| 1 | //! Background advisor watcher (#3982). |
| 2 | //! |
| 3 | //! When enabled, the advisor wakes on turn boundaries, reads a bounded slice |
| 4 | //! of recent tool calls from the session transcript, makes a concise LLM |
| 5 | //! advisory call (reusing the same `DeepSeekClient` as the parent turn), and |
| 6 | //! emits an [`Event::AdvisoryNote`] fire-and-forget. |
| 7 | //! |
| 8 | //! Key design properties: |
| 9 | //! - **Off by default** — enabled via `[advisor] enabled = true` or `/advisor on`. |
| 10 | //! - **Bounded input** — at most `max_tool_calls` tool-call/result pairs are |
| 11 | //! included; the rest are dropped oldest-first. |
| 12 | //! - **Rate-limited** — at most one emission per `rate_limit_secs` seconds. |
| 13 | //! - **Deduplicated** — notes whose content hash matches the previous note |
| 14 | //! within `dedup_window_secs` are silently dropped. |
| 15 | //! - **Child-failure isolated** — advisor errors are logged but never surface |
| 16 | //! as parent turn failures. |
| 17 | //! - **Policy-bounded** — the advisor uses a read-only reviewer prompt and |
| 18 | //! no tool access; it cannot exceed the parent session policy. |
| 19 | |
| 20 | use std::collections::hash_map::DefaultHasher; |
| 21 | use std::hash::{Hash, Hasher}; |
| 22 | use std::time::{Duration, Instant}; |
| 23 | |
| 24 | use codewhale_config::AdvisorConfigToml; |
| 25 | use tokio::sync::mpsc; |
| 26 | use tracing::debug; |
| 27 | |
| 28 | use crate::client::DeepSeekClient; |
| 29 | use crate::core::events::Event; |
| 30 | use crate::llm_client::LlmClient; |
| 31 | use crate::models::{ContentBlock, Message, MessageRequest, SystemPrompt}; |
| 32 | use crate::utils::truncate_with_ellipsis; |
| 33 | |
| 34 | /// Maximum tokens the advisor may generate. Kept short so the note stays |
| 35 | /// concise and does not compete with the parent turn's billing budget. |
| 36 | const ADVISOR_MAX_TOKENS: u32 = 256; |
| 37 | |
| 38 | /// Maximum characters of tool input + result to include per tool-call pair. |
| 39 | const MAX_CHARS_PER_PAIR: usize = 800; |
| 40 | |
| 41 | /// System prompt for the advisor LLM call. Read-only review posture — no |
| 42 | /// tool access, no code generation. |
| 43 | const ADVISOR_SYSTEM_PROMPT: &str = "You are a concise background advisor reviewing recent tool activity. \ |
| 44 | Your role: identify one or two concrete concerns (correctness, risk, or \ |
| 45 | missed alternatives) in the tool calls provided. \ |
| 46 | If nothing notable stands out, respond with exactly the word \"ok\". \ |
| 47 | Otherwise write one to three short sentences — no preamble, no markdown, \ |
| 48 | no praise. Focus on signal; omit noise."; |
| 49 | |
| 50 | /// A single tool-call/result pair extracted from the session transcript. |
| 51 | #[derive(Debug, Clone)] |
| 52 | pub struct ToolCallPair { |
| 53 | /// Tool name (e.g. `exec_shell`, `file_write`). |
| 54 | pub name: String, |
| 55 | /// Bounded serialization of the tool input. |
| 56 | pub input_preview: String, |
| 57 | /// Bounded serialization of the tool result. |
| 58 | pub result_preview: String, |
| 59 | } |
| 60 | |
| 61 | /// Resolved advisor configuration derived from [`AdvisorConfigToml`]. |
| 62 | #[derive(Debug, Clone)] |
| 63 | pub struct AdvisorConfig { |
| 64 | /// Whether the advisor is currently enabled (session-level toggle). |
| 65 | pub enabled: bool, |
| 66 | /// Max tool-call pairs to review per turn. |
| 67 | pub max_tool_calls: u32, |
| 68 | /// Min seconds between consecutive emissions. |
| 69 | pub rate_limit: Duration, |
| 70 | /// Window during which duplicate notes are suppressed. |
| 71 | pub dedup_window: Duration, |
| 72 | /// Optional model override (falls back to session model when `None`). |
| 73 | pub model: Option<String>, |
| 74 | } |
| 75 | |
| 76 | impl AdvisorConfig { |
| 77 | /// Build a resolved config from the TOML schema. |
| 78 | #[must_use] |
| 79 | pub fn from_toml(toml: &AdvisorConfigToml) -> Self { |
| 80 | Self { |
| 81 | enabled: toml.enabled, |
| 82 | max_tool_calls: toml.max_tool_calls.clamp(1, 50), |
| 83 | rate_limit: Duration::from_secs(toml.rate_limit_secs.clamp(5, 3600)), |
| 84 | dedup_window: Duration::from_secs(toml.dedup_window_secs), |
| 85 | model: toml.model.clone(), |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | /// Default disabled config (matches `[advisor]` absent from config.toml). |
| 90 | #[must_use] |
| 91 | pub fn disabled() -> Self { |
| 92 | Self { |
| 93 | enabled: false, |
| 94 | max_tool_calls: 10, |
| 95 | rate_limit: Duration::from_secs(60), |
| 96 | dedup_window: Duration::from_secs(300), |
| 97 | model: None, |
| 98 | } |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | /// Runtime emission guard: tracks the last emission time and the hash of the |
| 103 | /// last advisory note to enforce rate limiting and deduplication. |
| 104 | #[derive(Debug)] |
| 105 | pub struct EmissionGuard { |
| 106 | last_emission: Option<Instant>, |
| 107 | last_note_hash: Option<u64>, |
| 108 | last_note_hash_at: Option<Instant>, |
| 109 | } |
| 110 | |
| 111 | impl EmissionGuard { |
| 112 | /// Create a fresh guard with no emission history. |
| 113 | #[must_use] |
| 114 | pub fn new() -> Self { |
| 115 | Self { |
| 116 | last_emission: None, |
| 117 | last_note_hash: None, |
| 118 | last_note_hash_at: None, |
| 119 | } |
| 120 | } |
| 121 | |
| 122 | /// Check whether emitting `note` is allowed under `config`'s rate-limit |
| 123 | /// and dedup policy. Returns `true` when the note may be emitted. |
| 124 | #[must_use] |
| 125 | pub fn may_emit(&self, note: &str, config: &AdvisorConfig) -> bool { |
| 126 | // Suppress trivial "ok" responses from the model. |
| 127 | if note.trim().eq_ignore_ascii_case("ok") { |
| 128 | return false; |
| 129 | } |
| 130 | |
| 131 | let now = Instant::now(); |
| 132 | |
| 133 | // Rate limit: require at least `rate_limit` since last emission. |
| 134 | if let Some(last) = self.last_emission |
| 135 | && now.duration_since(last) < config.rate_limit |
| 136 | { |
| 137 | return false; |
| 138 | } |
| 139 | |
| 140 | // Dedup: suppress if the note content hash matches the previous note |
| 141 | // within the dedup window. |
| 142 | let note_hash = hash_str(note); |
| 143 | if let (Some(prev_hash), Some(prev_at)) = (self.last_note_hash, self.last_note_hash_at) |
| 144 | && prev_hash == note_hash |
| 145 | && now.duration_since(prev_at) < config.dedup_window |
| 146 | { |
| 147 | return false; |
| 148 | } |
| 149 | |
| 150 | true |
| 151 | } |
| 152 | |
| 153 | /// Record that `note` was emitted now. Must be called immediately after |
| 154 | /// sending the `AdvisoryNote` event. |
| 155 | pub fn record_emission(&mut self, note: &str) { |
| 156 | let now = Instant::now(); |
| 157 | self.last_emission = Some(now); |
| 158 | self.last_note_hash = Some(hash_str(note)); |
| 159 | self.last_note_hash_at = Some(now); |
| 160 | } |
| 161 | } |
| 162 | |
| 163 | impl Default for EmissionGuard { |
| 164 | fn default() -> Self { |
| 165 | Self::new() |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | /// Extract bounded tool-call/result pairs from a session message slice. |
| 170 | /// |
| 171 | /// Scans `messages` in reverse (newest first), collects up to `max_pairs` |
| 172 | /// `ToolUse`/`ToolResult` pairs, then returns them oldest-first. |
| 173 | #[must_use] |
| 174 | pub fn extract_tool_call_pairs(messages: &[Message], max_pairs: usize) -> Vec<ToolCallPair> { |
| 175 | // Collect ToolUse names+inputs (from assistant messages) and |
| 176 | // ToolResult texts (from user messages) into a pairing structure. |
| 177 | let mut uses: Vec<(String, String, String)> = Vec::new(); // (id, name, input) |
| 178 | let mut results: std::collections::HashMap<String, String> = std::collections::HashMap::new(); |
| 179 | |
| 180 | for msg in messages { |
| 181 | for block in &msg.content { |
| 182 | match block { |
| 183 | ContentBlock::ToolUse { |
| 184 | id, name, input, .. |
| 185 | } => { |
| 186 | let input_str = truncate_with_ellipsis( |
| 187 | &serde_json::to_string(input).unwrap_or_default(), |
| 188 | MAX_CHARS_PER_PAIR / 2, |
| 189 | "…", |
| 190 | ); |
| 191 | uses.push((id.clone(), name.clone(), input_str)); |
| 192 | } |
| 193 | ContentBlock::ToolResult { |
| 194 | tool_use_id, |
| 195 | content, |
| 196 | .. |
| 197 | } => { |
| 198 | results.insert( |
| 199 | tool_use_id.clone(), |
| 200 | truncate_with_ellipsis(content, MAX_CHARS_PER_PAIR / 2, "…"), |
| 201 | ); |
| 202 | } |
| 203 | _ => {} |
| 204 | } |
| 205 | } |
| 206 | } |
| 207 | |
| 208 | // Match uses with results and take the last `max_pairs`. |
| 209 | let start = uses.len().saturating_sub(max_pairs); |
| 210 | uses[start..] |
| 211 | .iter() |
| 212 | .map(|(id, name, input)| { |
| 213 | let result = results |
| 214 | .get(id.as_str()) |
| 215 | .cloned() |
| 216 | .unwrap_or_else(|| "(pending)".to_string()); |
| 217 | ToolCallPair { |
| 218 | name: name.clone(), |
| 219 | input_preview: input.clone(), |
| 220 | result_preview: result, |
| 221 | } |
| 222 | }) |
| 223 | .collect() |
| 224 | } |
| 225 | |
| 226 | /// Build the user prompt for the advisor from a slice of tool-call pairs. |
| 227 | #[must_use] |
| 228 | pub fn build_advisor_prompt(pairs: &[ToolCallPair]) -> String { |
| 229 | let mut out = String::from("Recent tool activity to review (oldest → newest):\n\n"); |
| 230 | for (i, pair) in pairs.iter().enumerate() { |
| 231 | out.push_str(&format!( |
| 232 | "{}. tool={}\n input: {}\n result: {}\n\n", |
| 233 | i + 1, |
| 234 | pair.name, |
| 235 | pair.input_preview, |
| 236 | pair.result_preview |
| 237 | )); |
| 238 | } |
| 239 | out.push_str( |
| 240 | "Provide your advisory in one to three sentences, or respond with \"ok\" if nothing notable.", |
| 241 | ); |
| 242 | out |
| 243 | } |
| 244 | |
| 245 | /// Run one advisor review cycle for a completed turn. |
| 246 | /// |
| 247 | /// This is the async work dispatched by `spawn_supervised` in the engine. It: |
| 248 | /// 1. Checks whether emission is allowed by `guard`. |
| 249 | /// 2. Extracts bounded tool-call pairs from `messages`. |
| 250 | /// 3. Makes a non-streaming LLM call with a short read-only prompt. |
| 251 | /// 4. Checks emission again (the LLM call may have taken time). |
| 252 | /// 5. Sends `Event::AdvisoryNote` if the note passes the guard. |
| 253 | /// |
| 254 | /// All errors are logged and swallowed — the advisor must never fail the |
| 255 | /// parent turn. |
| 256 | pub async fn run_advisor_for_turn( |
| 257 | turn_id: String, |
| 258 | messages: Vec<Message>, |
| 259 | config: AdvisorConfig, |
| 260 | client: DeepSeekClient, |
| 261 | session_model: String, |
| 262 | guard: std::sync::Arc<tokio::sync::Mutex<EmissionGuard>>, |
| 263 | tx_event: mpsc::Sender<Event>, |
| 264 | ) { |
| 265 | // Pre-flight: skip if the guard already blocks (avoids the LLM call when |
| 266 | // rate-limited, which is the common case for rapid turn sequences). |
| 267 | { |
| 268 | let g = guard.lock().await; |
| 269 | // We don't have the note content yet, so we only check the rate limit |
| 270 | // here by testing with a placeholder. The dedup check runs after the |
| 271 | // LLM call, when we have the actual content. |
| 272 | if let Some(last) = g.last_emission |
| 273 | && std::time::Instant::now().duration_since(last) < config.rate_limit |
| 274 | { |
| 275 | debug!(target: "advisor", "rate-limited, skipping advisor run for turn {turn_id}"); |
| 276 | return; |
| 277 | } |
| 278 | } |
| 279 | |
| 280 | // Extract a bounded slice of tool-call pairs. |
| 281 | let pairs = extract_tool_call_pairs(&messages, config.max_tool_calls as usize); |
| 282 | if pairs.is_empty() { |
| 283 | debug!(target: "advisor", "no tool calls found; skipping advisor for turn {turn_id}"); |
| 284 | return; |
| 285 | } |
| 286 | |
| 287 | let tool_call_count = pairs.len() as u32; |
| 288 | let prompt = build_advisor_prompt(&pairs); |
| 289 | let model = config |
| 290 | .model |
| 291 | .clone() |
| 292 | .unwrap_or_else(|| session_model.clone()); |
| 293 | |
| 294 | let request = MessageRequest { |
| 295 | model: model.clone(), |
| 296 | messages: vec![Message { |
| 297 | role: "user".to_string(), |
| 298 | content: vec![ContentBlock::Text { |
| 299 | text: prompt, |
| 300 | cache_control: None, |
| 301 | }], |
| 302 | }], |
| 303 | max_tokens: ADVISOR_MAX_TOKENS, |
| 304 | system: Some(SystemPrompt::Text(ADVISOR_SYSTEM_PROMPT.to_string())), |
| 305 | tools: None, |
| 306 | tool_choice: None, |
| 307 | metadata: None, |
| 308 | thinking: None, |
| 309 | reasoning_effort: None, |
| 310 | stream: Some(false), |
| 311 | temperature: Some(0.3), |
| 312 | top_p: Some(0.9), |
| 313 | }; |
| 314 | |
| 315 | let response = match client.create_message(request).await { |
| 316 | Ok(r) => r, |
| 317 | Err(e) => { |
| 318 | tracing::warn!(target: "advisor", "advisor LLM call failed for turn {turn_id}: {e}"); |
| 319 | return; |
| 320 | } |
| 321 | }; |
| 322 | |
| 323 | // Extract the text from the response. |
| 324 | let note: String = response |
| 325 | .content |
| 326 | .iter() |
| 327 | .filter_map(|block| { |
| 328 | if let ContentBlock::Text { text, .. } = block { |
| 329 | Some(text.as_str()) |
| 330 | } else { |
| 331 | None |
| 332 | } |
| 333 | }) |
| 334 | .collect::<Vec<_>>() |
| 335 | .join("\n") |
| 336 | .trim() |
| 337 | .to_string(); |
| 338 | |
| 339 | if note.is_empty() { |
| 340 | debug!(target: "advisor", "empty advisor response for turn {turn_id}; skipping"); |
| 341 | return; |
| 342 | } |
| 343 | |
| 344 | // Post-flight emission check (rate limit + dedup). |
| 345 | let mut guard_lock = guard.lock().await; |
| 346 | if !guard_lock.may_emit(¬e, &config) { |
| 347 | debug!(target: "advisor", "emission suppressed by guard for turn {turn_id}"); |
| 348 | return; |
| 349 | } |
| 350 | |
| 351 | guard_lock.record_emission(¬e); |
| 352 | drop(guard_lock); |
| 353 | |
| 354 | let _ = tx_event |
| 355 | .send(Event::AdvisoryNote { |
| 356 | turn_id: turn_id.clone(), |
| 357 | note: note.clone(), |
| 358 | tool_call_count, |
| 359 | }) |
| 360 | .await; |
| 361 | |
| 362 | debug!(target: "advisor", "advisory note emitted for turn {turn_id} ({tool_call_count} tool calls reviewed)"); |
| 363 | } |
| 364 | |
| 365 | fn hash_str(s: &str) -> u64 { |
| 366 | let mut h = DefaultHasher::new(); |
| 367 | s.hash(&mut h); |
| 368 | h.finish() |
| 369 | } |
| 370 | |
| 371 | // ── Tests ────────────────────────────────────────────────────────────────── |
| 372 | |
| 373 | #[cfg(test)] |
| 374 | mod tests { |
| 375 | use super::*; |
| 376 | use std::time::Duration; |
| 377 | |
| 378 | fn test_config() -> AdvisorConfig { |
| 379 | AdvisorConfig { |
| 380 | enabled: true, |
| 381 | max_tool_calls: 5, |
| 382 | rate_limit: Duration::from_secs(1), |
| 383 | dedup_window: Duration::from_secs(10), |
| 384 | model: None, |
| 385 | } |
| 386 | } |
| 387 | |
| 388 | // ── enable/disable ──────────────────────────────────────────────────── |
| 389 | |
| 390 | #[test] |
| 391 | fn disabled_config_has_enabled_false() { |
| 392 | let cfg = AdvisorConfig::disabled(); |
| 393 | assert!(!cfg.enabled); |
| 394 | } |
| 395 | |
| 396 | #[test] |
| 397 | fn from_toml_clamps_max_tool_calls() { |
| 398 | let toml = AdvisorConfigToml { |
| 399 | enabled: true, |
| 400 | max_tool_calls: 999, |
| 401 | rate_limit_secs: 60, |
| 402 | dedup_window_secs: 300, |
| 403 | model: None, |
| 404 | }; |
| 405 | let cfg = AdvisorConfig::from_toml(&toml); |
| 406 | assert_eq!( |
| 407 | cfg.max_tool_calls, 50, |
| 408 | "max_tool_calls must be clamped to 50" |
| 409 | ); |
| 410 | } |
| 411 | |
| 412 | #[test] |
| 413 | fn from_toml_clamps_rate_limit() { |
| 414 | let toml = AdvisorConfigToml { |
| 415 | enabled: true, |
| 416 | max_tool_calls: 10, |
| 417 | rate_limit_secs: 0, // below minimum of 5 |
| 418 | dedup_window_secs: 300, |
| 419 | model: None, |
| 420 | }; |
| 421 | let cfg = AdvisorConfig::from_toml(&toml); |
| 422 | assert!( |
| 423 | cfg.rate_limit >= Duration::from_secs(5), |
| 424 | "rate_limit must be at least 5s" |
| 425 | ); |
| 426 | } |
| 427 | |
| 428 | // ── bounded input ───────────────────────────────────────────────────── |
| 429 | |
| 430 | fn make_messages_with_n_tool_calls(n: usize) -> Vec<Message> { |
| 431 | let mut messages = Vec::new(); |
| 432 | for i in 0..n { |
| 433 | let id = format!("tool_{i}"); |
| 434 | // assistant message with ToolUse |
| 435 | messages.push(Message { |
| 436 | role: "assistant".to_string(), |
| 437 | content: vec![ContentBlock::ToolUse { |
| 438 | id: id.clone(), |
| 439 | name: "exec_shell".to_string(), |
| 440 | input: serde_json::json!({"command": format!("echo {i}")}), |
| 441 | caller: None, |
| 442 | }], |
| 443 | }); |
| 444 | // user message with ToolResult |
| 445 | messages.push(Message { |
| 446 | role: "user".to_string(), |
| 447 | content: vec![ContentBlock::ToolResult { |
| 448 | tool_use_id: id, |
| 449 | content: format!("{i}"), |
| 450 | is_error: None, |
| 451 | content_blocks: None, |
| 452 | }], |
| 453 | }); |
| 454 | } |
| 455 | messages |
| 456 | } |
| 457 | |
| 458 | #[test] |
| 459 | fn extract_tool_call_pairs_bounded_by_max() { |
| 460 | let messages = make_messages_with_n_tool_calls(20); |
| 461 | let pairs = extract_tool_call_pairs(&messages, 5); |
| 462 | assert_eq!(pairs.len(), 5, "must return at most max_pairs"); |
| 463 | // Should be the last 5 (newest). |
| 464 | assert_eq!(pairs[0].name, "exec_shell"); |
| 465 | } |
| 466 | |
| 467 | #[test] |
| 468 | fn extract_tool_call_pairs_empty_when_no_tool_calls() { |
| 469 | let messages = vec![Message { |
| 470 | role: "user".to_string(), |
| 471 | content: vec![ContentBlock::Text { |
| 472 | text: "hello".to_string(), |
| 473 | cache_control: None, |
| 474 | }], |
| 475 | }]; |
| 476 | let pairs = extract_tool_call_pairs(&messages, 5); |
| 477 | assert!(pairs.is_empty()); |
| 478 | } |
| 479 | |
| 480 | #[test] |
| 481 | fn extract_tool_call_pairs_fewer_than_max_returns_all() { |
| 482 | let messages = make_messages_with_n_tool_calls(3); |
| 483 | let pairs = extract_tool_call_pairs(&messages, 10); |
| 484 | assert_eq!(pairs.len(), 3); |
| 485 | } |
| 486 | |
| 487 | // ── rate limiting ───────────────────────────────────────────────────── |
| 488 | |
| 489 | #[test] |
| 490 | fn emission_guard_allows_first_emission() { |
| 491 | let guard = EmissionGuard::new(); |
| 492 | let config = test_config(); |
| 493 | assert!( |
| 494 | guard.may_emit("something concerning here", &config), |
| 495 | "first emission must be allowed" |
| 496 | ); |
| 497 | } |
| 498 | |
| 499 | #[test] |
| 500 | fn emission_guard_blocks_immediately_after_emission() { |
| 501 | let mut guard = EmissionGuard::new(); |
| 502 | let config = test_config(); |
| 503 | let note = "something concerning"; |
| 504 | guard.record_emission(note); |
| 505 | assert!( |
| 506 | !guard.may_emit("a completely different note", &config), |
| 507 | "emission must be blocked immediately after a prior emission (rate limit)" |
| 508 | ); |
| 509 | } |
| 510 | |
| 511 | #[test] |
| 512 | fn emission_guard_allows_after_rate_limit_expires() { |
| 513 | let mut guard = EmissionGuard::new(); |
| 514 | // Rate limit of 0ms — always expired. |
| 515 | let config = AdvisorConfig { |
| 516 | rate_limit: Duration::ZERO, |
| 517 | dedup_window: Duration::from_secs(300), |
| 518 | ..AdvisorConfig::disabled() |
| 519 | }; |
| 520 | let note = "first note"; |
| 521 | guard.record_emission(note); |
| 522 | assert!( |
| 523 | guard.may_emit("second different note", &config), |
| 524 | "emission must be allowed when rate limit duration is zero" |
| 525 | ); |
| 526 | } |
| 527 | |
| 528 | // ── deduplication ───────────────────────────────────────────────────── |
| 529 | |
| 530 | #[test] |
| 531 | fn emission_guard_suppresses_ok_response() { |
| 532 | let guard = EmissionGuard::new(); |
| 533 | let config = test_config(); |
| 534 | assert!(!guard.may_emit("ok", &config), "\"ok\" must be suppressed"); |
| 535 | assert!(!guard.may_emit("OK", &config), "\"OK\" must be suppressed"); |
| 536 | assert!( |
| 537 | !guard.may_emit(" ok ", &config), |
| 538 | "\" ok \" must be suppressed" |
| 539 | ); |
| 540 | } |
| 541 | |
| 542 | #[test] |
| 543 | fn emission_guard_dedup_blocks_identical_note_within_window() { |
| 544 | let mut guard = EmissionGuard::new(); |
| 545 | // Use a zero rate limit so only dedup is tested. |
| 546 | let config = AdvisorConfig { |
| 547 | rate_limit: Duration::ZERO, |
| 548 | dedup_window: Duration::from_secs(300), |
| 549 | ..AdvisorConfig::disabled() |
| 550 | }; |
| 551 | let note = "risky shell command with no error checking"; |
| 552 | guard.record_emission(note); |
| 553 | assert!( |
| 554 | !guard.may_emit(note, &config), |
| 555 | "identical note must be suppressed within the dedup window" |
| 556 | ); |
| 557 | } |
| 558 | |
| 559 | #[test] |
| 560 | fn emission_guard_allows_different_note_within_dedup_window() { |
| 561 | let mut guard = EmissionGuard::new(); |
| 562 | let config = AdvisorConfig { |
| 563 | rate_limit: Duration::ZERO, |
| 564 | dedup_window: Duration::from_secs(300), |
| 565 | ..AdvisorConfig::disabled() |
| 566 | }; |
| 567 | guard.record_emission("first note"); |
| 568 | assert!( |
| 569 | guard.may_emit("entirely different note", &config), |
| 570 | "a different note must be allowed even within the dedup window" |
| 571 | ); |
| 572 | } |
| 573 | |
| 574 | // ── child failure isolation ─────────────────────────────────────────── |
| 575 | |
| 576 | #[test] |
| 577 | fn advisor_prompt_is_non_empty_for_non_empty_pairs() { |
| 578 | let pairs = vec![ToolCallPair { |
| 579 | name: "exec_shell".to_string(), |
| 580 | input_preview: r#"{"command":"ls -la"}"#.to_string(), |
| 581 | result_preview: "total 4\ndrwxr-xr-x 2 user user 4096".to_string(), |
| 582 | }]; |
| 583 | let prompt = build_advisor_prompt(&pairs); |
| 584 | assert!( |
| 585 | prompt.contains("exec_shell"), |
| 586 | "prompt must include the tool name" |
| 587 | ); |
| 588 | assert!( |
| 589 | prompt.contains("ls -la"), |
| 590 | "prompt must include the tool input" |
| 591 | ); |
| 592 | } |
| 593 | } |
| 594 |