| 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 on an exactly resolved provider/model client, 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::CodewhaleClient; |
| 29 | use crate::config::Config; |
| 30 | use crate::core::events::Event; |
| 31 | use crate::llm_client::LlmClient; |
| 32 | use crate::utils::truncate_with_ellipsis; |
| 33 | use codewhale_models::Role; |
| 34 | use codewhale_models::{ContentBlock, Message, MessageRequest, SystemPrompt}; |
| 35 | |
| 36 | /// Maximum tokens the advisor may generate. Kept short so the note stays |
| 37 | /// concise and does not compete with the parent turn's billing budget. |
| 38 | const ADVISOR_MAX_TOKENS: u32 = 256; |
| 39 | |
| 40 | /// Maximum characters of tool input + result to include per tool-call pair. |
| 41 | const MAX_CHARS_PER_PAIR: usize = 800; |
| 42 | |
| 43 | /// System prompt for the advisor LLM call. Read-only review posture — no |
| 44 | /// tool access, no code generation. |
| 45 | const ADVISOR_SYSTEM_PROMPT: &str = "You are a concise background advisor reviewing recent tool activity. \ |
| 46 | Your role: identify one or two concrete concerns (correctness, risk, or \ |
| 47 | missed alternatives) in the tool calls provided. \ |
| 48 | If nothing notable stands out, respond with exactly the word \"ok\". \ |
| 49 | Otherwise write one to three short sentences — no preamble, no markdown, \ |
| 50 | no praise. Focus on signal; omit noise."; |
| 51 | |
| 52 | /// A single tool-call/result pair extracted from the session transcript. |
| 53 | #[derive(Debug, Clone)] |
| 54 | pub struct ToolCallPair { |
| 55 | /// Tool name (e.g. `exec_shell`, `file_write`). |
| 56 | pub name: String, |
| 57 | /// Bounded serialization of the tool input. |
| 58 | pub input_preview: String, |
| 59 | /// Bounded serialization of the tool result. |
| 60 | pub result_preview: String, |
| 61 | } |
| 62 | |
| 63 | /// Resolved advisor configuration derived from [`AdvisorConfigToml`]. |
| 64 | #[derive(Debug, Clone)] |
| 65 | pub struct AdvisorConfig { |
| 66 | /// Whether the advisor is currently enabled (session-level toggle). |
| 67 | pub enabled: bool, |
| 68 | /// Max tool-call pairs to review per turn. |
| 69 | pub max_tool_calls: u32, |
| 70 | /// Min seconds between consecutive emissions. |
| 71 | pub rate_limit: Duration, |
| 72 | /// Window during which duplicate notes are suppressed. |
| 73 | pub dedup_window: Duration, |
| 74 | /// Optional model override (falls back to session model when `None`). |
| 75 | pub model: Option<String>, |
| 76 | } |
| 77 | |
| 78 | impl AdvisorConfig { |
| 79 | /// Build a resolved config from the TOML schema. |
| 80 | #[must_use] |
| 81 | pub fn from_toml(toml: &AdvisorConfigToml) -> Self { |
| 82 | Self { |
| 83 | enabled: toml.enabled, |
| 84 | max_tool_calls: toml.max_tool_calls.clamp(1, 50), |
| 85 | rate_limit: Duration::from_secs(toml.rate_limit_secs.clamp(5, 3600)), |
| 86 | dedup_window: Duration::from_secs(toml.dedup_window_secs), |
| 87 | model: toml.model.clone(), |
| 88 | } |
| 89 | } |
| 90 | |
| 91 | /// Default disabled config (matches `[advisor]` absent from config.toml). |
| 92 | #[must_use] |
| 93 | pub fn disabled() -> Self { |
| 94 | Self { |
| 95 | enabled: false, |
| 96 | max_tool_calls: 10, |
| 97 | rate_limit: Duration::from_secs(60), |
| 98 | dedup_window: Duration::from_secs(300), |
| 99 | model: None, |
| 100 | } |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | /// Runtime emission guard: tracks the last emission time and the hash of the |
| 105 | /// last advisory note to enforce rate limiting and deduplication. |
| 106 | #[derive(Debug)] |
| 107 | pub struct EmissionGuard { |
| 108 | last_emission: Option<Instant>, |
| 109 | last_note_hash: Option<u64>, |
| 110 | last_note_hash_at: Option<Instant>, |
| 111 | } |
| 112 | |
| 113 | /// Accounting ownership captured while the originating turn is still live. |
| 114 | /// |
| 115 | /// Runtime turns retain their synchronous durable sink through the lease; |
| 116 | /// ordinary interactive turns fall back to the exact session cost generation |
| 117 | /// captured here. Neither path can spill into a later session. |
| 118 | #[derive(Debug)] |
| 119 | pub(crate) struct AdvisorUsageContext { |
| 120 | cost_scope: crate::cost_status::CostScopeToken, |
| 121 | runtime_usage_lease: Option<crate::cost_status::RuntimeUsageLease>, |
| 122 | } |
| 123 | |
| 124 | impl AdvisorUsageContext { |
| 125 | #[must_use] |
| 126 | pub(crate) fn capture(runtime_owner: Option<&str>) -> Self { |
| 127 | Self { |
| 128 | cost_scope: crate::cost_status::scope_token(), |
| 129 | runtime_usage_lease: runtime_owner |
| 130 | .and_then(crate::cost_status::acquire_runtime_usage_lease), |
| 131 | } |
| 132 | } |
| 133 | |
| 134 | fn report( |
| 135 | &self, |
| 136 | source_id: &str, |
| 137 | route: &crate::cost_status::EffectiveRouteEnvelope, |
| 138 | usage: &codewhale_models::Usage, |
| 139 | ) { |
| 140 | crate::cost_status::report_effective_route_for_runtime( |
| 141 | self.cost_scope, |
| 142 | self.runtime_usage_lease |
| 143 | .as_ref() |
| 144 | .map(crate::cost_status::RuntimeUsageLease::owner), |
| 145 | source_id, |
| 146 | route, |
| 147 | usage, |
| 148 | ); |
| 149 | } |
| 150 | |
| 151 | fn report_unreceipted( |
| 152 | &self, |
| 153 | source_id: &str, |
| 154 | route: &crate::cost_status::EffectiveRouteEnvelope, |
| 155 | ) { |
| 156 | crate::cost_status::report_unreceipted_provider_success( |
| 157 | self.cost_scope, |
| 158 | self.runtime_usage_lease |
| 159 | .as_ref() |
| 160 | .map(crate::cost_status::RuntimeUsageLease::owner), |
| 161 | source_id, |
| 162 | route, |
| 163 | ); |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | impl EmissionGuard { |
| 168 | /// Create a fresh guard with no emission history. |
| 169 | #[must_use] |
| 170 | pub fn new() -> Self { |
| 171 | Self { |
| 172 | last_emission: None, |
| 173 | last_note_hash: None, |
| 174 | last_note_hash_at: None, |
| 175 | } |
| 176 | } |
| 177 | |
| 178 | /// Check whether emitting `note` is allowed under `config`'s rate-limit |
| 179 | /// and dedup policy. Returns `true` when the note may be emitted. |
| 180 | #[must_use] |
| 181 | pub fn may_emit(&self, note: &str, config: &AdvisorConfig) -> bool { |
| 182 | // Suppress trivial "ok" responses from the model. |
| 183 | if note.trim().eq_ignore_ascii_case("ok") { |
| 184 | return false; |
| 185 | } |
| 186 | |
| 187 | let now = Instant::now(); |
| 188 | |
| 189 | // Rate limit: require at least `rate_limit` since last emission. |
| 190 | if let Some(last) = self.last_emission |
| 191 | && now.duration_since(last) < config.rate_limit |
| 192 | { |
| 193 | return false; |
| 194 | } |
| 195 | |
| 196 | // Dedup: suppress if the note content hash matches the previous note |
| 197 | // within the dedup window. |
| 198 | let note_hash = hash_str(note); |
| 199 | if let (Some(prev_hash), Some(prev_at)) = (self.last_note_hash, self.last_note_hash_at) |
| 200 | && prev_hash == note_hash |
| 201 | && now.duration_since(prev_at) < config.dedup_window |
| 202 | { |
| 203 | return false; |
| 204 | } |
| 205 | |
| 206 | true |
| 207 | } |
| 208 | |
| 209 | /// Record that `note` was emitted now. Must be called immediately after |
| 210 | /// sending the `AdvisoryNote` event. |
| 211 | pub fn record_emission(&mut self, note: &str) { |
| 212 | let now = Instant::now(); |
| 213 | self.last_emission = Some(now); |
| 214 | self.last_note_hash = Some(hash_str(note)); |
| 215 | self.last_note_hash_at = Some(now); |
| 216 | } |
| 217 | } |
| 218 | |
| 219 | impl Default for EmissionGuard { |
| 220 | fn default() -> Self { |
| 221 | Self::new() |
| 222 | } |
| 223 | } |
| 224 | |
| 225 | /// Extract bounded tool-call/result pairs from a session message slice. |
| 226 | /// |
| 227 | /// Scans `messages` in reverse (newest first), collects up to `max_pairs` |
| 228 | /// `ToolUse`/`ToolResult` pairs, then returns them oldest-first. |
| 229 | #[must_use] |
| 230 | pub fn extract_tool_call_pairs(messages: &[Message], max_pairs: usize) -> Vec<ToolCallPair> { |
| 231 | // Collect ToolUse names+inputs (from assistant messages) and |
| 232 | // ToolResult texts (from user messages) into a pairing structure. |
| 233 | let mut uses: Vec<(String, String, String)> = Vec::new(); // (id, name, input) |
| 234 | let mut results: std::collections::HashMap<String, String> = std::collections::HashMap::new(); |
| 235 | |
| 236 | for msg in messages { |
| 237 | for block in &msg.content { |
| 238 | match block { |
| 239 | ContentBlock::ToolUse { |
| 240 | id, name, input, .. |
| 241 | } => { |
| 242 | let input_str = truncate_with_ellipsis( |
| 243 | &serde_json::to_string(input).unwrap_or_default(), |
| 244 | MAX_CHARS_PER_PAIR / 2, |
| 245 | "…", |
| 246 | ); |
| 247 | uses.push((id.clone(), name.clone(), input_str)); |
| 248 | } |
| 249 | ContentBlock::ToolResult { |
| 250 | tool_use_id, |
| 251 | content, |
| 252 | .. |
| 253 | } => { |
| 254 | results.insert( |
| 255 | tool_use_id.clone(), |
| 256 | truncate_with_ellipsis(content, MAX_CHARS_PER_PAIR / 2, "…"), |
| 257 | ); |
| 258 | } |
| 259 | _ => {} |
| 260 | } |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | // Match uses with results and take the last `max_pairs`. |
| 265 | let start = uses.len().saturating_sub(max_pairs); |
| 266 | uses[start..] |
| 267 | .iter() |
| 268 | .map(|(id, name, input)| { |
| 269 | let result = results |
| 270 | .get(id.as_str()) |
| 271 | .cloned() |
| 272 | .unwrap_or_else(|| "(pending)".to_string()); |
| 273 | ToolCallPair { |
| 274 | name: name.clone(), |
| 275 | input_preview: input.clone(), |
| 276 | result_preview: result, |
| 277 | } |
| 278 | }) |
| 279 | .collect() |
| 280 | } |
| 281 | |
| 282 | /// Build the user prompt for the advisor from a slice of tool-call pairs. |
| 283 | #[must_use] |
| 284 | pub fn build_advisor_prompt(pairs: &[ToolCallPair]) -> String { |
| 285 | let mut out = String::from("Recent tool activity to review (oldest → newest):\n\n"); |
| 286 | for (i, pair) in pairs.iter().enumerate() { |
| 287 | out.push_str(&format!( |
| 288 | "{}. tool={}\n input: {}\n result: {}\n\n", |
| 289 | i + 1, |
| 290 | pair.name, |
| 291 | pair.input_preview, |
| 292 | pair.result_preview |
| 293 | )); |
| 294 | } |
| 295 | out.push_str( |
| 296 | "Provide your advisory in one to three sentences, or respond with \"ok\" if nothing notable.", |
| 297 | ); |
| 298 | out |
| 299 | } |
| 300 | |
| 301 | /// Run one advisor review cycle for a completed turn. |
| 302 | /// |
| 303 | /// This is the async work dispatched by `spawn_supervised` in the engine. It: |
| 304 | /// 1. Checks whether emission is allowed by `guard`. |
| 305 | /// 2. Extracts bounded tool-call pairs from `messages`. |
| 306 | /// 3. Makes a non-streaming LLM call with a short read-only prompt. |
| 307 | /// 4. Checks emission again (the LLM call may have taken time). |
| 308 | /// 5. Sends `Event::AdvisoryNote` if the note passes the guard. |
| 309 | /// |
| 310 | /// All errors are logged and swallowed — the advisor must never fail the |
| 311 | /// parent turn. |
| 312 | pub async fn run_advisor_for_turn( |
| 313 | turn_id: String, |
| 314 | messages: Vec<Message>, |
| 315 | config: AdvisorConfig, |
| 316 | client: CodewhaleClient, |
| 317 | route_config: Config, |
| 318 | session_model: String, |
| 319 | usage_context: AdvisorUsageContext, |
| 320 | guard: std::sync::Arc<tokio::sync::Mutex<EmissionGuard>>, |
| 321 | tx_event: mpsc::Sender<Event>, |
| 322 | ) { |
| 323 | // Pre-flight: skip if the guard already blocks (avoids the LLM call when |
| 324 | // rate-limited, which is the common case for rapid turn sequences). |
| 325 | { |
| 326 | let g = guard.lock().await; |
| 327 | // We don't have the note content yet, so we only check the rate limit |
| 328 | // here by testing with a placeholder. The dedup check runs after the |
| 329 | // LLM call, when we have the actual content. |
| 330 | if let Some(last) = g.last_emission |
| 331 | && std::time::Instant::now().duration_since(last) < config.rate_limit |
| 332 | { |
| 333 | debug!(target: "advisor", "rate-limited, skipping advisor run for turn {turn_id}"); |
| 334 | return; |
| 335 | } |
| 336 | } |
| 337 | |
| 338 | // Extract a bounded slice of tool-call pairs. |
| 339 | let pairs = extract_tool_call_pairs(&messages, config.max_tool_calls as usize); |
| 340 | if pairs.is_empty() { |
| 341 | debug!(target: "advisor", "no tool calls found; skipping advisor for turn {turn_id}"); |
| 342 | return; |
| 343 | } |
| 344 | |
| 345 | let tool_call_count = pairs.len() as u32; |
| 346 | let prompt = build_advisor_prompt(&pairs); |
| 347 | let model = config |
| 348 | .model |
| 349 | .clone() |
| 350 | .unwrap_or_else(|| session_model.clone()); |
| 351 | |
| 352 | let (client, model) = match exact_advisor_client(&route_config, client, &session_model, &model) |
| 353 | { |
| 354 | Ok(route) => route, |
| 355 | Err(error) => { |
| 356 | tracing::warn!(target: "advisor", "advisor route resolution failed for turn {turn_id}: {error}"); |
| 357 | return; |
| 358 | } |
| 359 | }; |
| 360 | let route = client.effective_route_envelope(&model, chrono::Utc::now()); |
| 361 | let request = MessageRequest { |
| 362 | model: model.clone(), |
| 363 | messages: vec![Message { |
| 364 | role: Role::User, |
| 365 | content: vec![ContentBlock::Text { |
| 366 | text: prompt, |
| 367 | cache_control: None, |
| 368 | }], |
| 369 | }], |
| 370 | max_tokens: ADVISOR_MAX_TOKENS, |
| 371 | system: Some(SystemPrompt::Text(ADVISOR_SYSTEM_PROMPT.to_string())), |
| 372 | tools: None, |
| 373 | tool_choice: None, |
| 374 | metadata: None, |
| 375 | thinking: None, |
| 376 | // The advisor has a deliberately tiny answer contract. Hidden |
| 377 | // reasoning would spend that allowance before the note is emitted. |
| 378 | reasoning_effort: Some("off".to_string()), |
| 379 | stream: Some(false), |
| 380 | temperature: None, |
| 381 | top_p: None, |
| 382 | }; |
| 383 | |
| 384 | let response = match client.create_message(request).await { |
| 385 | Ok(r) => r, |
| 386 | Err(e) => { |
| 387 | tracing::warn!(target: "advisor", "advisor LLM call failed for turn {turn_id}: {e}"); |
| 388 | return; |
| 389 | } |
| 390 | }; |
| 391 | |
| 392 | // A decoded provider response is billable even when its partial/empty |
| 393 | // content is rejected below or the emission guard suppresses a duplicate. |
| 394 | let usage_source_id = format!("advisor:{turn_id}:provider-response:0"); |
| 395 | if response.usage == codewhale_models::Usage::default() { |
| 396 | usage_context.report_unreceipted(&usage_source_id, &route); |
| 397 | tracing::warn!( |
| 398 | target: "advisor", |
| 399 | "advisor provider response omitted usage for turn {turn_id}; cost coverage is unknown" |
| 400 | ); |
| 401 | } else { |
| 402 | usage_context.report(&usage_source_id, &route, &response.usage); |
| 403 | } |
| 404 | |
| 405 | if codewhale_models::is_incomplete_stop_reason(response.stop_reason.as_deref()) { |
| 406 | tracing::warn!( |
| 407 | target: "advisor", |
| 408 | "advisor response incomplete for turn {turn_id} (stop reason `{}`); dropping partial note", |
| 409 | codewhale_models::stop_reason_detail(response.stop_reason.as_deref()) |
| 410 | ); |
| 411 | return; |
| 412 | } |
| 413 | |
| 414 | // Extract the text from the response. |
| 415 | let note: String = response |
| 416 | .content |
| 417 | .iter() |
| 418 | .filter_map(|block| { |
| 419 | if let ContentBlock::Text { text, .. } = block { |
| 420 | Some(text.as_str()) |
| 421 | } else { |
| 422 | None |
| 423 | } |
| 424 | }) |
| 425 | .collect::<Vec<_>>() |
| 426 | .join("\n") |
| 427 | .trim() |
| 428 | .to_string(); |
| 429 | |
| 430 | if note.is_empty() { |
| 431 | debug!(target: "advisor", "empty advisor response for turn {turn_id}; skipping"); |
| 432 | return; |
| 433 | } |
| 434 | |
| 435 | // Post-flight emission check (rate limit + dedup). |
| 436 | let mut guard_lock = guard.lock().await; |
| 437 | if !guard_lock.may_emit(¬e, &config) { |
| 438 | debug!(target: "advisor", "emission suppressed by guard for turn {turn_id}"); |
| 439 | return; |
| 440 | } |
| 441 | |
| 442 | guard_lock.record_emission(¬e); |
| 443 | drop(guard_lock); |
| 444 | |
| 445 | let _ = tx_event |
| 446 | .send(Event::AdvisoryNote { |
| 447 | turn_id: turn_id.clone(), |
| 448 | note: note.clone(), |
| 449 | tool_call_count, |
| 450 | }) |
| 451 | .await; |
| 452 | |
| 453 | debug!(target: "advisor", "advisory note emitted for turn {turn_id} ({tool_call_count} tool calls reviewed)"); |
| 454 | } |
| 455 | |
| 456 | fn exact_advisor_client( |
| 457 | config: &Config, |
| 458 | parent_client: CodewhaleClient, |
| 459 | session_model: &str, |
| 460 | requested_model: &str, |
| 461 | ) -> anyhow::Result<(CodewhaleClient, String)> { |
| 462 | if requested_model |
| 463 | .trim() |
| 464 | .eq_ignore_ascii_case(session_model.trim()) |
| 465 | { |
| 466 | return Ok((parent_client, session_model.trim().to_string())); |
| 467 | } |
| 468 | |
| 469 | if config.providers.as_ref().is_some_and(|providers| { |
| 470 | providers.custom.values().any(|provider| { |
| 471 | provider |
| 472 | .model |
| 473 | .as_deref() |
| 474 | .is_some_and(|model| model.trim().eq_ignore_ascii_case(requested_model.trim())) |
| 475 | }) |
| 476 | }) { |
| 477 | anyhow::bail!( |
| 478 | "advisor model `{}` belongs to a custom provider but no exact provider identity is carried", |
| 479 | requested_model.trim() |
| 480 | ); |
| 481 | } |
| 482 | |
| 483 | let selection = |
| 484 | crate::model_routing::resolve_explicit_route_with_inventory(config, requested_model); |
| 485 | let (provider, model) = if let Some(selection) = selection { |
| 486 | if selection.provider == crate::config::ApiProvider::Custom { |
| 487 | anyhow::bail!( |
| 488 | "advisor model `{}` resolved only to a custom provider kind without an exact provider identity", |
| 489 | requested_model.trim() |
| 490 | ); |
| 491 | } |
| 492 | (selection.provider, selection.model) |
| 493 | } else { |
| 494 | let candidates = |
| 495 | crate::model_routing::explicit_route_candidate_providers(config, requested_model); |
| 496 | if !candidates.is_empty() && !candidates.contains(&config.api_provider()) { |
| 497 | anyhow::bail!( |
| 498 | "advisor model `{}` is not owned by the originating provider and has no unique exact route", |
| 499 | requested_model.trim() |
| 500 | ); |
| 501 | } |
| 502 | (config.api_provider(), requested_model.trim().to_string()) |
| 503 | }; |
| 504 | let client = crate::route_runtime::resolve_runtime_route(config, provider, Some(&model)) |
| 505 | .map_err(anyhow::Error::msg)? |
| 506 | .validate() |
| 507 | .map(|route| route.client) |
| 508 | .map_err(anyhow::Error::msg)?; |
| 509 | Ok((client, model)) |
| 510 | } |
| 511 | |
| 512 | fn hash_str(s: &str) -> u64 { |
| 513 | let mut h = DefaultHasher::new(); |
| 514 | s.hash(&mut h); |
| 515 | h.finish() |
| 516 | } |
| 517 | |
| 518 | // ── Tests ────────────────────────────────────────────────────────────────── |
| 519 | |
| 520 | #[cfg(test)] |
| 521 | mod tests { |
| 522 | use super::*; |
| 523 | use crate::config::{ProviderConfig, ProvidersConfig}; |
| 524 | use std::time::Duration; |
| 525 | use wiremock::matchers::{method, path}; |
| 526 | use wiremock::{Mock, MockServer, ResponseTemplate}; |
| 527 | |
| 528 | fn test_config() -> AdvisorConfig { |
| 529 | AdvisorConfig { |
| 530 | enabled: true, |
| 531 | max_tool_calls: 5, |
| 532 | rate_limit: Duration::from_secs(1), |
| 533 | dedup_window: Duration::from_secs(10), |
| 534 | model: None, |
| 535 | } |
| 536 | } |
| 537 | |
| 538 | // ── enable/disable ──────────────────────────────────────────────────── |
| 539 | |
| 540 | #[test] |
| 541 | fn disabled_config_has_enabled_false() { |
| 542 | let cfg = AdvisorConfig::disabled(); |
| 543 | assert!(!cfg.enabled); |
| 544 | } |
| 545 | |
| 546 | #[test] |
| 547 | fn from_toml_clamps_max_tool_calls() { |
| 548 | let toml = AdvisorConfigToml { |
| 549 | enabled: true, |
| 550 | max_tool_calls: 999, |
| 551 | rate_limit_secs: 60, |
| 552 | dedup_window_secs: 300, |
| 553 | model: None, |
| 554 | }; |
| 555 | let cfg = AdvisorConfig::from_toml(&toml); |
| 556 | assert_eq!( |
| 557 | cfg.max_tool_calls, 50, |
| 558 | "max_tool_calls must be clamped to 50" |
| 559 | ); |
| 560 | } |
| 561 | |
| 562 | #[test] |
| 563 | fn from_toml_clamps_rate_limit() { |
| 564 | let toml = AdvisorConfigToml { |
| 565 | enabled: true, |
| 566 | max_tool_calls: 10, |
| 567 | rate_limit_secs: 0, // below minimum of 5 |
| 568 | dedup_window_secs: 300, |
| 569 | model: None, |
| 570 | }; |
| 571 | let cfg = AdvisorConfig::from_toml(&toml); |
| 572 | assert!( |
| 573 | cfg.rate_limit >= Duration::from_secs(5), |
| 574 | "rate_limit must be at least 5s" |
| 575 | ); |
| 576 | } |
| 577 | |
| 578 | // ── bounded input ───────────────────────────────────────────────────── |
| 579 | |
| 580 | fn make_messages_with_n_tool_calls(n: usize) -> Vec<Message> { |
| 581 | let mut messages = Vec::new(); |
| 582 | for i in 0..n { |
| 583 | let id = format!("tool_{i}"); |
| 584 | // assistant message with ToolUse |
| 585 | messages.push(Message { |
| 586 | role: Role::Assistant, |
| 587 | content: vec![ContentBlock::ToolUse { |
| 588 | id: id.clone(), |
| 589 | name: "exec_shell".to_string(), |
| 590 | input: serde_json::json!({"command": format!("echo {i}")}), |
| 591 | caller: None, |
| 592 | thought_signature: None, |
| 593 | }], |
| 594 | }); |
| 595 | // user message with ToolResult |
| 596 | messages.push(Message { |
| 597 | role: Role::User, |
| 598 | content: vec![ContentBlock::ToolResult { |
| 599 | tool_use_id: id, |
| 600 | content: format!("{i}"), |
| 601 | is_error: None, |
| 602 | content_blocks: None, |
| 603 | }], |
| 604 | }); |
| 605 | } |
| 606 | messages |
| 607 | } |
| 608 | |
| 609 | #[test] |
| 610 | fn extract_tool_call_pairs_bounded_by_max() { |
| 611 | let messages = make_messages_with_n_tool_calls(20); |
| 612 | let pairs = extract_tool_call_pairs(&messages, 5); |
| 613 | assert_eq!(pairs.len(), 5, "must return at most max_pairs"); |
| 614 | // Should be the last 5 (newest). |
| 615 | assert_eq!(pairs[0].name, "exec_shell"); |
| 616 | } |
| 617 | |
| 618 | #[test] |
| 619 | fn extract_tool_call_pairs_empty_when_no_tool_calls() { |
| 620 | let messages = vec![Message { |
| 621 | role: Role::User, |
| 622 | content: vec![ContentBlock::Text { |
| 623 | text: "hello".to_string(), |
| 624 | cache_control: None, |
| 625 | }], |
| 626 | }]; |
| 627 | let pairs = extract_tool_call_pairs(&messages, 5); |
| 628 | assert!(pairs.is_empty()); |
| 629 | } |
| 630 | |
| 631 | #[test] |
| 632 | fn extract_tool_call_pairs_fewer_than_max_returns_all() { |
| 633 | let messages = make_messages_with_n_tool_calls(3); |
| 634 | let pairs = extract_tool_call_pairs(&messages, 10); |
| 635 | assert_eq!(pairs.len(), 3); |
| 636 | } |
| 637 | |
| 638 | // ── rate limiting ───────────────────────────────────────────────────── |
| 639 | |
| 640 | #[test] |
| 641 | fn emission_guard_allows_first_emission() { |
| 642 | let guard = EmissionGuard::new(); |
| 643 | let config = test_config(); |
| 644 | assert!( |
| 645 | guard.may_emit("something concerning here", &config), |
| 646 | "first emission must be allowed" |
| 647 | ); |
| 648 | } |
| 649 | |
| 650 | #[test] |
| 651 | fn emission_guard_blocks_immediately_after_emission() { |
| 652 | let mut guard = EmissionGuard::new(); |
| 653 | let config = test_config(); |
| 654 | let note = "something concerning"; |
| 655 | guard.record_emission(note); |
| 656 | assert!( |
| 657 | !guard.may_emit("a completely different note", &config), |
| 658 | "emission must be blocked immediately after a prior emission (rate limit)" |
| 659 | ); |
| 660 | } |
| 661 | |
| 662 | #[test] |
| 663 | fn emission_guard_allows_after_rate_limit_expires() { |
| 664 | let mut guard = EmissionGuard::new(); |
| 665 | // Rate limit of 0ms — always expired. |
| 666 | let config = AdvisorConfig { |
| 667 | rate_limit: Duration::ZERO, |
| 668 | dedup_window: Duration::from_secs(300), |
| 669 | ..AdvisorConfig::disabled() |
| 670 | }; |
| 671 | let note = "first note"; |
| 672 | guard.record_emission(note); |
| 673 | assert!( |
| 674 | guard.may_emit("second different note", &config), |
| 675 | "emission must be allowed when rate limit duration is zero" |
| 676 | ); |
| 677 | } |
| 678 | |
| 679 | // ── deduplication ───────────────────────────────────────────────────── |
| 680 | |
| 681 | #[test] |
| 682 | fn emission_guard_suppresses_ok_response() { |
| 683 | let guard = EmissionGuard::new(); |
| 684 | let config = test_config(); |
| 685 | assert!(!guard.may_emit("ok", &config), "\"ok\" must be suppressed"); |
| 686 | assert!(!guard.may_emit("OK", &config), "\"OK\" must be suppressed"); |
| 687 | assert!( |
| 688 | !guard.may_emit(" ok ", &config), |
| 689 | "\" ok \" must be suppressed" |
| 690 | ); |
| 691 | } |
| 692 | |
| 693 | #[test] |
| 694 | fn emission_guard_dedup_blocks_identical_note_within_window() { |
| 695 | let mut guard = EmissionGuard::new(); |
| 696 | // Use a zero rate limit so only dedup is tested. |
| 697 | let config = AdvisorConfig { |
| 698 | rate_limit: Duration::ZERO, |
| 699 | dedup_window: Duration::from_secs(300), |
| 700 | ..AdvisorConfig::disabled() |
| 701 | }; |
| 702 | let note = "risky shell command with no error checking"; |
| 703 | guard.record_emission(note); |
| 704 | assert!( |
| 705 | !guard.may_emit(note, &config), |
| 706 | "identical note must be suppressed within the dedup window" |
| 707 | ); |
| 708 | } |
| 709 | |
| 710 | #[test] |
| 711 | fn emission_guard_allows_different_note_within_dedup_window() { |
| 712 | let mut guard = EmissionGuard::new(); |
| 713 | let config = AdvisorConfig { |
| 714 | rate_limit: Duration::ZERO, |
| 715 | dedup_window: Duration::from_secs(300), |
| 716 | ..AdvisorConfig::disabled() |
| 717 | }; |
| 718 | guard.record_emission("first note"); |
| 719 | assert!( |
| 720 | guard.may_emit("entirely different note", &config), |
| 721 | "a different note must be allowed even within the dedup window" |
| 722 | ); |
| 723 | } |
| 724 | |
| 725 | // ── child failure isolation ─────────────────────────────────────────── |
| 726 | |
| 727 | #[test] |
| 728 | fn advisor_prompt_is_non_empty_for_non_empty_pairs() { |
| 729 | let pairs = vec![ToolCallPair { |
| 730 | name: "exec_shell".to_string(), |
| 731 | input_preview: r#"{"command":"ls -la"}"#.to_string(), |
| 732 | result_preview: "total 4\ndrwxr-xr-x 2 user user 4096".to_string(), |
| 733 | }]; |
| 734 | let prompt = build_advisor_prompt(&pairs); |
| 735 | assert!( |
| 736 | prompt.contains("exec_shell"), |
| 737 | "prompt must include the tool name" |
| 738 | ); |
| 739 | assert!( |
| 740 | prompt.contains("ls -la"), |
| 741 | "prompt must include the tool input" |
| 742 | ); |
| 743 | } |
| 744 | |
| 745 | #[test] |
| 746 | fn advisor_model_override_builds_the_owning_provider_client() { |
| 747 | let config = Config { |
| 748 | provider: Some("deepseek".to_string()), |
| 749 | providers: Some(ProvidersConfig { |
| 750 | deepseek: ProviderConfig { |
| 751 | api_key: Some("sk-deepseek-advisor-test".to_string()), |
| 752 | model: Some("deepseek-chat".to_string()), |
| 753 | ..ProviderConfig::default() |
| 754 | }, |
| 755 | zai: ProviderConfig { |
| 756 | api_key: Some("zai-advisor-test-key".to_string()), |
| 757 | model: Some(crate::config::DEFAULT_ZAI_MODEL.to_string()), |
| 758 | ..ProviderConfig::default() |
| 759 | }, |
| 760 | ..ProvidersConfig::default() |
| 761 | }), |
| 762 | ..Config::default() |
| 763 | }; |
| 764 | let parent = CodewhaleClient::new(&config).expect("parent client"); |
| 765 | let (advisor, resolved_model) = exact_advisor_client( |
| 766 | &config, |
| 767 | parent, |
| 768 | "deepseek-chat", |
| 769 | crate::config::DEFAULT_ZAI_MODEL, |
| 770 | ) |
| 771 | .expect("cross-provider advisor route"); |
| 772 | let route = advisor.effective_route_envelope(&resolved_model, chrono::Utc::now()); |
| 773 | |
| 774 | assert_eq!(route.provider, crate::config::ApiProvider::Zai); |
| 775 | assert_eq!(route.provider_identity, "zai"); |
| 776 | assert_eq!(route.model, crate::config::DEFAULT_ZAI_MODEL); |
| 777 | } |
| 778 | |
| 779 | #[test] |
| 780 | fn advisor_foreign_custom_override_fails_closed_without_exact_identity() { |
| 781 | let config = Config { |
| 782 | provider: Some("deepseek".to_string()), |
| 783 | providers: Some(ProvidersConfig { |
| 784 | deepseek: ProviderConfig { |
| 785 | api_key: Some("sk-deepseek-advisor-test".to_string()), |
| 786 | model: Some("deepseek-chat".to_string()), |
| 787 | ..ProviderConfig::default() |
| 788 | }, |
| 789 | custom: [( |
| 790 | "private-route".to_string(), |
| 791 | ProviderConfig { |
| 792 | api_key: Some("custom-advisor-test-key".to_string()), |
| 793 | base_url: Some("https://custom.invalid/v1".to_string()), |
| 794 | model: Some("private-advisor-model".to_string()), |
| 795 | ..ProviderConfig::default() |
| 796 | }, |
| 797 | )] |
| 798 | .into_iter() |
| 799 | .collect(), |
| 800 | ..ProvidersConfig::default() |
| 801 | }), |
| 802 | ..Config::default() |
| 803 | }; |
| 804 | let parent = CodewhaleClient::new(&config).expect("parent client"); |
| 805 | let error = |
| 806 | match exact_advisor_client(&config, parent, "deepseek-chat", "private-advisor-model") { |
| 807 | Ok(_) => panic!("generic custom kind cannot identify the exact foreign route"), |
| 808 | Err(error) => error, |
| 809 | }; |
| 810 | assert!( |
| 811 | error.to_string().contains("exact provider identity"), |
| 812 | "{error}" |
| 813 | ); |
| 814 | } |
| 815 | |
| 816 | #[test] |
| 817 | fn advisor_named_custom_a_cannot_route_model_owned_by_custom_b() { |
| 818 | let config = Config { |
| 819 | provider: Some("custom-a".to_string()), |
| 820 | providers: Some(ProvidersConfig { |
| 821 | custom: [ |
| 822 | ( |
| 823 | "custom-a".to_string(), |
| 824 | ProviderConfig { |
| 825 | api_key: Some("custom-a-advisor-test-key".to_string()), |
| 826 | base_url: Some("https://custom-a.invalid/v1".to_string()), |
| 827 | model: Some("custom-a-model".to_string()), |
| 828 | kind: Some("openai-compatible".to_string()), |
| 829 | ..ProviderConfig::default() |
| 830 | }, |
| 831 | ), |
| 832 | ( |
| 833 | "custom-b".to_string(), |
| 834 | ProviderConfig { |
| 835 | api_key: Some("custom-b-advisor-test-key".to_string()), |
| 836 | base_url: Some("https://custom-b.invalid/v1".to_string()), |
| 837 | model: Some("custom-b-model".to_string()), |
| 838 | kind: Some("openai-compatible".to_string()), |
| 839 | ..ProviderConfig::default() |
| 840 | }, |
| 841 | ), |
| 842 | ] |
| 843 | .into_iter() |
| 844 | .collect(), |
| 845 | ..ProvidersConfig::default() |
| 846 | }), |
| 847 | ..Config::default() |
| 848 | }; |
| 849 | let parent = CodewhaleClient::new(&config).expect("active custom-a client"); |
| 850 | let error = match exact_advisor_client(&config, parent, "custom-a-model", "custom-b-model") |
| 851 | { |
| 852 | Ok(_) => panic!("custom-b must not reuse custom-a's endpoint or credential"), |
| 853 | Err(error) => error, |
| 854 | }; |
| 855 | assert!( |
| 856 | error.to_string().contains("exact provider identity"), |
| 857 | "{error}" |
| 858 | ); |
| 859 | } |
| 860 | |
| 861 | async fn run_billed_advisor_fixture( |
| 862 | note: &str, |
| 863 | stop_reason: &str, |
| 864 | suppress_as_duplicate: bool, |
| 865 | include_usage: bool, |
| 866 | ) -> (crate::cost_status::PendingBackgroundCost, Option<Event>) { |
| 867 | let _scope = crate::cost_status::test_scope(); |
| 868 | let server = MockServer::start().await; |
| 869 | let mut provider_response = serde_json::json!({ |
| 870 | "id": "advisor-provider-response", |
| 871 | "model": "deepseek-chat", |
| 872 | "choices": [{ |
| 873 | "index": 0, |
| 874 | "message": {"role": "assistant", "content": note}, |
| 875 | "finish_reason": stop_reason |
| 876 | }] |
| 877 | }); |
| 878 | if include_usage { |
| 879 | provider_response["usage"] = serde_json::json!({ |
| 880 | "prompt_tokens": 9, |
| 881 | "completion_tokens": 3, |
| 882 | "total_tokens": 12 |
| 883 | }); |
| 884 | } |
| 885 | Mock::given(method("POST")) |
| 886 | .and(path("/v1/chat/completions")) |
| 887 | .respond_with(ResponseTemplate::new(200).set_body_json(provider_response)) |
| 888 | .expect(1) |
| 889 | .mount(&server) |
| 890 | .await; |
| 891 | let route_config = Config { |
| 892 | provider: Some("deepseek".to_string()), |
| 893 | providers: Some(ProvidersConfig { |
| 894 | deepseek: ProviderConfig { |
| 895 | api_key: Some("sk-deepseek-advisor-test".to_string()), |
| 896 | model: Some("deepseek-chat".to_string()), |
| 897 | ..ProviderConfig::default() |
| 898 | }, |
| 899 | ..ProvidersConfig::default() |
| 900 | }), |
| 901 | ..Config::default() |
| 902 | }; |
| 903 | let mut client = CodewhaleClient::new(&route_config).expect("advisor client"); |
| 904 | client.set_test_chat_transport_base_url(server.uri()); |
| 905 | let mut emission_guard = EmissionGuard::new(); |
| 906 | if suppress_as_duplicate { |
| 907 | emission_guard.record_emission(note); |
| 908 | } |
| 909 | let guard = std::sync::Arc::new(tokio::sync::Mutex::new(emission_guard)); |
| 910 | let (tx, mut rx) = mpsc::channel(1); |
| 911 | run_advisor_for_turn( |
| 912 | "advisor-turn".to_string(), |
| 913 | make_messages_with_n_tool_calls(1), |
| 914 | AdvisorConfig { |
| 915 | enabled: true, |
| 916 | max_tool_calls: 5, |
| 917 | rate_limit: Duration::ZERO, |
| 918 | dedup_window: Duration::from_secs(60), |
| 919 | model: None, |
| 920 | }, |
| 921 | client, |
| 922 | route_config, |
| 923 | "deepseek-chat".to_string(), |
| 924 | AdvisorUsageContext::capture(None), |
| 925 | guard, |
| 926 | tx, |
| 927 | ) |
| 928 | .await; |
| 929 | (crate::cost_status::drain(), rx.try_recv().ok()) |
| 930 | } |
| 931 | |
| 932 | #[tokio::test] |
| 933 | async fn advisor_incomplete_and_dedup_suppressed_responses_are_each_billed_once() { |
| 934 | let (incomplete, incomplete_event) = |
| 935 | run_billed_advisor_fixture("partial note", "max_tokens", false, true).await; |
| 936 | assert!(incomplete_event.is_none()); |
| 937 | assert_eq!( |
| 938 | incomplete |
| 939 | .priced_turns |
| 940 | .saturating_add(incomplete.unpriced_turns), |
| 941 | 1 |
| 942 | ); |
| 943 | |
| 944 | let (dedup, dedup_event) = |
| 945 | run_billed_advisor_fixture("same advisory", "stop", true, true).await; |
| 946 | assert!(dedup_event.is_none()); |
| 947 | assert_eq!(dedup.priced_turns.saturating_add(dedup.unpriced_turns), 1); |
| 948 | } |
| 949 | |
| 950 | #[tokio::test] |
| 951 | async fn advisor_provider_success_without_usage_marks_unknown_once() { |
| 952 | let (pending, event) = |
| 953 | run_billed_advisor_fixture("use a smaller focused slice", "stop", false, false).await; |
| 954 | |
| 955 | assert!( |
| 956 | event.is_some(), |
| 957 | "the semantic advisor response remains usable" |
| 958 | ); |
| 959 | assert_eq!(pending.priced_turns, 0); |
| 960 | assert_eq!(pending.unpriced_turns, 1); |
| 961 | assert_eq!(pending.cny_unpriced_turns, 1); |
| 962 | assert!( |
| 963 | pending |
| 964 | .unpriced_reasons |
| 965 | .contains("provider_success_missing_usage") |
| 966 | ); |
| 967 | } |
| 968 | } |
| 969 |