| 1 | //! Deferred tool catalog and built-in advanced tool helpers. |
| 2 | //! |
| 3 | //! The streaming turn loop owns when tools are offered or executed. This module |
| 4 | //! owns the catalog-level policy around deferred loading, tool search, missing |
| 5 | //! tool suggestions, and the small set of built-in advanced tools that are not |
| 6 | //! registered by the normal runtime tool registry. |
| 7 | |
| 8 | use std::collections::HashSet; |
| 9 | use std::path::Path; |
| 10 | use std::time::Duration; |
| 11 | |
| 12 | use serde_json::{Value, json}; |
| 13 | |
| 14 | #[cfg(test)] |
| 15 | use crate::mcp::McpPool; |
| 16 | use crate::model_profile::ToolSurfaceBudget; |
| 17 | use crate::models::Tool; |
| 18 | use crate::tools::spec::{ToolError, ToolResult, optional_str, optional_u64, required_str}; |
| 19 | use crate::tui::app::AppMode; |
| 20 | |
| 21 | use crate::dependencies::ExternalTool; |
| 22 | use crate::regex_cache::compile_user_regex; |
| 23 | |
| 24 | pub(super) const MULTI_TOOL_PARALLEL_NAME: &str = "multi_tool_use.parallel"; |
| 25 | pub(super) const REQUEST_USER_INPUT_NAME: &str = "request_user_input"; |
| 26 | pub(super) const CODE_EXECUTION_TOOL_NAME: &str = "code_execution"; |
| 27 | const CODE_EXECUTION_TOOL_TYPE: &str = "code_execution_20250825"; |
| 28 | const CODE_EXECUTION_DESCRIPTION: &str = "Execute Python code with the local Python interpreter in the workspace and return stdout/stderr/return_code as JSON."; |
| 29 | pub(super) use crate::tools::js_execution::JS_EXECUTION_TOOL_NAME; |
| 30 | pub(super) const TOOL_SEARCH_NAME: &str = "tool_search"; |
| 31 | const TOOL_SEARCH_TYPE: &str = "tool_search_20251119"; |
| 32 | const LEGACY_TOOL_SEARCH_REGEX_NAME: &str = "tool_search_tool_regex"; |
| 33 | const LEGACY_TOOL_SEARCH_BM25_NAME: &str = "tool_search_tool_bm25"; |
| 34 | const TOOL_SEARCH_DEFAULT_MAX_RESULTS: usize = 20; |
| 35 | const TOOL_SEARCH_MAX_RESULTS_LIMIT: usize = 100; |
| 36 | |
| 37 | pub(super) fn is_tool_search_tool(name: &str) -> bool { |
| 38 | matches!( |
| 39 | name, |
| 40 | TOOL_SEARCH_NAME | LEGACY_TOOL_SEARCH_REGEX_NAME | LEGACY_TOOL_SEARCH_BM25_NAME |
| 41 | ) |
| 42 | } |
| 43 | |
| 44 | // Crate-visible rather than `pub(super)` so the hook gate's classifier test can |
| 45 | // assert it still recognises every default-active name. Without that anchor the |
| 46 | // test pins hardcoded strings and stays green through a tool rename, while the |
| 47 | // gate silently reclassifies the renamed tool as unknown. |
| 48 | pub(crate) const DEFAULT_ACTIVE_NATIVE_TOOLS: &[&str] = &[ |
| 49 | // #4625: the model-facing shell tool is `Bash`; legacy `exec_shell*` |
| 50 | // names are hidden compat aliases and must not be default-active. |
| 51 | "Bash", |
| 52 | "File", |
| 53 | "Git", |
| 54 | "Run", |
| 55 | "agent", |
| 56 | // Skills use a bounded ambient index. Keep the tiny discovery/load schema |
| 57 | // active so omitted skills remain one-call discoverable on the first turn. |
| 58 | "load_skill", |
| 59 | "remember", |
| 60 | // Piagent phase B: the model-facing durable-task tool is `tasks`; the |
| 61 | // legacy `task_create`/`task_list`/`task_read` names it replaces are |
| 62 | // hidden compat aliases and must not be default-active. |
| 63 | "tasks", |
| 64 | "work_update", |
| 65 | ]; |
| 66 | |
| 67 | const CORE_ACTION_TOOL_FALLBACKS: &[CoreActionToolFallback] = &[ |
| 68 | CoreActionToolFallback { |
| 69 | name: "Bash", |
| 70 | description: "Run shell commands in the workspace.", |
| 71 | unavailable_reason: "Not present in the current model-visible catalog. Interactive Agent sessions expose shell by default unless allow_shell = false; noninteractive and durable profiles require allow_shell = true. Plan mode hides shell, and command tool allow/deny gates can also block it.", |
| 72 | }, |
| 73 | CoreActionToolFallback { |
| 74 | name: "File", |
| 75 | description: "Read, search, and modify workspace files.", |
| 76 | unavailable_reason: "Not present in the current model-visible catalog. File reads are available in Plan and Agent modes; write and edit actions require an executable mode, while patch also requires the apply_patch feature.", |
| 77 | }, |
| 78 | ]; |
| 79 | |
| 80 | #[derive(Debug, Clone, Copy)] |
| 81 | struct CoreActionToolFallback { |
| 82 | name: &'static str, |
| 83 | description: &'static str, |
| 84 | unavailable_reason: &'static str, |
| 85 | } |
| 86 | |
| 87 | /// Pre-computed lowercased haystack + name for each fallback; built once. |
| 88 | struct CachedFallback { |
| 89 | fallback: CoreActionToolFallback, |
| 90 | haystack: String, |
| 91 | name_lower: String, |
| 92 | } |
| 93 | |
| 94 | static CACHED_FALLBACKS: std::sync::OnceLock<Vec<CachedFallback>> = std::sync::OnceLock::new(); |
| 95 | |
| 96 | fn cached_fallbacks() -> &'static [CachedFallback] { |
| 97 | CACHED_FALLBACKS.get_or_init(|| { |
| 98 | CORE_ACTION_TOOL_FALLBACKS |
| 99 | .iter() |
| 100 | .map(|f| CachedFallback { |
| 101 | fallback: *f, |
| 102 | haystack: format!( |
| 103 | "{}\n{}\n{}", |
| 104 | f.name.to_lowercase(), |
| 105 | f.description.to_lowercase(), |
| 106 | f.unavailable_reason.to_lowercase(), |
| 107 | ), |
| 108 | name_lower: f.name.to_lowercase(), |
| 109 | }) |
| 110 | .collect() |
| 111 | }) |
| 112 | } |
| 113 | |
| 114 | /// Membership index over [`DEFAULT_ACTIVE_NATIVE_TOOLS`], built once for the |
| 115 | /// process lifetime. The array stays the source of truth for ordered |
| 116 | /// inspection; this set only accelerates the |
| 117 | /// hot membership check in [`should_default_defer_tool`], which runs once per |
| 118 | /// catalog tool on every catalog rebuild (i.e. per turn) — an O(n·m) linear |
| 119 | /// scan over the array collapses to O(1) hashed lookups. |
| 120 | static DEFAULT_ACTIVE_NATIVE_TOOLS_SET: std::sync::OnceLock<HashSet<&'static str>> = |
| 121 | std::sync::OnceLock::new(); |
| 122 | |
| 123 | fn default_active_native_tools_set() -> &'static HashSet<&'static str> { |
| 124 | DEFAULT_ACTIVE_NATIVE_TOOLS_SET |
| 125 | .get_or_init(|| DEFAULT_ACTIVE_NATIVE_TOOLS.iter().copied().collect()) |
| 126 | } |
| 127 | |
| 128 | pub(super) fn should_default_defer_tool(name: &str, always_load: &HashSet<String>) -> bool { |
| 129 | if always_load.contains(name) { |
| 130 | return false; |
| 131 | } |
| 132 | |
| 133 | if is_tool_search_tool(name) { |
| 134 | return false; |
| 135 | } |
| 136 | |
| 137 | // Membership-only test (no ordering dependency): the side set built from |
| 138 | // DEFAULT_ACTIVE_NATIVE_TOOLS returns identical hit/miss results as the |
| 139 | // former `.iter().any(...)` linear scan. |
| 140 | !default_active_native_tools_set().contains(name) |
| 141 | } |
| 142 | |
| 143 | pub(super) fn apply_native_tool_deferral(catalog: &mut [Tool], always_load: &HashSet<String>) { |
| 144 | for tool in catalog { |
| 145 | tool.defer_loading = Some(should_default_defer_tool(&tool.name, always_load)); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | fn should_keep_mcp_tool_loaded(name: &str) -> bool { |
| 150 | matches!( |
| 151 | name, |
| 152 | "list_mcp_resources" |
| 153 | | "list_mcp_resource_templates" |
| 154 | | "mcp_read_resource" |
| 155 | | "read_mcp_resource" |
| 156 | | "mcp_get_prompt" |
| 157 | ) |
| 158 | } |
| 159 | |
| 160 | pub(super) fn apply_mcp_tool_deferral( |
| 161 | catalog: &mut [Tool], |
| 162 | mode: AppMode, |
| 163 | always_load: &HashSet<String>, |
| 164 | ) { |
| 165 | for tool in catalog { |
| 166 | if always_load.contains(&tool.name) { |
| 167 | tool.defer_loading = Some(false); |
| 168 | continue; |
| 169 | } |
| 170 | tool.defer_loading = |
| 171 | Some(mode != AppMode::Yolo && !should_keep_mcp_tool_loaded(&tool.name)); |
| 172 | } |
| 173 | } |
| 174 | |
| 175 | /// Build the model tool catalog from native and MCP tool lists. |
| 176 | /// |
| 177 | /// **Catalog-head stability invariant.** The head of the catalog (all |
| 178 | /// non-deferred tools) must remain byte-identical across mode toggles |
| 179 | /// (Plan ↔ Agent ↔ YOLO) for tools that are common to both modes. |
| 180 | /// Deferred tool activations append to the tail and never reorder the |
| 181 | /// head. This invariant is critical for DeepSeek's KV prefix cache: |
| 182 | /// the tools array is part of the immutable prefix, and any byte-level |
| 183 | /// change in the head forces a full re-prefill on the next turn. |
| 184 | #[cfg(test)] |
| 185 | pub(super) fn build_model_tool_catalog( |
| 186 | native_tools: Vec<Tool>, |
| 187 | mcp_tools: Vec<Tool>, |
| 188 | mode: AppMode, |
| 189 | always_load: &HashSet<String>, |
| 190 | ) -> Vec<Tool> { |
| 191 | build_model_tool_catalog_with_surface( |
| 192 | native_tools, |
| 193 | mcp_tools, |
| 194 | mode, |
| 195 | always_load, |
| 196 | ToolSurfaceBudget::Standard, |
| 197 | ) |
| 198 | } |
| 199 | |
| 200 | pub(super) fn build_model_tool_catalog_with_surface( |
| 201 | mut native_tools: Vec<Tool>, |
| 202 | mut mcp_tools: Vec<Tool>, |
| 203 | mode: AppMode, |
| 204 | always_load: &HashSet<String>, |
| 205 | surface_budget: ToolSurfaceBudget, |
| 206 | ) -> Vec<Tool> { |
| 207 | apply_native_tool_deferral(&mut native_tools, always_load); |
| 208 | apply_mcp_tool_deferral(&mut mcp_tools, mode, always_load); |
| 209 | apply_tool_surface_budget(&mut native_tools, surface_budget, always_load); |
| 210 | apply_tool_surface_budget(&mut mcp_tools, surface_budget, always_load); |
| 211 | // Sort each partition by name for prefix-cache stability (#263). The |
| 212 | // upstream `to_api_tools()` already sorts the registry's HashMap output; |
| 213 | // this catalog is built from caller-supplied Vecs which the test harness |
| 214 | // and (future) caller refactors may not pre-sort. Built-ins stay as a |
| 215 | // contiguous prefix ahead of MCP tools so adding/removing an MCP tool |
| 216 | // never shifts a built-in's position. |
| 217 | native_tools.sort_by(|a, b| a.name.cmp(&b.name)); |
| 218 | mcp_tools.sort_by(|a, b| a.name.cmp(&b.name)); |
| 219 | native_tools.extend(mcp_tools); |
| 220 | native_tools |
| 221 | } |
| 222 | |
| 223 | const REGISTRY_FIRST_SHELL_GUIDANCE: &str = "Before using this tool for a task whose core operation is a specialized capability (for example media or document conversion, data transformation, browser automation, database or service access, or a developer utility), call registry_sync first. If its complete catalog contains any plausible match, call start_registry_mcp_server and inspect the connected tools before using a shell alternative. Use the shell directly for ordinary repo-native work and simple file operations, or after every Registry entry is clearly irrelevant or the matching server fails to start."; |
| 224 | |
| 225 | /// Put the Registry-first decision at the point where the model considers its |
| 226 | /// strongest fallback. The discovery skill body is lazy-loaded, so relying on |
| 227 | /// it alone creates a loop: the model must already prefer discovery before it |
| 228 | /// can read the instruction that tells it to prefer discovery. |
| 229 | /// |
| 230 | /// This is applied only while MCP is enabled. It changes no dispatch order and |
| 231 | /// performs no task matching in the host; the model still compares the user's |
| 232 | /// context against the Registry catalog itself. |
| 233 | pub(super) fn apply_registry_first_shell_guidance(catalog: &mut [Tool]) { |
| 234 | let Some(shell) = catalog.iter_mut().find(|tool| tool.name == "exec_shell") else { |
| 235 | return; |
| 236 | }; |
| 237 | if shell.description.contains(REGISTRY_FIRST_SHELL_GUIDANCE) { |
| 238 | return; |
| 239 | } |
| 240 | if !shell.description.ends_with(char::is_whitespace) { |
| 241 | shell.description.push(' '); |
| 242 | } |
| 243 | shell.description.push_str(REGISTRY_FIRST_SHELL_GUIDANCE); |
| 244 | } |
| 245 | |
| 246 | fn apply_tool_surface_budget( |
| 247 | catalog: &mut [Tool], |
| 248 | surface_budget: ToolSurfaceBudget, |
| 249 | always_load: &HashSet<String>, |
| 250 | ) { |
| 251 | if !matches!(surface_budget, ToolSurfaceBudget::Compact) { |
| 252 | return; |
| 253 | } |
| 254 | for tool in catalog { |
| 255 | if always_load.contains(&tool.name) { |
| 256 | continue; |
| 257 | } |
| 258 | if matches!(tool.name.as_str(), "agent" | "Run" | "tasks" | "Web") { |
| 259 | tool.defer_loading = Some(true); |
| 260 | } |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | /// Whether two tool-surface budgets currently produce the same catalog. |
| 265 | /// |
| 266 | /// Runs [`apply_tool_surface_budget`] over `catalog` under both budgets and |
| 267 | /// compares the results. `/preview-request` publishes the Standard-vs-Full |
| 268 | /// answer as a derived field so the truthful "these are currently collapsed" |
| 269 | /// disclosure cannot drift from the code: the day the shaper narrows Standard |
| 270 | /// differently from Full, this starts returning `false` on its own. |
| 271 | pub(super) fn surface_budgets_produce_same_catalog( |
| 272 | catalog: &[Tool], |
| 273 | always_load: &HashSet<String>, |
| 274 | left_budget: ToolSurfaceBudget, |
| 275 | right_budget: ToolSurfaceBudget, |
| 276 | ) -> bool { |
| 277 | let mut left = catalog.to_vec(); |
| 278 | let mut right = catalog.to_vec(); |
| 279 | apply_tool_surface_budget(&mut left, left_budget, always_load); |
| 280 | apply_tool_surface_budget(&mut right, right_budget, always_load); |
| 281 | serde_json::to_string(&left).ok() == serde_json::to_string(&right).ok() |
| 282 | } |
| 283 | |
| 284 | pub(super) fn ensure_advanced_tooling( |
| 285 | catalog: &mut Vec<Tool>, |
| 286 | mode: AppMode, |
| 287 | always_load: &HashSet<String>, |
| 288 | ) { |
| 289 | // code_execution depends on a locally-installed Python interpreter |
| 290 | // (python3 / python / py -3). Before v0.8.31, the tool was always |
| 291 | // advertised and would fail at execution time on Windows where |
| 292 | // `python3` isn't on PATH — the model treated the tool as reliable |
| 293 | // once it appeared in the catalog. We now probe at catalog-build |
| 294 | // time and only advertise when an interpreter resolves. See |
| 295 | // `crate::dependencies::resolve_python_interpreter` for the probe. |
| 296 | if mode != AppMode::Plan |
| 297 | && !catalog.iter().any(|t| t.name == CODE_EXECUTION_TOOL_NAME) |
| 298 | && crate::dependencies::resolve_python_interpreter().is_some() |
| 299 | { |
| 300 | catalog.push(Tool { |
| 301 | tool_type: Some(CODE_EXECUTION_TOOL_TYPE.to_string()), |
| 302 | name: CODE_EXECUTION_TOOL_NAME.to_string(), |
| 303 | description: CODE_EXECUTION_DESCRIPTION.to_string(), |
| 304 | input_schema: json!({ |
| 305 | "type": "object", |
| 306 | "properties": { |
| 307 | "code": { "type": "string", "description": "Python source code to execute." } |
| 308 | }, |
| 309 | "required": ["code"] |
| 310 | }), |
| 311 | allowed_callers: Some(vec!["direct".to_string()]), |
| 312 | defer_loading: Some(should_default_defer_tool( |
| 313 | CODE_EXECUTION_TOOL_NAME, |
| 314 | always_load, |
| 315 | )), |
| 316 | input_examples: None, |
| 317 | strict: None, |
| 318 | cache_control: None, |
| 319 | }); |
| 320 | } |
| 321 | |
| 322 | // js_execution mirrors code_execution: gate on Node.js being |
| 323 | // present locally so the model never sees a runtime it can't |
| 324 | // actually use. Plan mode hides shell/exec surfaces (including |
| 325 | // both interpreter tools) by construction; Agent / YOLO advertise |
| 326 | // the tool only when `resolve_node()` succeeds. |
| 327 | if mode != AppMode::Plan |
| 328 | && !catalog.iter().any(|t| t.name == JS_EXECUTION_TOOL_NAME) |
| 329 | && crate::dependencies::resolve_node().is_some() |
| 330 | { |
| 331 | let mut tool = crate::tools::js_execution::js_execution_tool_definition(); |
| 332 | tool.defer_loading = Some(should_default_defer_tool(&tool.name, always_load)); |
| 333 | catalog.push(tool); |
| 334 | } |
| 335 | |
| 336 | if !catalog.iter().any(|t| t.name == TOOL_SEARCH_NAME) { |
| 337 | catalog.push(Tool { |
| 338 | tool_type: Some(TOOL_SEARCH_TYPE.to_string()), |
| 339 | name: TOOL_SEARCH_NAME.to_string(), |
| 340 | description: "Search deferred tool definitions and return matching tool references.".to_string(), |
| 341 | input_schema: json!({ |
| 342 | "type": "object", |
| 343 | "properties": { |
| 344 | "query": { "type": "string", "description": "Search query for tool discovery." }, |
| 345 | "match": { |
| 346 | "type": "string", |
| 347 | "enum": ["bm25", "regex"], |
| 348 | "default": "bm25", |
| 349 | "description": "Matching algorithm: bm25 for natural-language matching, regex for a regular expression over tool names/descriptions/schema." |
| 350 | }, |
| 351 | "max_results": { |
| 352 | "type": "integer", |
| 353 | "minimum": 1, |
| 354 | "maximum": TOOL_SEARCH_MAX_RESULTS_LIMIT, |
| 355 | "default": TOOL_SEARCH_DEFAULT_MAX_RESULTS, |
| 356 | "description": "Maximum number of matching tool references to return." |
| 357 | } |
| 358 | }, |
| 359 | "required": ["query"] |
| 360 | }), |
| 361 | allowed_callers: Some(vec!["direct".to_string()]), |
| 362 | defer_loading: Some(false), |
| 363 | input_examples: None, |
| 364 | strict: None, |
| 365 | cache_control: None, |
| 366 | }); |
| 367 | } |
| 368 | } |
| 369 | |
| 370 | pub(super) fn initial_active_tools(catalog: &[Tool]) -> HashSet<String> { |
| 371 | let mut active = HashSet::new(); |
| 372 | for tool in catalog { |
| 373 | if !tool.defer_loading.unwrap_or(false) || is_tool_search_tool(&tool.name) { |
| 374 | active.insert(tool.name.clone()); |
| 375 | } |
| 376 | } |
| 377 | if active.is_empty() |
| 378 | && !catalog.is_empty() |
| 379 | && let Some(first) = catalog.first() |
| 380 | { |
| 381 | active.insert(first.name.clone()); |
| 382 | } |
| 383 | active |
| 384 | } |
| 385 | |
| 386 | fn active_tool_list_from_catalog(catalog: &[Tool], active: &HashSet<String>) -> Vec<Tool> { |
| 387 | // Two-pass for prefix-cache stability (#263). Always-loaded tools come |
| 388 | // first in their stable catalog order; tools that started life deferred |
| 389 | // and were activated mid-conversation by ToolSearch get appended at the |
| 390 | // tail. Otherwise activating a deferred tool shifts every later tool's |
| 391 | // byte offset and busts the cached prefix from that point onwards. |
| 392 | let catalog_len = catalog.len(); |
| 393 | let mut head: Vec<Tool> = Vec::with_capacity(catalog_len); |
| 394 | let mut tail: Vec<Tool> = Vec::with_capacity(catalog_len); |
| 395 | for tool in catalog { |
| 396 | if !active.contains(&tool.name) { |
| 397 | continue; |
| 398 | } |
| 399 | if tool.defer_loading.unwrap_or(false) { |
| 400 | tail.push(tool.clone()); |
| 401 | } else { |
| 402 | head.push(tool.clone()); |
| 403 | } |
| 404 | } |
| 405 | head.extend(tail); |
| 406 | head |
| 407 | } |
| 408 | |
| 409 | pub(super) fn active_tools_for_step(catalog: &[Tool], active: &HashSet<String>) -> Vec<Tool> { |
| 410 | active_tool_list_from_catalog(catalog, active) |
| 411 | } |
| 412 | |
| 413 | /// One turn's executable and model-visible tool contract. |
| 414 | /// |
| 415 | /// The catalog is also the prompt's only availability taxonomy: stable prompt |
| 416 | /// prose deliberately does not enumerate tool names. Keeping the concrete |
| 417 | /// registry, searchable catalog, initial request subset, and command gates in |
| 418 | /// one value prevents preview, dispatch, and execution from reconstructing |
| 419 | /// different surfaces from mutable engine configuration. |
| 420 | pub(super) struct ToolSurfacePolicy { |
| 421 | /// Runtime registry that executes native and plugin tools. |
| 422 | pub(super) registry: crate::tools::ToolRegistry, |
| 423 | /// Full model-facing catalog, including deferred entries. |
| 424 | pub(super) catalog: Vec<Tool>, |
| 425 | /// Names active at the start of the turn. |
| 426 | pub(super) active_names: HashSet<String>, |
| 427 | /// Exact initial `tools` field. `None` means no field is sent. |
| 428 | pub(super) active: Option<Vec<Tool>>, |
| 429 | pub(super) mode: AppMode, |
| 430 | pub(super) strict_tool_mode: bool, |
| 431 | allowed_tools: Option<Vec<String>>, |
| 432 | disallowed_tools: Option<Vec<String>>, |
| 433 | /// Hard per-turn cap on admitted tool calls (#4415). The turn loop copies |
| 434 | /// this limit into its own admission counter at turn start; `None` means |
| 435 | /// unlimited (the default), which keeps the admission gate inert. |
| 436 | pub(super) max_tool_calls: Option<u32>, |
| 437 | questions_allowed: bool, |
| 438 | } |
| 439 | |
| 440 | impl ToolSurfacePolicy { |
| 441 | #[allow(clippy::too_many_arguments)] |
| 442 | pub(super) fn new( |
| 443 | registry: crate::tools::ToolRegistry, |
| 444 | tools: Option<Vec<Tool>>, |
| 445 | mode: AppMode, |
| 446 | always_load: &HashSet<String>, |
| 447 | dynamic_active_tools: &[&'static str], |
| 448 | strict_tool_mode: bool, |
| 449 | allowed_tools: Option<Vec<String>>, |
| 450 | disallowed_tools: Option<Vec<String>>, |
| 451 | max_tool_calls: Option<u32>, |
| 452 | approval_mode: crate::tui::approval::ApprovalMode, |
| 453 | ) -> Self { |
| 454 | let mut catalog = tools.unwrap_or_default(); |
| 455 | if !catalog.is_empty() { |
| 456 | ensure_advanced_tooling(&mut catalog, mode, always_load); |
| 457 | } |
| 458 | |
| 459 | // Synthetic tools are injected before narrowing. Doing this after the |
| 460 | // retain would re-advertise tool_search/code execution despite an |
| 461 | // explicit command gate. |
| 462 | catalog.retain(|tool| { |
| 463 | !tool_denied(disallowed_tools.as_deref(), &tool.name) |
| 464 | && tool_allowed(allowed_tools.as_deref(), &tool.name) |
| 465 | }); |
| 466 | let questions_allowed = |
| 467 | super::super::authority::permission_posture_allows_questions(approval_mode); |
| 468 | if !questions_allowed { |
| 469 | catalog.retain(|tool| tool.name != REQUEST_USER_INPUT_NAME); |
| 470 | } |
| 471 | |
| 472 | let mut active_names = initial_active_tools(&catalog); |
| 473 | active_names.extend(dynamic_active_tools.iter().map(|name| (*name).to_string())); |
| 474 | active_names.retain(|name| catalog.iter().any(|tool| tool.name == *name)); |
| 475 | let active = active_tools_for_request(&catalog, &active_names, strict_tool_mode); |
| 476 | |
| 477 | Self { |
| 478 | registry, |
| 479 | catalog, |
| 480 | active_names, |
| 481 | active, |
| 482 | mode, |
| 483 | strict_tool_mode, |
| 484 | allowed_tools, |
| 485 | disallowed_tools, |
| 486 | max_tool_calls, |
| 487 | questions_allowed, |
| 488 | } |
| 489 | } |
| 490 | |
| 491 | #[cfg(test)] |
| 492 | pub(super) fn allows_tool(&self, name: &str) -> bool { |
| 493 | !self.denies_tool(name) && self.passes_allow_list(name) |
| 494 | } |
| 495 | |
| 496 | pub(super) fn passes_allow_list(&self, name: &str) -> bool { |
| 497 | tool_allowed(self.allowed_tools.as_deref(), name) |
| 498 | } |
| 499 | |
| 500 | pub(super) fn denies_tool(&self, name: &str) -> bool { |
| 501 | tool_denied(self.disallowed_tools.as_deref(), name) |
| 502 | } |
| 503 | |
| 504 | pub(super) fn allows_questions(&self) -> bool { |
| 505 | self.questions_allowed |
| 506 | } |
| 507 | } |
| 508 | |
| 509 | pub(super) fn tool_allowed(allowed_tools: Option<&[String]>, tool_name: &str) -> bool { |
| 510 | let Some(allowed_tools) = allowed_tools else { |
| 511 | return true; |
| 512 | }; |
| 513 | tool_matches_any_rule(allowed_tools, tool_name) |
| 514 | } |
| 515 | |
| 516 | pub(super) fn tool_denied(disallowed_tools: Option<&[String]>, tool_name: &str) -> bool { |
| 517 | disallowed_tools.is_some_and(|rules| tool_matches_any_rule(rules, tool_name)) |
| 518 | } |
| 519 | |
| 520 | fn tool_matches_any_rule(rules: &[String], tool_name: &str) -> bool { |
| 521 | let tool_name = tool_name.to_ascii_lowercase(); |
| 522 | rules.iter().any(|rule| { |
| 523 | let rule = rule.to_ascii_lowercase(); |
| 524 | rule.strip_suffix('*') |
| 525 | .map_or_else(|| tool_name == rule, |prefix| tool_name.starts_with(prefix)) |
| 526 | }) |
| 527 | } |
| 528 | |
| 529 | /// The `tools` field of one outbound request, from a catalog and the set of |
| 530 | /// currently-active tool names. |
| 531 | /// |
| 532 | /// Shared by [`ToolSurfacePolicy`] and the per-step rebuild inside the turn |
| 533 | /// loop, so activating a deferred tool mid-turn goes through one code path. |
| 534 | pub(super) fn active_tools_for_request( |
| 535 | catalog: &[Tool], |
| 536 | active: &HashSet<String>, |
| 537 | strict_tool_mode: bool, |
| 538 | ) -> Option<Vec<Tool>> { |
| 539 | if catalog.is_empty() { |
| 540 | return None; |
| 541 | } |
| 542 | let mut tools = active_tools_for_step(catalog, active); |
| 543 | if strict_tool_mode { |
| 544 | crate::tools::schema_sanitize::prepare_tools_for_strict_mode(&mut tools); |
| 545 | } |
| 546 | Some(tools) |
| 547 | } |
| 548 | |
| 549 | fn tool_search_haystack(tool: &Tool) -> String { |
| 550 | format!( |
| 551 | "{}\n{}\n{}", |
| 552 | tool.name.to_lowercase(), |
| 553 | tool.description.to_lowercase(), |
| 554 | tool.input_schema.to_string().to_lowercase() |
| 555 | ) |
| 556 | } |
| 557 | |
| 558 | fn catalog_contains_tool(catalog: &[Tool], name: &str) -> bool { |
| 559 | catalog.iter().any(|tool| tool.name == name) |
| 560 | } |
| 561 | |
| 562 | fn unavailable_core_action_tools_with_regex( |
| 563 | catalog: &[Tool], |
| 564 | query: &str, |
| 565 | max_results: usize, |
| 566 | ) -> Result<Vec<CoreActionToolFallback>, ToolError> { |
| 567 | if max_results == 0 { |
| 568 | return Ok(Vec::new()); |
| 569 | } |
| 570 | let regex = compile_user_regex(query) |
| 571 | .map_err(|err| ToolError::invalid_input(format!("Invalid regex query: {err}")))?; |
| 572 | Ok(cached_fallbacks() |
| 573 | .iter() |
| 574 | .filter(|cf| !catalog_contains_tool(catalog, cf.fallback.name)) |
| 575 | .filter(|cf| regex.is_match(&cf.haystack)) |
| 576 | .take(max_results) |
| 577 | .map(|cf| cf.fallback) |
| 578 | .collect()) |
| 579 | } |
| 580 | |
| 581 | fn unavailable_core_action_tools_with_bm25_like( |
| 582 | catalog: &[Tool], |
| 583 | query: &str, |
| 584 | max_results: usize, |
| 585 | ) -> Vec<CoreActionToolFallback> { |
| 586 | if max_results == 0 { |
| 587 | return Vec::new(); |
| 588 | } |
| 589 | let terms: Vec<String> = query |
| 590 | .split_whitespace() |
| 591 | .map(|term| term.trim().to_lowercase()) |
| 592 | .filter(|term| !term.is_empty()) |
| 593 | .collect(); |
| 594 | if terms.is_empty() { |
| 595 | return Vec::new(); |
| 596 | } |
| 597 | |
| 598 | let mut scored: Vec<(i64, CoreActionToolFallback)> = Vec::new(); |
| 599 | for cf in cached_fallbacks() { |
| 600 | if catalog_contains_tool(catalog, cf.fallback.name) { |
| 601 | continue; |
| 602 | } |
| 603 | let hay = &cf.haystack; |
| 604 | let name = &cf.name_lower; |
| 605 | let mut score = 0i64; |
| 606 | for term in &terms { |
| 607 | if hay.contains(term) { |
| 608 | score += 1; |
| 609 | } |
| 610 | if name.contains(term) { |
| 611 | score += 2; |
| 612 | } |
| 613 | } |
| 614 | if score > 0 { |
| 615 | scored.push((score, cf.fallback)); |
| 616 | } |
| 617 | } |
| 618 | scored.sort_by(|a, b| b.0.cmp(&a.0).then_with(|| a.1.name.cmp(b.1.name))); |
| 619 | scored |
| 620 | .into_iter() |
| 621 | .take(max_results) |
| 622 | .map(|(_, fallback)| fallback) |
| 623 | .collect() |
| 624 | } |
| 625 | |
| 626 | fn discover_tools_with_regex( |
| 627 | catalog: &[Tool], |
| 628 | query: &str, |
| 629 | max_results: usize, |
| 630 | ) -> Result<Vec<String>, ToolError> { |
| 631 | let regex = compile_user_regex(query) |
| 632 | .map_err(|err| ToolError::invalid_input(format!("Invalid regex query: {err}")))?; |
| 633 | |
| 634 | let mut matches = Vec::new(); |
| 635 | for tool in catalog { |
| 636 | if is_tool_search_tool(&tool.name) { |
| 637 | continue; |
| 638 | } |
| 639 | let hay = tool_search_haystack(tool); |
| 640 | if regex.is_match(&hay) { |
| 641 | matches.push(tool.name.clone()); |
| 642 | } |
| 643 | if matches.len() >= max_results { |
| 644 | break; |
| 645 | } |
| 646 | } |
| 647 | Ok(matches) |
| 648 | } |
| 649 | |
| 650 | fn discover_tools_with_bm25_like(catalog: &[Tool], query: &str, max_results: usize) -> Vec<String> { |
| 651 | let terms: Vec<String> = query |
| 652 | .split_whitespace() |
| 653 | .map(|term| term.trim().to_lowercase()) |
| 654 | .filter(|term| !term.is_empty()) |
| 655 | .collect(); |
| 656 | if terms.is_empty() { |
| 657 | return Vec::new(); |
| 658 | } |
| 659 | |
| 660 | let mut scored: Vec<(i64, String)> = Vec::new(); |
| 661 | for tool in catalog { |
| 662 | if is_tool_search_tool(&tool.name) { |
| 663 | continue; |
| 664 | } |
| 665 | let hay = tool_search_haystack(tool); |
| 666 | let mut score = 0i64; |
| 667 | for term in &terms { |
| 668 | if hay.contains(term) { |
| 669 | score += 1; |
| 670 | } |
| 671 | if tool.name.to_lowercase().contains(term) { |
| 672 | score += 2; |
| 673 | } |
| 674 | } |
| 675 | if score > 0 { |
| 676 | scored.push((score, tool.name.clone())); |
| 677 | } |
| 678 | } |
| 679 | scored.sort_by(|a, b| b.0.cmp(&a.0).then_with(|| a.1.cmp(&b.1))); |
| 680 | scored |
| 681 | .into_iter() |
| 682 | .take(max_results) |
| 683 | .map(|(_, name)| name) |
| 684 | .collect() |
| 685 | } |
| 686 | |
| 687 | fn edit_distance(a: &str, b: &str) -> usize { |
| 688 | if a == b { |
| 689 | return 0; |
| 690 | } |
| 691 | if a.is_empty() { |
| 692 | return b.chars().count(); |
| 693 | } |
| 694 | if b.is_empty() { |
| 695 | return a.chars().count(); |
| 696 | } |
| 697 | |
| 698 | let b_chars: Vec<char> = b.chars().collect(); |
| 699 | let mut prev: Vec<usize> = (0..=b_chars.len()).collect(); |
| 700 | let mut curr = vec![0usize; b_chars.len() + 1]; |
| 701 | |
| 702 | for (i, a_ch) in a.chars().enumerate() { |
| 703 | curr[0] = i + 1; |
| 704 | for (j, b_ch) in b_chars.iter().enumerate() { |
| 705 | let cost = if a_ch == *b_ch { 0 } else { 1 }; |
| 706 | let delete = prev[j + 1] + 1; |
| 707 | let insert = curr[j] + 1; |
| 708 | let substitute = prev[j] + cost; |
| 709 | curr[j + 1] = delete.min(insert).min(substitute); |
| 710 | } |
| 711 | std::mem::swap(&mut prev, &mut curr); |
| 712 | } |
| 713 | |
| 714 | prev[b_chars.len()] |
| 715 | } |
| 716 | |
| 717 | fn suggest_tool_names(catalog: &[Tool], requested: &str, limit: usize) -> Vec<String> { |
| 718 | let requested = requested.trim().to_ascii_lowercase(); |
| 719 | if requested.is_empty() || limit == 0 { |
| 720 | return Vec::new(); |
| 721 | } |
| 722 | |
| 723 | let mut candidates: Vec<(u8, usize, String)> = Vec::new(); |
| 724 | for tool in catalog { |
| 725 | let candidate = tool.name.to_ascii_lowercase(); |
| 726 | let prefix_match = candidate.starts_with(&requested) || requested.starts_with(&candidate); |
| 727 | let contains_match = candidate.contains(&requested) || requested.contains(&candidate); |
| 728 | let distance = edit_distance(&candidate, &requested); |
| 729 | let close_typo = distance <= 3; |
| 730 | |
| 731 | if !(prefix_match || contains_match || close_typo) { |
| 732 | continue; |
| 733 | } |
| 734 | |
| 735 | let rank = if prefix_match { |
| 736 | 0 |
| 737 | } else if contains_match { |
| 738 | 1 |
| 739 | } else { |
| 740 | 2 |
| 741 | }; |
| 742 | candidates.push((rank, distance, tool.name.clone())); |
| 743 | } |
| 744 | |
| 745 | candidates.sort_by(|a, b| { |
| 746 | a.0.cmp(&b.0) |
| 747 | .then_with(|| a.1.cmp(&b.1)) |
| 748 | .then_with(|| a.2.cmp(&b.2)) |
| 749 | }); |
| 750 | candidates.dedup_by(|a, b| a.2 == b.2); |
| 751 | candidates |
| 752 | .into_iter() |
| 753 | .take(limit) |
| 754 | .map(|(_, _, name)| name) |
| 755 | .collect() |
| 756 | } |
| 757 | |
| 758 | /// Catalog tools the engine injects itself rather than registering, plus the |
| 759 | /// legacy tool-search spellings. Exposed so the read-only request projection |
| 760 | /// can label their provenance as `synthetic` from the same source of truth as |
| 761 | /// `is_synthetic_catalog_tool` test coverage instead of guessing. |
| 762 | /// |
| 763 | /// MCP-contributed names are deliberately *not* here: those resolve through the |
| 764 | /// real pool, and stay unknown when the pool did not resolve them. |
| 765 | /// [`MULTI_TOOL_PARALLEL_NAME`] is not here either — it is a call name the model |
| 766 | /// may emit, never a catalog entry, so it can never appear in a transmitted |
| 767 | /// tool array and has no catalog provenance to report. |
| 768 | pub(super) fn default_synthetic_catalog_tool_names() -> Vec<String> { |
| 769 | let mut names: Vec<String> = vec![ |
| 770 | TOOL_SEARCH_NAME.to_string(), |
| 771 | LEGACY_TOOL_SEARCH_REGEX_NAME.to_string(), |
| 772 | LEGACY_TOOL_SEARCH_BM25_NAME.to_string(), |
| 773 | CODE_EXECUTION_TOOL_NAME.to_string(), |
| 774 | JS_EXECUTION_TOOL_NAME.to_string(), |
| 775 | ]; |
| 776 | names.sort(); |
| 777 | names.dedup(); |
| 778 | names |
| 779 | } |
| 780 | |
| 781 | #[cfg(test)] |
| 782 | fn is_synthetic_catalog_tool(name: &str) -> bool { |
| 783 | is_tool_search_tool(name) |
| 784 | || matches!(name, CODE_EXECUTION_TOOL_NAME | JS_EXECUTION_TOOL_NAME) |
| 785 | || McpPool::is_mcp_tool(name) |
| 786 | } |
| 787 | |
| 788 | #[cfg(test)] |
| 789 | pub(super) fn tool_catalog_consistency_issues( |
| 790 | catalog: &[Tool], |
| 791 | registry: &crate::tools::ToolRegistry, |
| 792 | ) -> Vec<String> { |
| 793 | let catalog_names = catalog |
| 794 | .iter() |
| 795 | .map(|tool| tool.name.as_str()) |
| 796 | .collect::<HashSet<_>>(); |
| 797 | let registry_api_tools = registry.to_api_tools(); |
| 798 | let registry_model_visible_names = registry_api_tools |
| 799 | .iter() |
| 800 | .map(|tool| tool.name.as_str()) |
| 801 | .collect::<HashSet<_>>(); |
| 802 | let mut issues = Vec::new(); |
| 803 | |
| 804 | for tool in catalog { |
| 805 | if is_synthetic_catalog_tool(&tool.name) { |
| 806 | continue; |
| 807 | } |
| 808 | if !registry.contains(&tool.name) { |
| 809 | issues.push(format!( |
| 810 | "catalog advertises '{}' but no registered handler exists", |
| 811 | tool.name |
| 812 | )); |
| 813 | } |
| 814 | } |
| 815 | |
| 816 | for name in DEFAULT_ACTIVE_NATIVE_TOOLS { |
| 817 | if registry_model_visible_names.contains(name) && !catalog_names.contains(name) { |
| 818 | issues.push(format!( |
| 819 | "registered core tool '{name}' is missing from the model/search catalog" |
| 820 | )); |
| 821 | } |
| 822 | } |
| 823 | |
| 824 | issues.sort(); |
| 825 | issues |
| 826 | } |
| 827 | |
| 828 | pub(super) fn missing_tool_error_message(tool_name: &str, catalog: &[Tool]) -> String { |
| 829 | // Dogfood A5 (#4092): models mid-checklist sometimes emit each list entry |
| 830 | // as its own tool call named `item`/`todo`/... . Fuzzy suggestions are |
| 831 | // actively misleading there ("Did you mean: note, tts?"); name the actual |
| 832 | // fix instead. |
| 833 | if matches!( |
| 834 | tool_name, |
| 835 | "item" | "items" | "todo" | "todos" | "checklist" | "checklist_item" | "plan_item" |
| 836 | ) { |
| 837 | return format!( |
| 838 | "Tool '{tool_name}' is not available in the current tool catalog. \ |
| 839 | Checklist entries are not separate tool calls — write the whole list \ |
| 840 | in one `work_update` call with a `todos` array of \ |
| 841 | {{content, status}} objects." |
| 842 | ); |
| 843 | } |
| 844 | let suggestions = suggest_tool_names(catalog, tool_name, 3); |
| 845 | let shell_hint = if is_shell_tool_name(tool_name) { |
| 846 | Some(shell_tool_allow_shell_hint()) |
| 847 | } else { |
| 848 | None |
| 849 | }; |
| 850 | // #5123-class: `exec_shell` was renamed to `Bash`. Name the rename first — |
| 851 | // otherwise the error misdiagnoses a retired-name call as an allow_shell |
| 852 | // permission problem and sends the model fixing the wrong thing. |
| 853 | let renamed_action = match tool_name { |
| 854 | "exec_shell" => Some("run"), |
| 855 | "exec_shell_wait" => Some("wait"), |
| 856 | "exec_shell_interact" => Some("interact"), |
| 857 | "exec_shell_cancel" => Some("cancel"), |
| 858 | _ => None, |
| 859 | }; |
| 860 | if let Some(action) = renamed_action { |
| 861 | return format!( |
| 862 | "Tool '{tool_name}' is not available in the current tool catalog. \ |
| 863 | `exec_shell` was renamed to `Bash` — call `Bash` with action \"{action}\" instead. \ |
| 864 | If `Bash` is also absent: {shell_hint}.", |
| 865 | shell_hint = shell_tool_allow_shell_hint() |
| 866 | ); |
| 867 | } |
| 868 | if suggestions.is_empty() { |
| 869 | if let Some(shell_hint) = shell_hint { |
| 870 | return format!( |
| 871 | "Tool '{tool_name}' is not available in the current tool catalog. \ |
| 872 | {shell_hint}, or use {TOOL_SEARCH_NAME} with a short query." |
| 873 | ); |
| 874 | } |
| 875 | return format!( |
| 876 | "Tool '{tool_name}' is not available in the current tool catalog. \ |
| 877 | Verify mode/feature flags, or use {TOOL_SEARCH_NAME} with a short query." |
| 878 | ); |
| 879 | } |
| 880 | |
| 881 | let suggestion_text = format!("Did you mean: {}?", suggestions.join(", ")); |
| 882 | if let Some(shell_hint) = shell_hint { |
| 883 | return format!( |
| 884 | "Tool '{tool_name}' is not available in the current tool catalog. \ |
| 885 | {suggestion_text} {shell_hint}. \ |
| 886 | You can also use {TOOL_SEARCH_NAME} to discover tools." |
| 887 | ); |
| 888 | } |
| 889 | |
| 890 | format!( |
| 891 | "Tool '{tool_name}' is not available in the current tool catalog. \ |
| 892 | {suggestion_text} You can also use {TOOL_SEARCH_NAME} to discover tools." |
| 893 | ) |
| 894 | } |
| 895 | |
| 896 | fn shell_tool_allow_shell_hint() -> &'static str { |
| 897 | "Shell tools are absent because this session or profile disabled shell access, \ |
| 898 | commonly via top-level `allow_shell = false` or Plan mode. \ |
| 899 | Interactive Act mode exposes shell by default with approval gating unless disabled. \ |
| 900 | Run `/config allow_shell true` for this session or add `--save` for future sessions; \ |
| 901 | the next turn will expose shell again" |
| 902 | } |
| 903 | |
| 904 | fn is_shell_tool_name(tool_name: &str) -> bool { |
| 905 | matches!( |
| 906 | tool_name, |
| 907 | "exec_shell" |
| 908 | | "exec_shell_wait" |
| 909 | | "exec_shell_interact" |
| 910 | | "task_shell_start" |
| 911 | | "task_shell_wait" |
| 912 | ) |
| 913 | } |
| 914 | |
| 915 | #[cfg(test)] |
| 916 | pub(super) fn maybe_activate_requested_deferred_tool( |
| 917 | tool_name: &str, |
| 918 | catalog: &[Tool], |
| 919 | active_tools: &mut HashSet<String>, |
| 920 | ) -> bool { |
| 921 | let Some(def) = catalog.iter().find(|def| def.name == tool_name) else { |
| 922 | return false; |
| 923 | }; |
| 924 | |
| 925 | if !def.defer_loading.unwrap_or(false) || active_tools.contains(tool_name) { |
| 926 | return false; |
| 927 | } |
| 928 | |
| 929 | active_tools.insert(tool_name.to_string()) |
| 930 | } |
| 931 | |
| 932 | pub(super) fn maybe_hydrate_requested_deferred_tool( |
| 933 | tool_name: &str, |
| 934 | tool_input: &Value, |
| 935 | catalog: &[Tool], |
| 936 | active_tools_at_batch_start: &HashSet<String>, |
| 937 | hydrated_tools_this_batch: &mut HashSet<String>, |
| 938 | ) -> Option<ToolResult> { |
| 939 | let def = catalog.iter().find(|def| def.name == tool_name)?; |
| 940 | |
| 941 | if !def.defer_loading.unwrap_or(false) || active_tools_at_batch_start.contains(tool_name) { |
| 942 | return None; |
| 943 | } |
| 944 | |
| 945 | hydrated_tools_this_batch.insert(tool_name.to_string()); |
| 946 | Some(deferred_tool_schema_hydration_result(def, tool_input)) |
| 947 | } |
| 948 | |
| 949 | #[cfg(test)] |
| 950 | pub(super) fn preflight_requested_deferred_tool( |
| 951 | tool_name: &str, |
| 952 | tool_input: &Value, |
| 953 | catalog: &[Tool], |
| 954 | active_tools: &mut HashSet<String>, |
| 955 | ) -> Option<ToolResult> { |
| 956 | let active_tools_at_batch_start = active_tools.clone(); |
| 957 | let mut hydrated_tools_this_batch = HashSet::new(); |
| 958 | let result = maybe_hydrate_requested_deferred_tool( |
| 959 | tool_name, |
| 960 | tool_input, |
| 961 | catalog, |
| 962 | &active_tools_at_batch_start, |
| 963 | &mut hydrated_tools_this_batch, |
| 964 | ); |
| 965 | active_tools.extend(hydrated_tools_this_batch); |
| 966 | result |
| 967 | } |
| 968 | |
| 969 | fn deferred_tool_schema_hydration_result(tool: &Tool, tool_input: &Value) -> ToolResult { |
| 970 | let expected = schema_fields(&tool.input_schema); |
| 971 | let required = schema_required_fields(&tool.input_schema); |
| 972 | let received = received_field_names(tool_input); |
| 973 | let missing = required |
| 974 | .iter() |
| 975 | .filter(|field| !received.contains(field)) |
| 976 | .cloned() |
| 977 | .collect::<Vec<_>>(); |
| 978 | let unexpected = received |
| 979 | .iter() |
| 980 | .filter(|field| !expected.iter().any(|expected| &expected.name == *field)) |
| 981 | .cloned() |
| 982 | .collect::<Vec<_>>(); |
| 983 | let corrections = likely_field_corrections(&received, &expected, &tool.name); |
| 984 | |
| 985 | let mut lines = vec![ |
| 986 | format!("Tool `{}` was deferred and has now been loaded.", tool.name), |
| 987 | String::new(), |
| 988 | "The tool was not executed. Retry with the loaded schema.".to_string(), |
| 989 | String::new(), |
| 990 | "Expected fields:".to_string(), |
| 991 | ]; |
| 992 | if expected.is_empty() { |
| 993 | lines.push(" (none)".to_string()); |
| 994 | } else { |
| 995 | for field in &expected { |
| 996 | let required_marker = if required.contains(&field.name) { |
| 997 | " required" |
| 998 | } else { |
| 999 | "" |
| 1000 | }; |
| 1001 | lines.push(format!( |
| 1002 | " {}: {}{}", |
| 1003 | field.name, field.kind, required_marker |
| 1004 | )); |
| 1005 | } |
| 1006 | } |
| 1007 | lines.push(String::new()); |
| 1008 | lines.push("Received fields:".to_string()); |
| 1009 | if received.is_empty() { |
| 1010 | lines.push(" (none)".to_string()); |
| 1011 | } else { |
| 1012 | lines.push(format!(" {}", received.join(", "))); |
| 1013 | } |
| 1014 | if !missing.is_empty() { |
| 1015 | lines.push(String::new()); |
| 1016 | lines.push("Missing required fields:".to_string()); |
| 1017 | lines.push(format!(" {}", missing.join(", "))); |
| 1018 | } |
| 1019 | if !unexpected.is_empty() { |
| 1020 | lines.push(String::new()); |
| 1021 | lines.push("Unexpected fields:".to_string()); |
| 1022 | lines.push(format!(" {}", unexpected.join(", "))); |
| 1023 | } |
| 1024 | if !corrections.is_empty() { |
| 1025 | lines.push(String::new()); |
| 1026 | lines.push("Likely corrections:".to_string()); |
| 1027 | for correction in &corrections { |
| 1028 | lines.push(format!(" {correction}")); |
| 1029 | } |
| 1030 | } |
| 1031 | |
| 1032 | ToolResult::success(lines.join("\n")).with_metadata(json!({ |
| 1033 | "event": "tool.schema_hydrated", |
| 1034 | "tool": tool.name, |
| 1035 | "executed": false, |
| 1036 | "retry_required": true, |
| 1037 | "reason": "deferred_tool_first_use", |
| 1038 | "deferred_tool_loaded": true, |
| 1039 | "tool_name": tool.name, |
| 1040 | "expected_fields": expected.iter().map(|field| field.name.clone()).collect::<Vec<_>>(), |
| 1041 | "received_fields": received, |
| 1042 | "missing_required_fields": missing, |
| 1043 | "unexpected_fields": unexpected, |
| 1044 | "likely_corrections": corrections, |
| 1045 | })) |
| 1046 | } |
| 1047 | |
| 1048 | #[derive(Debug, Clone)] |
| 1049 | struct SchemaField { |
| 1050 | name: String, |
| 1051 | kind: String, |
| 1052 | } |
| 1053 | |
| 1054 | fn schema_fields(schema: &Value) -> Vec<SchemaField> { |
| 1055 | let Some(properties) = schema.get("properties").and_then(Value::as_object) else { |
| 1056 | return Vec::new(); |
| 1057 | }; |
| 1058 | let mut fields = properties |
| 1059 | .iter() |
| 1060 | .map(|(name, spec)| SchemaField { |
| 1061 | name: name.clone(), |
| 1062 | kind: schema_type_label(spec), |
| 1063 | }) |
| 1064 | .collect::<Vec<_>>(); |
| 1065 | fields.sort_by(|a, b| a.name.cmp(&b.name)); |
| 1066 | fields |
| 1067 | } |
| 1068 | |
| 1069 | fn schema_required_fields(schema: &Value) -> Vec<String> { |
| 1070 | let mut required = schema |
| 1071 | .get("required") |
| 1072 | .and_then(Value::as_array) |
| 1073 | .into_iter() |
| 1074 | .flatten() |
| 1075 | .filter_map(|value| value.as_str().map(str::to_string)) |
| 1076 | .collect::<Vec<_>>(); |
| 1077 | required.sort(); |
| 1078 | required |
| 1079 | } |
| 1080 | |
| 1081 | fn schema_type_label(spec: &Value) -> String { |
| 1082 | let Some(kind) = spec.get("type").and_then(Value::as_str) else { |
| 1083 | return "value".to_string(); |
| 1084 | }; |
| 1085 | if let Some(values) = spec.get("enum").and_then(Value::as_array) { |
| 1086 | let labels = values.iter().filter_map(Value::as_str).collect::<Vec<_>>(); |
| 1087 | if !labels.is_empty() { |
| 1088 | return format!("{kind} ({})", labels.join(" | ")); |
| 1089 | } |
| 1090 | } |
| 1091 | kind.to_string() |
| 1092 | } |
| 1093 | |
| 1094 | fn received_field_names(input: &Value) -> Vec<String> { |
| 1095 | let mut fields = input |
| 1096 | .as_object() |
| 1097 | .map(|object| object.keys().cloned().collect::<Vec<_>>()) |
| 1098 | .unwrap_or_default(); |
| 1099 | fields.sort(); |
| 1100 | fields |
| 1101 | } |
| 1102 | |
| 1103 | fn likely_field_corrections( |
| 1104 | received: &[String], |
| 1105 | expected: &[SchemaField], |
| 1106 | tool_name: &str, |
| 1107 | ) -> Vec<String> { |
| 1108 | let has_expected = |name: &str| expected.iter().any(|field| field.name == name); |
| 1109 | let has_received = |name: &str| received.iter().any(|field| field == name); |
| 1110 | let mut corrections = Vec::new(); |
| 1111 | |
| 1112 | if has_received("old_string") && has_expected("search") { |
| 1113 | corrections.push("old_string -> search".to_string()); |
| 1114 | } else if has_received("old_str") && has_expected("search") { |
| 1115 | corrections.push("old_str -> search".to_string()); |
| 1116 | } |
| 1117 | if has_received("new_string") && has_expected("replace") { |
| 1118 | corrections.push("new_string -> replace".to_string()); |
| 1119 | } else if has_received("new_str") && has_expected("replace") { |
| 1120 | corrections.push("new_str -> replace".to_string()); |
| 1121 | } else if has_received("replacement") && has_expected("replace") { |
| 1122 | corrections.push("replacement -> replace".to_string()); |
| 1123 | } |
| 1124 | if matches!(tool_name, "checklist_update" | "todo_update") && has_received("todos") { |
| 1125 | corrections.push( |
| 1126 | "Use work_update to replace the full list, or retry checklist_update/todo_update with id and status." |
| 1127 | .to_string(), |
| 1128 | ); |
| 1129 | } |
| 1130 | // RLM source fields are easy to misname (#2659). rlm_open takes exactly one |
| 1131 | // of file_path / content / url / session_object; nudge common wrong names |
| 1132 | // toward those. The unified `rlm` tool carries the same fields for |
| 1133 | // action=open, so it gets the same correction. |
| 1134 | if matches!(tool_name, "rlm_open" | "rlm") { |
| 1135 | for wrong in [ |
| 1136 | "prompt", |
| 1137 | "resident_file", |
| 1138 | "text", |
| 1139 | "body", |
| 1140 | "path", |
| 1141 | "file", |
| 1142 | "source", |
| 1143 | ] { |
| 1144 | if has_received(wrong) |
| 1145 | && !has_received("file_path") |
| 1146 | && !has_received("content") |
| 1147 | && !has_received("url") |
| 1148 | && !has_received("session_object") |
| 1149 | { |
| 1150 | corrections.push(format!("{wrong} -> file_path (local file), content (inline text), url, or session_object")); |
| 1151 | } |
| 1152 | } |
| 1153 | } |
| 1154 | corrections |
| 1155 | } |
| 1156 | |
| 1157 | pub(super) fn execute_tool_search( |
| 1158 | tool_name: &str, |
| 1159 | input: &serde_json::Value, |
| 1160 | catalog: &[Tool], |
| 1161 | active_tools: &mut HashSet<String>, |
| 1162 | ) -> Result<ToolResult, ToolError> { |
| 1163 | let query = required_str(input, "query")?; |
| 1164 | let match_kind = match tool_name { |
| 1165 | LEGACY_TOOL_SEARCH_REGEX_NAME => "regex", |
| 1166 | LEGACY_TOOL_SEARCH_BM25_NAME => "bm25", |
| 1167 | _ => optional_str(input, "match")?.unwrap_or("bm25"), |
| 1168 | }; |
| 1169 | if !matches!(match_kind, "bm25" | "regex") { |
| 1170 | return Err(ToolError::invalid_input(format!( |
| 1171 | "Unsupported match algorithm '{match_kind}'. Expected one of: bm25, regex" |
| 1172 | ))); |
| 1173 | } |
| 1174 | let max_results = usize::try_from(optional_u64( |
| 1175 | input, |
| 1176 | "max_results", |
| 1177 | TOOL_SEARCH_DEFAULT_MAX_RESULTS as u64, |
| 1178 | )?) |
| 1179 | .unwrap_or(TOOL_SEARCH_DEFAULT_MAX_RESULTS) |
| 1180 | .clamp(1, TOOL_SEARCH_MAX_RESULTS_LIMIT); |
| 1181 | let discovered = if match_kind == "regex" { |
| 1182 | discover_tools_with_regex(catalog, query, max_results)? |
| 1183 | } else { |
| 1184 | discover_tools_with_bm25_like(catalog, query, max_results) |
| 1185 | }; |
| 1186 | let remaining_results = max_results.saturating_sub(discovered.len()); |
| 1187 | let unavailable = if match_kind == "regex" { |
| 1188 | unavailable_core_action_tools_with_regex(catalog, query, remaining_results)? |
| 1189 | } else { |
| 1190 | unavailable_core_action_tools_with_bm25_like(catalog, query, remaining_results) |
| 1191 | }; |
| 1192 | |
| 1193 | for name in &discovered { |
| 1194 | active_tools.insert(name.clone()); |
| 1195 | } |
| 1196 | |
| 1197 | let references = discovered |
| 1198 | .iter() |
| 1199 | .map(|name| json!({"type": "tool_reference", "tool_name": name})) |
| 1200 | .collect::<Vec<_>>(); |
| 1201 | let unavailable_references = unavailable |
| 1202 | .iter() |
| 1203 | .map(|fallback| { |
| 1204 | json!({ |
| 1205 | "type": "unavailable_tool_reference", |
| 1206 | "tool_name": fallback.name, |
| 1207 | "reason": fallback.unavailable_reason, |
| 1208 | }) |
| 1209 | }) |
| 1210 | .collect::<Vec<_>>(); |
| 1211 | |
| 1212 | let payload = json!({ |
| 1213 | "type": "tool_search_tool_search_result", |
| 1214 | "tool_references": references, |
| 1215 | "unavailable_tool_references": unavailable_references.clone(), |
| 1216 | }); |
| 1217 | |
| 1218 | Ok(ToolResult { |
| 1219 | content: serde_json::to_string(&payload).unwrap_or_else(|_| payload.to_string()), |
| 1220 | success: true, |
| 1221 | metadata: Some(json!({ |
| 1222 | "tool_references": discovered, |
| 1223 | "unavailable_tool_references": unavailable_references, |
| 1224 | })), |
| 1225 | }) |
| 1226 | } |
| 1227 | |
| 1228 | pub(super) async fn execute_code_execution_tool( |
| 1229 | input: &serde_json::Value, |
| 1230 | workspace: &Path, |
| 1231 | ) -> Result<ToolResult, ToolError> { |
| 1232 | let code = required_str(input, "code")?; |
| 1233 | |
| 1234 | // Resolve the locally-installed Python interpreter we cached at |
| 1235 | // catalog-build time. If it's absent now (somehow registered but |
| 1236 | // disappeared between startup and this call — concurrent uninstall, |
| 1237 | // PATH change, etc.) the ExternalTool::tokio_command() will return |
| 1238 | // None and we fail fast with a clear message. |
| 1239 | // |
| 1240 | // Write the code to a temp file and execute it as a script rather |
| 1241 | // than passing it via `-c "<code>"`. Reasons: |
| 1242 | // * `-c` has length limits (argv) on Windows. |
| 1243 | // * Multiline code with quote nesting is brittle through `-c`. |
| 1244 | // * Tracebacks reference a real filename instead of `<string>`, |
| 1245 | // so the model can interpret line numbers correctly. |
| 1246 | // Tempfile lives only for the duration of this execution; Drop |
| 1247 | // removes it. We use `.py` so any shebang / encoding-sniffer |
| 1248 | // logic in the interpreter behaves normally. |
| 1249 | let temp_dir = tempfile::tempdir() |
| 1250 | .map_err(|e| ToolError::execution_failed(format!("tempdir failed: {e}")))?; |
| 1251 | let script_path = temp_dir.path().join("code_execution.py"); |
| 1252 | tokio::fs::write(&script_path, code) |
| 1253 | .await |
| 1254 | .map_err(|e| ToolError::execution_failed(format!("tempfile write failed: {e}")))?; |
| 1255 | |
| 1256 | let mut cmd = crate::dependencies::Python::tokio_command().ok_or_else(|| { |
| 1257 | ToolError::execution_failed( |
| 1258 | "code_execution: Python interpreter became unavailable".to_string(), |
| 1259 | ) |
| 1260 | })?; |
| 1261 | cmd.arg(&script_path).current_dir(workspace); |
| 1262 | |
| 1263 | let output = tokio::time::timeout(Duration::from_secs(120), cmd.output()) |
| 1264 | .await |
| 1265 | .map_err(|_| ToolError::Timeout { seconds: 120 }) |
| 1266 | .and_then(|res| res.map_err(|e| ToolError::execution_failed(e.to_string())))?; |
| 1267 | |
| 1268 | let stdout = String::from_utf8_lossy(&output.stdout).to_string(); |
| 1269 | let stderr = String::from_utf8_lossy(&output.stderr).to_string(); |
| 1270 | let return_code = output.status.code().unwrap_or(-1); |
| 1271 | let success = output.status.success(); |
| 1272 | let payload = json!({ |
| 1273 | "type": "code_execution_result", |
| 1274 | "stdout": stdout, |
| 1275 | "stderr": stderr, |
| 1276 | "return_code": return_code, |
| 1277 | "content": [], |
| 1278 | }); |
| 1279 | |
| 1280 | Ok(ToolResult { |
| 1281 | content: serde_json::to_string(&payload).unwrap_or_else(|_| payload.to_string()), |
| 1282 | success, |
| 1283 | metadata: Some(payload), |
| 1284 | }) |
| 1285 | } |
| 1286 | |
| 1287 | #[cfg(test)] |
| 1288 | mod synthetic_name_tests { |
| 1289 | use super::{ |
| 1290 | CODE_EXECUTION_DESCRIPTION, default_synthetic_catalog_tool_names, is_synthetic_catalog_tool, |
| 1291 | }; |
| 1292 | |
| 1293 | /// `code_execution` writes the script to a tempdir and runs it as a plain |
| 1294 | /// child process in the workspace — no seccomp, no jail, no container. The |
| 1295 | /// description is model-facing, so calling it a sandbox would tell the model |
| 1296 | /// it has isolation the runtime never provides. |
| 1297 | #[test] |
| 1298 | fn code_execution_description_does_not_claim_process_sandboxing() { |
| 1299 | assert!(CODE_EXECUTION_DESCRIPTION.contains("local Python interpreter")); |
| 1300 | assert!(!CODE_EXECUTION_DESCRIPTION.contains("sandbox")); |
| 1301 | } |
| 1302 | |
| 1303 | /// The published synthetic-name list and the predicate that classifies a |
| 1304 | /// catalog entry as synthetic must agree. A name that appears in the list |
| 1305 | /// but is not classified synthetic would let the request projection report |
| 1306 | /// a provenance the engine itself disputes. |
| 1307 | #[test] |
| 1308 | fn published_synthetic_names_agree_with_the_synthetic_predicate() { |
| 1309 | let names = default_synthetic_catalog_tool_names(); |
| 1310 | assert!(!names.is_empty()); |
| 1311 | for name in &names { |
| 1312 | assert!( |
| 1313 | is_synthetic_catalog_tool(name), |
| 1314 | "'{name}' is published as synthetic but the predicate disagrees" |
| 1315 | ); |
| 1316 | } |
| 1317 | let mut sorted = names.clone(); |
| 1318 | sorted.sort(); |
| 1319 | sorted.dedup(); |
| 1320 | assert_eq!(names, sorted, "the list must be sorted and deduplicated"); |
| 1321 | |
| 1322 | // MCP names resolve through the real pool, so they are deliberately |
| 1323 | // absent here even though the predicate accepts them. |
| 1324 | assert!(!names.iter().any(|name| name.starts_with("mcp_"))); |
| 1325 | |
| 1326 | // `multi_tool_use.parallel` is a call name, never a catalog entry, so |
| 1327 | // it has no catalog provenance and must not be published as synthetic. |
| 1328 | assert!( |
| 1329 | !names |
| 1330 | .iter() |
| 1331 | .any(|name| name == super::MULTI_TOOL_PARALLEL_NAME) |
| 1332 | ); |
| 1333 | } |
| 1334 | } |
| 1335 |