 223 | struct ThreadIdParams { |
| 224 | thread_id: String, |
| 225 | } |
| 226 | |
| 227 | #[derive(Debug, Deserialize)] |
| 228 | struct ThreadMessageParams { |
| 229 | thread_id: String, |
| 230 | input: String, |
| 231 | } |
| 232 | |
| 233 | #[derive(Debug, Deserialize)] |
| 234 | struct ThreadInterruptParams { |
| 235 | thread_id: String, |
| 236 | } |
| 237 | |
| 238 | pub async fn run(options: AppServerOptions) -> Result<()> { |
| 239 | let auth_token = resolve_auth_token(&options)?; |
| 240 | let state = build_state(options.config_path.clone(), auth_token)?; |
| 241 | let app = app_router(state, &options.cors_origins); |
| 242 | |
| 243 | let listener = tokio::net::TcpListener::bind(options.listen).await?; |
| 244 | axum::serve(listener, app) |
| 245 | .with_graceful_shutdown(shutdown_signal()) |
| 246 | .await?; |
| 247 | Ok(()) |
| 248 | } |
| 249 | |
| 250 | async fn shutdown_signal() { |
| 251 | let ctrl_c = async { |
| 252 | let _ = tokio::signal::ctrl_c().await; |
| 253 | }; |
| 254 | |
| 255 | #[cfg(unix)] |
| 256 | let terminate = async { |
| 257 | match tokio::signal::unix::signal(tokio::signal::unix::SignalKind::terminate()) { |
| 258 | Ok(mut signal) => { |
| 259 | signal.recv().await; |
| 260 | } |
| 261 | Err(_) => std::future::pending::<()>().await, |
| 262 | } |
| 263 | }; |
| 264 | |
| 265 | #[cfg(not(unix))] |
| 266 | let terminate = std::future::pending::<()>(); |
| 267 | |
| 268 | tokio::select! { |
| 269 | _ = ctrl_c => {} |
| 270 | _ = terminate => {} |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | fn app_router(state: AppState, cors_origins: &[String]) -> Router { |
| 275 | let protected_routes = Router::new() |
| 276 | .route("/thread", post(thread_handler)) |
| 277 | .route("/app", post(app_handler)) |
| 278 | .route("/prompt", post(prompt_handler)) |
| 279 | .route("/tool", post(tool_handler)) |
| 280 | .route("/jobs", get(jobs_handler)) |
| 281 | .route("/mcp/startup", post(mcp_startup_handler)) |
| 282 | .route( |
| 283 | "/v1/chat/completions", |
| 284 | post(chat_completions::chat_completions_handler), |
| 285 | ) |
| 286 | .route_layer(middleware::from_fn_with_state( |
| 287 | state.clone(), |
| 288 | require_app_server_token, |
| 289 | )); |
| 290 | |
| 291 | Router::new() |
| 292 | .route("/healthz", get(healthz)) |
| 293 | .merge(protected_routes) |
| 294 | .layer(DefaultBodyLimit::max(MAX_HTTP_BODY_BYTES)) |
| 295 | .layer(cors_layer(cors_origins)) |
| 296 | .with_state(state) |
| 297 | } |
| 298 | |
| 299 | pub async fn run_stdio(config_path: Option<PathBuf>) -> Result<()> { |
| 300 | let state = build_state_with_transport(config_path, None, AppTransport::Stdio)?; |
| 301 | let reader = BufReader::new(tokio::io::stdin()).lines(); |
| 302 | let writer = tokio::io::BufWriter::new(tokio::io::stdout()); |
| 303 | run_stdio_loop(&state, reader, writer).await |
| 304 | } |
| 305 | |
| 306 | /// The stdio JSON-RPC loop, generic over its transport so it can be driven by |
| 307 | /// a duplex pipe in tests rather than the process's real stdin/stdout. |
| 308 | async fn run_stdio_loop<R, W>( |
| 309 | state: &AppState, |
| 310 | mut reader: tokio::io::Lines<R>, |
| 311 | mut writer: W, |
| 312 | ) -> Result<()> |
| 313 | where |
| 314 | R: AsyncBufRead + Unpin, |
| 315 | W: AsyncWrite + Unpin, |
| 316 | { |
| 317 | // Work that arrived while a turn was streaming. The turn owns the writer |
| 318 | // for its whole duration, so these wait for it rather than interleaving |
| 319 | // into the middle of a response. |
| 320 | let mut pending: VecDeque<PendingStdioWork> = VecDeque::new(); |
| 321 | let mut stdin_open = true; |
| 322 | |
| 323 | loop { |
| 324 | let request = match pending.pop_front() { |
| 325 | Some(PendingStdioWork::Response(response)) => { |
| 326 | write_stdio_line(&mut writer, &response).await?; |
| 327 | continue; |
| 328 | } |
| 329 | Some(PendingStdioWork::Request(request)) => request, |
| 330 | None => { |
| 331 | if !stdin_open { |
| 332 | break; |
| 333 | } |
| 334 | let Some(line) = reader.next_line().await? else { |
| 335 | break; |
| 336 | }; |
| 337 | match parse_stdio_line(&line) { |
| 338 | ParsedStdioLine::Blank => continue, |
| 339 | ParsedStdioLine::Rejected(response) => { |
| 340 | write_stdio_line(&mut writer, &response).await?; |
| 341 | continue; |
| 342 | } |
| 343 | ParsedStdioLine::Request(request) => request, |
| 344 | } |
| 345 | } |
| 346 | }; |
| 347 | |
| 348 | let id = request.id.clone(); |
| 349 | let dispatched = if request.method == "thread/message" { |
| 350 | // A turn can run for minutes. Keep reading stdin while it streams |
| 351 | // so an interrupt (or a shutdown) can actually reach it — with a |
| 352 | // plain `await` here, nothing could be read until it finished. |
| 353 | let dispatch = dispatch_stdio_request_with_writer( |
| 354 | state, |
| 355 | &mut writer, |
| 356 | &request.method, |
| 357 | request.params, |
| 358 | ); |
| 359 | tokio::pin!(dispatch); |
| 360 | loop { |
| 361 | tokio::select! { |
| 362 | outcome = &mut dispatch => break outcome, |
| 363 | line = reader.next_line(), if stdin_open => { |
| 364 | match line? { |
| 365 | None => stdin_open = false, |
| 366 | Some(line) => { |
| 367 | handle_line_during_turn(state, &line, &mut pending).await; |
| 368 | } |
| 369 | } |
| 370 | } |
| 371 | } |
| 372 | } |
| 373 | } else { |
| 374 | dispatch_stdio_request_with_writer(state, &mut writer, &request.method, request.params) |
| 375 | .await |
| 376 | }; |
| 377 | |
| 378 | match dispatched { |
| 379 | Ok(dispatch) => { |
| 380 | write_stdio_line(&mut writer, &jsonrpc_result(id, dispatch.result)).await?; |
| 381 | if dispatch.should_exit { |
| 382 | break; |
| 383 | } |
| 384 | } |
| 385 | Err(err) => { |
| 386 | write_stdio_line(&mut writer, &jsonrpc_error(id, err)).await?; |
| 387 | } |
| 388 | } |
| 389 | } |
| 390 | |
| 391 | Ok(()) |
| 392 | } |
| 393 | |
| 394 | /// Work deferred until a streaming turn releases the writer. |
| 395 | enum PendingStdioWork { |
| 396 | /// Already answered (an interrupt acted immediately); just needs writing. |
| 397 | Response(Value), |
| 398 | /// Not started yet; runs normally once the turn is done. |
| 399 | Request(JsonRpcRequest), |
| 400 | } |
| 401 | |
| 402 | enum ParsedStdioLine { |
| 403 | Blank, |
| 404 | Request(JsonRpcRequest), |
| 405 | Rejected(Value), |
| 406 | } |
| 407 | |
| 408 | fn parse_stdio_line(line: &str) -> ParsedStdioLine { |
| 409 | if line.trim().is_empty() { |
| 410 | return ParsedStdioLine::Blank; |
| 411 | } |
| 412 | let request: JsonRpcRequest = match serde_json::from_str(line) { |
| 413 | Ok(value) => value, |
| 414 | Err(err) => { |
| 415 | return ParsedStdioLine::Rejected(jsonrpc_error( |
| 416 | None, |
| 417 | JsonRpcError::parse_error(format!("invalid json: {err}")), |
| 418 | )); |
| 419 | } |
| 420 | }; |
| 421 | if request |
| 422 | .jsonrpc |
| 423 | .as_deref() |
| 424 | .is_some_and(|version| version != "2.0") |
| 425 | { |
| 426 | return ParsedStdioLine::Rejected(jsonrpc_error( |
| 427 | request.id, |
| 428 | JsonRpcError::invalid_request("jsonrpc version must be 2.0"), |
| 429 | )); |
| 430 | } |
| 431 | ParsedStdioLine::Request(request) |
| 432 | } |
| 433 | |
| 434 | /// Triage a request that arrived mid-turn. |
| 435 | /// |
| 436 | /// Cancellation is the whole point of reading here, so `thread/interrupt` |
| 437 | /// runs immediately and only its reply waits for the writer. `shutdown` also |
| 438 | /// interrupts immediately — otherwise it would block on the bridge mutex the |
| 439 | /// turn is holding — and then queues so the turn can unwind first. Everything |
| 440 | /// else simply queues: it was never urgent, and running it now would race the |
| 441 | /// turn for the writer. |
| 442 | async fn handle_line_during_turn( |
| 443 | state: &AppState, |
| 444 | line: &str, |
| 445 | pending: &mut VecDeque<PendingStdioWork>, |
| 446 | ) { |
| 447 | let request = match parse_stdio_line(line) { |
| 448 | ParsedStdioLine::Blank => return, |
| 449 | ParsedStdioLine::Rejected(response) => { |
| 450 | pending.push_back(PendingStdioWork::Response(response)); |
| 451 | return; |
| 452 | } |
| 453 | ParsedStdioLine::Request(request) => request, |
| 454 | }; |
| 455 | |
| 456 | match request.method.as_str() { |
| 457 | "thread/interrupt" => { |
| 458 | let id = request.id.clone(); |
| 459 | let response = match parse_params::<ThreadInterruptParams>(params_or_object( |
| 460 | request.params.clone(), |
| 461 | )) { |
| 462 | Ok(parsed) => match interrupt_stdio_turn(state, &parsed.thread_id).await { |
| 463 | Ok(interrupted) => jsonrpc_result( |
| 464 | id, |
| 465 | json!({ "thread_id": parsed.thread_id, "interrupted": interrupted }), |
| 466 | ), |
| 467 | Err(err) => jsonrpc_error(id, err), |
| 468 | }, |
| 469 | Err(err) => jsonrpc_error(id, err), |
| 470 | }; |
| 471 | pending.push_back(PendingStdioWork::Response(response)); |
| 472 | } |
| 473 | "shutdown" => { |
| 474 | let live: Vec<String> = state.in_flight_turns.lock().await.keys().cloned().collect(); |
| 475 | for thread_id in live { |
| 476 | let _ = interrupt_stdio_turn(state, &thread_id).await; |
| 477 | } |
| 478 | pending.push_back(PendingStdioWork::Request(request)); |
| 479 | } |
| 480 | _ => pending.push_back(PendingStdioWork::Request(request)), |
| 481 | } |
| 482 | } |
| 483 | |
| 484 | async fn write_stdio_line<W: AsyncWrite + Unpin>(writer: &mut W, response: &Value) -> Result<()> { |
| 485 | writer.write_all(&serde_json::to_vec(response)?).await?; |
| 486 | writer.write_all(b"\n").await?; |
| 487 | writer.flush().await?; |
| 488 | Ok(()) |
| 489 | } |
| 490 | |
| 491 | async fn healthz() -> Json<Value> { |
| 492 | Json(json!({ |
| 493 | "status": "ok", |
| 494 | "protocol": "v2", |
| 495 | "service": legacy_deepseek_compat::SERVICE_NAME |
| 496 | })) |
| 497 | } |
| 498 | |
| 499 | async fn thread_handler( |
| 500 | State(state): State<AppState>, |
| 501 | Json(req): Json<ThreadRequest>, |
| 502 | ) -> (StatusCode, Json<ThreadResponse>) { |
| 503 | let mut runtime = state.runtime.write().await; |
| 504 | match runtime.handle_thread(req).await { |
| 505 | Ok(res) => (StatusCode::OK, Json(res)), |
| 506 | Err(err) => ( |
| 507 | StatusCode::INTERNAL_SERVER_ERROR, |
| 508 | Json(ThreadResponse { |
| 509 | thread_id: "error".to_string(), |
| 510 | status: format!("error:{err}"), |
| 511 | thread: None, |
| 512 | threads: Vec::new(), |
| 513 | goal: None, |
| 514 | model: None, |
| 515 | model_provider: None, |
| 516 | cwd: None, |
| 517 | approval_policy: None, |
| 518 | sandbox: None, |
| 519 | events: Vec::new(), |
| 520 | data: json!({}), |
| 521 | }), |
| 522 | ), |
| 523 | } |
| 524 | } |
| 525 | |
| 526 | async fn prompt_handler( |
| 527 | State(state): State<AppState>, |
| 528 | Json(req): Json<PromptRequest>, |
| 529 | ) -> (StatusCode, Json<PromptResponse>) { |
| 530 | let mut runtime = state.runtime.write().await; |
| 531 | let overrides = CliRuntimeOverrides::default(); |
| 532 | match runtime.handle_prompt(req, &overrides).await { |
| 533 | Ok(res) => (StatusCode::OK, Json(res)), |
| 534 | Err(err) => ( |
| 535 | StatusCode::INTERNAL_SERVER_ERROR, |
| 536 | Json(PromptResponse { |
| 537 | output: err.to_string(), |
| 538 | model: "unknown".to_string(), |
| 539 | events: Vec::new(), |
| 540 | }), |
| 541 | ), |
| 542 | } |
| 543 | } |
| 544 | |
| 545 | async fn tool_handler( |
| 546 | State(state): State<AppState>, |
| 547 | Json(req): Json<ToolCallRequest>, |
| 548 | ) -> (StatusCode, Json<Value>) { |
| 549 | let cwd = req |
| 550 | .cwd |
| 551 | .unwrap_or_else(|| std::env::current_dir().unwrap_or_else(|_| PathBuf::from("."))); |
| 552 | // Resolve approval policy from config instead of hardcoding. |
| 553 | let approval_mode = { |
| 554 | let cfg = state.config.read().await; |
| 555 | cfg.approval_policy |
| 556 | .as_deref() |
| 557 | .and_then(|p| match p.trim().to_ascii_lowercase().as_str() { |
| 558 | "auto" | "yolo" => Some(codewhale_execpolicy::AskForApproval::UnlessTrusted), |
| 559 | "never" | "deny" => Some(codewhale_execpolicy::AskForApproval::Never), |
| 560 | _ => None, |
| 561 | }) |
| 562 | .unwrap_or(codewhale_execpolicy::AskForApproval::OnRequest) |
| 563 | }; |
| 564 | // `invoke_tool` takes `&self`, so long-running tool executions share a |
| 565 | // read guard: they run concurrently with each other and with status |
| 566 | // reads instead of serializing every request behind one Mutex. |
| 567 | let runtime = state.runtime.read().await; |
| 568 | match runtime.invoke_tool(req.call, approval_mode, &cwd).await { |
| 569 | Ok(value) => (StatusCode::OK, Json(value)), |
| 570 | Err(err) => ( |
| 571 | StatusCode::INTERNAL_SERVER_ERROR, |
| 572 | Json(json!({ "ok": false, "error": err.to_string() })), |
| 573 | ), |
| 574 | } |
| 575 | } |
| 576 | |
| 577 | async fn jobs_handler(State(state): State<AppState>) -> Json<AppResponse> { |
| 578 | let runtime = state.runtime.read().await; |
| 579 | Json(runtime.app_status()) |
| 580 | } |
| 581 | |
| 582 | async fn mcp_startup_handler(State(state): State<AppState>) -> Json<Value> { |
| 583 | let runtime = state.runtime.read().await; |
| 584 | let summary = runtime.mcp_startup().await; |
| 585 | Json(json!({ |
| 586 | "ok": true, |
| 587 | "summary": summary |
| 588 | })) |
| 589 | } |
| 590 | |
| 591 | async fn app_handler( |
| 592 | State(state): State<AppState>, |
| 593 | Json(req): Json<AppRequest>, |
| 594 | ) -> (StatusCode, Json<AppResponse>) { |
| 595 | let response = process_app_request(&state, req, AppTransport::Http).await; |
| 596 | (app_response_status(&response), Json(response)) |
| 597 | } |
| 598 | |
| 599 | fn app_response_status(response: &AppResponse) -> StatusCode { |
| 600 | if response.ok { |
| 601 | return StatusCode::OK; |
| 602 | } |
| 603 | if response.data.get("request_id").is_some() { |
| 604 | StatusCode::CONFLICT |
| 605 | } else if response |
| 606 | .data |
| 607 | .get("error") |
| 608 | .and_then(Value::as_str) |
| 609 | .is_some_and(|err| err.contains("failed to load config")) |
| 610 | { |
| 611 | StatusCode::INTERNAL_SERVER_ERROR |
| 612 | } else { |
| 613 | StatusCode::BAD_REQUEST |
| 614 | } |
| 615 | } |
| 616 | |
| 617 | fn build_state(config_path: Option<PathBuf>, auth_token: Option<String>) -> Result<AppState> { |
| 618 | build_state_with_transport(config_path, auth_token, AppTransport::Http) |
| 619 | } |
| 620 | |
| 621 | fn build_state_with_transport( |
| 622 | config_path: Option<PathBuf>, |
| 623 | auth_token: Option<String>, |
| 624 | transport: AppTransport, |
| 625 | ) -> Result<AppState> { |
| 626 | let has_explicit_config_path = config_path.is_some(); |
| 627 | let store = ConfigStore::load(config_path)?; |
| 628 | let config_path = has_explicit_config_path.then(|| store.path().to_path_buf()); |
| 629 | let config = store.config.clone(); |
| 630 | let exec_policy = store.exec_policy_engine(); |
| 631 | let registry = ModelRegistry::default(); |
| 632 | |
| 633 | let state_db_path = config_path |
| 634 | .as_ref() |
| 635 | .and_then(|p| p.parent().map(|parent| parent.join("state.db"))); |
| 636 | let state_store = StateStore::open(state_db_path)?; |
| 637 | |
| 638 | let mut hooks = HookDispatcher::default(); |
| 639 | // Stdio carries JSON-RPC on stdout: printing raw hook events there |
| 640 | // corrupts the protocol stream (#5165). HTTP mode keeps the stdout |
| 641 | // sink for local development visibility. |
| 642 | if transport == AppTransport::Http { |
| 643 | hooks.add_sink(Arc::new(StdoutHookSink)); |
| 644 | } |
| 645 | let hook_log_path = config_path |
| 646 | .as_ref() |
| 647 | .and_then(|p| p.parent().map(|parent| parent.join("events.jsonl"))) |
| 648 | .unwrap_or_else(legacy_deepseek_compat::default_events_log_path); |
| 649 | hooks.add_sink(Arc::new(JsonlHookSink::new(hook_log_path))); |
| 650 | |
| 651 | if let Some(socket_path) = config |
| 652 | .hook_sinks |
| 653 | .as_ref() |
| 654 | .and_then(|sinks| sinks.unix_socket_path.as_ref()) |
| 655 | .filter(|path| !path.as_os_str().is_empty()) |
| 656 | { |
| 657 | hooks.add_sink(Arc::new(UnixSocketHookSink::new(socket_path.clone()))); |
| 658 | } |
| 659 | |
| 660 | let runtime = Runtime::new( |
| 661 | config.clone(), |
| 662 | registry.clone(), |
| 663 | state_store, |
| 664 | Arc::new(ToolRegistry::default()), |
| 665 | Arc::new(McpManager::default()), |
| 666 | exec_policy, |
| 667 | hooks, |
| 668 | ); |
| 669 | |
| 670 | Ok(AppState { |
| 671 | config_path, |
| 672 | config: Arc::new(RwLock::new(config)), |
| 673 | runtime: Arc::new(RwLock::new(runtime)), |
| 674 | registry, |
| 675 | auth_token, |
| 676 | stdio_bridge: Arc::new(Mutex::new(None)), |
| 677 | stdio_thread_hints: Arc::new(Mutex::new(HashMap::new())), |
| 678 | pending_user_input: Arc::new(Mutex::new(std::collections::HashMap::new())), |
| 679 | in_flight_turns: Arc::new(Mutex::new(HashMap::new())), |
| 680 | }) |
| 681 | } |
| 682 | |
| 683 | fn resolve_auth_token(options: &AppServerOptions) -> Result<Option<String>> { |
| 684 | let configured = options.auth_token.as_ref().map(|token| token.trim()); |
| 685 | if let Some(token) = configured |
| 686 | && token.is_empty() |
| 687 | { |
| 688 | bail!("app-server auth token cannot be empty"); |
| 689 | } |
| 690 | let has_explicit_token = configured.is_some(); |
| 691 | |
| 692 | if options.insecure_no_auth { |
| 693 | if !options.listen.ip().is_loopback() { |
| 694 | bail!("refusing unauthenticated app-server bind on non-loopback address"); |
| 695 | } |
| 696 | eprintln!("warning: app-server HTTP auth disabled by --insecure-no-auth"); |
| 697 | return Ok(None); |
| 698 | } |
| 699 | |
| 700 | if !has_explicit_token && !options.listen.ip().is_loopback() { |
| 701 | bail!( |
| 702 | "refusing non-loopback app-server bind without explicit auth token; pass --auth-token or set CODEWHALE_APP_SERVER_TOKEN" |
| 703 | ); |
| 704 | } |
| 705 | |
| 706 | let token = configured |
| 707 | .map(str::to_string) |
| 708 | .unwrap_or_else(|| format!("cwapp_{}", Uuid::new_v4().simple())); |
| 709 | for line in app_server_auth_status_lines(has_explicit_token) { |
| 710 | eprintln!("{line}"); |
| 711 | } |
| 712 | Ok(Some(token)) |
| 713 | } |
| 714 | |
| 715 | fn app_server_auth_status_lines(has_explicit_token: bool) -> Vec<&'static str> { |
| 716 | if has_explicit_token { |
| 717 | return vec!["app-server auth: bearer token required for HTTP routes."]; |
| 718 | } |
| 719 | vec![ |
| 720 | "app-server auth: generated bearer token for this process (not printed).", |
| 721 | " Pass --auth-token or set CODEWHALE_APP_SERVER_TOKEN when another client needs to connect.", |
| 722 | ] |
| 723 | } |
| 724 | |
| 725 | fn cors_layer(extra_origins: &[String]) -> CorsLayer { |
| 726 | let mut origins: Vec<HeaderValue> = DEFAULT_CORS_ORIGINS |
| 727 | .iter() |
| 728 | .filter_map(|origin| HeaderValue::from_str(origin).ok()) |
| 729 | .collect(); |
| 730 | for raw in extra_origins { |
| 731 | let trimmed = raw.trim(); |
| 732 | if trimmed.is_empty() { |
| 733 | continue; |
| 734 | } |
| 735 | match HeaderValue::from_str(trimmed) { |
| 736 | Ok(value) if !origins.contains(&value) => origins.push(value), |
| 737 | Ok(_) => {} |
| 738 | Err(err) => { |
| 739 | eprintln!("warning: ignoring invalid app-server CORS origin `{trimmed}`: {err}") |
| 740 | } |
| 741 | } |
| 742 | } |
| 743 | |
| 744 | CorsLayer::new() |
| 745 | .allow_origin(origins) |
| 746 | .allow_methods([Method::GET, Method::POST, Method::OPTIONS]) |
| 747 | .allow_headers([header::AUTHORIZATION, header::CONTENT_TYPE]) |
| 748 | } |
| 749 | |
| 750 | async fn require_app_server_token( |
| 751 | State(state): State<AppState>, |
| 752 | req: Request, |
| 753 | next: Next, |
| 754 | ) -> Response { |
| 755 | let Some(expected) = state.auth_token.as_deref() else { |
| 756 | return next.run(req).await; |
| 757 | }; |
| 758 | let authorized = req |
| 759 | .headers() |
| 760 | .get(header::AUTHORIZATION) |
| 761 | .and_then(|value| value.to_str().ok()) |
| 762 | .and_then(|raw| raw.strip_prefix("Bearer ")) |
| 763 | .is_some_and(|token| constant_time_eq(token.as_bytes(), expected.as_bytes())); |
| 764 | |
| 765 | if authorized { |
| 766 | next.run(req).await |
| 767 | } else { |
| 768 | ( |
| 769 | StatusCode::UNAUTHORIZED, |
| 770 | Json(json!({ |
| 771 | "error": { |
| 772 | "message": "app-server bearer token required", |
| 773 | "status": StatusCode::UNAUTHORIZED.as_u16(), |
| 774 | } |
| 775 | })), |
| 776 | ) |
| 777 | .into_response() |
| 778 | } |
| 779 | } |
| 780 | |
| 781 | /// Compares the full length of both inputs regardless of where they first |
| 782 | /// differ, so auth failures don't leak the matching prefix length via timing. |
| 783 | fn constant_time_eq(a: &[u8], b: &[u8]) -> bool { |
| 784 | let mut diff = a.len() ^ b.len(); |
| 785 | for i in 0..a.len().max(b.len()) { |
| 786 | let x = a.get(i).copied().unwrap_or(0); |
| 787 | let y = b.get(i).copied().unwrap_or(0); |
| 788 | diff |= usize::from(x ^ y); |
| 789 | } |
| 790 | diff == 0 |
| 791 | } |
| 792 | |
| 793 | fn params_or_object(params: Value) -> Value { |
| 794 | if params.is_null() { json!({}) } else { params } |
| 795 | } |
| 796 | |
| 797 | fn parse_params<T: DeserializeOwned>(params: Value) -> std::result::Result<T, JsonRpcError> { |
| 798 | serde_json::from_value(params).map_err(|err| JsonRpcError::invalid_params(err.to_string())) |
| 799 | } |
| 800 | |
| 801 | fn jsonrpc_result(id: Option<Value>, result: Value) -> Value { |
| 802 | json!({ |
| 803 | "jsonrpc": "2.0", |
| 804 | "id": id.unwrap_or(Value::Null), |
| 805 | "result": result |
| 806 | }) |
| 807 | } |
| 808 | |
| 809 | fn jsonrpc_error(id: Option<Value>, err: JsonRpcError) -> Value { |
| 810 | json!({ |
| 811 | "jsonrpc": "2.0", |
| 812 | "id": id.unwrap_or(Value::Null), |
| 813 | "error": { |
| 814 | "code": err.code, |
| 815 | "message": err.message, |
| 816 | "data": err.data |
| 817 | } |
| 818 | }) |
| 819 | } |
| 820 | |
| 821 | impl JsonRpcError { |
| 822 | fn parse_error(message: impl Into<String>) -> Self { |
| 823 | Self { |
| 824 | code: -32700, |
| 825 | message: message.into(), |
| 826 | data: None, |
| 827 | } |
| 828 | } |
| 829 | |
| 830 | fn invalid_request(message: impl Into<String>) -> Self { |
| 831 | Self { |
| 832 | code: -32600, |
| 833 | message: message.into(), |
| 834 | data: None, |
| 835 | } |
| 836 | } |
| 837 | |
| 838 | fn method_not_found(method: &str) -> Self { |
| 839 | Self { |
| 840 | code: -32601, |
| 841 | message: format!("unsupported method: {method}"), |
| 842 | data: None, |
| 843 | } |
| 844 | } |
| 845 | |
| 846 | fn invalid_params(message: impl Into<String>) -> Self { |
| 847 | Self { |
| 848 | code: -32602, |
| 849 | message: message.into(), |
| 850 | data: None, |
| 851 | } |
| 852 | } |
| 853 | |
| 854 | /// Server error (-32000..-32099): the named thread does not exist. |
| 855 | fn thread_not_found(thread_id: &str) -> Self { |
| 856 | Self { |
| 857 | code: -32004, |
| 858 | message: format!("thread not found: {thread_id}"), |
| 859 | data: Some(json!({ |
| 860 | "error": "thread_not_found", |
| 861 | "thread_id": thread_id, |
| 862 | })), |
| 863 | } |
| 864 | } |
| 865 | |
| 866 | fn internal(message: impl Into<String>) -> Self { |
| 867 | Self { |
| 868 | code: -32603, |
| 869 | message: message.into(), |
| 870 | data: None, |
| 871 | } |
| 872 | } |
| 873 | } |
| 874 | |
| 875 | async fn handle_thread_request( |
| 876 | state: &AppState, |
| 877 | req: ThreadRequest, |
| 878 | ) -> std::result::Result<ThreadResponse, JsonRpcError> { |
| 879 | let mut runtime = state.runtime.write().await; |
| 880 | runtime |
| 881 | .handle_thread(req) |
| 882 | .await |
| 883 | .map_err(|err| JsonRpcError::internal(err.to_string())) |
| 884 | } |
| 885 | |
| 886 | async fn handle_prompt_request( |
| 887 | state: &AppState, |
| 888 | req: PromptRequest, |
| 889 | ) -> std::result::Result<PromptResponse, JsonRpcError> { |
| 890 | let mut runtime = state.runtime.write().await; |
| 891 | runtime |
| 892 | .handle_prompt(req, &CliRuntimeOverrides::default()) |
| 893 | .await |
| 894 | .map_err(|err| JsonRpcError::internal(err.to_string())) |
| 895 | } |
| 896 | |
| 897 | async fn handle_stdio_thread_message<W: AsyncWrite + Unpin>( |
| 898 | state: &AppState, |
| 899 | writer: &mut W, |
| 900 | parsed: ThreadMessageParams, |
| 901 | ) -> std::result::Result<Value, JsonRpcError> { |
| 902 | let hint = { |
| 903 | let hints = state.stdio_thread_hints.lock().await; |
| 904 | hints.get(&parsed.thread_id).cloned() |
| 905 | }; |
| 906 | let bridge = acquire_stdio_bridge(state).await?; |
| 907 | // The inner bridge lock is held for the whole turn: one child process |
| 908 | // serves all threads and per-thread seq tracking requires ordered |
| 909 | // access. The cache slot itself stays unlocked, so config updates and |
| 910 | // bridge invalidation are never queued behind a streaming turn. |
| 911 | let mut bridge = bridge.lock().await; |
| 912 | let runtime_thread_id = bridge |
| 913 | .ensure_runtime_thread(&parsed.thread_id, hint) |
| 914 | .await |
| 915 | .map_err(|err| JsonRpcError::internal(err.to_string()))?; |
| 916 | let mut result = bridge |
| 917 | .message_thread( |
| 918 | &runtime_thread_id, |
| 919 | &parsed.input, |
| 920 | writer, |
| 921 | Some((state.in_flight_turns.clone(), parsed.thread_id.clone())), |
| 922 | ) |
| 923 | .await |
| 924 | .map_err(|err| JsonRpcError::internal(err.to_string()))?; |
| 925 | if let Some(object) = result.as_object_mut() { |
| 926 | object.insert("thread_id".to_string(), Value::String(parsed.thread_id)); |
| 927 | } |
| 928 | Ok(result) |
| 929 | } |
| 930 | |
| 931 | /// Resuming or forking a thread the runtime reports as `missing` must fail |
| 932 | /// with a named not-found error. Recording the null model/workspace of that |
| 933 | /// response as a stdio hint would clobber any previously cached hint for |
| 934 | /// the same thread id (#5171). |
| 935 | fn ensure_thread_found(response: &ThreadResponse) -> std::result::Result<(), JsonRpcError> { |
| 936 | if response.status == "missing" { |
| 937 | return Err(JsonRpcError::thread_not_found(&response.thread_id)); |
| 938 | } |
| 939 | Ok(()) |
| 940 | } |
| 941 | |
| 942 | async fn record_stdio_thread_hint(state: &AppState, response: &ThreadResponse) { |
| 943 | let mut hints = state.stdio_thread_hints.lock().await; |
| 944 | hints.insert( |
| 945 | response.thread_id.clone(), |
| 946 | RuntimeThreadHint { |
| 947 | model: response.model.clone(), |
| 948 | workspace: response.cwd.clone(), |
| 949 | }, |
| 950 | ); |
| 951 | } |
| 952 | |
| 953 | /// Fetch the cached stdio→runtime bridge, spawning one on first use. |
| 954 | /// |
| 955 | /// The cache-slot lock is held only for the lookup/insert — never across |
| 956 | /// the child spawn or any request traffic — so [`invalidate_stdio_bridge`] |
| 957 | /// and other slot users are never blocked behind a slow bridge operation. |
| 958 | async fn acquire_stdio_bridge( |
| 959 | state: &AppState, |
| 960 | ) -> std::result::Result<SharedRuntimeBridge, JsonRpcError> { |
| 961 | if let Some(bridge) = state.stdio_bridge.lock().await.as_ref() { |
| 962 | return Ok(bridge.clone()); |
| 963 | } |
| 964 | let bridge = Arc::new(Mutex::new( |
| 965 | RuntimeBridge::start(state.config_path.as_deref()) |
| 966 | .await |
| 967 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 968 | )); |
| 969 | let mut slot = state.stdio_bridge.lock().await; |
| 970 | // Prefer a bridge cached by a concurrent caller while we were spawning; |
| 971 | // dropping our unused one kills the extra child via `Drop`. |
| 972 | Ok(slot.get_or_insert_with(|| bridge.clone()).clone()) |
| 973 | } |
| 974 | |
| 975 | /// Ask the runtime to interrupt a turn that is streaming right now. |
| 976 | /// |
| 977 | /// Everything this needs was copied out of the bridge when the turn started, |
| 978 | /// so it never touches the bridge mutex the turn is holding. Returns whether |
| 979 | /// a live turn was found for `thread_id`. |
| 980 | async fn interrupt_stdio_turn( |
| 981 | state: &AppState, |
| 982 | thread_id: &str, |
| 983 | ) -> std::result::Result<bool, JsonRpcError> { |
| 984 | let Some(turn) = state.in_flight_turns.lock().await.get(thread_id).cloned() else { |
| 985 | return Ok(false); |
| 986 | }; |
| 987 | let mut request = codewhale_release::platform_http_client_builder() |
| 988 | .timeout(Duration::from_secs(10)) |
| 989 | .build() |
| 990 | .map_err(|err| JsonRpcError::internal(err.to_string()))? |
| 991 | .post(format!( |
| 992 | "{}/v1/threads/{}/turns/{}/interrupt", |
| 993 | turn.base_url, turn.runtime_thread_id, turn.turn_id |
| 994 | )); |
| 995 | if let Some(token) = turn.auth_token.as_deref() { |
| 996 | request = request.bearer_auth(token); |
| 997 | } |
| 998 | request |
| 999 | .send() |
| 1000 | .await |
| 1001 | .and_then(reqwest::Response::error_for_status) |
| 1002 | .map_err(|err| JsonRpcError::internal(format!("interrupt failed: {err}")))?; |
| 1003 | Ok(true) |
| 1004 | } |
| 1005 | |
| 1006 | /// Drop the cached runtime bridge so the next stdio thread message spawns a |
| 1007 | /// fresh child that re-reads the persisted config. An in-flight message |
| 1008 | /// keeps its own [`SharedRuntimeBridge`] clone and finishes against the old |
| 1009 | /// child, which is killed when the last clone drops. |
| 1010 | async fn invalidate_stdio_bridge(state: &AppState) { |
| 1011 | let mut bridge = state.stdio_bridge.lock().await; |
| 1012 | *bridge = None; |
| 1013 | } |
| 1014 | |
| 1015 | impl RuntimeBridge { |
| 1016 | async fn start(config_path: Option<&Path>) -> Result<Self> { |
| 1017 | install_rustls_crypto_provider(); |
| 1018 | let port = reserve_runtime_port()?; |
| 1019 | let auth_token = format!("cwrt_{}", Uuid::new_v4().simple()); |
| 1020 | let child = Self::runtime_command(config_path, port, &auth_token)? |
| 1021 | .spawn() |
| 1022 | .context("failed to start runtime API bridge")?; |
| 1023 | let mut bridge = Self { |
| 1024 | base_url: format!("http://127.0.0.1:{port}"), |
| 1025 | client: codewhale_release::platform_http_client_builder() |
| 1026 | .build() |
| 1027 | .context("failed to build runtime API client")?, |
| 1028 | auth_token: Some(auth_token), |
| 1029 | child: Some(child), |
| 1030 | thread_map: HashMap::new(), |
| 1031 | last_seq_by_thread: HashMap::new(), |
| 1032 | }; |
| 1033 | bridge.wait_until_ready().await?; |
| 1034 | Ok(bridge) |
| 1035 | } |
| 1036 | |
| 1037 | fn runtime_command(config_path: Option<&Path>, port: u16, auth_token: &str) -> Result<Command> { |
| 1038 | let current_exe = std::env::current_exe().ok(); |
| 1039 | let mut command = if let Some(path) = current_exe { |
| 1040 | Command::new(path) |
| 1041 | } else { |
| 1042 | Command::new("codewhale") |
| 1043 | }; |
| 1044 | // Pass the runtime auth token out-of-band via env (not argv) so local |
| 1045 | // `ps` cannot read credential material from the child command line. |
| 1046 | // The TUI/runtime server already accepts CODEWHALE_RUNTIME_TOKEN / |
| 1047 | // DEEPSEEK_RUNTIME_TOKEN when --auth-token is absent. |
| 1048 | command |
| 1049 | .arg("app-server") |
| 1050 | .arg("--http") |
| 1051 | .arg("--host") |
| 1052 | .arg("127.0.0.1") |
| 1053 | .arg("--port") |
| 1054 | .arg(port.to_string()) |
| 1055 | .env("CODEWHALE_RUNTIME_TOKEN", auth_token) |
| 1056 | .env("DEEPSEEK_RUNTIME_TOKEN", auth_token) |
| 1057 | .stdin(Stdio::null()) |
| 1058 | .stdout(Stdio::null()) |
| 1059 | .stderr(Stdio::null()); |
| 1060 | if let Some(config_path) = config_path { |
| 1061 | command.arg("--config").arg(config_path); |
| 1062 | } |
| 1063 | Ok(command) |
| 1064 | } |
| 1065 | |
| 1066 | async fn wait_until_ready(&mut self) -> Result<()> { |
| 1067 | let deadline = Instant::now() + Duration::from_secs(15); |
| 1068 | loop { |
| 1069 | if let Some(child) = self.child.as_mut() |
| 1070 | && let Some(status) = child.try_wait()? |
| 1071 | { |
| 1072 | return Err(anyhow!( |
| 1073 | "runtime API bridge exited before becoming ready (status {status})" |
| 1074 | )); |
| 1075 | } |
| 1076 | |
| 1077 | match self |
| 1078 | .client |
| 1079 | .get(format!("{}/health", self.base_url)) |
| 1080 | .send() |
| 1081 | .await |
| 1082 | { |
| 1083 | Ok(response) if response.status().is_success() => return Ok(()), |
| 1084 | _ if Instant::now() >= deadline => { |
| 1085 | bail!( |
| 1086 | "timed out waiting for runtime API bridge at {}/health", |
| 1087 | self.base_url |
| 1088 | ) |
| 1089 | } |
| 1090 | _ => tokio::time::sleep(Duration::from_millis(50)).await, |
| 1091 | } |
| 1092 | } |
| 1093 | } |
| 1094 | |
| 1095 | fn authed(&self, builder: reqwest::RequestBuilder) -> reqwest::RequestBuilder { |
| 1096 | match self.auth_token.as_deref() { |
| 1097 | Some(token) => builder.bearer_auth(token), |
| 1098 | None => builder, |
| 1099 | } |
| 1100 | } |
| 1101 | |
| 1102 | async fn request_json(&self, builder: reqwest::RequestBuilder) -> Result<Value> { |
| 1103 | let response = builder.send().await?; |
| 1104 | let status = response.status(); |
| 1105 | let body = response.text().await?; |
| 1106 | if !status.is_success() { |
| 1107 | let detail = body.trim(); |
| 1108 | if detail.is_empty() { |
| 1109 | bail!("runtime API returned {status}"); |
| 1110 | } |
| 1111 | bail!("runtime API returned {status}: {detail}"); |
| 1112 | } |
| 1113 | serde_json::from_str(&body).with_context(|| format!("invalid runtime API json: {body}")) |
| 1114 | } |
| 1115 | |
| 1116 | async fn ensure_runtime_thread( |
| 1117 | &mut self, |
| 1118 | stdio_thread_id: &str, |
| 1119 | hint: Option<RuntimeThreadHint>, |
| 1120 | ) -> Result<String> { |
| 1121 | if let Some(runtime_thread_id) = self.thread_map.get(stdio_thread_id) { |
| 1122 | return Ok(runtime_thread_id.clone()); |
| 1123 | } |
| 1124 | let hint = hint.unwrap_or_default(); |
| 1125 | let runtime_thread_id = self |
| 1126 | .create_runtime_thread(hint.model, hint.workspace) |
| 1127 | .await?; |
| 1128 | self.thread_map |
| 1129 | .insert(stdio_thread_id.to_string(), runtime_thread_id.clone()); |
| 1130 | Ok(runtime_thread_id) |
| 1131 | } |
| 1132 | |
| 1133 | async fn create_runtime_thread( |
| 1134 | &mut self, |
| 1135 | model: Option<String>, |
| 1136 | workspace: Option<PathBuf>, |
| 1137 | ) -> Result<String> { |
| 1138 | let record = self |
| 1139 | .request_json( |
| 1140 | self.authed(self.client.post(format!("{}/v1/threads", self.base_url))) |
| 1141 | .json(&json!({ |
| 1142 | "model": model, |
| 1143 | "workspace": workspace, |
| 1144 | "mode": "agent", |
| 1145 | "archived": false, |
| 1146 | })), |
| 1147 | ) |
| 1148 | .await?; |
| 1149 | let thread_id = extract_runtime_thread_id(&record)?.to_string(); |
| 1150 | self.last_seq_by_thread |
| 1151 | .entry(thread_id.clone()) |
| 1152 | .or_insert(0); |
| 1153 | Ok(thread_id) |
| 1154 | } |
| 1155 | |
| 1156 | /// Run one turn to completion, streaming its events to `writer`. |
| 1157 | /// |
| 1158 | /// `registration` is `Some` on the stdio path: it publishes the live turn |
| 1159 | /// so an `thread/interrupt` arriving mid-stream can reach the runtime |
| 1160 | /// without waiting on the bridge mutex this call holds. |
| 1161 | async fn message_thread<W: AsyncWrite + Unpin>( |
| 1162 | &mut self, |
| 1163 | thread_id: &str, |
| 1164 | input: &str, |
| 1165 | writer: &mut W, |
| 1166 | registration: Option<(TurnRegistry, String)>, |
| 1167 | ) -> Result<Value> { |
| 1168 | let turn = self |
| 1169 | .request_json( |
| 1170 | self.authed( |
| 1171 | self.client |
| 1172 | .post(format!("{}/v1/threads/{thread_id}/turns", self.base_url)), |
| 1173 | ) |
| 1174 | .json(&json!({ "prompt": input })), |
| 1175 | ) |
| 1176 | .await?; |
| 1177 | let turn_id = turn |
| 1178 | .pointer("/turn/id") |
| 1179 | .and_then(Value::as_str) |
| 1180 | .ok_or_else(|| anyhow!("runtime API turn response missing turn.id"))? |
| 1181 | .to_string(); |
| 1182 | let response_id = format!("{thread_id}:{turn_id}"); |
| 1183 | |
| 1184 | emit_stdio_event( |
| 1185 | writer, |
| 1186 | json!({ |
| 1187 | "type": "response_start", |
| 1188 | "response_id": response_id, |
| 1189 | }), |
| 1190 | ) |
| 1191 | .await?; |
| 1192 | |
| 1193 | // Publish the turn only for the streaming window, and take it back |
| 1194 | // before any `?` below: a turn that has already finished must never |
| 1195 | // look cancellable. |
| 1196 | if let Some((registry, key)) = registration.as_ref() { |
| 1197 | registry.lock().await.insert( |
| 1198 | key.clone(), |
| 1199 | InFlightTurn { |
| 1200 | base_url: self.base_url.clone(), |
| 1201 | auth_token: self.auth_token.clone(), |
| 1202 | runtime_thread_id: thread_id.to_string(), |
| 1203 | turn_id: turn_id.clone(), |
| 1204 | }, |
| 1205 | ); |
| 1206 | } |
| 1207 | |
| 1208 | let since_seq = self.last_seq_by_thread.get(thread_id).copied().unwrap_or(0); |
| 1209 | let stream_result = self |
| 1210 | .stream_turn_events(thread_id, &turn_id, &response_id, writer, since_seq) |
| 1211 | .await; |
| 1212 | |
| 1213 | if let Some((registry, key)) = registration.as_ref() { |
| 1214 | registry.lock().await.remove(key); |
| 1215 | } |
| 1216 | |
| 1217 | let _ = emit_stdio_event( |
| 1218 | writer, |
| 1219 | json!({ |
| 1220 | "type": "response_end", |
| 1221 | "response_id": response_id, |
| 1222 | }), |
| 1223 | ) |
| 1224 | .await; |
| 1225 | |
| 1226 | let (last_seq, status, error) = stream_result?; |
| 1227 | self.last_seq_by_thread |
| 1228 | .insert(thread_id.to_string(), last_seq); |
| 1229 | |
| 1230 | match status { |
| 1231 | TurnTerminalStatus::Completed => Ok(json!({ |
| 1232 | "thread_id": thread_id, |
| 1233 | "status": "accepted", |
| 1234 | "thread": Value::Null, |
| 1235 | "threads": [], |
| 1236 | "model": Value::Null, |
| 1237 | "model_provider": Value::Null, |
| 1238 | "cwd": Value::Null, |
| 1239 | "approval_policy": Value::Null, |
| 1240 | "sandbox": Value::Null, |
| 1241 | "events": [], |
| 1242 | "data": { "turn_id": turn_id }, |
| 1243 | })), |
| 1244 | TurnTerminalStatus::Failed => Err(anyhow!( |
| 1245 | "{}", |
| 1246 | error.unwrap_or_else(|| "turn failed".to_string()) |
| 1247 | )), |
| 1248 | TurnTerminalStatus::Interrupted => Err(anyhow!( |
| 1249 | "{}", |
| 1250 | error.unwrap_or_else(|| "turn interrupted".to_string()) |
| 1251 | )), |
| 1252 | TurnTerminalStatus::Canceled => Err(anyhow!( |
| 1253 | "{}", |
| 1254 | error.unwrap_or_else(|| "turn canceled".to_string()) |
| 1255 | )), |
| 1256 | } |
| 1257 | } |
| 1258 | |
| 1259 | async fn stream_turn_events<W: AsyncWrite + Unpin>( |
| 1260 | &self, |
| 1261 | thread_id: &str, |
| 1262 | turn_id: &str, |
| 1263 | response_id: &str, |
| 1264 | writer: &mut W, |
| 1265 | since_seq: u64, |
| 1266 | ) -> Result<(u64, TurnTerminalStatus, Option<String>)> { |
| 1267 | let mut response = self |
| 1268 | .authed(self.client.get(format!( |
| 1269 | "{}/v1/threads/{thread_id}/events?since_seq={since_seq}", |
| 1270 | self.base_url |
| 1271 | ))) |
| 1272 | .send() |
| 1273 | .await? |
| 1274 | .error_for_status()?; |
| 1275 | |
| 1276 | let mut buffer = Vec::new(); |
| 1277 | let mut last_seq = since_seq; |
| 1278 | |
| 1279 | while let Some(chunk) = response.chunk().await? { |
| 1280 | buffer.extend_from_slice(&chunk); |
| 1281 | if buffer.len() > MAX_SSE_FRAME_BYTES { |
| 1282 | bail!( |
| 1283 | "runtime SSE frame exceeded {MAX_SSE_FRAME_BYTES} bytes without a frame delimiter" |
| 1284 | ); |
| 1285 | } |
| 1286 | while let Some(frame_bytes) = take_sse_frame(&mut buffer) { |
| 1287 | let Some((event_name, frame_data)) = parse_sse_frame(&frame_bytes) else { |
| 1288 | continue; |
| 1289 | }; |
| 1290 | let envelope: Value = serde_json::from_str(&frame_data) |
| 1291 | .with_context(|| format!("invalid SSE json for {event_name}: {frame_data}"))?; |
| 1292 | if let Some(seq) = envelope.get("seq").and_then(Value::as_u64) { |
| 1293 | last_seq = last_seq.max(seq); |
| 1294 | } |
| 1295 | if envelope.get("turn_id").and_then(Value::as_str) != Some(turn_id) { |
| 1296 | continue; |
| 1297 | } |
| 1298 | let payload = envelope.get("payload").cloned().unwrap_or(Value::Null); |
| 1299 | match event_name.as_str() { |
| 1300 | "item.delta" => { |
| 1301 | let kind = payload |
| 1302 | .get("kind") |
| 1303 | .and_then(Value::as_str) |
| 1304 | .unwrap_or_default(); |
| 1305 | if kind == "agent_message" |
| 1306 | && let Some(delta) = payload.get("delta").and_then(Value::as_str) |
| 1307 | && !delta.is_empty() |
| 1308 | { |
| 1309 | emit_stdio_event( |
| 1310 | writer, |
| 1311 | json!({ |
| 1312 | "type": "response_delta", |
| 1313 | "response_id": response_id, |
| 1314 | "delta": delta, |
| 1315 | }), |
| 1316 | ) |
| 1317 | .await?; |
| 1318 | } |
| 1319 | } |
| 1320 | "turn.completed" => { |
| 1321 | let status = turn_terminal_status(&payload); |
| 1322 | let error = payload |
| 1323 | .pointer("/turn/error") |
| 1324 | .and_then(Value::as_str) |
| 1325 | .map(str::to_string); |
| 1326 | return Ok((last_seq, status, error)); |
| 1327 | } |
| 1328 | _ => {} |
| 1329 | } |
| 1330 | } |
| 1331 | } |
| 1332 | |
| 1333 | bail!("runtime event stream ended before turn.completed") |
| 1334 | } |
| 1335 | |
| 1336 | #[cfg(test)] |
| 1337 | fn from_base_url_for_test(base_url: String) -> Self { |
| 1338 | install_rustls_crypto_provider(); |
| 1339 | Self { |
| 1340 | base_url, |
| 1341 | client: codewhale_release::platform_http_client_builder() |
| 1342 | .timeout(Duration::from_secs(5)) |
| 1343 | .build() |
| 1344 | .expect("build reqwest test client"), |
| 1345 | auth_token: None, |
| 1346 | child: None, |
| 1347 | thread_map: HashMap::new(), |
| 1348 | last_seq_by_thread: HashMap::new(), |
| 1349 | } |
| 1350 | } |
| 1351 | } |
| 1352 | |
| 1353 | impl RuntimeBridge { |
| 1354 | /// Kills the managed runtime child and reaps it on a detached thread so |
| 1355 | /// neither an explicit shutdown nor Drop blocks a Tokio runtime thread. |
| 1356 | fn shutdown_child(&mut self) { |
| 1357 | if let Some(mut child) = self.child.take() { |
| 1358 | let _ = child.kill(); |
| 1359 | std::thread::spawn(move || { |
| 1360 | let _ = child.wait(); |
| 1361 | }); |
| 1362 | } |
| 1363 | } |
| 1364 | } |
| 1365 | |
| 1366 | impl Drop for RuntimeBridge { |
| 1367 | fn drop(&mut self) { |
| 1368 | self.shutdown_child(); |
| 1369 | } |
| 1370 | } |
| 1371 | |
| 1372 | fn reserve_runtime_port() -> Result<u16> { |
| 1373 | let listener = std::net::TcpListener::bind("127.0.0.1:0")?; |
| 1374 | Ok(listener.local_addr()?.port()) |
| 1375 | } |
| 1376 | |
| 1377 | fn install_rustls_crypto_provider() { |
| 1378 | let _ = rustls::crypto::ring::default_provider().install_default(); |
| 1379 | } |
| 1380 | |
| 1381 | fn extract_runtime_thread_id(record: &Value) -> Result<&str> { |
| 1382 | record |
| 1383 | .get("id") |
| 1384 | .and_then(Value::as_str) |
| 1385 | .ok_or_else(|| anyhow!("runtime API thread response missing id")) |
| 1386 | } |
| 1387 | |
| 1388 | fn turn_terminal_status(payload: &Value) -> TurnTerminalStatus { |
| 1389 | match payload |
| 1390 | .pointer("/turn/status") |
| 1391 | .and_then(Value::as_str) |
| 1392 | .unwrap_or("completed") |
| 1393 | .to_ascii_lowercase() |
| 1394 | .as_str() |
| 1395 | { |
| 1396 | "failed" => TurnTerminalStatus::Failed, |
| 1397 | "interrupted" => TurnTerminalStatus::Interrupted, |
| 1398 | "canceled" | "cancelled" => TurnTerminalStatus::Canceled, |
| 1399 | _ => TurnTerminalStatus::Completed, |
| 1400 | } |
| 1401 | } |
| 1402 | |
| 1403 | async fn emit_stdio_event<W: AsyncWrite + Unpin>(writer: &mut W, event: Value) -> Result<()> { |
| 1404 | writer.write_all(&serde_json::to_vec(&event)?).await?; |
| 1405 | writer.write_all(b"\n").await?; |
| 1406 | writer.flush().await?; |
| 1407 | Ok(()) |
| 1408 | } |
| 1409 | |
| 1410 | fn take_sse_frame(buffer: &mut Vec<u8>) -> Option<Vec<u8>> { |
| 1411 | if let Some(pos) = buffer.windows(4).position(|window| window == b"\r\n\r\n") { |
| 1412 | return Some(buffer.drain(..pos + 4).collect()); |
| 1413 | } |
| 1414 | buffer |
| 1415 | .windows(2) |
| 1416 | .position(|window| window == b"\n\n") |
| 1417 | .map(|pos| buffer.drain(..pos + 2).collect()) |
| 1418 | } |
| 1419 | |
| 1420 | fn parse_sse_frame(frame_bytes: &[u8]) -> Option<(String, String)> { |
| 1421 | let text = String::from_utf8(frame_bytes.to_vec()).ok()?; |
| 1422 | let mut event_name = None; |
| 1423 | let mut data_lines = Vec::new(); |
| 1424 | for raw_line in text.lines() { |
| 1425 | let line = raw_line.trim_end_matches('\r'); |
| 1426 | if let Some(value) = line.strip_prefix("event:") { |
| 1427 | event_name = Some(value.trim().to_string()); |
| 1428 | } else if let Some(value) = line.strip_prefix("data:") { |
| 1429 | data_lines.push(value.trim_start().to_string()); |
| 1430 | } |
| 1431 | } |
| 1432 | match (event_name, data_lines.is_empty()) { |
| 1433 | (Some(event), false) => Some((event, data_lines.join("\n"))), |
| 1434 | _ => None, |
| 1435 | } |
| 1436 | } |
| 1437 | |
| 1438 | #[cfg(test)] |
| 1439 | async fn dispatch_stdio_request( |
| 1440 | state: &AppState, |
| 1441 | method: &str, |
| 1442 | params: Value, |
| 1443 | ) -> std::result::Result<StdioDispatchResult, JsonRpcError> { |
| 1444 | let mut sink = tokio::io::sink(); |
| 1445 | dispatch_stdio_request_with_writer(state, &mut sink, method, params).await |
| 1446 | } |
| 1447 | |
| 1448 | async fn dispatch_stdio_app_request( |
| 1449 | state: &AppState, |
| 1450 | request: AppRequest, |
| 1451 | ) -> std::result::Result<StdioDispatchResult, JsonRpcError> { |
| 1452 | let response = Box::pin(process_app_request(state, request, AppTransport::Stdio)).await; |
| 1453 | Ok(StdioDispatchResult { |
| 1454 | result: serde_json::to_value(response) |
| 1455 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1456 | should_exit: false, |
| 1457 | }) |
| 1458 | } |
| 1459 | |
| 1460 | async fn dispatch_stdio_request_with_writer<W: AsyncWrite + Unpin>( |
| 1461 | state: &AppState, |
| 1462 | writer: &mut W, |
| 1463 | method: &str, |
| 1464 | params: Value, |
| 1465 | ) -> std::result::Result<StdioDispatchResult, JsonRpcError> { |
| 1466 | let outcome = match method { |
| 1467 | "healthz" | "app/healthz" => StdioDispatchResult { |
| 1468 | result: json!({ |
| 1469 | "status": "ok", |
| 1470 | "service": legacy_deepseek_compat::SERVICE_NAME, |
| 1471 | "transport": "stdio" |
| 1472 | }), |
| 1473 | should_exit: false, |
| 1474 | }, |
| 1475 | "capabilities" => StdioDispatchResult { |
| 1476 | result: json!({ |
| 1477 | "transport": "stdio", |
| 1478 | "families": ["thread/*", "app/*", "prompt/*"], |
| 1479 | "methods": [ |
| 1480 | "healthz", |
| 1481 | "thread/capabilities", |
| 1482 | "thread/request", |
| 1483 | "thread/create", |
| 1484 | "thread/start", |
| 1485 | "thread/resume", |
| 1486 | "thread/fork", |
| 1487 | "thread/list", |
| 1488 | "thread/read", |
| 1489 | "thread/set_name", |
| 1490 | "thread/goal/set", |
| 1491 | "thread/goal/get", |
| 1492 | "thread/goal/clear", |
| 1493 | "thread/archive", |
| 1494 | "thread/unarchive", |
| 1495 | "thread/message", |
| 1496 | "thread/interrupt", |
| 1497 | "app/capabilities", |
| 1498 | "app/request", |
| 1499 | "app/config/get", |
| 1500 | "app/config/set", |
| 1501 | "app/config/unset", |
| 1502 | "app/config/list", |
| 1503 | "app/config/reload", |
| 1504 | "app/models", |
| 1505 | "app/thread_loaded_list", |
| 1506 | "prompt/capabilities", |
| 1507 | "prompt/request", |
| 1508 | "prompt/run", |
| 1509 | "shutdown" |
| 1510 | ] |
| 1511 | }), |
| 1512 | should_exit: false, |
| 1513 | }, |
| 1514 | "thread/capabilities" => StdioDispatchResult { |
| 1515 | result: json!({ |
| 1516 | "methods": [ |
| 1517 | "thread/request", |
| 1518 | "thread/create", |
| 1519 | "thread/start", |
| 1520 | "thread/resume", |
| 1521 | "thread/fork", |
| 1522 | "thread/list", |
| 1523 | "thread/read", |
| 1524 | "thread/set_name", |
| 1525 | "thread/goal/set", |
| 1526 | "thread/goal/get", |
| 1527 | "thread/goal/clear", |
| 1528 | "thread/archive", |
| 1529 | "thread/unarchive", |
| 1530 | "thread/message", |
| 1531 | "thread/interrupt" |
| 1532 | ] |
| 1533 | }), |
| 1534 | should_exit: false, |
| 1535 | }, |
| 1536 | "thread/request" => { |
| 1537 | let request: ThreadRequest = parse_params(params)?; |
| 1538 | if let ThreadRequest::Message { thread_id, input } = request { |
| 1539 | let response = handle_stdio_thread_message( |
| 1540 | state, |
| 1541 | writer, |
| 1542 | ThreadMessageParams { thread_id, input }, |
| 1543 | ) |
| 1544 | .await?; |
| 1545 | return Ok(StdioDispatchResult { |
| 1546 | result: response, |
| 1547 | should_exit: false, |
| 1548 | }); |
| 1549 | } |
| 1550 | let should_record_hint = matches!( |
| 1551 | &request, |
| 1552 | ThreadRequest::Create { .. } |
| 1553 | | ThreadRequest::Start(_) |
| 1554 | | ThreadRequest::Resume(_) |
| 1555 | | ThreadRequest::Fork(_) |
| 1556 | ); |
| 1557 | let response = handle_thread_request(state, request).await?; |
| 1558 | if should_record_hint { |
| 1559 | record_stdio_thread_hint(state, &response).await; |
| 1560 | } |
| 1561 | StdioDispatchResult { |
| 1562 | result: serde_json::to_value(response) |
| 1563 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1564 | should_exit: false, |
| 1565 | } |
| 1566 | } |
| 1567 | "thread/create" => { |
| 1568 | #[derive(Debug, Deserialize)] |
| 1569 | struct CreateParams { |
| 1570 | #[serde(default)] |
| 1571 | metadata: Value, |
| 1572 | } |
| 1573 | let parsed: CreateParams = parse_params(params_or_object(params))?; |
| 1574 | let response = handle_thread_request( |
| 1575 | state, |
| 1576 | ThreadRequest::Create { |
| 1577 | metadata: parsed.metadata, |
| 1578 | }, |
| 1579 | ) |
| 1580 | .await?; |
| 1581 | record_stdio_thread_hint(state, &response).await; |
| 1582 | StdioDispatchResult { |
| 1583 | result: serde_json::to_value(response) |
| 1584 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1585 | should_exit: false, |
| 1586 | } |
| 1587 | } |
| 1588 | "thread/start" => { |
| 1589 | let request = ThreadRequest::Start(parse_params(params_or_object(params))?); |
| 1590 | let response = handle_thread_request(state, request).await?; |
| 1591 | record_stdio_thread_hint(state, &response).await; |
| 1592 | StdioDispatchResult { |
| 1593 | result: serde_json::to_value(response) |
| 1594 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1595 | should_exit: false, |
| 1596 | } |
| 1597 | } |
| 1598 | "thread/resume" => { |
| 1599 | let request = ThreadRequest::Resume(parse_params(params_or_object(params))?); |
| 1600 | let response = handle_thread_request(state, request).await?; |
| 1601 | ensure_thread_found(&response)?; |
| 1602 | record_stdio_thread_hint(state, &response).await; |
| 1603 | StdioDispatchResult { |
| 1604 | result: serde_json::to_value(response) |
| 1605 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1606 | should_exit: false, |
| 1607 | } |
| 1608 | } |
| 1609 | "thread/fork" => { |
| 1610 | let request = ThreadRequest::Fork(parse_params(params_or_object(params))?); |
| 1611 | let response = handle_thread_request(state, request).await?; |
| 1612 | ensure_thread_found(&response)?; |
| 1613 | record_stdio_thread_hint(state, &response).await; |
| 1614 | StdioDispatchResult { |
| 1615 | result: serde_json::to_value(response) |
| 1616 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1617 | should_exit: false, |
| 1618 | } |
| 1619 | } |
| 1620 | "thread/list" => { |
| 1621 | let request = ThreadRequest::List(parse_params(params_or_object(params))?); |
| 1622 | let response = handle_thread_request(state, request).await?; |
| 1623 | StdioDispatchResult { |
| 1624 | result: serde_json::to_value(response) |
| 1625 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1626 | should_exit: false, |
| 1627 | } |
| 1628 | } |
| 1629 | "thread/read" => { |
| 1630 | let request = ThreadRequest::Read(parse_params(params_or_object(params))?); |
| 1631 | let response = handle_thread_request(state, request).await?; |
| 1632 | StdioDispatchResult { |
| 1633 | result: serde_json::to_value(response) |
| 1634 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1635 | should_exit: false, |
| 1636 | } |
| 1637 | } |
| 1638 | "thread/set_name" | "thread/set-name" => { |
| 1639 | let request = ThreadRequest::SetName(parse_params(params_or_object(params))?); |
| 1640 | let response = handle_thread_request(state, request).await?; |
| 1641 | StdioDispatchResult { |
| 1642 | result: serde_json::to_value(response) |
| 1643 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1644 | should_exit: false, |
| 1645 | } |
| 1646 | } |
| 1647 | "thread/goal/set" | "thread/goal_set" | "thread/goal-set" => { |
| 1648 | let request = ThreadRequest::GoalSet(parse_params::<ThreadGoalSetParams>( |
| 1649 | params_or_object(params), |
| 1650 | )?); |
| 1651 | let response = handle_thread_request(state, request).await?; |
| 1652 | StdioDispatchResult { |
| 1653 | result: serde_json::to_value(response) |
| 1654 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1655 | should_exit: false, |
| 1656 | } |
| 1657 | } |
| 1658 | "thread/goal/get" | "thread/goal_get" | "thread/goal-get" => { |
| 1659 | let request = ThreadRequest::GoalGet(parse_params::<ThreadGoalGetParams>( |
| 1660 | params_or_object(params), |
| 1661 | )?); |
| 1662 | let response = handle_thread_request(state, request).await?; |
| 1663 | StdioDispatchResult { |
| 1664 | result: serde_json::to_value(response) |
| 1665 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1666 | should_exit: false, |
| 1667 | } |
| 1668 | } |
| 1669 | "thread/goal/clear" | "thread/goal_clear" | "thread/goal-clear" => { |
| 1670 | let request = ThreadRequest::GoalClear(parse_params::<ThreadGoalClearParams>( |
| 1671 | params_or_object(params), |
| 1672 | )?); |
| 1673 | let response = handle_thread_request(state, request).await?; |
| 1674 | StdioDispatchResult { |
| 1675 | result: serde_json::to_value(response) |
| 1676 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1677 | should_exit: false, |
| 1678 | } |
| 1679 | } |
| 1680 | "thread/archive" => { |
| 1681 | let parsed: ThreadIdParams = parse_params(params_or_object(params))?; |
| 1682 | let response = handle_thread_request( |
| 1683 | state, |
| 1684 | ThreadRequest::Archive { |
| 1685 | thread_id: parsed.thread_id, |
| 1686 | }, |
| 1687 | ) |
| 1688 | .await?; |
| 1689 | StdioDispatchResult { |
| 1690 | result: serde_json::to_value(response) |
| 1691 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1692 | should_exit: false, |
| 1693 | } |
| 1694 | } |
| 1695 | "thread/unarchive" => { |
| 1696 | let parsed: ThreadIdParams = parse_params(params_or_object(params))?; |
| 1697 | let response = handle_thread_request( |
| 1698 | state, |
| 1699 | ThreadRequest::Unarchive { |
| 1700 | thread_id: parsed.thread_id, |
| 1701 | }, |
| 1702 | ) |
| 1703 | .await?; |
| 1704 | StdioDispatchResult { |
| 1705 | result: serde_json::to_value(response) |
| 1706 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1707 | should_exit: false, |
| 1708 | } |
| 1709 | } |
| 1710 | "thread/message" => { |
| 1711 | let parsed: ThreadMessageParams = parse_params(params_or_object(params))?; |
| 1712 | let response = handle_stdio_thread_message(state, writer, parsed).await?; |
| 1713 | StdioDispatchResult { |
| 1714 | result: response, |
| 1715 | should_exit: false, |
| 1716 | } |
| 1717 | } |
| 1718 | "app/capabilities" => dispatch_stdio_app_request(state, AppRequest::Capabilities).await?, |
| 1719 | "app/request" => { |
| 1720 | let request: AppRequest = parse_params(params)?; |
| 1721 | dispatch_stdio_app_request(state, request).await? |
| 1722 | } |
| 1723 | "app/config/get" => { |
| 1724 | let parsed: ConfigGetParams = parse_params(params_or_object(params))?; |
| 1725 | dispatch_stdio_app_request(state, AppRequest::ConfigGet { key: parsed.key }).await? |
| 1726 | } |
| 1727 | "app/config/set" => { |
| 1728 | let parsed: ConfigSetParams = parse_params(params_or_object(params))?; |
| 1729 | dispatch_stdio_app_request( |
| 1730 | state, |
| 1731 | AppRequest::ConfigSet { |
| 1732 | key: parsed.key, |
| 1733 | value: parsed.value, |
| 1734 | }, |
| 1735 | ) |
| 1736 | .await? |
| 1737 | } |
| 1738 | "app/config/unset" => { |
| 1739 | let parsed: ConfigGetParams = parse_params(params_or_object(params))?; |
| 1740 | dispatch_stdio_app_request(state, AppRequest::ConfigUnset { key: parsed.key }).await? |
| 1741 | } |
| 1742 | "app/config/list" => dispatch_stdio_app_request(state, AppRequest::ConfigList).await?, |
| 1743 | "app/config/reload" => dispatch_stdio_app_request(state, AppRequest::ConfigReload).await?, |
| 1744 | "app/models" => dispatch_stdio_app_request(state, AppRequest::Models).await?, |
| 1745 | "app/thread_loaded_list" | "app/thread-loaded-list" => { |
| 1746 | dispatch_stdio_app_request(state, AppRequest::ThreadLoadedList).await? |
| 1747 | } |
| 1748 | "prompt/capabilities" => StdioDispatchResult { |
| 1749 | result: json!({ |
| 1750 | "methods": ["prompt/request", "prompt/run"] |
| 1751 | }), |
| 1752 | should_exit: false, |
| 1753 | }, |
| 1754 | "prompt/request" | "prompt/run" => { |
| 1755 | let request: PromptRequest = parse_params(params)?; |
| 1756 | let response = handle_prompt_request(state, request).await?; |
| 1757 | StdioDispatchResult { |
| 1758 | result: serde_json::to_value(response) |
| 1759 | .map_err(|err| JsonRpcError::internal(err.to_string()))?, |
| 1760 | should_exit: false, |
| 1761 | } |
| 1762 | } |
| 1763 | "thread/interrupt" => { |
| 1764 | let parsed: ThreadInterruptParams = parse_params(params_or_object(params))?; |
| 1765 | let interrupted = interrupt_stdio_turn(state, &parsed.thread_id).await?; |
| 1766 | StdioDispatchResult { |
| 1767 | result: json!({ |
| 1768 | "thread_id": parsed.thread_id, |
| 1769 | "interrupted": interrupted, |
| 1770 | }), |
| 1771 | should_exit: false, |
| 1772 | } |
| 1773 | } |
| 1774 | "shutdown" => { |
| 1775 | // A turn streaming right now holds the bridge mutex, so taking it |
| 1776 | // to kill the child would block until that turn ends — the exact |
| 1777 | // deadlock that made shutdown useless against a runaway turn. |
| 1778 | // Interrupt live turns first; they release the mutex promptly. |
| 1779 | let live: Vec<String> = state.in_flight_turns.lock().await.keys().cloned().collect(); |
| 1780 | for thread_id in live { |
| 1781 | let _ = interrupt_stdio_turn(state, &thread_id).await; |
| 1782 | } |
| 1783 | if let Some(bridge) = state.stdio_bridge.lock().await.take() { |
| 1784 | bridge.lock().await.shutdown_child(); |
| 1785 | } |
| 1786 | StdioDispatchResult { |
| 1787 | result: json!({"ok": true, "status": "stopped"}), |
| 1788 | should_exit: true, |
| 1789 | } |
| 1790 | } |
| 1791 | _ => return Err(JsonRpcError::method_not_found(method)), |
| 1792 | }; |
| 1793 | Ok(outcome) |
| 1794 | } |
| 1795 | |
| 1796 | async fn process_app_request( |
| 1797 | state: &AppState, |
| 1798 | req: AppRequest, |
| 1799 | _transport: AppTransport, |
| 1800 | ) -> AppResponse { |
| 1801 | match req { |
| 1802 | AppRequest::Capabilities => AppResponse { |
| 1803 | ok: true, |
| 1804 | data: json!({ |
| 1805 | "routes": ["/thread", "/app", "/prompt", "/tool", "/jobs", "/mcp/startup"], |
| 1806 | "config": ["get", "set", "unset", "list", "reload"], |
| 1807 | "events": ["response_start", "response_delta", "response_end", "tool_call_start", "tool_call_result", "mcp_startup_update", "mcp_startup_complete"], |
| 1808 | "transport": "stdio+http", |
| 1809 | "config_path": state.config_path.as_ref().map(|p| p.display().to_string()), |
| 1810 | }), |
| 1811 | events: Vec::new(), |
| 1812 | }, |
| 1813 | AppRequest::ConfigGet { key } => { |
| 1814 | let cfg = state.config.read().await; |
| 1815 | let value = cfg.get_display_value(&key); |
| 1816 | AppResponse { |
| 1817 | ok: true, |
| 1818 | data: json!({ "key": key, "value": value }), |
| 1819 | events: Vec::new(), |
| 1820 | } |
| 1821 | } |
| 1822 | AppRequest::ConfigSet { key, value } => { |
| 1823 | let (result, snapshot) = { |
| 1824 | let mut cfg = state.config.write().await; |
| 1825 | let result = cfg.set_value(&key, &value); |
| 1826 | (result, cfg.clone()) |
| 1827 | }; |
| 1828 | let ok = result.is_ok(); |
| 1829 | let message = result.err().map(|e| e.to_string()); |
| 1830 | // Only propagate a mutation that actually happened. `set_value` |
| 1831 | // leaves the config untouched on an unknown key or invalid value, |
| 1832 | // so this is a no-op from the caller's point of view — but |
| 1833 | // `apply_config_update` invalidates the cached stdio bridge |
| 1834 | // regardless, and dropping the last reference kills the running |
| 1835 | // child runtime along with its thread map. A single typo'd key |
| 1836 | // would orphan every in-flight thread on that bridge. |
| 1837 | if ok { |
| 1838 | apply_config_update(state, snapshot, None, true).await; |
| 1839 | } |
| 1840 | AppResponse { |
| 1841 | ok, |
| 1842 | data: json!({ "key": key, "value": value, "error": message }), |
| 1843 | events: Vec::new(), |
| 1844 | } |
| 1845 | } |
| 1846 | AppRequest::ConfigUnset { key } => { |
| 1847 | let (result, snapshot) = { |
| 1848 | let mut cfg = state.config.write().await; |
| 1849 | let result = cfg.unset_value(&key); |
| 1850 | (result, cfg.clone()) |
| 1851 | }; |
| 1852 | let ok = result.is_ok(); |
| 1853 | let message = result.err().map(|e| e.to_string()); |
| 1854 | // See ConfigSet: a failed unset changed nothing and must not tear |
| 1855 | // down the runtime bridge. |
| 1856 | if ok { |
| 1857 | apply_config_update(state, snapshot, None, true).await; |
| 1858 | } |
| 1859 | AppResponse { |
| 1860 | ok, |
| 1861 | data: json!({ "key": key, "error": message }), |
| 1862 | events: Vec::new(), |
| 1863 | } |
| 1864 | } |
| 1865 | AppRequest::ConfigList => { |
| 1866 | let cfg = state.config.read().await; |
| 1867 | AppResponse { |
| 1868 | ok: true, |
| 1869 | data: json!({ "values": cfg.list_values() }), |
| 1870 | events: Vec::new(), |
| 1871 | } |
| 1872 | } |
| 1873 | AppRequest::ConfigReload => { |
| 1874 | // Re-read both `config.toml` and the sibling `permissions.toml` |
| 1875 | // from disk (the headless equivalent of the TUI |
| 1876 | // `reload_runtime_config` codepath) and push the fresh |
| 1877 | // snapshots into `state.config` and the live `Runtime`. |
| 1878 | // |
| 1879 | // `ConfigStore::load` resolves the same default config path |
| 1880 | // that `build_state` used at startup when `config_path` is |
| 1881 | // `None`, so a `None` here reloads from the same on-disk file |
| 1882 | // the server booted from. |
| 1883 | let store = match ConfigStore::load(state.config_path.clone()) { |
| 1884 | Ok(store) => store, |
| 1885 | Err(e) => { |
| 1886 | return AppResponse { |
| 1887 | ok: false, |
| 1888 | data: json!({ "error": format!("failed to load config: {e}") }), |
| 1889 | events: Vec::new(), |
| 1890 | }; |
| 1891 | } |
| 1892 | }; |
| 1893 | let new_config = store.config.clone(); |
| 1894 | let new_exec_policy = store.exec_policy_engine(); |
| 1895 | |
| 1896 | // Disk is already the source of truth here, so nothing to |
| 1897 | // persist; the exec policy rides along so the runtime picks up |
| 1898 | // external `permissions.toml` edits too. |
| 1899 | apply_config_update(state, new_config, Some(new_exec_policy), false).await; |
| 1900 | |
| 1901 | AppResponse { |
| 1902 | ok: true, |
| 1903 | data: json!({ "reloaded": true }), |
| 1904 | events: Vec::new(), |
| 1905 | } |
| 1906 | } |
| 1907 | AppRequest::Models => AppResponse { |
| 1908 | ok: true, |
| 1909 | data: json!({ "models": state.registry.list() }), |
| 1910 | events: Vec::new(), |
| 1911 | }, |
| 1912 | AppRequest::ThreadLoadedList => { |
| 1913 | let mut runtime = state.runtime.write().await; |
| 1914 | let response = runtime |
| 1915 | .handle_thread(codewhale_protocol::ThreadRequest::List( |
| 1916 | codewhale_protocol::ThreadListParams { |
| 1917 | include_archived: false, |
| 1918 | limit: Some(50), |
| 1919 | }, |
| 1920 | )) |
| 1921 | .await; |
| 1922 | match response { |
| 1923 | Ok(thread_resp) => AppResponse { |
| 1924 | ok: true, |
| 1925 | data: json!({ "threads": thread_resp.threads }), |
| 1926 | events: thread_resp.events, |
| 1927 | }, |
| 1928 | Err(err) => AppResponse { |
| 1929 | ok: false, |
| 1930 | data: json!({ "error": err.to_string() }), |
| 1931 | events: Vec::new(), |
| 1932 | }, |
| 1933 | } |
| 1934 | } |
| 1935 | AppRequest::SubmitUserInput { |
| 1936 | request_id, |
| 1937 | answers, |
| 1938 | } => { |
| 1939 | // Record the user's answers against the pending clarification |
| 1940 | // request so a driver can retrieve them. The headless runtime does |
| 1941 | // not block on `request_user_input` (fire-and-return, like |
| 1942 | // approval), so there is no in-flight turn to resume here — the |
| 1943 | // caller is expected to feed these answers into the next turn. |
| 1944 | let mut pending = state.pending_user_input.lock().await; |
| 1945 | if pending.contains_key(&request_id) { |
| 1946 | return AppResponse { |
| 1947 | ok: false, |
| 1948 | data: json!({ |
| 1949 | "error": "request_id already resolved", |
| 1950 | "request_id": request_id, |
| 1951 | }), |
| 1952 | events: Vec::new(), |
| 1953 | }; |
| 1954 | } |
| 1955 | pending.insert(request_id.clone(), answers); |
| 1956 | AppResponse { |
| 1957 | ok: true, |
| 1958 | data: json!({ "request_id": request_id, "resolved": true }), |
| 1959 | events: Vec::new(), |
| 1960 | } |
| 1961 | } |
| 1962 | } |
| 1963 | } |
| 1964 | |
| 1965 | /// Propagate a new config snapshot to every place that must observe it: |
| 1966 | /// optionally persist it to disk, install it in the shared `state.config`, |
| 1967 | /// push it into the live [`Runtime`], and invalidate the cached stdio |
| 1968 | /// bridge so the next stdio request spawns a fresh child that reads the |
| 1969 | /// new on-disk config. Shared by `ConfigSet` / `ConfigUnset` / `ConfigReload`. |
| 1970 | /// |
| 1971 | /// `exec_policy` is `Some` only on the reload path, which re-reads |
| 1972 | /// `permissions.toml` from disk; set/unset intentionally leave the live |
| 1973 | /// exec policy alone (use `ConfigReload` to pick up external permission |
| 1974 | /// edits). `persist` is false on the reload path because disk is already |
| 1975 | /// the source of truth there. |
| 1976 | async fn apply_config_update( |
| 1977 | state: &AppState, |
| 1978 | snapshot: codewhale_config::ConfigToml, |
| 1979 | exec_policy: Option<codewhale_execpolicy::ExecPolicyEngine>, |
| 1980 | persist: bool, |
| 1981 | ) { |
| 1982 | if persist && let Err(e) = persist_config(state, snapshot.clone()).await { |
| 1983 | tracing::error!("Failed to persist config update: {e}"); |
| 1984 | } |
| 1985 | { |
| 1986 | let mut cfg = state.config.write().await; |
| 1987 | *cfg = snapshot.clone(); |
| 1988 | } |
| 1989 | // Sync into the live Runtime so the next turn picks up the change |
| 1990 | // without a restart. MCP server connections are NOT refreshed here — |
| 1991 | // see `Runtime::reload_config_and_policy` for the headless boundary; |
| 1992 | // the TUI's explicit `/mcp reload` operation is a separate path. |
| 1993 | { |
| 1994 | let mut runtime = state.runtime.write().await; |
| 1995 | match exec_policy { |
| 1996 | Some(policy) => runtime.reload_config_and_policy(snapshot, policy), |
| 1997 | None => runtime.update_config(snapshot), |
| 1998 | } |
| 1999 | } |
| 2000 | invalidate_stdio_bridge(state).await; |
| 2001 | } |
| 2002 | |
| 2003 | async fn persist_config(state: &AppState, config: codewhale_config::ConfigToml) -> Result<()> { |
| 2004 | if state.config_path.is_none() { |
| 2005 | return Ok(()); |
| 2006 | } |
| 2007 | let mut store = ConfigStore::load(state.config_path.clone())?; |
| 2008 | store.config = config; |
| 2009 | store.save() |
| 2010 | } |
| 2011 | |
| 2012 | #[cfg(test)] |
| 2013 | mod tests { |
| 2014 | use super::*; |
| 2015 | use axum::body::{Body, to_bytes}; |
| 2016 | use axum::extract::{Path as AxumPath, Query}; |
| 2017 | use axum::http::header; |
| 2018 | use codewhale_protocol::AppRequest; |
| 2019 | use std::collections::HashMap; |
| 2020 | use std::fs; |
| 2021 | use tokio::io::AsyncReadExt; |
| 2022 | use tower::ServiceExt; |
| 2023 | |
| 2024 | fn app_with_config(auth_token: Option<&str>) -> (Router, tempfile::TempDir) { |
| 2025 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2026 | let config_path = tmp.path().join("config.toml"); |
| 2027 | fs::write(&config_path, "api_key = \"sk-deepseek-secret\"\n").expect("write config"); |
| 2028 | let state = build_state( |
| 2029 | Some(config_path), |
| 2030 | auth_token.map(std::string::ToString::to_string), |
| 2031 | ) |
| 2032 | .expect("state"); |
| 2033 | (app_router(state, &[]), tmp) |
| 2034 | } |
| 2035 | |
| 2036 | #[test] |
| 2037 | fn build_state_keeps_resolved_explicit_config_path() { |
| 2038 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2039 | let config_dir = tmp.path().join("config-dir"); |
| 2040 | fs::create_dir_all(&config_dir).expect("config dir"); |
| 2041 | let config_path = config_dir.join("config.toml"); |
| 2042 | fs::write(&config_path, "api_key = \"sk-deepseek-secret\"\n").expect("write config"); |
| 2043 | |
| 2044 | let state = build_state(Some(config_path.clone()), None).expect("state"); |
| 2045 | |
| 2046 | assert_eq!( |
| 2047 | state.config_path.as_deref(), |
| 2048 | Some( |
| 2049 | config_path |
| 2050 | .canonicalize() |
| 2051 | .expect("canonical config") |
| 2052 | .as_path() |
| 2053 | ) |
| 2054 | ); |
| 2055 | } |
| 2056 | |
| 2057 | #[tokio::test] |
| 2058 | async fn stdio_transport_never_registers_the_stdout_hook_sink() { |
| 2059 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2060 | let config_path = tmp.path().join("config.toml"); |
| 2061 | fs::write(&config_path, "api_key = \"sk-deepseek-secret\"\n").expect("write config"); |
| 2062 | |
| 2063 | let http_state = |
| 2064 | build_state_with_transport(Some(config_path.clone()), None, AppTransport::Http) |
| 2065 | .expect("http state"); |
| 2066 | let stdio_state = build_state_with_transport(Some(config_path), None, AppTransport::Stdio) |
| 2067 | .expect("stdio state"); |
| 2068 | |
| 2069 | let http_sinks = http_state.runtime.read().await.hooks.sink_count(); |
| 2070 | let stdio_sinks = stdio_state.runtime.read().await.hooks.sink_count(); |
| 2071 | assert_eq!( |
| 2072 | http_sinks, |
| 2073 | stdio_sinks + 1, |
| 2074 | "HTTP mode keeps StdoutHookSink + JsonlHookSink; stdio must drop the stdout sink (#5165)" |
| 2075 | ); |
| 2076 | } |
| 2077 | |
| 2078 | async fn response_body_json(response: Response) -> Value { |
| 2079 | let bytes = to_bytes(response.into_body(), usize::MAX) |
| 2080 | .await |
| 2081 | .expect("body bytes"); |
| 2082 | serde_json::from_slice(&bytes).expect("json response") |
| 2083 | } |
| 2084 | |
| 2085 | #[tokio::test] |
| 2086 | async fn http_app_routes_require_bearer_token_when_auth_enabled() { |
| 2087 | let (app, _tmp) = app_with_config(Some("test-token")); |
| 2088 | let response = app |
| 2089 | .oneshot( |
| 2090 | Request::builder() |
| 2091 | .method(Method::POST) |
| 2092 | .uri("/app") |
| 2093 | .header(header::CONTENT_TYPE, "application/json") |
| 2094 | .body(Body::from( |
| 2095 | serde_json::to_vec(&AppRequest::ConfigGet { |
| 2096 | key: "api_key".to_string(), |
| 2097 | }) |
| 2098 | .expect("request json"), |
| 2099 | )) |
| 2100 | .expect("request"), |
| 2101 | ) |
| 2102 | .await |
| 2103 | .expect("response"); |
| 2104 | |
| 2105 | assert_eq!(response.status(), StatusCode::UNAUTHORIZED); |
| 2106 | } |
| 2107 | |
| 2108 | #[tokio::test] |
| 2109 | async fn http_config_get_redacts_sensitive_values_after_auth() { |
| 2110 | let (app, _tmp) = app_with_config(Some("test-token")); |
| 2111 | let response = app |
| 2112 | .oneshot( |
| 2113 | Request::builder() |
| 2114 | .method(Method::POST) |
| 2115 | .uri("/app") |
| 2116 | .header(header::AUTHORIZATION, "Bearer test-token") |
| 2117 | .header(header::CONTENT_TYPE, "application/json") |
| 2118 | .body(Body::from( |
| 2119 | serde_json::to_vec(&AppRequest::ConfigGet { |
| 2120 | key: "api_key".to_string(), |
| 2121 | }) |
| 2122 | .expect("request json"), |
| 2123 | )) |
| 2124 | .expect("request"), |
| 2125 | ) |
| 2126 | .await |
| 2127 | .expect("response"); |
| 2128 | |
| 2129 | assert_eq!(response.status(), StatusCode::OK); |
| 2130 | let body = response_body_json(response).await; |
| 2131 | assert_eq!(body["data"]["value"], "sk-d***cret"); |
| 2132 | } |
| 2133 | |
| 2134 | #[tokio::test] |
| 2135 | async fn cors_does_not_allow_arbitrary_origins() { |
| 2136 | let (app, _tmp) = app_with_config(Some("test-token")); |
| 2137 | let response = app |
| 2138 | .oneshot( |
| 2139 | Request::builder() |
| 2140 | .method(Method::GET) |
| 2141 | .uri("/healthz") |
| 2142 | .header(header::ORIGIN, "https://attacker.example") |
| 2143 | .body(Body::empty()) |
| 2144 | .expect("request"), |
| 2145 | ) |
| 2146 | .await |
| 2147 | .expect("response"); |
| 2148 | |
| 2149 | assert_eq!(response.status(), StatusCode::OK); |
| 2150 | assert!( |
| 2151 | response |
| 2152 | .headers() |
| 2153 | .get(header::ACCESS_CONTROL_ALLOW_ORIGIN) |
| 2154 | .is_none() |
| 2155 | ); |
| 2156 | } |
| 2157 | |
| 2158 | #[tokio::test] |
| 2159 | async fn build_state_loads_permissions_into_runtime_policy() { |
| 2160 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2161 | let config_path = tmp.path().join("config.toml"); |
| 2162 | fs::write(&config_path, "api_key = \"sk-deepseek-secret\"\n").expect("write config"); |
| 2163 | fs::write( |
| 2164 | tmp.path().join("permissions.toml"), |
| 2165 | r#" |
| 2166 | [[rules]] |
| 2167 | tool = "exec_shell" |
| 2168 | command = "cargo test" |
| 2169 | "#, |
| 2170 | ) |
| 2171 | .expect("write permissions"); |
| 2172 | |
| 2173 | let state = build_state(Some(config_path), None).expect("state"); |
| 2174 | let runtime = state.runtime.read().await; |
| 2175 | let decision = runtime |
| 2176 | .exec_policy |
| 2177 | .check(codewhale_execpolicy::ExecPolicyContext { |
| 2178 | command: "cargo test --workspace", |
| 2179 | cwd: "/workspace", |
| 2180 | tool: Some("exec_shell"), |
| 2181 | path: None, |
| 2182 | ask_for_approval: codewhale_execpolicy::AskForApproval::UnlessTrusted, |
| 2183 | sandbox_mode: Some("workspace-write"), |
| 2184 | }) |
| 2185 | .expect("policy check"); |
| 2186 | |
| 2187 | assert!(decision.allow); |
| 2188 | assert!(decision.requires_approval); |
| 2189 | assert_eq!( |
| 2190 | decision.matched_rule.as_deref(), |
| 2191 | Some("tool=exec_shell command=cargo test") |
| 2192 | ); |
| 2193 | } |
| 2194 | |
| 2195 | #[tokio::test] |
| 2196 | async fn config_reload_refreshes_runtime_config_and_exec_policy_from_disk() { |
| 2197 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2198 | let config_path = tmp.path().join("config.toml"); |
| 2199 | fs::write( |
| 2200 | &config_path, |
| 2201 | "api_key = \"sk-deepseek-secret\"\nmodel = \"deepseek-chat\"\n", |
| 2202 | ) |
| 2203 | .expect("write config"); |
| 2204 | // No permissions.toml at startup → exec_policy starts empty. |
| 2205 | let state = build_state(Some(config_path.clone()), None).expect("state"); |
| 2206 | |
| 2207 | // Sanity: initial runtime sees the on-disk model and has no rule. |
| 2208 | { |
| 2209 | let runtime = state.runtime.read().await; |
| 2210 | assert_eq!(runtime.config.model.as_deref(), Some("deepseek-chat")); |
| 2211 | let decision = runtime |
| 2212 | .exec_policy |
| 2213 | .check(codewhale_execpolicy::ExecPolicyContext { |
| 2214 | command: "cargo test", |
| 2215 | cwd: "/workspace", |
| 2216 | tool: Some("exec_shell"), |
| 2217 | path: None, |
| 2218 | ask_for_approval: codewhale_execpolicy::AskForApproval::UnlessTrusted, |
| 2219 | sandbox_mode: Some("workspace-write"), |
| 2220 | }) |
| 2221 | .expect("policy check"); |
| 2222 | assert!(decision.matched_rule.is_none()); |
| 2223 | } |
| 2224 | |
| 2225 | // Edit both files on disk: new model + a permission rule. |
| 2226 | fs::write( |
| 2227 | &config_path, |
| 2228 | "api_key = \"sk-deepseek-secret\"\nmodel = \"deepseek-reasoner\"\n", |
| 2229 | ) |
| 2230 | .expect("rewrite config"); |
| 2231 | fs::write( |
| 2232 | tmp.path().join("permissions.toml"), |
| 2233 | r#" |
| 2234 | [[rules]] |
| 2235 | tool = "exec_shell" |
| 2236 | command = "cargo test" |
| 2237 | "#, |
| 2238 | ) |
| 2239 | .expect("write permissions"); |
| 2240 | |
| 2241 | // ConfigReload must re-read both files and push them into the |
| 2242 | // live Runtime without a restart. |
| 2243 | let response = |
| 2244 | process_app_request(&state, AppRequest::ConfigReload, AppTransport::Stdio).await; |
| 2245 | assert!(response.ok, "reload should succeed"); |
| 2246 | assert_eq!(response.data["reloaded"], true); |
| 2247 | |
| 2248 | // The shared config lock reflects the new model. |
| 2249 | { |
| 2250 | let cfg = state.config.read().await; |
| 2251 | assert_eq!(cfg.model.as_deref(), Some("deepseek-reasoner")); |
| 2252 | } |
| 2253 | // The live Runtime reflects both the new model and the new rule. |
| 2254 | { |
| 2255 | let runtime = state.runtime.read().await; |
| 2256 | assert_eq!(runtime.config.model.as_deref(), Some("deepseek-reasoner")); |
| 2257 | let decision = runtime |
| 2258 | .exec_policy |
| 2259 | .check(codewhale_execpolicy::ExecPolicyContext { |
| 2260 | command: "cargo test --workspace", |
| 2261 | cwd: "/workspace", |
| 2262 | tool: Some("exec_shell"), |
| 2263 | path: None, |
| 2264 | ask_for_approval: codewhale_execpolicy::AskForApproval::UnlessTrusted, |
| 2265 | sandbox_mode: Some("workspace-write"), |
| 2266 | }) |
| 2267 | .expect("policy check"); |
| 2268 | assert!(decision.allow); |
| 2269 | assert!(decision.requires_approval); |
| 2270 | assert_eq!( |
| 2271 | decision.matched_rule.as_deref(), |
| 2272 | Some("tool=exec_shell command=cargo test") |
| 2273 | ); |
| 2274 | } |
| 2275 | } |
| 2276 | |
| 2277 | #[tokio::test] |
| 2278 | async fn config_set_propagates_to_runtime_config_without_touching_exec_policy() { |
| 2279 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2280 | let config_path = tmp.path().join("config.toml"); |
| 2281 | fs::write( |
| 2282 | &config_path, |
| 2283 | "api_key = \"sk-deepseek-secret\"\nmodel = \"deepseek-chat\"\n", |
| 2284 | ) |
| 2285 | .expect("write config"); |
| 2286 | let state = build_state(Some(config_path.clone()), None).expect("state"); |
| 2287 | |
| 2288 | // Set a new model via the API. Only config.toml is touched; no |
| 2289 | // permissions.toml exists, so exec_policy must stay empty. |
| 2290 | let response = process_app_request( |
| 2291 | &state, |
| 2292 | AppRequest::ConfigSet { |
| 2293 | key: "model".to_string(), |
| 2294 | value: "deepseek-reasoner".to_string(), |
| 2295 | }, |
| 2296 | AppTransport::Stdio, |
| 2297 | ) |
| 2298 | .await; |
| 2299 | assert!(response.ok, "set should succeed"); |
| 2300 | |
| 2301 | // Live runtime sees the new model. |
| 2302 | { |
| 2303 | let runtime = state.runtime.read().await; |
| 2304 | assert_eq!(runtime.config.model.as_deref(), Some("deepseek-reasoner")); |
| 2305 | // exec_policy was empty at startup and must remain empty. |
| 2306 | let decision = runtime |
| 2307 | .exec_policy |
| 2308 | .check(codewhale_execpolicy::ExecPolicyContext { |
| 2309 | command: "cargo test", |
| 2310 | cwd: "/workspace", |
| 2311 | tool: Some("exec_shell"), |
| 2312 | path: None, |
| 2313 | ask_for_approval: codewhale_execpolicy::AskForApproval::UnlessTrusted, |
| 2314 | sandbox_mode: Some("workspace-write"), |
| 2315 | }) |
| 2316 | .expect("policy check"); |
| 2317 | assert!(decision.matched_rule.is_none()); |
| 2318 | } |
| 2319 | // The on-disk file was persisted. |
| 2320 | let persisted = fs::read_to_string(&config_path).expect("read config"); |
| 2321 | assert!(persisted.contains("deepseek-reasoner")); |
| 2322 | } |
| 2323 | |
| 2324 | /// A bridge stand-in with no child process: this test only cares about |
| 2325 | /// whether the cache slot survives, not about talking to a runtime. |
| 2326 | fn sentinel_bridge() -> SharedRuntimeBridge { |
| 2327 | Arc::new(Mutex::new(RuntimeBridge { |
| 2328 | base_url: "http://127.0.0.1:0".to_string(), |
| 2329 | client: reqwest::Client::new(), |
| 2330 | auth_token: None, |
| 2331 | child: None, |
| 2332 | thread_map: HashMap::from([("stdio-1".to_string(), "runtime-1".to_string())]), |
| 2333 | last_seq_by_thread: HashMap::new(), |
| 2334 | })) |
| 2335 | } |
| 2336 | |
| 2337 | #[tokio::test] |
| 2338 | async fn failed_config_set_keeps_the_stdio_bridge() { |
| 2339 | // #4737: `set_value` rejects an invalid value before assigning, so the |
| 2340 | // request is a no-op — but `apply_config_update` ran anyway and |
| 2341 | // invalidated the cached bridge, dropping the child runtime along with |
| 2342 | // its thread map. A single bad value orphaned every in-flight stdio |
| 2343 | // thread, behind a response that correctly reported `ok: false`. |
| 2344 | // |
| 2345 | // Only `set_value` is exercised: an unknown key lands in `extras` and |
| 2346 | // succeeds, and `unset_value` has no failing input today, so its |
| 2347 | // identical guard has nothing to assert against. |
| 2348 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2349 | let config_path = tmp.path().join("config.toml"); |
| 2350 | fs::write(&config_path, "model = \"deepseek-chat\"\n").expect("write config"); |
| 2351 | let state = build_state(Some(config_path.clone()), None).expect("state"); |
| 2352 | *state.stdio_bridge.lock().await = Some(sentinel_bridge()); |
| 2353 | |
| 2354 | let response = process_app_request( |
| 2355 | &state, |
| 2356 | AppRequest::ConfigSet { |
| 2357 | key: "telemetry".to_string(), |
| 2358 | value: "not-a-bool".to_string(), |
| 2359 | }, |
| 2360 | AppTransport::Stdio, |
| 2361 | ) |
| 2362 | .await; |
| 2363 | assert!(!response.ok, "invalid value must fail: {response:?}"); |
| 2364 | |
| 2365 | let slot = state.stdio_bridge.lock().await; |
| 2366 | let kept = slot |
| 2367 | .as_ref() |
| 2368 | .expect("bridge must survive a failed config/set"); |
| 2369 | assert_eq!( |
| 2370 | kept.lock() |
| 2371 | .await |
| 2372 | .thread_map |
| 2373 | .get("stdio-1") |
| 2374 | .map(String::as_str), |
| 2375 | Some("runtime-1"), |
| 2376 | "the live thread map must be intact", |
| 2377 | ); |
| 2378 | } |
| 2379 | |
| 2380 | #[tokio::test] |
| 2381 | async fn successful_config_set_still_invalidates_the_stdio_bridge() { |
| 2382 | // The other half of #4737: a mutation that *did* happen must still |
| 2383 | // rebuild the bridge, or the runtime keeps serving the old config. |
| 2384 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2385 | let config_path = tmp.path().join("config.toml"); |
| 2386 | fs::write(&config_path, "model = \"deepseek-chat\"\n").expect("write config"); |
| 2387 | let state = build_state(Some(config_path.clone()), None).expect("state"); |
| 2388 | *state.stdio_bridge.lock().await = Some(sentinel_bridge()); |
| 2389 | |
| 2390 | let response = process_app_request( |
| 2391 | &state, |
| 2392 | AppRequest::ConfigSet { |
| 2393 | key: "model".to_string(), |
| 2394 | value: "deepseek-reasoner".to_string(), |
| 2395 | }, |
| 2396 | AppTransport::Stdio, |
| 2397 | ) |
| 2398 | .await; |
| 2399 | assert!(response.ok, "valid set should succeed: {response:?}"); |
| 2400 | assert!( |
| 2401 | state.stdio_bridge.lock().await.is_none(), |
| 2402 | "a successful config change must invalidate the cached bridge", |
| 2403 | ); |
| 2404 | } |
| 2405 | |
| 2406 | #[tokio::test] |
| 2407 | async fn config_unset_propagates_to_runtime_config() { |
| 2408 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2409 | let config_path = tmp.path().join("config.toml"); |
| 2410 | fs::write( |
| 2411 | &config_path, |
| 2412 | "api_key = \"sk-deepseek-secret\"\nmodel = \"deepseek-chat\"\n", |
| 2413 | ) |
| 2414 | .expect("write config"); |
| 2415 | let state = build_state(Some(config_path.clone()), None).expect("state"); |
| 2416 | |
| 2417 | // Sanity: runtime starts with the on-disk model. |
| 2418 | { |
| 2419 | let runtime = state.runtime.read().await; |
| 2420 | assert_eq!(runtime.config.model.as_deref(), Some("deepseek-chat")); |
| 2421 | } |
| 2422 | |
| 2423 | // Unset the model via the API. This walks a separate code path |
| 2424 | // from ConfigSet (unset_value + update_config), so it needs its |
| 2425 | // own regression coverage. |
| 2426 | let response = process_app_request( |
| 2427 | &state, |
| 2428 | AppRequest::ConfigUnset { |
| 2429 | key: "model".to_string(), |
| 2430 | }, |
| 2431 | AppTransport::Stdio, |
| 2432 | ) |
| 2433 | .await; |
| 2434 | assert!(response.ok, "unset should succeed"); |
| 2435 | |
| 2436 | // Live runtime sees the cleared model. |
| 2437 | { |
| 2438 | let runtime = state.runtime.read().await; |
| 2439 | assert!(runtime.config.model.is_none()); |
| 2440 | } |
| 2441 | // Shared config lock agrees. |
| 2442 | { |
| 2443 | let cfg = state.config.read().await; |
| 2444 | assert!(cfg.model.is_none()); |
| 2445 | } |
| 2446 | // The on-disk file no longer carries the model value. |
| 2447 | let persisted = fs::read_to_string(&config_path).expect("read config"); |
| 2448 | assert!(!persisted.contains("deepseek-chat")); |
| 2449 | } |
| 2450 | |
| 2451 | #[tokio::test] |
| 2452 | async fn config_reload_returns_error_when_disk_config_is_invalid() { |
| 2453 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2454 | let config_path = tmp.path().join("config.toml"); |
| 2455 | fs::write( |
| 2456 | &config_path, |
| 2457 | "api_key = \"sk-deepseek-secret\"\nmodel = \"deepseek-chat\"\n", |
| 2458 | ) |
| 2459 | .expect("write config"); |
| 2460 | let state = build_state(Some(config_path.clone()), None).expect("state"); |
| 2461 | |
| 2462 | // Corrupt the on-disk config so ConfigStore::load fails to parse. |
| 2463 | fs::write(&config_path, "api_key = \"unterminated\n").expect("corrupt config"); |
| 2464 | |
| 2465 | let response = |
| 2466 | process_app_request(&state, AppRequest::ConfigReload, AppTransport::Stdio).await; |
| 2467 | assert!(!response.ok, "reload of corrupt config must fail"); |
| 2468 | let err = response.data["error"] |
| 2469 | .as_str() |
| 2470 | .expect("error message present") |
| 2471 | .to_string(); |
| 2472 | assert!( |
| 2473 | err.contains("failed to load config"), |
| 2474 | "error should mention load failure, got: {err}" |
| 2475 | ); |
| 2476 | |
| 2477 | // Live state is untouched: the early-return on load error must |
| 2478 | // not have clobbered runtime.config or state.config. |
| 2479 | { |
| 2480 | let runtime = state.runtime.read().await; |
| 2481 | assert_eq!(runtime.config.model.as_deref(), Some("deepseek-chat")); |
| 2482 | } |
| 2483 | { |
| 2484 | let cfg = state.config.read().await; |
| 2485 | assert_eq!(cfg.model.as_deref(), Some("deepseek-chat")); |
| 2486 | } |
| 2487 | } |
| 2488 | |
| 2489 | async fn seed_test_bridge(state: &AppState) -> SharedRuntimeBridge { |
| 2490 | let bridge = Arc::new(Mutex::new(RuntimeBridge::from_base_url_for_test( |
| 2491 | "http://127.0.0.1:9".to_string(), |
| 2492 | ))); |
| 2493 | *state.stdio_bridge.lock().await = Some(bridge.clone()); |
| 2494 | bridge |
| 2495 | } |
| 2496 | |
| 2497 | #[tokio::test] |
| 2498 | async fn config_set_invalidates_cached_stdio_bridge() { |
| 2499 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2500 | let config_path = tmp.path().join("config.toml"); |
| 2501 | fs::write(&config_path, "model = \"deepseek-chat\"\n").expect("write config"); |
| 2502 | let state = build_state(Some(config_path), None).expect("state"); |
| 2503 | seed_test_bridge(&state).await; |
| 2504 | |
| 2505 | let response = process_app_request( |
| 2506 | &state, |
| 2507 | AppRequest::ConfigSet { |
| 2508 | key: "model".to_string(), |
| 2509 | value: "deepseek-reasoner".to_string(), |
| 2510 | }, |
| 2511 | AppTransport::Stdio, |
| 2512 | ) |
| 2513 | .await; |
| 2514 | assert!(response.ok, "set should succeed"); |
| 2515 | |
| 2516 | // The cached bridge child must be dropped so the next stdio request |
| 2517 | // spawns a fresh runtime that reads the persisted config. |
| 2518 | assert!(state.stdio_bridge.lock().await.is_none()); |
| 2519 | } |
| 2520 | |
| 2521 | #[tokio::test] |
| 2522 | async fn config_reload_invalidates_cached_stdio_bridge() { |
| 2523 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2524 | let config_path = tmp.path().join("config.toml"); |
| 2525 | fs::write(&config_path, "model = \"deepseek-chat\"\n").expect("write config"); |
| 2526 | let state = build_state(Some(config_path), None).expect("state"); |
| 2527 | seed_test_bridge(&state).await; |
| 2528 | |
| 2529 | let response = |
| 2530 | process_app_request(&state, AppRequest::ConfigReload, AppTransport::Stdio).await; |
| 2531 | assert!(response.ok, "reload should succeed"); |
| 2532 | |
| 2533 | assert!(state.stdio_bridge.lock().await.is_none()); |
| 2534 | } |
| 2535 | |
| 2536 | #[tokio::test] |
| 2537 | async fn stdio_bridge_invalidation_not_blocked_by_in_flight_turn() { |
| 2538 | let (state, _tmp) = capability_test_state(); |
| 2539 | let bridge = seed_test_bridge(&state).await; |
| 2540 | |
| 2541 | // Simulate a long streaming turn holding the inner bridge lock. |
| 2542 | let _in_flight = bridge.lock().await; |
| 2543 | |
| 2544 | // Invalidation only touches the cache slot, so it must complete |
| 2545 | // without waiting for the in-flight turn to release the bridge. |
| 2546 | tokio::time::timeout(Duration::from_secs(1), invalidate_stdio_bridge(&state)) |
| 2547 | .await |
| 2548 | .expect("invalidation must not wait on bridge traffic"); |
| 2549 | assert!(state.stdio_bridge.lock().await.is_none()); |
| 2550 | } |
| 2551 | |
| 2552 | #[tokio::test] |
| 2553 | async fn runtime_read_paths_run_concurrently() { |
| 2554 | // Tool/status/mcp handlers take read guards; two must coexist so a |
| 2555 | // long-running tool call cannot serialize unrelated requests. With |
| 2556 | // the old `Mutex<Runtime>` this pattern would deadlock. |
| 2557 | let (state, _tmp) = capability_test_state(); |
| 2558 | let first = state.runtime.read().await; |
| 2559 | let second = state.runtime.read().await; |
| 2560 | assert!(first.app_status().ok); |
| 2561 | assert!(second.app_status().ok); |
| 2562 | } |
| 2563 | |
| 2564 | #[tokio::test] |
| 2565 | async fn health_probes_advertise_legacy_deepseek_service_name() { |
| 2566 | // External probes still key off the DeepSeek-era service name; both |
| 2567 | // transports must serve it from the single compat shim. |
| 2568 | let (app, _tmp) = app_with_config(None); |
| 2569 | let response = app |
| 2570 | .oneshot( |
| 2571 | Request::builder() |
| 2572 | .method(Method::GET) |
| 2573 | .uri("/healthz") |
| 2574 | .body(Body::empty()) |
| 2575 | .expect("request"), |
| 2576 | ) |
| 2577 | .await |
| 2578 | .expect("response"); |
| 2579 | let body = response_body_json(response).await; |
| 2580 | assert_eq!(body["service"], legacy_deepseek_compat::SERVICE_NAME); |
| 2581 | assert_eq!(body["service"], "deepseek-app-server"); |
| 2582 | |
| 2583 | let (state, _tmp) = capability_test_state(); |
| 2584 | let stdio = dispatch_stdio_request(&state, "healthz", json!({})) |
| 2585 | .await |
| 2586 | .expect("stdio healthz"); |
| 2587 | assert_eq!( |
| 2588 | stdio.result["service"], |
| 2589 | legacy_deepseek_compat::SERVICE_NAME |
| 2590 | ); |
| 2591 | } |
| 2592 | |
| 2593 | #[test] |
| 2594 | fn non_loopback_bind_without_auth_fails_fast() { |
| 2595 | let options = AppServerOptions { |
| 2596 | listen: "0.0.0.0:8787".parse().expect("socket addr"), |
| 2597 | config_path: None, |
| 2598 | auth_token: None, |
| 2599 | insecure_no_auth: false, |
| 2600 | cors_origins: Vec::new(), |
| 2601 | }; |
| 2602 | |
| 2603 | let err = |
| 2604 | resolve_auth_token(&options).expect_err("non-loopback generated auth should fail"); |
| 2605 | assert!(err.to_string().contains("without explicit auth token")); |
| 2606 | } |
| 2607 | |
| 2608 | #[tokio::test] |
| 2609 | async fn stdio_transport_redacts_config_get_secrets() { |
| 2610 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2611 | let config_path = tmp.path().join("config.toml"); |
| 2612 | fs::write(&config_path, "").expect("write config"); |
| 2613 | let state = build_state(Some(config_path), None).expect("state"); |
| 2614 | { |
| 2615 | let mut cfg = state.config.write().await; |
| 2616 | cfg.api_key = Some("sk-deepseek-secret".to_string()); |
| 2617 | } |
| 2618 | |
| 2619 | let response = process_app_request( |
| 2620 | &state, |
| 2621 | AppRequest::ConfigGet { |
| 2622 | key: "api_key".to_string(), |
| 2623 | }, |
| 2624 | AppTransport::Stdio, |
| 2625 | ) |
| 2626 | .await; |
| 2627 | |
| 2628 | assert_eq!(response.data["value"], "sk-d***cret"); |
| 2629 | } |
| 2630 | |
| 2631 | #[tokio::test] |
| 2632 | async fn stdio_thread_goal_methods_round_trip_persisted_goal() { |
| 2633 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2634 | let config_path = tmp.path().join("config.toml"); |
| 2635 | fs::write(&config_path, "").expect("write config"); |
| 2636 | let state = build_state(Some(config_path), None).expect("state"); |
| 2637 | |
| 2638 | let capabilities = dispatch_stdio_request(&state, "thread/capabilities", json!({})) |
| 2639 | .await |
| 2640 | .expect("thread capabilities"); |
| 2641 | assert!( |
| 2642 | capabilities.result["methods"] |
| 2643 | .as_array() |
| 2644 | .expect("methods") |
| 2645 | .iter() |
| 2646 | .any(|method| method == "thread/goal/set") |
| 2647 | ); |
| 2648 | |
| 2649 | let started = dispatch_stdio_request(&state, "thread/start", json!({})) |
| 2650 | .await |
| 2651 | .expect("start thread"); |
| 2652 | let thread_id = started.result["thread_id"] |
| 2653 | .as_str() |
| 2654 | .expect("thread id") |
| 2655 | .to_string(); |
| 2656 | |
| 2657 | let set = dispatch_stdio_request( |
| 2658 | &state, |
| 2659 | "thread/goal/set", |
| 2660 | json!({ |
| 2661 | "thread_id": thread_id, |
| 2662 | "objective": "Release 0.8.59", |
| 2663 | "token_budget": 59000 |
| 2664 | }), |
| 2665 | ) |
| 2666 | .await |
| 2667 | .expect("set goal"); |
| 2668 | assert_eq!(set.result["status"], "ok"); |
| 2669 | assert_eq!(set.result["goal"]["objective"], "Release 0.8.59"); |
| 2670 | assert_eq!(set.result["goal"]["status"], "active"); |
| 2671 | |
| 2672 | let got = dispatch_stdio_request( |
| 2673 | &state, |
| 2674 | "thread/goal/get", |
| 2675 | json!({ |
| 2676 | "thread_id": thread_id |
| 2677 | }), |
| 2678 | ) |
| 2679 | .await |
| 2680 | .expect("get goal"); |
| 2681 | assert_eq!(got.result["goal"]["token_budget"], 59000); |
| 2682 | |
| 2683 | let cleared = dispatch_stdio_request( |
| 2684 | &state, |
| 2685 | "thread/goal/clear", |
| 2686 | json!({ |
| 2687 | "thread_id": thread_id |
| 2688 | }), |
| 2689 | ) |
| 2690 | .await |
| 2691 | .expect("clear goal"); |
| 2692 | assert_eq!(cleared.result["status"], "cleared"); |
| 2693 | assert_eq!(cleared.result["data"]["cleared"], true); |
| 2694 | } |
| 2695 | |
| 2696 | #[tokio::test] |
| 2697 | async fn stdio_resume_of_missing_thread_fails_without_clobbering_the_hint() { |
| 2698 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 2699 | let config_path = tmp.path().join("config.toml"); |
| 2700 | fs::write(&config_path, "").expect("write config"); |
| 2701 | let state = build_state(Some(config_path), None).expect("state"); |
| 2702 | |
| 2703 | // A cached hint for a thread the runtime no longer knows: the exact |
| 2704 | // clobber scenario from #5171. |
| 2705 | let workspace = tmp.path().join("ws"); |
| 2706 | { |
| 2707 | let mut hints = state.stdio_thread_hints.lock().await; |
| 2708 | hints.insert( |
| 2709 | "ghost-thread".to_string(), |
| 2710 | RuntimeThreadHint { |
| 2711 | model: Some("deepseek-v4-pro".to_string()), |
| 2712 | workspace: Some(workspace.clone()), |
| 2713 | }, |
| 2714 | ); |
| 2715 | } |
| 2716 | |
| 2717 | let err = dispatch_stdio_request( |
| 2718 | &state, |
| 2719 | "thread/resume", |
| 2720 | json!({ "thread_id": "ghost-thread" }), |
| 2721 | ) |
| 2722 | .await |
| 2723 | .expect_err("resuming a missing thread must fail with a named not-found error"); |
| 2724 | assert_eq!(err.code, -32004); |
| 2725 | assert!(err.message.contains("ghost-thread"), "{}", err.message); |
| 2726 | |
| 2727 | let fork_err = dispatch_stdio_request( |
| 2728 | &state, |
| 2729 | "thread/fork", |
| 2730 | json!({ "thread_id": "ghost-thread" }), |
| 2731 | ) |
| 2732 | .await |
| 2733 | .expect_err("forking a missing thread must fail with a named not-found error"); |
| 2734 | assert_eq!(fork_err.code, -32004); |
| 2735 | |
| 2736 | let hints = state.stdio_thread_hints.lock().await; |
| 2737 | let hint = hints.get("ghost-thread").expect("cached hint survives"); |
| 2738 | assert_eq!(hint.model.as_deref(), Some("deepseek-v4-pro")); |
| 2739 | assert_eq!(hint.workspace.as_deref(), Some(workspace.as_path())); |
| 2740 | } |
| 2741 | |
| 2742 | fn sse_frame(event: &str, payload: Value) -> String { |
| 2743 | format!("event: {event}\ndata: {payload}\n\n") |
| 2744 | } |
| 2745 | /// A runtime whose turn never ends on its own — only an interrupt stops |
| 2746 | /// it. That is the shape of the runaway turn this protects against. |
| 2747 | async fn spawn_uninterruptible_until_asked_runtime() -> ( |
| 2748 | String, |
| 2749 | Arc<tokio::sync::Notify>, |
| 2750 | tokio::task::JoinHandle<()>, |
| 2751 | ) { |
| 2752 | use axum::body::Body; |
| 2753 | use axum::extract::Path as AxumPath; |
| 2754 | |
| 2755 | let interrupted = Arc::new(tokio::sync::Notify::new()); |
| 2756 | |
| 2757 | async fn create_turn(AxumPath(_thread_id): AxumPath<String>) -> Json<Value> { |
| 2758 | Json(json!({ "turn": { "id": "turn_runaway" } })) |
| 2759 | } |
| 2760 | async fn create_thread() -> Json<Value> { |
| 2761 | Json(json!({ "id": "thr_runaway" })) |
| 2762 | } |
| 2763 | async fn interrupt( |
| 2764 | State(notify): State<Arc<tokio::sync::Notify>>, |
| 2765 | AxumPath((_thread_id, _turn_id)): AxumPath<(String, String)>, |
| 2766 | ) -> Json<Value> { |
| 2767 | notify.notify_waiters(); |
| 2768 | Json(json!({ "ok": true })) |
| 2769 | } |
| 2770 | async fn thread_events( |
| 2771 | State(notify): State<Arc<tokio::sync::Notify>>, |
| 2772 | AxumPath(_thread_id): AxumPath<String>, |
| 2773 | ) -> ([(header::HeaderName, &'static str); 1], Body) { |
| 2774 | // Hold the event response open until something interrupts the |
| 2775 | // turn. Nothing else can end it, which is the point. |
| 2776 | notify.notified().await; |
| 2777 | let body = [ |
| 2778 | sse_frame( |
| 2779 | "item.delta", |
| 2780 | json!({ |
| 2781 | "seq": 1, |
| 2782 | "turn_id": "turn_runaway", |
| 2783 | "payload": { "kind": "agent_message", "delta": "thinking" } |
| 2784 | }), |
| 2785 | ), |
| 2786 | sse_frame( |
| 2787 | "turn.completed", |
| 2788 | json!({ |
| 2789 | "seq": 2, |
| 2790 | "turn_id": "turn_runaway", |
| 2791 | "payload": { "turn": { "status": "interrupted" } } |
| 2792 | }), |
| 2793 | ), |
| 2794 | ] |
| 2795 | .concat(); |
| 2796 | ( |
| 2797 | [(header::CONTENT_TYPE, "text/event-stream")], |
| 2798 | Body::from(body), |
| 2799 | ) |
| 2800 | } |
| 2801 | |
| 2802 | let listener = tokio::net::TcpListener::bind("127.0.0.1:0") |
| 2803 | .await |
| 2804 | .expect("bind test listener"); |
| 2805 | let addr = listener.local_addr().expect("listener addr"); |
| 2806 | let app = Router::new() |
| 2807 | .route("/v1/threads", post(create_thread)) |
| 2808 | .route("/v1/threads/{thread_id}/turns", post(create_turn)) |
| 2809 | .route( |
| 2810 | "/v1/threads/{thread_id}/turns/{turn_id}/interrupt", |
| 2811 | post(interrupt), |
| 2812 | ) |
| 2813 | .route("/v1/threads/{thread_id}/events", get(thread_events)) |
| 2814 | .with_state(interrupted.clone()); |
| 2815 | let server = tokio::spawn(async move { |
| 2816 | axum::serve(listener, app) |
| 2817 | .await |
| 2818 | .expect("serve test runtime"); |
| 2819 | }); |
| 2820 | (format!("http://{addr}"), interrupted, server) |
| 2821 | } |
| 2822 | |
| 2823 | #[tokio::test] |
| 2824 | async fn interrupt_stops_a_turn_that_would_otherwise_stream_forever() { |
| 2825 | let (base_url, _notify, server) = spawn_uninterruptible_until_asked_runtime().await; |
| 2826 | let (state, _tmp) = capability_test_state(); |
| 2827 | *state.stdio_bridge.lock().await = Some(Arc::new(Mutex::new( |
| 2828 | RuntimeBridge::from_base_url_for_test(base_url), |
| 2829 | ))); |
| 2830 | |
| 2831 | let (client, server_side) = tokio::io::duplex(16 * 1024); |
| 2832 | let (client_reader, mut client_writer) = tokio::io::split(client); |
| 2833 | |
| 2834 | let loop_state = state.clone(); |
| 2835 | let loop_handle = tokio::spawn(async move { |
| 2836 | let (rx, tx) = tokio::io::split(server_side); |
| 2837 | run_stdio_loop(&loop_state, BufReader::new(rx).lines(), tx).await |
| 2838 | }); |
| 2839 | |
| 2840 | // Start the runaway turn. |
| 2841 | client_writer |
| 2842 | .write_all( |
| 2843 | b"{\"jsonrpc\":\"2.0\",\"id\":1,\"method\":\"thread/message\",\ |
| 2844 | \"params\":{\"thread_id\":\"thr_a\",\"input\":\"go\"}}\n", |
| 2845 | ) |
| 2846 | .await |
| 2847 | .expect("send thread/message"); |
| 2848 | |
| 2849 | // Wait until the turn is genuinely in flight before cancelling, so the |
| 2850 | // test exercises mid-stream cancellation rather than a race. |
| 2851 | tokio::time::timeout(Duration::from_secs(10), async { |
| 2852 | loop { |
| 2853 | if state.in_flight_turns.lock().await.contains_key("thr_a") { |
| 2854 | return; |
| 2855 | } |
| 2856 | tokio::time::sleep(Duration::from_millis(10)).await; |
| 2857 | } |
| 2858 | }) |
| 2859 | .await |
| 2860 | .expect("turn should register itself as in flight"); |
| 2861 | |
| 2862 | // The read loop must accept this while the turn holds the bridge. |
| 2863 | client_writer |
| 2864 | .write_all( |
| 2865 | b"{\"jsonrpc\":\"2.0\",\"id\":2,\"method\":\"thread/interrupt\",\ |
| 2866 | \"params\":{\"thread_id\":\"thr_a\"}}\n", |
| 2867 | ) |
| 2868 | .await |
| 2869 | .expect("send thread/interrupt"); |
| 2870 | client_writer |
| 2871 | .write_all(b"{\"jsonrpc\":\"2.0\",\"id\":3,\"method\":\"shutdown\"}\n") |
| 2872 | .await |
| 2873 | .expect("send shutdown"); |
| 2874 | |
| 2875 | let finished = tokio::time::timeout(Duration::from_secs(20), loop_handle) |
| 2876 | .await |
| 2877 | .expect("the loop must exit rather than hang on the runaway turn"); |
| 2878 | finished.expect("join loop").expect("loop result"); |
| 2879 | |
| 2880 | let mut output = String::new(); |
| 2881 | let mut lines = BufReader::new(client_reader); |
| 2882 | lines |
| 2883 | .read_to_string(&mut output) |
| 2884 | .await |
| 2885 | .expect("read stdio output"); |
| 2886 | |
| 2887 | let responses: Vec<Value> = output |
| 2888 | .lines() |
| 2889 | .filter_map(|line| serde_json::from_str::<Value>(line).ok()) |
| 2890 | .collect(); |
| 2891 | let by_id = |id: u64| { |
| 2892 | responses |
| 2893 | .iter() |
| 2894 | .find(|value| value["id"] == json!(id)) |
| 2895 | .unwrap_or_else(|| panic!("no response for id {id} in {output}")) |
| 2896 | .clone() |
| 2897 | }; |
| 2898 | |
| 2899 | // The turn ended as interrupted rather than running to completion. |
| 2900 | assert!( |
| 2901 | by_id(1)["error"].is_object(), |
| 2902 | "the interrupted turn should report an error, got: {}", |
| 2903 | by_id(1) |
| 2904 | ); |
| 2905 | assert_eq!(by_id(2)["result"]["interrupted"], json!(true)); |
| 2906 | assert_eq!(by_id(3)["result"]["status"], json!("stopped")); |
| 2907 | |
| 2908 | server.abort(); |
| 2909 | let _ = server.await; |
| 2910 | } |
| 2911 | |
| 2912 | #[tokio::test] |
| 2913 | async fn interrupting_an_idle_thread_is_not_an_error() { |
| 2914 | let (state, _tmp) = capability_test_state(); |
| 2915 | let response = dispatch_stdio_request( |
| 2916 | &state, |
| 2917 | "thread/interrupt", |
| 2918 | json!({ "thread_id": "thr_nothing_running" }), |
| 2919 | ) |
| 2920 | .await |
| 2921 | .expect("interrupt dispatch"); |
| 2922 | assert_eq!(response.result["interrupted"], json!(false)); |
| 2923 | } |
| 2924 | |
| 2925 | #[tokio::test] |
| 2926 | async fn stdio_runtime_bridge_streams_response_delta_events() { |
| 2927 | async fn create_turn(AxumPath(thread_id): AxumPath<String>) -> Json<Value> { |
| 2928 | Json(json!({ |
| 2929 | "thread": { "id": thread_id }, |
| 2930 | "turn": { "id": "turn_test" }, |
| 2931 | })) |
| 2932 | } |
| 2933 | |
| 2934 | async fn thread_events( |
| 2935 | AxumPath(thread_id): AxumPath<String>, |
| 2936 | Query(query): Query<HashMap<String, String>>, |
| 2937 | ) -> ([(header::HeaderName, &'static str); 1], String) { |
| 2938 | assert_eq!(thread_id, "thr_test"); |
| 2939 | assert_eq!(query.get("since_seq").map(String::as_str), Some("0")); |
| 2940 | |
| 2941 | let body = [ |
| 2942 | sse_frame( |
| 2943 | "item.delta", |
| 2944 | json!({ |
| 2945 | "seq": 1, |
| 2946 | "turn_id": "turn_test", |
| 2947 | "payload": { |
| 2948 | "kind": "agent_message", |
| 2949 | "delta": "hello" |
| 2950 | } |
| 2951 | }), |
| 2952 | ), |
| 2953 | sse_frame( |
| 2954 | "turn.completed", |
| 2955 | json!({ |
| 2956 | "seq": 2, |
| 2957 | "turn_id": "turn_test", |
| 2958 | "payload": { |
| 2959 | "turn": { |
| 2960 | "status": "completed" |
| 2961 | } |
| 2962 | } |
| 2963 | }), |
| 2964 | ), |
| 2965 | ] |
| 2966 | .concat(); |
| 2967 | |
| 2968 | ([(header::CONTENT_TYPE, "text/event-stream")], body) |
| 2969 | } |
| 2970 | |
| 2971 | let listener = tokio::net::TcpListener::bind("127.0.0.1:0") |
| 2972 | .await |
| 2973 | .expect("bind test listener"); |
| 2974 | let addr = listener.local_addr().expect("listener addr"); |
| 2975 | let app = Router::new() |
| 2976 | .route("/v1/threads/{thread_id}/turns", post(create_turn)) |
| 2977 | .route("/v1/threads/{thread_id}/events", get(thread_events)); |
| 2978 | |
| 2979 | let server = tokio::spawn(async move { |
| 2980 | axum::serve(listener, app) |
| 2981 | .await |
| 2982 | .expect("serve test runtime"); |
| 2983 | }); |
| 2984 | |
| 2985 | let mut bridge = RuntimeBridge::from_base_url_for_test(format!("http://{addr}")); |
| 2986 | let (mut reader, mut writer) = tokio::io::duplex(4096); |
| 2987 | |
| 2988 | let result = bridge |
| 2989 | .message_thread("thr_test", "hello", &mut writer, None) |
| 2990 | .await |
| 2991 | .expect("message_thread should succeed"); |
| 2992 | drop(writer); |
| 2993 | |
| 2994 | let mut stdout = Vec::new(); |
| 2995 | reader |
| 2996 | .read_to_end(&mut stdout) |
| 2997 | .await |
| 2998 | .expect("read stdio output"); |
| 2999 | server.abort(); |
| 3000 | let _ = server.await; |
| 3001 | |
| 3002 | let lines: Vec<Value> = String::from_utf8(stdout) |
| 3003 | .expect("utf8 output") |
| 3004 | .lines() |
| 3005 | .map(|line| serde_json::from_str(line).expect("json line")) |
| 3006 | .collect(); |
| 3007 | |
| 3008 | assert_eq!( |
| 3009 | result.get("status").and_then(Value::as_str), |
| 3010 | Some("accepted") |
| 3011 | ); |
| 3012 | assert_eq!( |
| 3013 | result.pointer("/data/turn_id").and_then(Value::as_str), |
| 3014 | Some("turn_test") |
| 3015 | ); |
| 3016 | assert_eq!(bridge.last_seq_by_thread.get("thr_test"), Some(&2)); |
| 3017 | |
| 3018 | let event_types: Vec<&str> = lines |
| 3019 | .iter() |
| 3020 | .map(|line| { |
| 3021 | line.get("type") |
| 3022 | .and_then(Value::as_str) |
| 3023 | .expect("event type") |
| 3024 | }) |
| 3025 | .collect(); |
| 3026 | assert_eq!( |
| 3027 | event_types, |
| 3028 | vec!["response_start", "response_delta", "response_end"] |
| 3029 | ); |
| 3030 | assert_eq!(lines[1]["delta"], "hello"); |
| 3031 | } |
| 3032 | |
| 3033 | #[tokio::test] |
| 3034 | async fn stdio_runtime_bridge_applies_thread_start_hints() { |
| 3035 | async fn create_thread(Json(body): Json<Value>) -> Json<Value> { |
| 3036 | assert_eq!(body["model"], "deepseek-v4"); |
| 3037 | assert_eq!(body["workspace"], "/tmp/codewhale-stdio"); |
| 3038 | Json(json!({ |
| 3039 | "id": "thr_runtime", |
| 3040 | "model": body["model"].clone(), |
| 3041 | "workspace": body["workspace"].clone(), |
| 3042 | })) |
| 3043 | } |
| 3044 | |
| 3045 | let listener = tokio::net::TcpListener::bind("127.0.0.1:0") |
| 3046 | .await |
| 3047 | .expect("bind test listener"); |
| 3048 | let addr = listener.local_addr().expect("listener addr"); |
| 3049 | let app = Router::new().route("/v1/threads", post(create_thread)); |
| 3050 | |
| 3051 | let server = tokio::spawn(async move { |
| 3052 | axum::serve(listener, app) |
| 3053 | .await |
| 3054 | .expect("serve test runtime"); |
| 3055 | }); |
| 3056 | |
| 3057 | let mut bridge = RuntimeBridge::from_base_url_for_test(format!("http://{addr}")); |
| 3058 | let runtime_id = bridge |
| 3059 | .ensure_runtime_thread( |
| 3060 | "legacy_thread", |
| 3061 | Some(RuntimeThreadHint { |
| 3062 | model: Some("deepseek-v4".to_string()), |
| 3063 | workspace: Some(PathBuf::from("/tmp/codewhale-stdio")), |
| 3064 | }), |
| 3065 | ) |
| 3066 | .await |
| 3067 | .expect("runtime thread"); |
| 3068 | server.abort(); |
| 3069 | let _ = server.await; |
| 3070 | |
| 3071 | assert_eq!(runtime_id, "thr_runtime"); |
| 3072 | assert_eq!( |
| 3073 | bridge.thread_map.get("legacy_thread").map(String::as_str), |
| 3074 | Some("thr_runtime") |
| 3075 | ); |
| 3076 | } |
| 3077 | |
| 3078 | // ── capability drift guard ───────────────────────────────────────── |
| 3079 | // |
| 3080 | // The stdio `capabilities` method is the benchmark/SDK contract: external |
| 3081 | // harnesses probe it (without spending model tokens) to learn what the |
| 3082 | // app-server can do. Pin the advertised method set so any change forces a |
| 3083 | // deliberate update here, in the dispatcher, and in docs/RUNTIME_API.md. |
| 3084 | |
| 3085 | /// Methods advertised by the top-level `capabilities` probe, in order. |
| 3086 | const EXPECTED_CAPABILITY_METHODS: &[&str] = &[ |
| 3087 | "healthz", |
| 3088 | "thread/capabilities", |
| 3089 | "thread/request", |
| 3090 | "thread/create", |
| 3091 | "thread/start", |
| 3092 | "thread/resume", |
| 3093 | "thread/fork", |
| 3094 | "thread/list", |
| 3095 | "thread/read", |
| 3096 | "thread/set_name", |
| 3097 | "thread/goal/set", |
| 3098 | "thread/goal/get", |
| 3099 | "thread/goal/clear", |
| 3100 | "thread/archive", |
| 3101 | "thread/unarchive", |
| 3102 | "thread/message", |
| 3103 | "thread/interrupt", |
| 3104 | "app/capabilities", |
| 3105 | "app/request", |
| 3106 | "app/config/get", |
| 3107 | "app/config/set", |
| 3108 | "app/config/unset", |
| 3109 | "app/config/list", |
| 3110 | "app/config/reload", |
| 3111 | "app/models", |
| 3112 | "app/thread_loaded_list", |
| 3113 | "prompt/capabilities", |
| 3114 | "prompt/request", |
| 3115 | "prompt/run", |
| 3116 | "shutdown", |
| 3117 | ]; |
| 3118 | |
| 3119 | fn capability_test_state() -> (AppState, tempfile::TempDir) { |
| 3120 | let tmp = tempfile::tempdir().expect("tempdir"); |
| 3121 | let config_path = tmp.path().join("config.toml"); |
| 3122 | fs::write(&config_path, "").expect("write config"); |
| 3123 | let state = build_state(Some(config_path), None).expect("state"); |
| 3124 | (state, tmp) |
| 3125 | } |
| 3126 | |
| 3127 | #[tokio::test] |
| 3128 | async fn capabilities_method_set_is_stable() { |
| 3129 | let (state, _tmp) = capability_test_state(); |
| 3130 | let caps = dispatch_stdio_request(&state, "capabilities", json!({})) |
| 3131 | .await |
| 3132 | .expect("capabilities dispatch"); |
| 3133 | let methods: Vec<String> = caps.result["methods"] |
| 3134 | .as_array() |
| 3135 | .expect("methods array") |
| 3136 | .iter() |
| 3137 | .map(|m| m.as_str().expect("method string").to_string()) |
| 3138 | .collect(); |
| 3139 | assert_eq!( |
| 3140 | methods, EXPECTED_CAPABILITY_METHODS, |
| 3141 | "app-server stdio capability set drifted; update the dispatcher, this \ |
| 3142 | snapshot, and docs/RUNTIME_API.md together" |
| 3143 | ); |
| 3144 | } |
| 3145 | |
| 3146 | #[tokio::test] |
| 3147 | async fn every_advertised_capability_is_dispatchable() { |
| 3148 | let (state, _tmp) = capability_test_state(); |
| 3149 | // Empty params: methods may fail validation (-32602), but none may report |
| 3150 | // method-not-found (-32601). Required fields (e.g. PromptRequest.prompt) |
| 3151 | // make the prompt routes fail at parse time, so no model tokens are spent. |
| 3152 | for method in EXPECTED_CAPABILITY_METHODS { |
| 3153 | if let Err(err) = dispatch_stdio_request(&state, method, json!({})).await { |
| 3154 | assert_ne!( |
| 3155 | err.code, |
| 3156 | JsonRpcError::method_not_found(method).code, |
| 3157 | "advertised capability `{method}` is not dispatchable" |
| 3158 | ); |
| 3159 | } |
| 3160 | } |
| 3161 | } |
| 3162 | |
| 3163 | // ── resolve_auth_token ───────────────────────────────────────────── |
| 3164 | |
| 3165 | #[test] |
| 3166 | fn auth_token_empty_string_fails() { |
| 3167 | let options = AppServerOptions { |
| 3168 | listen: "127.0.0.1:0".parse().expect("addr"), |
| 3169 | config_path: None, |
| 3170 | auth_token: Some(" ".to_string()), |
| 3171 | insecure_no_auth: false, |
| 3172 | cors_origins: Vec::new(), |
| 3173 | }; |
| 3174 | let err = resolve_auth_token(&options).expect_err("empty token should fail"); |
| 3175 | assert!(err.to_string().contains("cannot be empty")); |
| 3176 | } |
| 3177 | |
| 3178 | #[test] |
| 3179 | fn auth_token_generated_when_none_provided() { |
| 3180 | let options = AppServerOptions { |
| 3181 | listen: "127.0.0.1:0".parse().expect("addr"), |
| 3182 | config_path: None, |
| 3183 | auth_token: None, |
| 3184 | insecure_no_auth: false, |
| 3185 | cors_origins: Vec::new(), |
| 3186 | }; |
| 3187 | let token = resolve_auth_token(&options).unwrap(); |
| 3188 | assert!(token.is_some()); |
| 3189 | assert!(token.unwrap().starts_with("cwapp_")); |
| 3190 | } |
| 3191 | |
| 3192 | #[test] |
| 3193 | fn runtime_bridge_command_keeps_auth_token_out_of_argv() { |
| 3194 | // FR001-C001: runtime auth token must not appear on the child argv |
| 3195 | // (visible via local `ps`); pass it via env instead. |
| 3196 | let token = "cwrt_unit_test_secret_token_not_for_argv"; |
| 3197 | let cmd = RuntimeBridge::runtime_command(None, 18787, token).expect("command"); |
| 3198 | let argv: Vec<String> = cmd |
| 3199 | .get_args() |
| 3200 | .map(|a| a.to_string_lossy().into_owned()) |
| 3201 | .collect(); |
| 3202 | assert!( |
| 3203 | !argv |
| 3204 | .iter() |
| 3205 | .any(|a| a.contains(token) || a == "--auth-token"), |
| 3206 | "auth token must not be present in child argv: {argv:?}" |
| 3207 | ); |
| 3208 | let envs: Vec<(String, String)> = cmd |
| 3209 | .get_envs() |
| 3210 | .filter_map(|(k, v)| { |
| 3211 | Some(( |
| 3212 | k.to_string_lossy().into_owned(), |
| 3213 | v?.to_string_lossy().into_owned(), |
| 3214 | )) |
| 3215 | }) |
| 3216 | .collect(); |
| 3217 | assert!( |
| 3218 | envs.iter() |
| 3219 | .any(|(k, v)| k == "CODEWHALE_RUNTIME_TOKEN" && v == token), |
| 3220 | "token must be carried via CODEWHALE_RUNTIME_TOKEN: {envs:?}" |
| 3221 | ); |
| 3222 | assert!( |
| 3223 | envs.iter() |
| 3224 | .any(|(k, v)| k == "DEEPSEEK_RUNTIME_TOKEN" && v == token), |
| 3225 | "legacy alias DEEPSEEK_RUNTIME_TOKEN must also carry the token: {envs:?}" |
| 3226 | ); |
| 3227 | } |
| 3228 | |
| 3229 | #[test] |
| 3230 | fn generated_auth_status_does_not_render_token() { |
| 3231 | let rendered = app_server_auth_status_lines(false).join("\n"); |
| 3232 | |
| 3233 | assert!(!rendered.contains("Authorization: Bearer")); |
| 3234 | assert!(rendered.contains("not printed")); |
| 3235 | assert!(rendered.contains("CODEWHALE_APP_SERVER_TOKEN")); |
| 3236 | } |
| 3237 | |
| 3238 | #[test] |
| 3239 | fn auth_token_explicit_is_preserved() { |
| 3240 | let options = AppServerOptions { |
| 3241 | listen: "127.0.0.1:0".parse().expect("addr"), |
| 3242 | config_path: None, |
| 3243 | auth_token: Some("my-secret".to_string()), |
| 3244 | insecure_no_auth: false, |
| 3245 | cors_origins: Vec::new(), |
| 3246 | }; |
| 3247 | let token = resolve_auth_token(&options).unwrap(); |
| 3248 | assert_eq!(token.as_deref(), Some("my-secret")); |
| 3249 | } |
| 3250 | |
| 3251 | #[test] |
| 3252 | fn auth_token_explicit_allows_non_loopback_bind() { |
| 3253 | let options = AppServerOptions { |
| 3254 | listen: "0.0.0.0:8787".parse().expect("socket addr"), |
| 3255 | config_path: None, |
| 3256 | auth_token: Some("my-secret".to_string()), |
| 3257 | insecure_no_auth: false, |
| 3258 | cors_origins: Vec::new(), |
| 3259 | }; |
| 3260 | let token = resolve_auth_token(&options).unwrap(); |
| 3261 | assert_eq!(token.as_deref(), Some("my-secret")); |
| 3262 | } |
| 3263 | |
| 3264 | #[test] |
| 3265 | fn insecure_no_auth_on_loopback_returns_none() { |
| 3266 | let options = AppServerOptions { |
| 3267 | listen: "127.0.0.1:0".parse().expect("addr"), |
| 3268 | config_path: None, |
| 3269 | auth_token: None, |
| 3270 | insecure_no_auth: true, |
| 3271 | cors_origins: Vec::new(), |
| 3272 | }; |
| 3273 | let token = resolve_auth_token(&options).unwrap(); |
| 3274 | assert!(token.is_none()); |
| 3275 | } |
| 3276 | |
| 3277 | #[test] |
| 3278 | fn insecure_no_auth_on_non_loopback_fails_fast() { |
| 3279 | let options = AppServerOptions { |
| 3280 | listen: "0.0.0.0:8787".parse().expect("socket addr"), |
| 3281 | config_path: None, |
| 3282 | auth_token: None, |
| 3283 | insecure_no_auth: true, |
| 3284 | cors_origins: Vec::new(), |
| 3285 | }; |
| 3286 | |
| 3287 | let err = resolve_auth_token(&options).expect_err("non-loopback unauth should fail"); |
| 3288 | assert!( |
| 3289 | err.to_string() |
| 3290 | .contains("refusing unauthenticated app-server bind") |
| 3291 | ); |
| 3292 | } |
| 3293 | |
| 3294 | // ── cors_layer ───────────────────────────────────────────────────── |
| 3295 | |
| 3296 | #[test] |
| 3297 | fn cors_layer_includes_default_origins() { |
| 3298 | let layer = cors_layer(&[]); |
| 3299 | // Just verify it doesn't panic and creates successfully |
| 3300 | let _ = layer; |
| 3301 | } |
| 3302 | |
| 3303 | #[test] |
| 3304 | fn cors_layer_adds_extra_origins() { |
| 3305 | let extras = vec!["https://example.com".to_string()]; |
| 3306 | let layer = cors_layer(&extras); |
| 3307 | let _ = layer; |
| 3308 | } |
| 3309 | |
| 3310 | #[test] |
| 3311 | fn cors_layer_skips_empty_origins() { |
| 3312 | let extras = vec!["".to_string(), " ".to_string()]; |
| 3313 | let layer = cors_layer(&extras); |
| 3314 | let _ = layer; |
| 3315 | } |
| 3316 | |
| 3317 | // ── JsonRpc helpers ──────────────────────────────────────────────── |
| 3318 | |
| 3319 | #[test] |
| 3320 | fn params_or_object_returns_object_for_null() { |
| 3321 | let result = params_or_object(Value::Null); |
| 3322 | assert_eq!(result, json!({})); |
| 3323 | } |
| 3324 | |
| 3325 | #[test] |
| 3326 | fn params_or_object_passthrough_for_non_null() { |
| 3327 | let input = json!({"key": "value"}); |
| 3328 | let result = params_or_object(input.clone()); |
| 3329 | assert_eq!(result, input); |
| 3330 | } |
| 3331 | |
| 3332 | #[test] |
| 3333 | fn jsonrpc_result_format() { |
| 3334 | let result = jsonrpc_result(Some(json!(1)), json!({"ok": true})); |
| 3335 | assert_eq!(result["jsonrpc"], "2.0"); |
| 3336 | assert_eq!(result["id"], 1); |
| 3337 | assert_eq!(result["result"]["ok"], true); |
| 3338 | } |
| 3339 | |
| 3340 | #[test] |
| 3341 | fn jsonrpc_result_null_id() { |
| 3342 | let result = jsonrpc_result(None, json!(null)); |
| 3343 | assert_eq!(result["id"], Value::Null); |
| 3344 | } |
| 3345 | |
| 3346 | #[test] |
| 3347 | fn jsonrpc_error_format() { |
| 3348 | let err = jsonrpc_error(Some(json!(2)), JsonRpcError::internal("oops")); |
| 3349 | assert_eq!(err["jsonrpc"], "2.0"); |
| 3350 | assert_eq!(err["id"], 2); |
| 3351 | assert_eq!(err["error"]["code"], -32603); |
| 3352 | assert_eq!(err["error"]["message"], "oops"); |
| 3353 | } |
| 3354 | |
| 3355 | #[test] |
| 3356 | fn jsonrpc_error_codes() { |
| 3357 | assert_eq!(JsonRpcError::parse_error("").code, -32700); |
| 3358 | assert_eq!(JsonRpcError::invalid_request("").code, -32600); |
| 3359 | assert_eq!(JsonRpcError::method_not_found("x").code, -32601); |
| 3360 | assert_eq!(JsonRpcError::invalid_params("").code, -32602); |
| 3361 | assert_eq!(JsonRpcError::internal("").code, -32603); |
| 3362 | } |
| 3363 | |
| 3364 | // ── AppServerOptions ─────────────────────────────────────────────── |
| 3365 | |
| 3366 | #[test] |
| 3367 | fn app_server_options_debug_does_not_leak_token() { |
| 3368 | let options = AppServerOptions { |
| 3369 | listen: "127.0.0.1:8080".parse().expect("addr"), |
| 3370 | config_path: None, |
| 3371 | auth_token: Some("secret-token".to_string()), |
| 3372 | insecure_no_auth: false, |
| 3373 | cors_origins: vec!["https://example.com".to_string()], |
| 3374 | }; |
| 3375 | let debug = format!("{options:?}"); |
| 3376 | assert!(!debug.contains("secret-token")); |
| 3377 | assert!(debug.contains("<redacted>")); |
| 3378 | assert!(debug.contains("8080")); |
| 3379 | } |
| 3380 | |
| 3381 | // ── Default CORS origins ────────────────────────────────────────── |
| 3382 | |
| 3383 | #[test] |
| 3384 | fn default_cors_origins_include_common_dev_ports() { |
| 3385 | assert!(DEFAULT_CORS_ORIGINS.contains(&"http://localhost:3000")); |
| 3386 | assert!(DEFAULT_CORS_ORIGINS.contains(&"http://localhost:5173")); |
| 3387 | assert!(DEFAULT_CORS_ORIGINS.contains(&"tauri://localhost")); |
| 3388 | } |
| 3389 | } |
| 3390 |