| 1 | use super::*; |
| 2 | |
| 3 | use crate::models::SystemBlock; |
| 4 | use serde_json::json; |
| 5 | use std::collections::HashSet; |
| 6 | use std::ffi::OsString; |
| 7 | use std::fs; |
| 8 | use std::path::{Path, PathBuf}; |
| 9 | use std::sync::LazyLock; |
| 10 | use std::time::Instant; |
| 11 | use tempfile::tempdir; |
| 12 | |
| 13 | const WORKING_SET_SUMMARY_MARKER: &str = "## Repo Working Set"; |
| 14 | static CAPACITY_MEMORY_ENV_LOCK: LazyLock<tokio::sync::Mutex<()>> = |
| 15 | LazyLock::new(|| tokio::sync::Mutex::new(())); |
| 16 | static API_KEY_ENV_LOCK: LazyLock<std::sync::Mutex<()>> = |
| 17 | LazyLock::new(|| std::sync::Mutex::new(())); |
| 18 | |
| 19 | struct ScopedCapacityMemoryDir { |
| 20 | previous: Option<OsString>, |
| 21 | } |
| 22 | |
| 23 | impl ScopedCapacityMemoryDir { |
| 24 | fn set(path: &Path) -> Self { |
| 25 | let previous = std::env::var_os("DEEPSEEK_CAPACITY_MEMORY_DIR"); |
| 26 | // Safety: capacity-memory tests serialize access with CAPACITY_MEMORY_ENV_LOCK |
| 27 | // and restore the original value in Drop. |
| 28 | unsafe { |
| 29 | std::env::set_var("DEEPSEEK_CAPACITY_MEMORY_DIR", path); |
| 30 | } |
| 31 | Self { previous } |
| 32 | } |
| 33 | } |
| 34 | |
| 35 | impl Drop for ScopedCapacityMemoryDir { |
| 36 | fn drop(&mut self) { |
| 37 | // Safety: capacity-memory tests serialize access with CAPACITY_MEMORY_ENV_LOCK. |
| 38 | unsafe { |
| 39 | if let Some(previous) = self.previous.take() { |
| 40 | std::env::set_var("DEEPSEEK_CAPACITY_MEMORY_DIR", previous); |
| 41 | } else { |
| 42 | std::env::remove_var("DEEPSEEK_CAPACITY_MEMORY_DIR"); |
| 43 | } |
| 44 | } |
| 45 | } |
| 46 | } |
| 47 | |
| 48 | struct ScopedDeepSeekApiKey { |
| 49 | previous: Option<OsString>, |
| 50 | } |
| 51 | |
| 52 | impl ScopedDeepSeekApiKey { |
| 53 | fn set(value: &str) -> Self { |
| 54 | let previous = std::env::var_os("DEEPSEEK_API_KEY"); |
| 55 | // Safety: tests using this helper serialize with API_KEY_ENV_LOCK and |
| 56 | // restore the original value in Drop. |
| 57 | unsafe { |
| 58 | std::env::set_var("DEEPSEEK_API_KEY", value); |
| 59 | } |
| 60 | Self { previous } |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | impl Drop for ScopedDeepSeekApiKey { |
| 65 | fn drop(&mut self) { |
| 66 | // Safety: tests using this helper serialize with API_KEY_ENV_LOCK. |
| 67 | unsafe { |
| 68 | if let Some(previous) = self.previous.take() { |
| 69 | std::env::set_var("DEEPSEEK_API_KEY", previous); |
| 70 | } else { |
| 71 | std::env::remove_var("DEEPSEEK_API_KEY"); |
| 72 | } |
| 73 | } |
| 74 | } |
| 75 | } |
| 76 | |
| 77 | fn build_engine_with_capacity(capacity: CapacityControllerConfig) -> Engine { |
| 78 | let engine_config = EngineConfig { |
| 79 | capacity, |
| 80 | ..Default::default() |
| 81 | }; |
| 82 | let (engine, _handle) = Engine::new(engine_config, &Config::default()); |
| 83 | engine |
| 84 | } |
| 85 | |
| 86 | #[test] |
| 87 | fn env_only_auth_error_gets_recovery_hint() { |
| 88 | let _guard = API_KEY_ENV_LOCK.lock().unwrap_or_else(|e| e.into_inner()); |
| 89 | let _env = ScopedDeepSeekApiKey::set("stale-env-key"); |
| 90 | let (engine, _handle) = Engine::new(EngineConfig::default(), &Config::default()); |
| 91 | |
| 92 | let message = |
| 93 | engine.decorate_auth_error_message("Authentication failed: invalid API key".to_string()); |
| 94 | |
| 95 | assert!(message.contains("DEEPSEEK_API_KEY")); |
| 96 | assert!(message.contains("no saved config key is present")); |
| 97 | assert!(message.contains("deepseek auth set --provider deepseek")); |
| 98 | } |
| 99 | |
| 100 | #[test] |
| 101 | fn config_auth_error_does_not_blame_env() { |
| 102 | let _guard = API_KEY_ENV_LOCK.lock().unwrap_or_else(|e| e.into_inner()); |
| 103 | let _env = ScopedDeepSeekApiKey::set("stale-env-key"); |
| 104 | let cfg = Config { |
| 105 | api_key: Some("fresh-config-key".to_string()), |
| 106 | ..Config::default() |
| 107 | }; |
| 108 | let (engine, _handle) = Engine::new(EngineConfig::default(), &cfg); |
| 109 | |
| 110 | let message = |
| 111 | engine.decorate_auth_error_message("Authentication failed: invalid API key".to_string()); |
| 112 | |
| 113 | assert_eq!(message, "Authentication failed: invalid API key"); |
| 114 | } |
| 115 | |
| 116 | fn make_plan( |
| 117 | read_only: bool, |
| 118 | supports_parallel: bool, |
| 119 | approval_required: bool, |
| 120 | interactive: bool, |
| 121 | ) -> ToolExecutionPlan { |
| 122 | ToolExecutionPlan { |
| 123 | index: 0, |
| 124 | id: "tool-1".to_string(), |
| 125 | name: "grep_files".to_string(), |
| 126 | input: json!({"pattern": "test"}), |
| 127 | caller: None, |
| 128 | interactive, |
| 129 | approval_required, |
| 130 | approval_description: "desc".to_string(), |
| 131 | supports_parallel, |
| 132 | read_only, |
| 133 | blocked_error: None, |
| 134 | guard_result: None, |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | fn api_tool(name: &str) -> Tool { |
| 139 | Tool { |
| 140 | tool_type: Some("function".to_string()), |
| 141 | name: name.to_string(), |
| 142 | description: format!("Test tool {name}"), |
| 143 | input_schema: json!({"type": "object"}), |
| 144 | allowed_callers: Some(vec!["direct".to_string()]), |
| 145 | defer_loading: None, |
| 146 | input_examples: None, |
| 147 | strict: None, |
| 148 | cache_control: None, |
| 149 | } |
| 150 | } |
| 151 | |
| 152 | #[test] |
| 153 | fn engine_handle_cancel_tracks_latest_turn_token() { |
| 154 | let (mut engine, handle) = Engine::new(EngineConfig::default(), &Config::default()); |
| 155 | let stale_token = engine.cancel_token.clone(); |
| 156 | |
| 157 | engine.reset_cancel_token(); |
| 158 | handle.cancel(); |
| 159 | |
| 160 | assert!(engine.cancel_token.is_cancelled()); |
| 161 | assert!(handle.is_cancelled()); |
| 162 | assert!(!stale_token.is_cancelled()); |
| 163 | } |
| 164 | |
| 165 | #[test] |
| 166 | fn engine_initial_prompt_includes_configured_goal() { |
| 167 | let config = EngineConfig { |
| 168 | goal_objective: Some("Fix goal handoff".to_string()), |
| 169 | ..Default::default() |
| 170 | }; |
| 171 | let (engine, _handle) = Engine::new(config, &Config::default()); |
| 172 | let prompt = match engine.session.system_prompt { |
| 173 | Some(SystemPrompt::Text(text)) => text, |
| 174 | Some(SystemPrompt::Blocks(blocks)) => blocks |
| 175 | .into_iter() |
| 176 | .map(|block| block.text) |
| 177 | .collect::<Vec<_>>() |
| 178 | .join("\n"), |
| 179 | None => panic!("expected system prompt"), |
| 180 | }; |
| 181 | |
| 182 | assert!(prompt.contains("<session_goal>")); |
| 183 | assert!(prompt.contains("Fix goal handoff")); |
| 184 | } |
| 185 | |
| 186 | #[test] |
| 187 | fn parallel_batch_requires_read_only_parallel_tools() { |
| 188 | let plans = vec![make_plan(true, true, false, false)]; |
| 189 | assert!(should_parallelize_tool_batch(&plans)); |
| 190 | |
| 191 | let plans = vec![ |
| 192 | make_plan(true, true, false, false), |
| 193 | make_plan(true, true, false, false), |
| 194 | ]; |
| 195 | assert!(should_parallelize_tool_batch(&plans)); |
| 196 | |
| 197 | let plans = vec![make_plan(false, true, false, false)]; |
| 198 | assert!(!should_parallelize_tool_batch(&plans)); |
| 199 | |
| 200 | let plans = vec![make_plan(true, false, false, false)]; |
| 201 | assert!(!should_parallelize_tool_batch(&plans)); |
| 202 | |
| 203 | let plans = vec![make_plan(true, true, true, false)]; |
| 204 | assert!(!should_parallelize_tool_batch(&plans)); |
| 205 | |
| 206 | let plans = vec![make_plan(true, true, false, true)]; |
| 207 | assert!(!should_parallelize_tool_batch(&plans)); |
| 208 | } |
| 209 | |
| 210 | #[test] |
| 211 | fn successful_update_plan_ends_plan_mode_turn_immediately() { |
| 212 | assert!(should_stop_after_plan_tool( |
| 213 | AppMode::Plan, |
| 214 | "update_plan", |
| 215 | &Ok(ToolResult::success("planned")) |
| 216 | )); |
| 217 | assert!(!should_stop_after_plan_tool( |
| 218 | AppMode::Agent, |
| 219 | "update_plan", |
| 220 | &Ok(ToolResult::success("planned")) |
| 221 | )); |
| 222 | assert!(!should_stop_after_plan_tool( |
| 223 | AppMode::Plan, |
| 224 | "request_user_input", |
| 225 | &Ok(ToolResult::success("input")) |
| 226 | )); |
| 227 | assert!(!should_stop_after_plan_tool( |
| 228 | AppMode::Plan, |
| 229 | "update_plan", |
| 230 | &Err(ToolError::execution_failed("failed".to_string())) |
| 231 | )); |
| 232 | } |
| 233 | |
| 234 | #[test] |
| 235 | fn quick_plan_requests_force_update_plan_on_first_step() { |
| 236 | assert!(should_force_update_plan_first( |
| 237 | AppMode::Plan, |
| 238 | "Give me a quick 3-step plan to verify the UI changes." |
| 239 | )); |
| 240 | assert!(should_force_update_plan_first( |
| 241 | AppMode::Plan, |
| 242 | "Make a high-level plan for the footer work." |
| 243 | )); |
| 244 | assert!(!should_force_update_plan_first( |
| 245 | AppMode::Plan, |
| 246 | "Inspect the repo and then give me a quick plan." |
| 247 | )); |
| 248 | assert!(!should_force_update_plan_first( |
| 249 | AppMode::Agent, |
| 250 | "Give me a quick 3-step plan." |
| 251 | )); |
| 252 | } |
| 253 | |
| 254 | #[test] |
| 255 | fn quick_plan_turn_can_narrow_first_step_tools_to_update_plan() { |
| 256 | let catalog = vec![ |
| 257 | Tool { |
| 258 | tool_type: Some("function".to_string()), |
| 259 | name: "read_file".to_string(), |
| 260 | description: "Read a file".to_string(), |
| 261 | input_schema: json!({"type": "object"}), |
| 262 | allowed_callers: Some(vec!["direct".to_string()]), |
| 263 | defer_loading: Some(false), |
| 264 | input_examples: None, |
| 265 | strict: None, |
| 266 | cache_control: None, |
| 267 | }, |
| 268 | Tool { |
| 269 | tool_type: Some("function".to_string()), |
| 270 | name: "update_plan".to_string(), |
| 271 | description: "Publish a plan".to_string(), |
| 272 | input_schema: json!({"type": "object"}), |
| 273 | allowed_callers: Some(vec!["direct".to_string()]), |
| 274 | defer_loading: Some(false), |
| 275 | input_examples: None, |
| 276 | strict: None, |
| 277 | cache_control: None, |
| 278 | }, |
| 279 | ]; |
| 280 | let active = initial_active_tools(&catalog); |
| 281 | |
| 282 | let forced = active_tools_for_step(&catalog, &active, true); |
| 283 | assert_eq!(forced.len(), 1); |
| 284 | assert_eq!(forced[0].name, "update_plan"); |
| 285 | |
| 286 | let default = active_tools_for_step(&catalog, &active, false); |
| 287 | assert_eq!(default.len(), 2); |
| 288 | } |
| 289 | |
| 290 | #[test] |
| 291 | fn tool_error_messages_include_actionable_hints() { |
| 292 | let path_error = ToolError::path_escape(PathBuf::from("../escape.txt")); |
| 293 | let formatted = format_tool_error(&path_error, "read_file"); |
| 294 | assert!(formatted.contains("escapes workspace")); |
| 295 | |
| 296 | let missing_field = ToolError::missing_field("path"); |
| 297 | let formatted = format_tool_error(&missing_field, "read_file"); |
| 298 | assert!(formatted.contains("missing required field")); |
| 299 | |
| 300 | let timeout = ToolError::Timeout { seconds: 5 }; |
| 301 | let formatted = format_tool_error(&timeout, "exec_shell"); |
| 302 | assert!(formatted.contains("timed out")); |
| 303 | } |
| 304 | |
| 305 | #[test] |
| 306 | fn tool_exec_outcome_tracks_duration() { |
| 307 | let outcome = ToolExecOutcome { |
| 308 | index: 0, |
| 309 | id: "tool-1".to_string(), |
| 310 | name: "grep_files".to_string(), |
| 311 | input: json!({"pattern": "test"}), |
| 312 | started_at: Instant::now(), |
| 313 | result: Ok(ToolResult::success("ok")), |
| 314 | }; |
| 315 | |
| 316 | assert!(outcome.started_at.elapsed().as_nanos() > 0); |
| 317 | } |
| 318 | |
| 319 | #[test] |
| 320 | fn yolo_mode_keeps_tools_preloaded() { |
| 321 | assert!(!should_default_defer_tool("exec_shell", AppMode::Yolo)); |
| 322 | assert!(!should_default_defer_tool( |
| 323 | "mcp_read_resource", |
| 324 | AppMode::Yolo |
| 325 | )); |
| 326 | } |
| 327 | |
| 328 | #[test] |
| 329 | fn non_yolo_mode_retains_default_defer_policy() { |
| 330 | // Shell tools are kept loaded in action modes so the model can verify |
| 331 | // work without an extra ToolSearch round-trip; non-action tools (e.g. |
| 332 | // MCP) still defer. |
| 333 | assert!(!should_default_defer_tool("exec_shell", AppMode::Agent)); |
| 334 | assert!(should_default_defer_tool("exec_shell", AppMode::Plan)); |
| 335 | assert!(!should_default_defer_tool("read_file", AppMode::Agent)); |
| 336 | assert!(should_default_defer_tool( |
| 337 | "mcp_read_resource", |
| 338 | AppMode::Agent |
| 339 | )); |
| 340 | } |
| 341 | |
| 342 | #[test] |
| 343 | fn model_tool_catalog_applies_native_and_mcp_deferral() { |
| 344 | let catalog = build_model_tool_catalog( |
| 345 | vec![ |
| 346 | api_tool("read_file"), |
| 347 | api_tool("exec_shell"), |
| 348 | api_tool("project_map"), |
| 349 | ], |
| 350 | vec![api_tool("list_mcp_resources"), api_tool("mcp_server_write")], |
| 351 | AppMode::Agent, |
| 352 | ); |
| 353 | |
| 354 | let defer_loading = |name: &str| { |
| 355 | catalog |
| 356 | .iter() |
| 357 | .find(|tool| tool.name == name) |
| 358 | .and_then(|tool| tool.defer_loading) |
| 359 | }; |
| 360 | |
| 361 | assert_eq!(defer_loading("read_file"), Some(false)); |
| 362 | assert_eq!(defer_loading("exec_shell"), Some(false)); |
| 363 | assert_eq!(defer_loading("project_map"), Some(true)); |
| 364 | assert_eq!(defer_loading("list_mcp_resources"), Some(false)); |
| 365 | assert_eq!(defer_loading("mcp_server_write"), Some(true)); |
| 366 | } |
| 367 | |
| 368 | #[test] |
| 369 | fn model_tool_catalog_keeps_everything_loaded_in_yolo_mode() { |
| 370 | let catalog = build_model_tool_catalog( |
| 371 | vec![api_tool("project_map")], |
| 372 | vec![api_tool("mcp_server_write")], |
| 373 | AppMode::Yolo, |
| 374 | ); |
| 375 | |
| 376 | assert!(catalog.iter().all(|tool| tool.defer_loading == Some(false))); |
| 377 | } |
| 378 | |
| 379 | #[test] |
| 380 | fn model_tool_catalog_sorts_each_partition_for_prefix_cache_stability() { |
| 381 | // Regression for #263: deterministic byte order of the tools array is a |
| 382 | // hard requirement for DeepSeek's KV prefix cache. Built-ins stay as a |
| 383 | // contiguous prefix; MCP tools follow. Within each partition: alphabetical. |
| 384 | let catalog = build_model_tool_catalog( |
| 385 | vec![ |
| 386 | api_tool("read_file"), |
| 387 | api_tool("apply_patch"), |
| 388 | api_tool("exec_shell"), |
| 389 | ], |
| 390 | vec![api_tool("mcp_zoo_b"), api_tool("mcp_aardvark_a")], |
| 391 | AppMode::Yolo, |
| 392 | ); |
| 393 | |
| 394 | let names: Vec<&str> = catalog.iter().map(|t| t.name.as_str()).collect(); |
| 395 | assert_eq!( |
| 396 | names, |
| 397 | vec![ |
| 398 | "apply_patch", |
| 399 | "exec_shell", |
| 400 | "read_file", |
| 401 | "mcp_aardvark_a", |
| 402 | "mcp_zoo_b", |
| 403 | ], |
| 404 | "built-ins must be alphabetical and contiguous; MCP tools follow, alphabetical", |
| 405 | ); |
| 406 | } |
| 407 | |
| 408 | #[test] |
| 409 | fn active_tool_list_pushes_deferred_activations_to_the_tail() { |
| 410 | // Regression for #263: when ToolSearch activates a deferred tool mid- |
| 411 | // session, it must NOT be inserted at its catalog index — that would |
| 412 | // shift every later tool's byte offset and bust the cached prefix. |
| 413 | // Deferred-but-now-active tools belong at the tail. |
| 414 | let mut a = api_tool("a_load_now"); |
| 415 | a.defer_loading = Some(false); |
| 416 | let mut search = api_tool("search_via_toolsearch"); |
| 417 | search.defer_loading = Some(true); |
| 418 | let mut b = api_tool("b_load_now"); |
| 419 | b.defer_loading = Some(false); |
| 420 | |
| 421 | let catalog = vec![a, search, b]; |
| 422 | let active: HashSet<String> = ["a_load_now", "search_via_toolsearch", "b_load_now"] |
| 423 | .into_iter() |
| 424 | .map(String::from) |
| 425 | .collect(); |
| 426 | |
| 427 | let listed = active_tools_for_step(&catalog, &active, false); |
| 428 | let names: Vec<&str> = listed.iter().map(|t| t.name.as_str()).collect(); |
| 429 | assert_eq!( |
| 430 | names, |
| 431 | vec!["a_load_now", "b_load_now", "search_via_toolsearch"], |
| 432 | "deferred-but-active tools must come after always-loaded tools", |
| 433 | ); |
| 434 | } |
| 435 | |
| 436 | #[test] |
| 437 | fn turn_tool_registry_builder_keeps_plan_mode_read_only_for_files() { |
| 438 | let (engine, _handle) = Engine::new(EngineConfig::default(), &Config::default()); |
| 439 | let registry = engine |
| 440 | .build_turn_tool_registry_builder( |
| 441 | AppMode::Plan, |
| 442 | engine.config.todos.clone(), |
| 443 | engine.config.plan_state.clone(), |
| 444 | ) |
| 445 | .build(engine.build_tool_context(AppMode::Plan, false)); |
| 446 | |
| 447 | assert!(registry.contains("read_file")); |
| 448 | assert!(registry.contains("list_dir")); |
| 449 | assert!(!registry.contains("write_file")); |
| 450 | assert!(!registry.contains("edit_file")); |
| 451 | assert!(registry.contains("update_plan")); |
| 452 | assert!(registry.contains("task_create")); |
| 453 | } |
| 454 | |
| 455 | #[test] |
| 456 | fn agent_mode_can_build_auto_approved_tool_context() { |
| 457 | let (engine, _handle) = Engine::new(EngineConfig::default(), &Config::default()); |
| 458 | |
| 459 | assert!( |
| 460 | !engine |
| 461 | .build_tool_context(AppMode::Agent, false) |
| 462 | .auto_approve |
| 463 | ); |
| 464 | assert!(engine.build_tool_context(AppMode::Agent, true).auto_approve); |
| 465 | assert!(engine.build_tool_context(AppMode::Yolo, false).auto_approve); |
| 466 | } |
| 467 | |
| 468 | #[test] |
| 469 | fn agent_and_yolo_modes_elevate_shell_sandbox_to_allow_network() { |
| 470 | // Regression for #273: the seatbelt-default policy denies all outbound |
| 471 | // network (including DNS), which broke `curl`, `yt-dlp`, package managers, |
| 472 | // and similar shell commands in Agent mode. Elevation must include |
| 473 | // network access so the application-level NetworkPolicy stays the only |
| 474 | // outbound boundary. |
| 475 | let (engine, _handle) = Engine::new(EngineConfig::default(), &Config::default()); |
| 476 | |
| 477 | let agent_ctx = engine.build_tool_context(AppMode::Agent, false); |
| 478 | let agent_policy = agent_ctx |
| 479 | .elevated_sandbox_policy |
| 480 | .as_ref() |
| 481 | .expect("Agent mode should elevate the sandbox policy"); |
| 482 | assert!( |
| 483 | agent_policy.has_network_access(), |
| 484 | "Agent mode must allow shell network access; got {agent_policy:?}", |
| 485 | ); |
| 486 | |
| 487 | let yolo_ctx = engine.build_tool_context(AppMode::Yolo, false); |
| 488 | let yolo_policy = yolo_ctx |
| 489 | .elevated_sandbox_policy |
| 490 | .as_ref() |
| 491 | .expect("Yolo mode should elevate the sandbox policy"); |
| 492 | assert!(yolo_policy.has_network_access()); |
| 493 | // v0.8.11: YOLO drops to DangerFullAccess (no sandbox) so the user |
| 494 | // is not bounced through approval round-trips for legitimate |
| 495 | // outside-workspace writes (package installs, sub-agent |
| 496 | // workspaces, ~/.cache mutations, etc.). YOLO is opt-in and |
| 497 | // already enables trust mode + auto-approve; the sandbox was the |
| 498 | // last guardrail and contradicts the contract. |
| 499 | assert!( |
| 500 | matches!(yolo_policy, crate::sandbox::SandboxPolicy::DangerFullAccess), |
| 501 | "Yolo mode must use DangerFullAccess (no sandbox); got {yolo_policy:?}", |
| 502 | ); |
| 503 | |
| 504 | // Plan mode is read-only investigation and does not register the shell |
| 505 | // tool, so it intentionally leaves the policy at the strict default. |
| 506 | assert!( |
| 507 | engine |
| 508 | .build_tool_context(AppMode::Plan, false) |
| 509 | .elevated_sandbox_policy |
| 510 | .is_none(), |
| 511 | ); |
| 512 | } |
| 513 | |
| 514 | #[tokio::test] |
| 515 | async fn session_update_preserves_reasoning_tool_only_turn() { |
| 516 | let (mut engine, handle) = Engine::new(EngineConfig::default(), &Config::default()); |
| 517 | let assistant = Message { |
| 518 | role: "assistant".to_string(), |
| 519 | content: vec![ |
| 520 | ContentBlock::Thinking { |
| 521 | thinking: "Need a tool before answering.".to_string(), |
| 522 | }, |
| 523 | ContentBlock::ToolUse { |
| 524 | id: "tool-1".to_string(), |
| 525 | name: "read_file".to_string(), |
| 526 | input: json!({"path": "Cargo.toml"}), |
| 527 | caller: None, |
| 528 | }, |
| 529 | ], |
| 530 | }; |
| 531 | |
| 532 | engine.add_session_message(assistant.clone()).await; |
| 533 | |
| 534 | let event = { |
| 535 | let mut rx = handle.rx_event.write().await; |
| 536 | rx.recv().await.expect("session update event") |
| 537 | }; |
| 538 | let Event::SessionUpdated { messages, .. } = event else { |
| 539 | panic!("expected session update event"); |
| 540 | }; |
| 541 | |
| 542 | assert_eq!(messages, vec![assistant]); |
| 543 | } |
| 544 | |
| 545 | #[test] |
| 546 | fn detects_context_length_errors_from_provider_payloads() { |
| 547 | let msg = r#"SSE stream request failed: HTTP 400 Bad Request: {"error":{"message":"This model's maximum context length is 131072 tokens. However, you requested 153056 tokens (148960 in the messages, 4096 in the completion).","type":"invalid_request_error"}}"#; |
| 548 | assert!(is_context_length_error_message(msg)); |
| 549 | assert!(!is_context_length_error_message( |
| 550 | "SSE stream request failed: HTTP 400 Bad Request: model not found" |
| 551 | )); |
| 552 | } |
| 553 | |
| 554 | #[test] |
| 555 | fn context_budget_reserves_output_and_headroom() { |
| 556 | // V4 has a 1M context window — the only family that comfortably hosts |
| 557 | // a 256K output reservation without saturating the input budget to 0. |
| 558 | let budget = context_input_budget("deepseek-v4-pro", TURN_MAX_OUTPUT_TOKENS) |
| 559 | .expect("deepseek-v4-pro should have a known context window"); |
| 560 | let v4_window: usize = 1_000_000; |
| 561 | let expected = v4_window - (TURN_MAX_OUTPUT_TOKENS as usize) - 1_024usize; |
| 562 | assert_eq!(budget, expected); |
| 563 | } |
| 564 | |
| 565 | #[test] |
| 566 | fn effective_max_output_tokens_caps_api_request_for_large_window_models() { |
| 567 | // V4 models have a 1M context window but the API request cap must stay |
| 568 | // well below common provider limits (e.g., 131K total on self-hosted |
| 569 | // vLLM/SGLang). The cap should never exceed 65K. |
| 570 | let v4_cap = effective_max_output_tokens("deepseek-v4-pro"); |
| 571 | assert!( |
| 572 | v4_cap <= 65_536, |
| 573 | "V4 API request cap should be ≤64K, got {v4_cap}" |
| 574 | ); |
| 575 | assert!( |
| 576 | v4_cap > 0, |
| 577 | "V4 API request cap should be positive, got {v4_cap}" |
| 578 | ); |
| 579 | |
| 580 | let flash_cap = effective_max_output_tokens("deepseek-v4-flash"); |
| 581 | assert_eq!(v4_cap, flash_cap); |
| 582 | } |
| 583 | |
| 584 | #[test] |
| 585 | fn internal_context_budget_unaffected_by_api_request_cap() { |
| 586 | // The internal context budget (used for compaction/preflight/recovery) |
| 587 | // must still use the full TURN_MAX_OUTPUT_TOKENS headroom, NOT the |
| 588 | // smaller API request cap. This ensures long-context V4 sessions don't |
| 589 | // compact prematurely. |
| 590 | let internal_budget = context_input_budget("deepseek-v4-pro", TURN_MAX_OUTPUT_TOKENS) |
| 591 | .expect("V4 should have a known context window"); |
| 592 | let api_cap_budget = context_input_budget( |
| 593 | "deepseek-v4-pro", |
| 594 | effective_max_output_tokens("deepseek-v4-pro"), |
| 595 | ) |
| 596 | .expect("V4 should have a known context window"); |
| 597 | |
| 598 | // Internal budget reserves 262K for output; API-cap budget would only |
| 599 | // reserve 64K. Internal budget must be smaller (more conservative). |
| 600 | assert!( |
| 601 | internal_budget < api_cap_budget, |
| 602 | "Internal budget ({internal_budget}) should be smaller than API-cap budget ({api_cap_budget}) \ |
| 603 | because it reserves more headroom for output" |
| 604 | ); |
| 605 | |
| 606 | // Verify the internal budget is what the compaction logic actually uses. |
| 607 | let v4_window: usize = 1_000_000; |
| 608 | let expected_internal = v4_window - (TURN_MAX_OUTPUT_TOKENS as usize) - 1_024usize; |
| 609 | assert_eq!(internal_budget, expected_internal); |
| 610 | } |
| 611 | |
| 612 | #[test] |
| 613 | fn v4_tool_outputs_keep_large_file_reads_in_context() { |
| 614 | let content = "0123456789abcdef\n".repeat(2_000); |
| 615 | let output = ToolResult::success(content.clone()); |
| 616 | |
| 617 | let v4_context = compact_tool_result_for_context("deepseek-v4-pro", "exec_shell", &output); |
| 618 | assert_eq!(v4_context, content.trim()); |
| 619 | |
| 620 | let legacy_context = |
| 621 | compact_tool_result_for_context("deepseek-v3.2-128k", "exec_shell", &output); |
| 622 | assert!(legacy_context.contains("output compacted to protect context")); |
| 623 | assert!(legacy_context.len() < v4_context.len()); |
| 624 | } |
| 625 | |
| 626 | #[test] |
| 627 | fn subagent_results_are_summarized_before_parent_context_insertion() { |
| 628 | let long_result = "verified detail\n".repeat(1_000); |
| 629 | let output = ToolResult::success( |
| 630 | json!({ |
| 631 | "agent_id": "agent_1234abcd", |
| 632 | "agent_type": "explore", |
| 633 | "assignment": { |
| 634 | "objective": "Inspect the RLM rendering path and report the smallest fix." |
| 635 | }, |
| 636 | "model": "deepseek-v4-flash", |
| 637 | "status": "Completed", |
| 638 | "result": long_result, |
| 639 | "steps_taken": 12, |
| 640 | "duration_ms": 3456 |
| 641 | }) |
| 642 | .to_string(), |
| 643 | ); |
| 644 | |
| 645 | let context = compact_tool_result_for_context("deepseek-v4-pro", "agent_result", &output); |
| 646 | |
| 647 | assert!(context.contains("[sub-agent result summarized for parent context]")); |
| 648 | assert!(context.contains("agent_1234abcd (explore) status=Completed")); |
| 649 | assert!(context.contains("Inspect the RLM rendering path")); |
| 650 | assert!(context.contains("steps=12")); |
| 651 | assert!(context.len() < output.content.len()); |
| 652 | } |
| 653 | |
| 654 | #[test] |
| 655 | fn refresh_system_prompt_leaves_working_set_out_of_system_prompt() { |
| 656 | let tmp = tempdir().expect("tempdir"); |
| 657 | fs::create_dir_all(tmp.path().join("src")).expect("mkdir"); |
| 658 | fs::write(tmp.path().join("src/lib.rs"), "pub fn sample() {}").expect("write"); |
| 659 | |
| 660 | let config = EngineConfig { |
| 661 | workspace: tmp.path().to_path_buf(), |
| 662 | ..Default::default() |
| 663 | }; |
| 664 | let (mut engine, _handle) = Engine::new(config, &Config::default()); |
| 665 | engine |
| 666 | .session |
| 667 | .working_set |
| 668 | .observe_user_message("please inspect src/lib.rs", tmp.path()); |
| 669 | |
| 670 | engine.refresh_system_prompt(AppMode::Agent); |
| 671 | |
| 672 | let prompt = match &engine.session.system_prompt { |
| 673 | Some(SystemPrompt::Text(text)) => text.clone(), |
| 674 | Some(SystemPrompt::Blocks(blocks)) => blocks |
| 675 | .iter() |
| 676 | .map(|block| block.text.as_str()) |
| 677 | .collect::<Vec<_>>() |
| 678 | .join("\n"), |
| 679 | None => panic!("expected system prompt"), |
| 680 | }; |
| 681 | assert!(!prompt.contains(WORKING_SET_SUMMARY_MARKER)); |
| 682 | } |
| 683 | |
| 684 | #[test] |
| 685 | fn working_set_reaches_model_as_turn_metadata() { |
| 686 | let tmp = tempdir().expect("tempdir"); |
| 687 | fs::create_dir_all(tmp.path().join("src")).expect("mkdir"); |
| 688 | fs::write(tmp.path().join("src/lib.rs"), "pub fn sample() {}").expect("write"); |
| 689 | |
| 690 | let config = EngineConfig { |
| 691 | workspace: tmp.path().to_path_buf(), |
| 692 | ..Default::default() |
| 693 | }; |
| 694 | let (mut engine, _handle) = Engine::new(config, &Config::default()); |
| 695 | engine |
| 696 | .session |
| 697 | .working_set |
| 698 | .observe_user_message("please inspect src/lib.rs", tmp.path()); |
| 699 | engine.session.add_message(Message { |
| 700 | role: "user".to_string(), |
| 701 | content: vec![ContentBlock::Text { |
| 702 | text: "please inspect src/lib.rs".to_string(), |
| 703 | cache_control: None, |
| 704 | }], |
| 705 | }); |
| 706 | |
| 707 | let messages = engine.messages_with_turn_metadata(); |
| 708 | let first_block = messages |
| 709 | .last() |
| 710 | .and_then(|message| message.content.first()) |
| 711 | .expect("turn metadata block"); |
| 712 | let ContentBlock::Text { text, .. } = first_block else { |
| 713 | panic!("expected text metadata block"); |
| 714 | }; |
| 715 | assert!(text.starts_with("<turn_meta>\n")); |
| 716 | assert!(text.contains(WORKING_SET_SUMMARY_MARKER)); |
| 717 | assert!(text.contains("src/lib.rs")); |
| 718 | } |
| 719 | |
| 720 | #[test] |
| 721 | fn turn_metadata_includes_current_local_date_without_working_set() { |
| 722 | let tmp = tempdir().expect("tempdir"); |
| 723 | let config = EngineConfig { |
| 724 | workspace: tmp.path().to_path_buf(), |
| 725 | ..Default::default() |
| 726 | }; |
| 727 | let (mut engine, _handle) = Engine::new(config, &Config::default()); |
| 728 | engine.session.add_message(Message { |
| 729 | role: "user".to_string(), |
| 730 | content: vec![ContentBlock::Text { |
| 731 | text: "what is today's date?".to_string(), |
| 732 | cache_control: None, |
| 733 | }], |
| 734 | }); |
| 735 | |
| 736 | let messages = engine.messages_with_turn_metadata(); |
| 737 | let first_block = messages |
| 738 | .last() |
| 739 | .and_then(|message| message.content.first()) |
| 740 | .expect("turn metadata block"); |
| 741 | let ContentBlock::Text { text, .. } = first_block else { |
| 742 | panic!("expected text metadata block"); |
| 743 | }; |
| 744 | |
| 745 | let today = chrono::Local::now().format("%Y-%m-%d").to_string(); |
| 746 | assert!(text.starts_with("<turn_meta>\n")); |
| 747 | assert!(text.contains(&format!("Current local date: {today}"))); |
| 748 | } |
| 749 | |
| 750 | /// v0.8.11 regression: tool-result messages serialize to role="tool" on |
| 751 | /// the wire but are stored as role="user" internally. Prepending |
| 752 | /// `<turn_meta>` text onto a tool-result message broke the |
| 753 | /// assistant→tool_result invariant and caused HTTP 400 from DeepSeek's |
| 754 | /// API ("insufficient tool messages following tool_calls"). The fix: |
| 755 | /// inject only into messages that have a Text content block and no |
| 756 | /// ToolResult blocks; mid-turn (tool-result is the trailing user |
| 757 | /// message) the injection skips. |
| 758 | #[test] |
| 759 | fn turn_metadata_skips_tool_result_messages() { |
| 760 | let tmp = tempdir().expect("tempdir"); |
| 761 | fs::create_dir_all(tmp.path().join("src")).expect("mkdir"); |
| 762 | fs::write(tmp.path().join("src/lib.rs"), "pub fn sample() {}").expect("write"); |
| 763 | |
| 764 | let config = EngineConfig { |
| 765 | workspace: tmp.path().to_path_buf(), |
| 766 | ..Default::default() |
| 767 | }; |
| 768 | let (mut engine, _handle) = Engine::new(config, &Config::default()); |
| 769 | engine |
| 770 | .session |
| 771 | .working_set |
| 772 | .observe_user_message("inspect src/lib.rs", tmp.path()); |
| 773 | |
| 774 | // Real user message — should be eligible for injection. |
| 775 | engine.session.add_message(Message { |
| 776 | role: "user".to_string(), |
| 777 | content: vec![ContentBlock::Text { |
| 778 | text: "inspect src/lib.rs".to_string(), |
| 779 | cache_control: None, |
| 780 | }], |
| 781 | }); |
| 782 | // Assistant tool-call. |
| 783 | engine.session.add_message(Message { |
| 784 | role: "assistant".to_string(), |
| 785 | content: vec![ContentBlock::ToolUse { |
| 786 | id: "call_42".to_string(), |
| 787 | name: "read_file".to_string(), |
| 788 | input: serde_json::json!({"path": "src/lib.rs"}), |
| 789 | caller: None, |
| 790 | }], |
| 791 | }); |
| 792 | // Tool result, stored as role="user" internally. |
| 793 | engine.session.add_message(Message { |
| 794 | role: "user".to_string(), |
| 795 | content: vec![ContentBlock::ToolResult { |
| 796 | tool_use_id: "call_42".to_string(), |
| 797 | content: "pub fn sample() {}".to_string(), |
| 798 | is_error: None, |
| 799 | content_blocks: None, |
| 800 | }], |
| 801 | }); |
| 802 | |
| 803 | let messages = engine.messages_with_turn_metadata(); |
| 804 | |
| 805 | // The trailing message is the tool result and MUST be untouched — |
| 806 | // no Text block sneaking in front of the ToolResult block. |
| 807 | let trailing = messages.last().expect("trailing message"); |
| 808 | assert_eq!(trailing.role, "user"); |
| 809 | assert_eq!(trailing.content.len(), 1); |
| 810 | assert!(matches!( |
| 811 | trailing.content.first(), |
| 812 | Some(ContentBlock::ToolResult { .. }) |
| 813 | )); |
| 814 | |
| 815 | // The earlier real user message receives the turn_meta prefix. |
| 816 | let real_user = messages.first().expect("first user message"); |
| 817 | assert_eq!(real_user.role, "user"); |
| 818 | let ContentBlock::Text { text, .. } = real_user.content.first().expect("user text content") |
| 819 | else { |
| 820 | panic!("expected Text block on real user message"); |
| 821 | }; |
| 822 | assert!(text.starts_with("<turn_meta>\n")); |
| 823 | assert!(text.contains("src/lib.rs")); |
| 824 | } |
| 825 | |
| 826 | /// When the turn is mid-execution and the trailing user message is a |
| 827 | /// tool result, no turn_meta is injected at all (rather than landing on |
| 828 | /// some earlier user message and confusing the API's tool-call |
| 829 | /// continuity check). The working_set surfaces again on the next |
| 830 | /// genuine user prompt. |
| 831 | #[test] |
| 832 | fn turn_metadata_skips_when_only_tool_results_trail() { |
| 833 | let tmp = tempdir().expect("tempdir"); |
| 834 | fs::create_dir_all(tmp.path().join("src")).expect("mkdir"); |
| 835 | fs::write(tmp.path().join("src/lib.rs"), "pub fn sample() {}").expect("write"); |
| 836 | |
| 837 | let config = EngineConfig { |
| 838 | workspace: tmp.path().to_path_buf(), |
| 839 | ..Default::default() |
| 840 | }; |
| 841 | let (mut engine, _handle) = Engine::new(config, &Config::default()); |
| 842 | engine |
| 843 | .session |
| 844 | .working_set |
| 845 | .observe_user_message("inspect src/lib.rs", tmp.path()); |
| 846 | |
| 847 | // Only a tool-result message in history — simulates the corner case |
| 848 | // where the prior real user message has already been compacted away |
| 849 | // but a tool-result is still pending. We must not retroactively |
| 850 | // inject. |
| 851 | engine.session.add_message(Message { |
| 852 | role: "user".to_string(), |
| 853 | content: vec![ContentBlock::ToolResult { |
| 854 | tool_use_id: "call_42".to_string(), |
| 855 | content: "pub fn sample() {}".to_string(), |
| 856 | is_error: None, |
| 857 | content_blocks: None, |
| 858 | }], |
| 859 | }); |
| 860 | |
| 861 | let messages = engine.messages_with_turn_metadata(); |
| 862 | |
| 863 | // Returned unchanged: the single tool-result message, no Text |
| 864 | // prefix, content length == 1. |
| 865 | let only = messages.last().expect("trailing message"); |
| 866 | assert_eq!(only.content.len(), 1); |
| 867 | assert!(matches!( |
| 868 | only.content.first(), |
| 869 | Some(ContentBlock::ToolResult { .. }) |
| 870 | )); |
| 871 | } |
| 872 | |
| 873 | #[test] |
| 874 | fn refresh_system_prompt_is_noop_when_unchanged() { |
| 875 | let tmp = tempdir().expect("tempdir"); |
| 876 | let config = EngineConfig { |
| 877 | workspace: tmp.path().to_path_buf(), |
| 878 | ..Default::default() |
| 879 | }; |
| 880 | let (mut engine, _handle) = Engine::new(config, &Config::default()); |
| 881 | |
| 882 | engine.refresh_system_prompt(AppMode::Agent); |
| 883 | let first_hash = engine.session.last_system_prompt_hash; |
| 884 | let first_prompt = engine.session.system_prompt.clone(); |
| 885 | engine.refresh_system_prompt(AppMode::Agent); |
| 886 | |
| 887 | assert_eq!(engine.session.last_system_prompt_hash, first_hash); |
| 888 | assert_eq!(engine.session.system_prompt, first_prompt); |
| 889 | } |
| 890 | |
| 891 | #[test] |
| 892 | fn compaction_summary_stays_in_stable_system_prompt() { |
| 893 | let tmp = tempdir().expect("tempdir"); |
| 894 | fs::create_dir_all(tmp.path().join("src")).expect("mkdir"); |
| 895 | fs::write(tmp.path().join("src/main.rs"), "fn main() {}").expect("write"); |
| 896 | |
| 897 | let config = EngineConfig { |
| 898 | workspace: tmp.path().to_path_buf(), |
| 899 | ..Default::default() |
| 900 | }; |
| 901 | let (mut engine, _handle) = Engine::new(config, &Config::default()); |
| 902 | engine |
| 903 | .session |
| 904 | .working_set |
| 905 | .observe_user_message("continue in src/main.rs", tmp.path()); |
| 906 | engine.refresh_system_prompt(AppMode::Agent); |
| 907 | engine.merge_compaction_summary(Some(SystemPrompt::Blocks(vec![SystemBlock { |
| 908 | block_type: "text".to_string(), |
| 909 | text: format!("{COMPACTION_SUMMARY_MARKER}\nsummary"), |
| 910 | cache_control: None, |
| 911 | }]))); |
| 912 | |
| 913 | let prompt = match &engine.session.system_prompt { |
| 914 | Some(SystemPrompt::Text(text)) => text.clone(), |
| 915 | Some(SystemPrompt::Blocks(blocks)) => blocks |
| 916 | .iter() |
| 917 | .map(|block| block.text.as_str()) |
| 918 | .collect::<Vec<_>>() |
| 919 | .join("\n"), |
| 920 | None => panic!("expected system prompt"), |
| 921 | }; |
| 922 | |
| 923 | assert!(prompt.contains(COMPACTION_SUMMARY_MARKER)); |
| 924 | assert!(!prompt.contains(WORKING_SET_SUMMARY_MARKER)); |
| 925 | } |
| 926 | |
| 927 | #[tokio::test] |
| 928 | async fn pre_request_refresh_skips_compaction_below_normal_threshold() { |
| 929 | let capacity = CapacityControllerConfig { |
| 930 | enabled: true, |
| 931 | low_risk_max: 0.0, |
| 932 | medium_risk_max: 1.0, |
| 933 | min_turns_before_guardrail: 0, |
| 934 | ..Default::default() |
| 935 | }; |
| 936 | |
| 937 | let mut engine = build_engine_with_capacity(capacity.clone()); |
| 938 | engine.config.capacity = capacity.clone(); |
| 939 | engine.capacity_controller = CapacityController::new(capacity); |
| 940 | engine.turn_counter = 5; |
| 941 | engine |
| 942 | .capacity_controller |
| 943 | .mark_turn_start(engine.turn_counter); |
| 944 | engine.session.model = "deepseek-v4-pro".to_string(); |
| 945 | engine.config.model = "deepseek-v4-pro".to_string(); |
| 946 | |
| 947 | for i in 0..20 { |
| 948 | engine.session.messages.push(Message { |
| 949 | role: "user".to_string(), |
| 950 | content: vec![ContentBlock::Text { |
| 951 | text: format!("small message {i}"), |
| 952 | cache_control: None, |
| 953 | }], |
| 954 | }); |
| 955 | } |
| 956 | |
| 957 | let before = engine.estimated_input_tokens(); |
| 958 | let before_len = engine.session.messages.len(); |
| 959 | let turn = TurnContext::new(10); |
| 960 | let applied = engine |
| 961 | .run_capacity_pre_request_checkpoint(&turn, None, AppMode::Agent) |
| 962 | .await; |
| 963 | let after = engine.estimated_input_tokens(); |
| 964 | |
| 965 | assert!(!applied); |
| 966 | assert_eq!(after, before); |
| 967 | assert_eq!(engine.session.messages.len(), before_len); |
| 968 | } |
| 969 | |
| 970 | #[tokio::test] |
| 971 | async fn pre_request_refresh_invoked_when_medium_risk() { |
| 972 | let capacity = CapacityControllerConfig { |
| 973 | enabled: true, |
| 974 | low_risk_max: 0.0, |
| 975 | medium_risk_max: 1.0, |
| 976 | min_turns_before_guardrail: 0, |
| 977 | ..Default::default() |
| 978 | }; |
| 979 | |
| 980 | let mut engine = build_engine_with_capacity(capacity.clone()); |
| 981 | engine.config.capacity = capacity.clone(); |
| 982 | engine.capacity_controller = CapacityController::new(capacity); |
| 983 | engine.turn_counter = 5; |
| 984 | engine |
| 985 | .capacity_controller |
| 986 | .mark_turn_start(engine.turn_counter); |
| 987 | |
| 988 | // Pin the model to an explicit 128k-context variant so the pressure ratio stays |
| 989 | // stable regardless of changes to the workspace-wide default model. |
| 990 | engine.session.model = "deepseek-v3.2-128k".to_string(); |
| 991 | engine.config.model = "deepseek-v3.2-128k".to_string(); |
| 992 | |
| 993 | let long = "x".repeat(5_000); |
| 994 | for _ in 0..900 { |
| 995 | engine.session.messages.push(Message { |
| 996 | role: "user".to_string(), |
| 997 | content: vec![ContentBlock::Text { |
| 998 | text: long.clone(), |
| 999 | cache_control: None, |
| 1000 | }], |
| 1001 | }); |
| 1002 | } |
| 1003 | |
| 1004 | let before = engine.estimated_input_tokens(); |
| 1005 | let turn = TurnContext::new(10); |
| 1006 | let applied = engine |
| 1007 | .run_capacity_pre_request_checkpoint(&turn, None, AppMode::Agent) |
| 1008 | .await; |
| 1009 | let after = engine.estimated_input_tokens(); |
| 1010 | |
| 1011 | assert!(applied); |
| 1012 | assert!(after < before); |
| 1013 | } |
| 1014 | |
| 1015 | #[tokio::test] |
| 1016 | async fn post_tool_replay_invoked_when_high_non_severe_risk() { |
| 1017 | let tmp = tempdir().expect("tempdir"); |
| 1018 | fs::write(tmp.path().join("sample.txt"), "hello replay").expect("write"); |
| 1019 | |
| 1020 | let capacity = CapacityControllerConfig { |
| 1021 | enabled: true, |
| 1022 | low_risk_max: 0.0, |
| 1023 | medium_risk_max: 0.0, |
| 1024 | severe_min_slack: -10.0, |
| 1025 | severe_violation_ratio: 2.0, |
| 1026 | min_turns_before_guardrail: 0, |
| 1027 | ..Default::default() |
| 1028 | }; |
| 1029 | |
| 1030 | let mut engine = build_engine_with_capacity(capacity.clone()); |
| 1031 | engine.session.workspace = tmp.path().to_path_buf(); |
| 1032 | engine.config.workspace = tmp.path().to_path_buf(); |
| 1033 | engine.config.capacity = capacity.clone(); |
| 1034 | engine.capacity_controller = CapacityController::new(capacity); |
| 1035 | engine.turn_counter = 4; |
| 1036 | engine |
| 1037 | .capacity_controller |
| 1038 | .mark_turn_start(engine.turn_counter); |
| 1039 | |
| 1040 | let mut turn = TurnContext::new(10); |
| 1041 | let mut tool_call = TurnToolCall::new( |
| 1042 | "tool_read_1".to_string(), |
| 1043 | "read_file".to_string(), |
| 1044 | json!({ "path": "sample.txt" }), |
| 1045 | ); |
| 1046 | tool_call.set_result( |
| 1047 | "hello replay".to_string(), |
| 1048 | std::time::Duration::from_millis(1), |
| 1049 | ); |
| 1050 | turn.record_tool_call(tool_call); |
| 1051 | |
| 1052 | let registry = ToolRegistryBuilder::new() |
| 1053 | .with_read_only_file_tools() |
| 1054 | .build(engine.build_tool_context(AppMode::Agent, false)); |
| 1055 | |
| 1056 | let restarted = engine |
| 1057 | .run_capacity_post_tool_checkpoint( |
| 1058 | &turn, |
| 1059 | AppMode::Agent, |
| 1060 | Some(®istry), |
| 1061 | Arc::new(RwLock::new(())), |
| 1062 | None, |
| 1063 | 0, |
| 1064 | 0, |
| 1065 | ) |
| 1066 | .await; |
| 1067 | |
| 1068 | assert!(!restarted); |
| 1069 | let has_verification_note = engine.session.messages.iter().any(|msg| { |
| 1070 | msg.content.iter().any(|block| match block { |
| 1071 | ContentBlock::ToolResult { content, .. } => content.contains("[verification replay]"), |
| 1072 | _ => false, |
| 1073 | }) |
| 1074 | }); |
| 1075 | assert!(has_verification_note); |
| 1076 | } |
| 1077 | |
| 1078 | #[tokio::test] |
| 1079 | async fn error_escalation_triggers_replan_when_severe_or_repeated_failures() { |
| 1080 | let _env_lock = CAPACITY_MEMORY_ENV_LOCK.lock().await; |
| 1081 | let tmp = tempdir().expect("tempdir"); |
| 1082 | let _env = ScopedCapacityMemoryDir::set(tmp.path()); |
| 1083 | |
| 1084 | let capacity = CapacityControllerConfig { |
| 1085 | enabled: true, |
| 1086 | low_risk_max: 0.0, |
| 1087 | medium_risk_max: 0.0, |
| 1088 | min_turns_before_guardrail: 0, |
| 1089 | ..Default::default() |
| 1090 | }; |
| 1091 | |
| 1092 | let mut engine = build_engine_with_capacity(capacity.clone()); |
| 1093 | engine.config.capacity = capacity.clone(); |
| 1094 | engine.capacity_controller = CapacityController::new(capacity); |
| 1095 | engine.turn_counter = 6; |
| 1096 | engine |
| 1097 | .capacity_controller |
| 1098 | .mark_turn_start(engine.turn_counter); |
| 1099 | |
| 1100 | for i in 0..10 { |
| 1101 | engine.session.messages.push(Message { |
| 1102 | role: if i % 2 == 0 { "user" } else { "assistant" }.to_string(), |
| 1103 | content: vec![ContentBlock::Text { |
| 1104 | text: format!("noise message {i}"), |
| 1105 | cache_control: None, |
| 1106 | }], |
| 1107 | }); |
| 1108 | } |
| 1109 | engine.session.messages.push(Message { |
| 1110 | role: "user".to_string(), |
| 1111 | content: vec![ContentBlock::Text { |
| 1112 | text: "Please finish task".to_string(), |
| 1113 | cache_control: None, |
| 1114 | }], |
| 1115 | }); |
| 1116 | |
| 1117 | let before_len = engine.session.messages.len(); |
| 1118 | let turn = TurnContext::new(10); |
| 1119 | let restarted = engine |
| 1120 | .run_capacity_error_escalation_checkpoint(&turn, AppMode::Agent, 2, 2, &[]) |
| 1121 | .await; |
| 1122 | |
| 1123 | assert!(restarted); |
| 1124 | assert!(engine.session.messages.len() < before_len); |
| 1125 | assert!(engine.session.messages.len() <= 2); |
| 1126 | |
| 1127 | let records = load_last_k_capacity_records(&engine.session.id, 1).expect("load memory"); |
| 1128 | assert!(!records.is_empty()); |
| 1129 | assert!(!records[0].canonical_state.goal.is_empty()); |
| 1130 | } |
| 1131 | |
| 1132 | /// v0.8.11: `CapacityControllerConfig::default()` ships with |
| 1133 | /// `enabled = false`. The capacity controller's destructive |
| 1134 | /// interventions (TargetedContextRefresh silently runs compaction; |
| 1135 | /// VerifyAndReplan clears the session message log) silently rewrote |
| 1136 | /// or nuked the user's transcript ("resetting plan" footer + |
| 1137 | /// black-screen symptom). v0.8.11 commits to "trust the model with |
| 1138 | /// the full 1M-token context, only compact on explicit user |
| 1139 | /// /compact" — auto-managing the prefix contradicts that posture. |
| 1140 | /// Power users can still opt in via `capacity.enabled = true`. |
| 1141 | #[tokio::test] |
| 1142 | async fn capacity_disabled_by_default_keeps_messages_intact() { |
| 1143 | let _env_lock = CAPACITY_MEMORY_ENV_LOCK.lock().await; |
| 1144 | let tmp = tempdir().expect("tempdir"); |
| 1145 | let _env = ScopedCapacityMemoryDir::set(tmp.path()); |
| 1146 | |
| 1147 | // Default config — what real users get. |
| 1148 | let mut engine = build_engine_with_capacity(CapacityControllerConfig::default()); |
| 1149 | assert!( |
| 1150 | !engine.config.capacity.enabled, |
| 1151 | "capacity controller must be off by default in v0.8.11+" |
| 1152 | ); |
| 1153 | engine.turn_counter = 6; |
| 1154 | engine |
| 1155 | .capacity_controller |
| 1156 | .mark_turn_start(engine.turn_counter); |
| 1157 | |
| 1158 | for i in 0..10 { |
| 1159 | engine.session.messages.push(Message { |
| 1160 | role: if i % 2 == 0 { "user" } else { "assistant" }.to_string(), |
| 1161 | content: vec![ContentBlock::Text { |
| 1162 | text: format!("noise message {i}"), |
| 1163 | cache_control: None, |
| 1164 | }], |
| 1165 | }); |
| 1166 | } |
| 1167 | engine.session.messages.push(Message { |
| 1168 | role: "user".to_string(), |
| 1169 | content: vec![ContentBlock::Text { |
| 1170 | text: "Please finish task".to_string(), |
| 1171 | cache_control: None, |
| 1172 | }], |
| 1173 | }); |
| 1174 | |
| 1175 | let before_len = engine.session.messages.len(); |
| 1176 | let turn = TurnContext::new(10); |
| 1177 | let restarted = engine |
| 1178 | .run_capacity_error_escalation_checkpoint(&turn, AppMode::Agent, 2, 2, &[]) |
| 1179 | .await; |
| 1180 | |
| 1181 | // Capacity is disabled → no replan, no message clear. |
| 1182 | assert!(!restarted); |
| 1183 | assert_eq!(engine.session.messages.len(), before_len); |
| 1184 | } |
| 1185 | |
| 1186 | #[tokio::test] |
| 1187 | async fn controller_disabled_keeps_behavior_unchanged() { |
| 1188 | let capacity = CapacityControllerConfig { |
| 1189 | enabled: false, |
| 1190 | ..Default::default() |
| 1191 | }; |
| 1192 | |
| 1193 | let mut engine = build_engine_with_capacity(capacity.clone()); |
| 1194 | engine.config.capacity = capacity.clone(); |
| 1195 | engine.capacity_controller = CapacityController::new(capacity); |
| 1196 | engine.turn_counter = 3; |
| 1197 | engine |
| 1198 | .capacity_controller |
| 1199 | .mark_turn_start(engine.turn_counter); |
| 1200 | |
| 1201 | let long = "y".repeat(5_000); |
| 1202 | for _ in 0..120 { |
| 1203 | engine.session.messages.push(Message { |
| 1204 | role: "user".to_string(), |
| 1205 | content: vec![ContentBlock::Text { |
| 1206 | text: long.clone(), |
| 1207 | cache_control: None, |
| 1208 | }], |
| 1209 | }); |
| 1210 | } |
| 1211 | |
| 1212 | let before = engine.estimated_input_tokens(); |
| 1213 | let before_len = engine.session.messages.len(); |
| 1214 | let turn = TurnContext::new(10); |
| 1215 | let applied = engine |
| 1216 | .run_capacity_pre_request_checkpoint(&turn, None, AppMode::Agent) |
| 1217 | .await; |
| 1218 | let after = engine.estimated_input_tokens(); |
| 1219 | let after_len = engine.session.messages.len(); |
| 1220 | |
| 1221 | assert!(!applied); |
| 1222 | assert_eq!(before, after); |
| 1223 | assert_eq!(before_len, after_len); |
| 1224 | } |
| 1225 | |
| 1226 | #[test] |
| 1227 | fn caller_policy_defaults_to_direct() { |
| 1228 | let tool = Tool { |
| 1229 | tool_type: None, |
| 1230 | name: "read_file".to_string(), |
| 1231 | description: "Read".to_string(), |
| 1232 | input_schema: json!({"type":"object"}), |
| 1233 | allowed_callers: Some(vec!["direct".to_string()]), |
| 1234 | defer_loading: Some(false), |
| 1235 | input_examples: None, |
| 1236 | strict: None, |
| 1237 | cache_control: None, |
| 1238 | }; |
| 1239 | let direct = ToolCaller { |
| 1240 | caller_type: "direct".to_string(), |
| 1241 | tool_id: None, |
| 1242 | }; |
| 1243 | let code = ToolCaller { |
| 1244 | caller_type: "code_execution_20250825".to_string(), |
| 1245 | tool_id: Some("srvtoolu_1".to_string()), |
| 1246 | }; |
| 1247 | assert!(caller_allowed_for_tool(Some(&direct), Some(&tool))); |
| 1248 | assert!(!caller_allowed_for_tool(Some(&code), Some(&tool))); |
| 1249 | assert!(caller_allowed_for_tool(None, Some(&tool))); |
| 1250 | } |
| 1251 | |
| 1252 | #[test] |
| 1253 | fn tool_search_activates_discovered_deferred_tools() { |
| 1254 | let mut catalog = vec![ |
| 1255 | Tool { |
| 1256 | tool_type: None, |
| 1257 | name: "read_file".to_string(), |
| 1258 | description: "Read files".to_string(), |
| 1259 | input_schema: json!({"type":"object","properties":{"path":{"type":"string"}}}), |
| 1260 | allowed_callers: Some(vec!["direct".to_string()]), |
| 1261 | defer_loading: Some(true), |
| 1262 | input_examples: None, |
| 1263 | strict: None, |
| 1264 | cache_control: None, |
| 1265 | }, |
| 1266 | Tool { |
| 1267 | tool_type: None, |
| 1268 | name: "grep_files".to_string(), |
| 1269 | description: "Search files".to_string(), |
| 1270 | input_schema: json!({"type":"object","properties":{"pattern":{"type":"string"}}}), |
| 1271 | allowed_callers: Some(vec!["direct".to_string()]), |
| 1272 | defer_loading: Some(true), |
| 1273 | input_examples: None, |
| 1274 | strict: None, |
| 1275 | cache_control: None, |
| 1276 | }, |
| 1277 | ]; |
| 1278 | ensure_advanced_tooling(&mut catalog); |
| 1279 | let mut active = initial_active_tools(&catalog); |
| 1280 | let result = execute_tool_search( |
| 1281 | TOOL_SEARCH_BM25_NAME, |
| 1282 | &json!({"query":"read file"}), |
| 1283 | &catalog, |
| 1284 | &mut active, |
| 1285 | ) |
| 1286 | .expect("search succeeds"); |
| 1287 | assert!(result.success); |
| 1288 | assert!(active.contains("read_file")); |
| 1289 | } |
| 1290 | |
| 1291 | #[tokio::test] |
| 1292 | async fn code_execution_runs_python_and_returns_result_payload() { |
| 1293 | let tmp = tempdir().expect("tempdir"); |
| 1294 | let result = |
| 1295 | execute_code_execution_tool(&json!({"code":"print('hello from code exec')"}), tmp.path()) |
| 1296 | .await |
| 1297 | .expect("code execution should run"); |
| 1298 | assert!(result.content.contains("hello from code exec")); |
| 1299 | assert!(result.content.contains("return_code")); |
| 1300 | } |
| 1301 | |
| 1302 | #[test] |
| 1303 | fn deferred_tool_requests_are_auto_activated() { |
| 1304 | use std::collections::HashSet; |
| 1305 | |
| 1306 | let catalog = vec![Tool { |
| 1307 | tool_type: None, |
| 1308 | name: "exec_shell".to_string(), |
| 1309 | description: "Run shell commands".to_string(), |
| 1310 | input_schema: json!({"type":"object","properties":{"cmd":{"type":"string"}}}), |
| 1311 | allowed_callers: Some(vec!["direct".to_string()]), |
| 1312 | defer_loading: Some(true), |
| 1313 | input_examples: None, |
| 1314 | strict: None, |
| 1315 | cache_control: None, |
| 1316 | }]; |
| 1317 | |
| 1318 | let mut active = HashSet::new(); |
| 1319 | assert!(!active.contains("exec_shell")); |
| 1320 | assert!(maybe_activate_requested_deferred_tool( |
| 1321 | "exec_shell", |
| 1322 | &catalog, |
| 1323 | &mut active |
| 1324 | )); |
| 1325 | assert!(active.contains("exec_shell")); |
| 1326 | } |
| 1327 | |
| 1328 | #[test] |
| 1329 | fn missing_tool_error_message_offers_suggestions() { |
| 1330 | let catalog = vec![ |
| 1331 | Tool { |
| 1332 | tool_type: None, |
| 1333 | name: "read_file".to_string(), |
| 1334 | description: "Read file contents".to_string(), |
| 1335 | input_schema: json!({"type":"object","properties":{"path":{"type":"string"}}}), |
| 1336 | allowed_callers: Some(vec!["direct".to_string()]), |
| 1337 | defer_loading: Some(false), |
| 1338 | input_examples: None, |
| 1339 | strict: None, |
| 1340 | cache_control: None, |
| 1341 | }, |
| 1342 | Tool { |
| 1343 | tool_type: None, |
| 1344 | name: "grep_files".to_string(), |
| 1345 | description: "Search file contents".to_string(), |
| 1346 | input_schema: json!({"type":"object","properties":{"pattern":{"type":"string"}}}), |
| 1347 | allowed_callers: Some(vec!["direct".to_string()]), |
| 1348 | defer_loading: Some(false), |
| 1349 | input_examples: None, |
| 1350 | strict: None, |
| 1351 | cache_control: None, |
| 1352 | }, |
| 1353 | ]; |
| 1354 | |
| 1355 | let message = missing_tool_error_message("reed_file", &catalog); |
| 1356 | assert!(message.contains("Did you mean:")); |
| 1357 | assert!(message.contains("read_file")); |
| 1358 | assert!(message.contains(TOOL_SEARCH_BM25_NAME)); |
| 1359 | } |
| 1360 | |
| 1361 | #[test] |
| 1362 | fn missing_tool_error_message_includes_discovery_guidance_when_no_match() { |
| 1363 | let catalog = vec![Tool { |
| 1364 | tool_type: None, |
| 1365 | name: "read_file".to_string(), |
| 1366 | description: "Read file contents".to_string(), |
| 1367 | input_schema: json!({"type":"object","properties":{"path":{"type":"string"}}}), |
| 1368 | allowed_callers: Some(vec!["direct".to_string()]), |
| 1369 | defer_loading: Some(false), |
| 1370 | input_examples: None, |
| 1371 | strict: None, |
| 1372 | cache_control: None, |
| 1373 | }]; |
| 1374 | |
| 1375 | let message = missing_tool_error_message("totally_unknown_tool", &catalog); |
| 1376 | assert!(message.contains("not available in the current tool catalog")); |
| 1377 | assert!(message.contains(TOOL_SEARCH_BM25_NAME)); |
| 1378 | } |
| 1379 | |
| 1380 | #[test] |
| 1381 | fn filter_tool_call_delta_strips_bracket_marker() { |
| 1382 | let mut in_block = false; |
| 1383 | let visible = filter_tool_call_delta( |
| 1384 | "intro [TOOL_CALL]\n{\"tool\":\"x\"}\n[/TOOL_CALL] outro", |
| 1385 | &mut in_block, |
| 1386 | ); |
| 1387 | assert!(!in_block); |
| 1388 | assert!(!visible.contains("[TOOL_CALL]")); |
| 1389 | assert!(!visible.contains("[/TOOL_CALL]")); |
| 1390 | assert!(!visible.contains("\"tool\":\"x\"")); |
| 1391 | assert!(visible.contains("intro")); |
| 1392 | assert!(visible.contains("outro")); |
| 1393 | } |
| 1394 | |
| 1395 | #[test] |
| 1396 | fn filter_tool_call_delta_strips_deepseek_xml_marker() { |
| 1397 | let mut in_block = false; |
| 1398 | let visible = filter_tool_call_delta( |
| 1399 | "before <deepseek:tool_call name=\"x\">payload</deepseek:tool_call> after", |
| 1400 | &mut in_block, |
| 1401 | ); |
| 1402 | assert!(!in_block); |
| 1403 | for marker in TOOL_CALL_START_MARKERS { |
| 1404 | assert!( |
| 1405 | !visible.contains(marker), |
| 1406 | "visible text leaked start marker `{marker}`: {visible:?}" |
| 1407 | ); |
| 1408 | } |
| 1409 | assert!(visible.contains("before")); |
| 1410 | assert!(visible.contains("after")); |
| 1411 | } |
| 1412 | |
| 1413 | #[test] |
| 1414 | fn filter_tool_call_delta_strips_generic_tool_call_marker() { |
| 1415 | let mut in_block = false; |
| 1416 | let visible = filter_tool_call_delta( |
| 1417 | "lead <tool_call>\n{\"name\":\"do\"}\n</tool_call> tail", |
| 1418 | &mut in_block, |
| 1419 | ); |
| 1420 | assert!(!in_block); |
| 1421 | assert!(!visible.contains("<tool_call")); |
| 1422 | assert!(!visible.contains("</tool_call>")); |
| 1423 | assert!(visible.contains("lead")); |
| 1424 | assert!(visible.contains("tail")); |
| 1425 | } |
| 1426 | |
| 1427 | #[test] |
| 1428 | fn filter_tool_call_delta_strips_invoke_marker() { |
| 1429 | let mut in_block = false; |
| 1430 | let visible = filter_tool_call_delta( |
| 1431 | "alpha <invoke name=\"x\"><parameter name=\"k\">v</parameter></invoke> beta", |
| 1432 | &mut in_block, |
| 1433 | ); |
| 1434 | assert!(!in_block); |
| 1435 | assert!(!visible.contains("<invoke ")); |
| 1436 | assert!(!visible.contains("</invoke>")); |
| 1437 | assert!(visible.contains("alpha")); |
| 1438 | assert!(visible.contains("beta")); |
| 1439 | } |
| 1440 | |
| 1441 | #[test] |
| 1442 | fn filter_tool_call_delta_strips_function_calls_marker() { |
| 1443 | let mut in_block = false; |
| 1444 | let visible = filter_tool_call_delta( |
| 1445 | "head <function_calls>\n{\"name\":\"x\"}\n</function_calls> tail", |
| 1446 | &mut in_block, |
| 1447 | ); |
| 1448 | assert!(!in_block); |
| 1449 | assert!(!visible.contains("<function_calls>")); |
| 1450 | assert!(!visible.contains("</function_calls>")); |
| 1451 | assert!(visible.contains("head")); |
| 1452 | assert!(visible.contains("tail")); |
| 1453 | } |
| 1454 | |
| 1455 | #[test] |
| 1456 | fn filter_tool_call_delta_handles_chunk_split_marker() { |
| 1457 | let mut in_block = false; |
| 1458 | // First chunk opens the wrapper but does not close it. |
| 1459 | let visible_a = filter_tool_call_delta("hello <tool_call>partial", &mut in_block); |
| 1460 | assert!(in_block, "filter must remember it is mid-wrapper"); |
| 1461 | assert_eq!(visible_a, "hello "); |
| 1462 | |
| 1463 | // Second chunk continues inside the wrapper, then closes it and adds tail. |
| 1464 | let visible_b = filter_tool_call_delta("payload</tool_call> tail", &mut in_block); |
| 1465 | assert!(!in_block); |
| 1466 | assert_eq!(visible_b, " tail"); |
| 1467 | } |
| 1468 | |
| 1469 | #[test] |
| 1470 | fn filter_tool_call_delta_unmatched_open_suppresses_remainder() { |
| 1471 | let mut in_block = false; |
| 1472 | let visible = filter_tool_call_delta("ok [TOOL_CALL]rest of stream", &mut in_block); |
| 1473 | assert_eq!(visible, "ok "); |
| 1474 | assert!( |
| 1475 | in_block, |
| 1476 | "unmatched open must leave filter in tool-call mode" |
| 1477 | ); |
| 1478 | } |
| 1479 | |
| 1480 | #[test] |
| 1481 | fn filter_tool_call_delta_passes_through_clean_text() { |
| 1482 | let mut in_block = false; |
| 1483 | let input = "no markers here, just prose with code `<not a tag>`."; |
| 1484 | let visible = filter_tool_call_delta(input, &mut in_block); |
| 1485 | assert!(!in_block); |
| 1486 | assert_eq!(visible, input); |
| 1487 | } |
| 1488 | |
| 1489 | #[test] |
| 1490 | fn contains_fake_tool_wrapper_detects_each_marker() { |
| 1491 | for marker in TOOL_CALL_START_MARKERS { |
| 1492 | let needle = format!("noise {marker} more noise"); |
| 1493 | assert!( |
| 1494 | contains_fake_tool_wrapper(&needle), |
| 1495 | "marker `{marker}` should be detected" |
| 1496 | ); |
| 1497 | } |
| 1498 | } |
| 1499 | |
| 1500 | #[test] |
| 1501 | fn contains_fake_tool_wrapper_returns_false_on_clean_text() { |
| 1502 | assert!(!contains_fake_tool_wrapper( |
| 1503 | "plain assistant text without wrappers" |
| 1504 | )); |
| 1505 | assert!(!contains_fake_tool_wrapper( |
| 1506 | "`<tool` lookalike but not a real start marker" |
| 1507 | )); |
| 1508 | } |
| 1509 | |
| 1510 | #[test] |
| 1511 | fn fake_wrapper_notice_is_compact_and_actionable() { |
| 1512 | // Keep this short so it fits cleanly in a single status line. |
| 1513 | assert!(FAKE_WRAPPER_NOTICE.len() < 120); |
| 1514 | assert!(FAKE_WRAPPER_NOTICE.contains("API tool channel")); |
| 1515 | } |
| 1516 | |
| 1517 | // ---- final_tool_input: bug-class regression for "<command>" placeholder ---- |
| 1518 | // |
| 1519 | // Background: a streamed tool block carries its `input` in two pieces — an |
| 1520 | // initial value at `ContentBlockStart` (often `{}`), then `InputJsonDelta` |
| 1521 | // chunks that build up `input_buffer`. The TUI used to fire `ToolCallStarted` |
| 1522 | // from `ContentBlockStart` with the empty initial input and never re-emit |
| 1523 | // once args were known, so cells rendered the literal text `<command>` / |
| 1524 | // `<file>` placeholders. The fix relocates the emission to `ContentBlockStop` |
| 1525 | // and routes the input through `final_tool_input`, which prefers the parsed |
| 1526 | // buffer over a stale empty placeholder. |
| 1527 | fn tool_state(initial: serde_json::Value, buffer: &str) -> ToolUseState { |
| 1528 | ToolUseState { |
| 1529 | id: "t1".into(), |
| 1530 | name: "exec_shell".into(), |
| 1531 | input: initial, |
| 1532 | caller: None, |
| 1533 | input_buffer: buffer.into(), |
| 1534 | } |
| 1535 | } |
| 1536 | |
| 1537 | #[test] |
| 1538 | fn final_tool_input_prefers_parsed_buffer_over_empty_initial() { |
| 1539 | // The exact regression: ContentBlockStart delivered `{}`, then args |
| 1540 | // streamed in via InputJsonDelta. The emitted ToolCallStarted must |
| 1541 | // carry the parsed buffer, not the placeholder. |
| 1542 | let state = tool_state(json!({}), r#"{"command": "ls -la"}"#); |
| 1543 | assert_eq!(final_tool_input(&state), json!({"command": "ls -la"})); |
| 1544 | } |
| 1545 | |
| 1546 | #[test] |
| 1547 | fn final_tool_input_falls_back_to_initial_when_buffer_empty() { |
| 1548 | // Models occasionally embed args directly in the start frame and never |
| 1549 | // send any InputJsonDelta. We must still report those args. |
| 1550 | let state = tool_state(json!({"command": "echo hi"}), ""); |
| 1551 | assert_eq!(final_tool_input(&state), json!({"command": "echo hi"})); |
| 1552 | } |
| 1553 | |
| 1554 | #[test] |
| 1555 | fn final_tool_input_repairs_unparseable_buffer() { |
| 1556 | // The arg_repair module converts unparseable input to an empty object |
| 1557 | // {} so dispatch always proceeds. The buffer wins over the initial input. |
| 1558 | let state = tool_state(json!({"command": "echo hi"}), "{not json"); |
| 1559 | assert_eq!(final_tool_input(&state), json!({})); |
| 1560 | } |
| 1561 | |
| 1562 | // === #103 transparent stream-retry policy ===================================== |
| 1563 | |
| 1564 | #[test] |
| 1565 | fn stream_retry_zero_content_then_error_is_transparently_retried() { |
| 1566 | // Case 2 from issue #103: stream yielded ZERO content then errored. |
| 1567 | // The decoder hit Err on the very first poll → engine should retry |
| 1568 | // because DeepSeek hasn't billed and the user has seen nothing. |
| 1569 | assert!( |
| 1570 | super::should_transparently_retry_stream(false, 0, false), |
| 1571 | "first attempt with no content must be eligible for transparent retry" |
| 1572 | ); |
| 1573 | assert!( |
| 1574 | super::should_transparently_retry_stream(false, 1, false), |
| 1575 | "second attempt (one prior retry) with no content must still be eligible" |
| 1576 | ); |
| 1577 | } |
| 1578 | |
| 1579 | #[test] |
| 1580 | fn stream_retry_after_content_received_surfaces_error() { |
| 1581 | // Case 3 from issue #103: stream yielded content then errored. We must |
| 1582 | // NOT transparently retry — the model has emitted billed output tokens |
| 1583 | // and the UI has streamed deltas; resending would double-bill and the |
| 1584 | // user would see the same prefix twice. |
| 1585 | assert!( |
| 1586 | !super::should_transparently_retry_stream(true, 0, false), |
| 1587 | "any content received → no transparent retry, even with full budget" |
| 1588 | ); |
| 1589 | assert!( |
| 1590 | !super::should_transparently_retry_stream(true, 1, false), |
| 1591 | "any content received → no transparent retry on subsequent attempts" |
| 1592 | ); |
| 1593 | } |
| 1594 | |
| 1595 | #[test] |
| 1596 | fn stream_retry_budget_caps_transparent_retries_at_two() { |
| 1597 | // Case 4 from issue #103: after MAX_TRANSPARENT_STREAM_RETRIES attempts |
| 1598 | // we stop trying transparently and let the outer error path surface. |
| 1599 | // (The outer per-turn `stream_retry_attempts` retry is a separate layer |
| 1600 | // and is still in effect at the whole-turn level.) |
| 1601 | assert!( |
| 1602 | super::should_transparently_retry_stream( |
| 1603 | false, |
| 1604 | super::MAX_TRANSPARENT_STREAM_RETRIES - 1, |
| 1605 | false, |
| 1606 | ), |
| 1607 | "one short of the cap should still retry" |
| 1608 | ); |
| 1609 | assert!( |
| 1610 | !super::should_transparently_retry_stream( |
| 1611 | false, |
| 1612 | super::MAX_TRANSPARENT_STREAM_RETRIES, |
| 1613 | false, |
| 1614 | ), |
| 1615 | "at the cap, no further transparent retries" |
| 1616 | ); |
| 1617 | assert!( |
| 1618 | !super::should_transparently_retry_stream( |
| 1619 | false, |
| 1620 | super::MAX_TRANSPARENT_STREAM_RETRIES + 5, |
| 1621 | false, |
| 1622 | ), |
| 1623 | "well past the cap, definitely no transparent retries" |
| 1624 | ); |
| 1625 | } |
| 1626 | |
| 1627 | #[test] |
| 1628 | fn stream_retry_respects_cancellation() { |
| 1629 | // Cancellation overrides every other condition. If the user pressed |
| 1630 | // Esc / Ctrl-C, do not silently re-issue the request behind their back. |
| 1631 | assert!( |
| 1632 | !super::should_transparently_retry_stream(false, 0, true), |
| 1633 | "cancelled turn must not be transparently retried" |
| 1634 | ); |
| 1635 | assert!( |
| 1636 | !super::should_transparently_retry_stream(false, 1, true), |
| 1637 | "cancelled turn must not be transparently retried even with budget" |
| 1638 | ); |
| 1639 | } |
| 1640 | |
| 1641 | #[test] |
| 1642 | fn stream_retry_threshold_relaxed_to_five() { |
| 1643 | // Case 1+4 from issue #103: the consecutive-error threshold for marking |
| 1644 | // the turn failed was relaxed from 3 → 5 in v0.6.7 because the new |
| 1645 | // HTTP/2 keepalive defaults make spurious decode errors rarer. |
| 1646 | // This test pins the constant so a future regression to 3 fails loudly. |
| 1647 | assert_eq!( |
| 1648 | super::MAX_STREAM_ERRORS_BEFORE_FAIL, |
| 1649 | 5, |
| 1650 | "the consecutive-stream-error threshold should be 5; \ |
| 1651 | lowering it back to 3 will fail mid-turn under transient flakiness" |
| 1652 | ); |
| 1653 | // And a regression guard on the transparent-retry cap. |
| 1654 | assert_eq!( |
| 1655 | super::MAX_TRANSPARENT_STREAM_RETRIES, |
| 1656 | 2, |
| 1657 | "transparent-retry cap should be 2; raising it risks hammering the \ |
| 1658 | provider on real outages" |
| 1659 | ); |
| 1660 | } |
| 1661 | |
| 1662 | // === Issue #66: error taxonomy wired through engine + audit + capacity === |
| 1663 | |
| 1664 | /// A failed-tool audit entry must carry the typed `category` and `severity` |
| 1665 | /// fields derived from the underlying `ToolError`. This is what makes |
| 1666 | /// downstream tooling able to bucket failures without scraping the message |
| 1667 | /// string. |
| 1668 | #[test] |
| 1669 | fn tool_failure_audit_payload_carries_category_and_severity() { |
| 1670 | use crate::error_taxonomy::ErrorEnvelope; |
| 1671 | use crate::tools::spec::ToolError; |
| 1672 | |
| 1673 | let error = ToolError::Timeout { seconds: 30 }; |
| 1674 | let envelope: ErrorEnvelope = error.clone().into(); |
| 1675 | let payload = json!({ |
| 1676 | "event": "tool.result", |
| 1677 | "tool_id": "tool-1", |
| 1678 | "tool_name": "exec_shell", |
| 1679 | "success": false, |
| 1680 | "error": error.to_string(), |
| 1681 | "category": envelope.category.to_string(), |
| 1682 | "severity": envelope.severity.to_string(), |
| 1683 | }); |
| 1684 | |
| 1685 | assert_eq!(payload["category"], "timeout"); |
| 1686 | assert_eq!(payload["severity"], "warning"); |
| 1687 | assert_eq!(payload["success"], false); |
| 1688 | } |
| 1689 | |
| 1690 | /// Capacity escalation sees `ErrorCategory::InvalidInput` as a context-overflow |
| 1691 | /// signal that must escalate even on the first failure (no consecutive |
| 1692 | /// requirement). The previous string-matching path scanned the message for |
| 1693 | /// "context length" — categories give us a typed contract instead. |
| 1694 | #[test] |
| 1695 | fn capacity_escalation_treats_invalid_input_as_overflow_signal() { |
| 1696 | use crate::error_taxonomy::ErrorCategory; |
| 1697 | |
| 1698 | // Replays the categorization branches inside |
| 1699 | // `run_capacity_error_escalation_checkpoint`. Keeping the assertions on |
| 1700 | // the typed surface (slice of `ErrorCategory`) means this test fails |
| 1701 | // loudly if a future refactor reverts to substring matching. |
| 1702 | let categories: &[ErrorCategory] = &[ErrorCategory::InvalidInput]; |
| 1703 | let has_context_overflow = categories.contains(&ErrorCategory::InvalidInput); |
| 1704 | assert!(has_context_overflow); |
| 1705 | |
| 1706 | let only_transient = !categories.is_empty() |
| 1707 | && categories.iter().all(|c| { |
| 1708 | matches!( |
| 1709 | c, |
| 1710 | ErrorCategory::Network | ErrorCategory::RateLimit | ErrorCategory::Timeout |
| 1711 | ) |
| 1712 | }); |
| 1713 | assert!(!only_transient); |
| 1714 | } |
| 1715 | |
| 1716 | /// Transient categories (network / rate limit / timeout) must NOT escalate by |
| 1717 | /// themselves — those resolve via the existing retry loop and shouldn't |
| 1718 | /// trigger a capacity-driven replan. |
| 1719 | #[test] |
| 1720 | fn capacity_escalation_skips_pure_transient_categories() { |
| 1721 | use crate::error_taxonomy::ErrorCategory; |
| 1722 | |
| 1723 | let categories: &[ErrorCategory] = &[ |
| 1724 | ErrorCategory::Network, |
| 1725 | ErrorCategory::RateLimit, |
| 1726 | ErrorCategory::Timeout, |
| 1727 | ]; |
| 1728 | let has_context_overflow = categories.contains(&ErrorCategory::InvalidInput); |
| 1729 | assert!(!has_context_overflow); |
| 1730 | |
| 1731 | let only_transient = !categories.is_empty() |
| 1732 | && categories.iter().all(|c| { |
| 1733 | matches!( |
| 1734 | c, |
| 1735 | ErrorCategory::Network | ErrorCategory::RateLimit | ErrorCategory::Timeout |
| 1736 | ) |
| 1737 | }); |
| 1738 | assert!(only_transient); |
| 1739 | } |
| 1740 | |
| 1741 | // ── #136: post-edit LSP diagnostics hook ───────────────────────────────── |
| 1742 | |
| 1743 | #[test] |
| 1744 | fn edited_paths_for_edit_file_returns_path() { |
| 1745 | let input = json!({ "path": "src/foo.rs", "search": "x", "replace": "y" }); |
| 1746 | let paths = edited_paths_for_tool("edit_file", &input); |
| 1747 | assert_eq!(paths, vec![PathBuf::from("src/foo.rs")]); |
| 1748 | } |
| 1749 | |
| 1750 | #[test] |
| 1751 | fn edited_paths_for_write_file_returns_path() { |
| 1752 | let input = json!({ "path": "src/bar.rs", "content": "fn main() {}" }); |
| 1753 | let paths = edited_paths_for_tool("write_file", &input); |
| 1754 | assert_eq!(paths, vec![PathBuf::from("src/bar.rs")]); |
| 1755 | } |
| 1756 | |
| 1757 | #[test] |
| 1758 | fn edited_paths_for_apply_patch_with_files_returns_each_path() { |
| 1759 | let input = json!({ |
| 1760 | "files": [ |
| 1761 | { "path": "a.rs", "content": "" }, |
| 1762 | { "path": "b.rs", "content": "" } |
| 1763 | ] |
| 1764 | }); |
| 1765 | let paths = edited_paths_for_tool("apply_patch", &input); |
| 1766 | assert_eq!(paths, vec![PathBuf::from("a.rs"), PathBuf::from("b.rs")]); |
| 1767 | } |
| 1768 | |
| 1769 | #[test] |
| 1770 | fn edited_paths_for_apply_patch_with_diff_text_extracts_paths() { |
| 1771 | let input = json!({ |
| 1772 | "patch": "--- a/foo.rs\n+++ b/foo.rs\n@@ -1 +1 @@\n-let x: i32 = 0;\n+let x: i32 = \"oops\";\n" |
| 1773 | }); |
| 1774 | let paths = edited_paths_for_tool("apply_patch", &input); |
| 1775 | assert_eq!(paths, vec![PathBuf::from("foo.rs")]); |
| 1776 | } |
| 1777 | |
| 1778 | #[test] |
| 1779 | fn edited_paths_for_unknown_tool_returns_empty() { |
| 1780 | let input = json!({ "path": "irrelevant.rs" }); |
| 1781 | let paths = edited_paths_for_tool("read_file", &input); |
| 1782 | assert!(paths.is_empty()); |
| 1783 | let paths = edited_paths_for_tool("grep_files", &input); |
| 1784 | assert!(paths.is_empty()); |
| 1785 | } |
| 1786 | |
| 1787 | #[test] |
| 1788 | fn parse_patch_paths_skips_dev_null() { |
| 1789 | let patch = "--- a/keep.rs\n+++ b/keep.rs\n--- a/deleted.rs\n+++ /dev/null\n"; |
| 1790 | let paths = parse_patch_paths(patch); |
| 1791 | assert_eq!(paths, vec![PathBuf::from("keep.rs")]); |
| 1792 | } |
| 1793 | |
| 1794 | #[tokio::test] |
| 1795 | async fn post_edit_hook_injects_diagnostics_message_before_next_request() { |
| 1796 | use crate::lsp::{Diagnostic, Language, Severity}; |
| 1797 | use std::sync::Arc; |
| 1798 | |
| 1799 | let tmp = tempdir().expect("tempdir"); |
| 1800 | let workspace = tmp.path().to_path_buf(); |
| 1801 | let target = workspace.join("src").join("main.rs"); |
| 1802 | fs::create_dir_all(workspace.join("src")).unwrap(); |
| 1803 | fs::write(&target, "let x: i32 = \"not a number\";").unwrap(); |
| 1804 | |
| 1805 | let lsp_config = crate::lsp::LspConfig::default(); |
| 1806 | let engine_config = EngineConfig { |
| 1807 | workspace: workspace.clone(), |
| 1808 | lsp_config: Some(lsp_config), |
| 1809 | ..Default::default() |
| 1810 | }; |
| 1811 | let (mut engine, _handle) = Engine::new(engine_config, &Config::default()); |
| 1812 | |
| 1813 | // Install a fake transport that always reports a type error. |
| 1814 | let fake = Arc::new(crate::lsp::tests::FakeTransport::new(vec![Diagnostic { |
| 1815 | line: 1, |
| 1816 | column: 14, |
| 1817 | severity: Severity::Error, |
| 1818 | message: "expected i32, found &str".to_string(), |
| 1819 | }])); |
| 1820 | engine |
| 1821 | .lsp_manager |
| 1822 | .install_test_transport(Language::Rust, fake) |
| 1823 | .await; |
| 1824 | |
| 1825 | // Simulate the success path of an edit_file tool call. |
| 1826 | let input = json!({ "path": "src/main.rs", "search": "0", "replace": "\"not a number\"" }); |
| 1827 | engine.run_post_edit_lsp_hook("edit_file", &input).await; |
| 1828 | assert_eq!(engine.pending_lsp_blocks.len(), 1); |
| 1829 | |
| 1830 | // Flush prepares the synthetic message. |
| 1831 | let messages_before = engine.session.messages.len(); |
| 1832 | engine.flush_pending_lsp_diagnostics().await; |
| 1833 | assert_eq!(engine.session.messages.len(), messages_before + 1); |
| 1834 | |
| 1835 | let last = engine.session.messages.last().expect("message appended"); |
| 1836 | assert_eq!(last.role, "user"); |
| 1837 | let text = match &last.content[0] { |
| 1838 | crate::models::ContentBlock::Text { text, .. } => text.clone(), |
| 1839 | other => panic!("expected text block, got {other:?}"), |
| 1840 | }; |
| 1841 | assert!(text.contains("<diagnostics file=\"")); |
| 1842 | assert!(text.contains("ERROR [1:14] expected i32, found &str")); |
| 1843 | } |
| 1844 | |
| 1845 | #[tokio::test] |
| 1846 | async fn post_edit_hook_is_silent_when_lsp_disabled() { |
| 1847 | let tmp = tempdir().expect("tempdir"); |
| 1848 | let workspace = tmp.path().to_path_buf(); |
| 1849 | let target = workspace.join("src").join("main.rs"); |
| 1850 | fs::create_dir_all(workspace.join("src")).unwrap(); |
| 1851 | fs::write(&target, "fn main() {}").unwrap(); |
| 1852 | |
| 1853 | let lsp_config = crate::lsp::LspConfig { |
| 1854 | enabled: false, |
| 1855 | ..Default::default() |
| 1856 | }; |
| 1857 | let engine_config = EngineConfig { |
| 1858 | workspace: workspace.clone(), |
| 1859 | lsp_config: Some(lsp_config), |
| 1860 | ..Default::default() |
| 1861 | }; |
| 1862 | let (mut engine, _handle) = Engine::new(engine_config, &Config::default()); |
| 1863 | |
| 1864 | let input = json!({ "path": "src/main.rs", "search": "x", "replace": "y" }); |
| 1865 | engine.run_post_edit_lsp_hook("edit_file", &input).await; |
| 1866 | assert!(engine.pending_lsp_blocks.is_empty()); |
| 1867 | |
| 1868 | let messages_before = engine.session.messages.len(); |
| 1869 | engine.flush_pending_lsp_diagnostics().await; |
| 1870 | assert_eq!(engine.session.messages.len(), messages_before); |
| 1871 | } |
| 1872 | |
| 1873 | #[tokio::test] |
| 1874 | async fn post_edit_hook_skips_unknown_tool_names() { |
| 1875 | use crate::lsp::{Diagnostic, Language, Severity}; |
| 1876 | use std::sync::Arc; |
| 1877 | |
| 1878 | let tmp = tempdir().expect("tempdir"); |
| 1879 | let engine_config = EngineConfig { |
| 1880 | workspace: tmp.path().to_path_buf(), |
| 1881 | lsp_config: Some(crate::lsp::LspConfig::default()), |
| 1882 | ..Default::default() |
| 1883 | }; |
| 1884 | let (mut engine, _handle) = Engine::new(engine_config, &Config::default()); |
| 1885 | let fake = Arc::new(crate::lsp::tests::FakeTransport::new(vec![Diagnostic { |
| 1886 | line: 1, |
| 1887 | column: 1, |
| 1888 | severity: Severity::Error, |
| 1889 | message: "should not be reported".to_string(), |
| 1890 | }])); |
| 1891 | engine |
| 1892 | .lsp_manager |
| 1893 | .install_test_transport(Language::Rust, fake.clone()) |
| 1894 | .await; |
| 1895 | |
| 1896 | let input = json!({ "path": "src/main.rs" }); |
| 1897 | engine.run_post_edit_lsp_hook("read_file", &input).await; |
| 1898 | assert!(engine.pending_lsp_blocks.is_empty()); |
| 1899 | assert_eq!(fake.call_count(), 0); |
| 1900 | } |
| 1901 |