| 1 | //! Truthful, bounded inspection of a prepared model-client request's tool field. |
| 2 | //! |
| 3 | //! Capture happens at the request-construction seam and retains only a bounded |
| 4 | //! projection. It never claims that the prepared request was delivered. |
| 5 | //! |
| 6 | //! Two kinds of fact live here, and they are kept apart on purpose: |
| 7 | //! |
| 8 | //! * **Wire facts** come from the prepared request itself — names, schemas, |
| 9 | //! descriptions, per-tool transport flags, byte accounting, and the |
| 10 | //! active-tool-catalog digest. The digest is not defined here; it is |
| 11 | //! [`crate::core::engine::preview::active_tool_catalog_sha256`], the same |
| 12 | //! function the request manifest publishes, so `/tools` and `/request` cannot |
| 13 | //! report two different hashes of one catalog. |
| 14 | //! * **Surface facts** come from a [`ToolSurfaceContext`] the engine resolves |
| 15 | //! once per turn: flattened registry facts, the MCP pool's resolved server |
| 16 | //! attributions, the engine-injected catalog names, and the provider receipt |
| 17 | //! taken from the *resolved model client*. When that context is present, |
| 18 | //! provenance, MCP server identity, capabilities, approval requirement, and |
| 19 | //! model visibility become available and true. When it is absent they stay |
| 20 | //! explicitly unknown — the context is optional, never faked. |
| 21 | //! |
| 22 | //! What stays unknowable stays unknown regardless: nothing here observes the |
| 23 | //! provider adapter's wire payload, so it is always reported as unavailable. |
| 24 | |
| 25 | mod portable_projection; |
| 26 | pub(crate) use portable_projection::tool_snapshot as project_snapshot; |
| 27 | |
| 28 | use std::collections::BTreeMap; |
| 29 | use std::io::{self, Write}; |
| 30 | |
| 31 | use serde::Serialize; |
| 32 | use serde_json::Value; |
| 33 | |
| 34 | use codewhale_models::Tool; |
| 35 | |
| 36 | const MAX_RENDERED_TOOLS: usize = 32; |
| 37 | const MAX_NAME_CHARS: usize = 256; |
| 38 | const MAX_DESCRIPTION_CHARS: usize = 512; |
| 39 | const MAX_SCHEMA_BYTES: usize = 2_048; |
| 40 | const MAX_AUXILIARY_CHARS: usize = 512; |
| 41 | const MAX_ALLOWED_CALLERS: usize = 16; |
| 42 | const MAX_ALLOWED_CALLER_CHARS: usize = 128; |
| 43 | const MAX_PAYLOAD_MEASUREMENT_BYTES: usize = 1_048_576; |
| 44 | |
| 45 | #[derive(Debug, Clone, PartialEq, Serialize)] |
| 46 | pub struct BoundedString { |
| 47 | pub value: String, |
| 48 | pub truncated: bool, |
| 49 | } |
| 50 | |
| 51 | /// Observed engine exit boundary. This never classifies the assistant's prose. |
| 52 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize)] |
| 53 | #[serde(rename_all = "snake_case")] |
| 54 | pub enum TurnStopReason { |
| 55 | ProviderNoToolCall, |
| 56 | ProviderToolCallMissing, |
| 57 | StepBudgetExhausted, |
| 58 | NoProgress, |
| 59 | Interrupted, |
| 60 | Failed, |
| 61 | } |
| 62 | |
| 63 | /// Terminal facts attached to the existing request inspector, without adding |
| 64 | /// conversation input or a notice to an ordinary successful response. |
| 65 | #[derive(Debug, Clone, Default, PartialEq, Serialize)] |
| 66 | pub struct TurnStopDiagnostics { |
| 67 | pub status: Option<crate::core::events::TurnOutcomeStatus>, |
| 68 | /// None means the precise runtime exit boundary was not observed. |
| 69 | pub reason: Option<TurnStopReason>, |
| 70 | /// None means the caller did not install a model-step ceiling. |
| 71 | pub effective_max_steps: Option<u32>, |
| 72 | pub step_budget_source: &'static str, |
| 73 | /// Existing zero-based scheduler step; transport retries do not advance it. |
| 74 | pub model_step_index: u32, |
| 75 | /// Parent streaming ModelClient calls, including stream retries. Excludes |
| 76 | /// HTTP retries inside the client, compaction and child calls; not invoices. |
| 77 | pub model_requests_started: u32, |
| 78 | /// HTTP retries actually entered inside the captured parent request. |
| 79 | /// This is dispatch evidence, not provider usage or a bill. |
| 80 | pub transport_retries: u32, |
| 81 | pub transparent_stream_retries: u32, |
| 82 | pub stream_resumes: u32, |
| 83 | pub reasoning_only_reprompts: u32, |
| 84 | /// Re-requests after a clean terminal stop that carried no text, no |
| 85 | /// reasoning and no tool call (#6310): an exact-prefix retry, then a |
| 86 | /// nudged one, before the turn fails visibly. |
| 87 | pub empty_stop_retries: u32, |
| 88 | pub soft_landing_sent: bool, |
| 89 | pub final_report_requested: bool, |
| 90 | pub permission_strategy_switches: u32, |
| 91 | /// Denied provider-response batches since the latest useful progress. |
| 92 | pub permission_denial_rounds_without_progress: u32, |
| 93 | pub last_provider_finish_reason: Option<BoundedString>, |
| 94 | /// Structured calls decoded from the stream, before legacy text-call parsing. |
| 95 | pub last_response_tool_calls: Option<usize>, |
| 96 | /// None means suppression was not counted at this exit boundary. |
| 97 | pub last_response_tool_calls_suppressed: Option<usize>, |
| 98 | /// Last parent response's reported input tokens, not cumulative billing. |
| 99 | pub last_reported_input_tokens: Option<u32>, |
| 100 | pub route_context_window_tokens: Option<u64>, |
| 101 | /// Engine-prepared output allowance. A transport may omit the field; |
| 102 | /// this is budget evidence, not a provider-published capability ceiling. |
| 103 | pub last_prepared_output_limit_tokens: Option<u32>, |
| 104 | pub automatic_compaction_attempts: u32, |
| 105 | pub emergency_compaction_attempts: u32, |
| 106 | } |
| 107 | |
| 108 | impl TurnStopDiagnostics { |
| 109 | pub(crate) fn observe_provider_response( |
| 110 | &mut self, |
| 111 | finish_reason: Option<&str>, |
| 112 | tool_calls: usize, |
| 113 | ) { |
| 114 | self.last_provider_finish_reason = |
| 115 | finish_reason.map(|reason| bounded_chars(reason, MAX_NAME_CHARS)); |
| 116 | self.last_response_tool_calls = Some(tool_calls); |
| 117 | self.last_response_tool_calls_suppressed = None; |
| 118 | } |
| 119 | } |
| 120 | |
| 121 | #[derive(Debug, Clone, PartialEq, Serialize)] |
| 122 | #[serde(tag = "status", rename_all = "snake_case")] |
| 123 | pub enum Evidence<T> { |
| 124 | Known { value: T }, |
| 125 | Unknown { reason: String }, |
| 126 | } |
| 127 | |
| 128 | #[derive(Debug, Clone, PartialEq, Serialize)] |
| 129 | pub struct BoundedList { |
| 130 | pub count: usize, |
| 131 | pub rendered: Vec<BoundedString>, |
| 132 | pub omitted: usize, |
| 133 | } |
| 134 | |
| 135 | #[derive(Debug, Clone, PartialEq, Serialize)] |
| 136 | pub struct CountOnly { |
| 137 | pub count: usize, |
| 138 | pub values: &'static str, |
| 139 | } |
| 140 | |
| 141 | /// One registry tool flattened to the exact facts this projection may report. |
| 142 | /// |
| 143 | /// The engine fills this from `ToolSpec`, so this module never holds a tool |
| 144 | /// object and therefore cannot execute one. |
| 145 | #[derive(Debug, Clone, PartialEq, Eq, Serialize)] |
| 146 | pub struct RegistryFacts { |
| 147 | pub name: String, |
| 148 | pub description: String, |
| 149 | pub model_visible: bool, |
| 150 | pub capabilities: Vec<String>, |
| 151 | pub approval: String, |
| 152 | /// `true` when the tool came from the plugin surface rather than the |
| 153 | /// built-in registry builder. |
| 154 | pub plugin: bool, |
| 155 | } |
| 156 | |
| 157 | /// Where a tool in the prepared request came from. |
| 158 | /// |
| 159 | /// `Unknown` is a real answer, not a fallback guess: it means the surface |
| 160 | /// context resolved no origin for that name. |
| 161 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize)] |
| 162 | #[serde(rename_all = "snake_case")] |
| 163 | pub enum ToolProvenance { |
| 164 | /// Registered by the built-in registry builder. |
| 165 | Builtin, |
| 166 | /// Loaded from the plugin/tools surface or `config.toml` overrides. |
| 167 | Plugin, |
| 168 | /// Contributed by the MCP pool, as attributed by the pool itself. |
| 169 | Mcp, |
| 170 | /// Injected into the request catalog by the engine rather than registered |
| 171 | /// (`tool_search` and its legacy spellings, `code_execution`, |
| 172 | /// `js_execution`). |
| 173 | Synthetic, |
| 174 | /// Present in the request with no resolved origin. |
| 175 | Unknown, |
| 176 | } |
| 177 | |
| 178 | /// A tool's state relative to the request that was prepared for this step. |
| 179 | #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize)] |
| 180 | #[serde(rename_all = "snake_case")] |
| 181 | pub enum ToolVisibility { |
| 182 | /// In this step's request with its schema included. |
| 183 | Active, |
| 184 | /// In this step's request, marked deferred (schema loads on demand). |
| 185 | Deferred, |
| 186 | /// In this step's request, with no transport flag to say which. |
| 187 | InRequest, |
| 188 | } |
| 189 | |
| 190 | impl ToolVisibility { |
| 191 | /// Whether this state means the tool's bytes are carried by the prepared |
| 192 | /// request. The only honest source of this answer is the request itself. |
| 193 | #[must_use] |
| 194 | #[cfg(test)] |
| 195 | pub const fn in_request(self) -> bool { |
| 196 | matches!(self, Self::Active | Self::Deferred | Self::InRequest) |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | /// Provider/route availability, derived from the resolved model client taken at |
| 201 | /// the request seam — never from the existence of a tool registry. |
| 202 | #[derive(Debug, Clone, Default, PartialEq, Eq, Serialize)] |
| 203 | #[serde(tag = "status", rename_all = "snake_case")] |
| 204 | pub enum ProviderAvailability { |
| 205 | /// No client receipt was taken, because no surface context was captured. |
| 206 | #[default] |
| 207 | Unknown, |
| 208 | /// A model client was resolved and this request was built for it. |
| 209 | Available { provider: String, model: String }, |
| 210 | /// The seam was reached with no resolved model client. The reason is a safe |
| 211 | /// label, never a URL or credential. |
| 212 | Unavailable { reason: String }, |
| 213 | } |
| 214 | |
| 215 | impl ProviderAvailability { |
| 216 | #[must_use] |
| 217 | pub const fn is_available(&self) -> bool { |
| 218 | matches!(self, Self::Available { .. }) |
| 219 | } |
| 220 | |
| 221 | #[must_use] |
| 222 | #[cfg(test)] |
| 223 | pub const fn label(&self) -> &'static str { |
| 224 | match self { |
| 225 | Self::Unknown => "unknown", |
| 226 | Self::Available { .. } => "available", |
| 227 | Self::Unavailable { .. } => "unavailable", |
| 228 | } |
| 229 | } |
| 230 | } |
| 231 | |
| 232 | /// Everything outside the prepared request that this projection is allowed to |
| 233 | /// report, resolved once per turn by the engine. |
| 234 | /// |
| 235 | /// Plain data only: no registry, no client, no credentials. Resolving it once |
| 236 | /// per turn is what keeps the per-step seam from re-locking the MCP pool. |
| 237 | #[derive(Debug, Clone, Default)] |
| 238 | pub struct ToolSurfaceContext { |
| 239 | /// The real registry, flattened. Sorted by name by the producer. |
| 240 | pub registry: Vec<RegistryFacts>, |
| 241 | /// Model tool name -> MCP server name, only for names the real pool |
| 242 | /// resolved. Absent names stay unknown rather than being split apart. |
| 243 | pub mcp_servers: BTreeMap<String, String>, |
| 244 | /// Request-catalog names the engine injects rather than registering. |
| 245 | pub synthetic_names: Vec<String>, |
| 246 | /// Receipt from the resolved model client for this turn. |
| 247 | pub provider: ProviderAvailability, |
| 248 | } |
| 249 | |
| 250 | impl ToolSurfaceContext { |
| 251 | fn provenance(&self, name: &str) -> ToolProvenance { |
| 252 | if let Some(facts) = self.registry.iter().find(|facts| facts.name == name) { |
| 253 | if facts.plugin { |
| 254 | return ToolProvenance::Plugin; |
| 255 | } |
| 256 | if self.mcp_servers.contains_key(name) { |
| 257 | return ToolProvenance::Mcp; |
| 258 | } |
| 259 | return ToolProvenance::Builtin; |
| 260 | } |
| 261 | if self.mcp_servers.contains_key(name) { |
| 262 | return ToolProvenance::Mcp; |
| 263 | } |
| 264 | if self.synthetic_names.iter().any(|entry| entry == name) { |
| 265 | return ToolProvenance::Synthetic; |
| 266 | } |
| 267 | ToolProvenance::Unknown |
| 268 | } |
| 269 | } |
| 270 | |
| 271 | #[derive(Debug, Clone, PartialEq, Serialize)] |
| 272 | pub struct ToolProjection { |
| 273 | pub ordinal: usize, |
| 274 | pub name: BoundedString, |
| 275 | pub tool_type: Evidence<BoundedString>, |
| 276 | pub description: BoundedString, |
| 277 | pub input_schema_json: BoundedString, |
| 278 | pub allowed_callers: Evidence<BoundedList>, |
| 279 | pub defer_loading: Evidence<bool>, |
| 280 | pub input_examples: Evidence<CountOnly>, |
| 281 | pub strict: Evidence<bool>, |
| 282 | pub cache_control_type: Evidence<BoundedString>, |
| 283 | /// Where this tool came from. Known only when a surface context was |
| 284 | /// captured; `ToolProvenance::Unknown` inside `Known` means the context was |
| 285 | /// captured and still resolved no origin. |
| 286 | pub provenance: Evidence<ToolProvenance>, |
| 287 | /// Owning MCP server, only when the real pool attributed this exact model |
| 288 | /// tool name. |
| 289 | pub mcp_server: Evidence<BoundedString>, |
| 290 | /// Declared capabilities from the registry, sorted. Known-and-empty means |
| 291 | /// "declares none"; unknown means "not in the registry". |
| 292 | pub capabilities: Evidence<BoundedList>, |
| 293 | /// Declared approval requirement from the registry. |
| 294 | pub approval: Evidence<BoundedString>, |
| 295 | /// Registry model visibility. A tool can be registered but hidden. |
| 296 | pub model_visible: Evidence<bool>, |
| 297 | /// State relative to this prepared request. Always known: the request is |
| 298 | /// the evidence. |
| 299 | pub visibility: ToolVisibility, |
| 300 | } |
| 301 | |
| 302 | /// Bounded evidence from a prepared model-client request. |
| 303 | #[derive(Debug, Clone, PartialEq, Serialize)] |
| 304 | pub struct ToolInspectionSnapshot { |
| 305 | pub schema_version: u32, |
| 306 | pub capture_source: &'static str, |
| 307 | pub delivery_status: &'static str, |
| 308 | pub turn_id: BoundedString, |
| 309 | pub step: u32, |
| 310 | #[serde(skip_serializing_if = "Option::is_none")] |
| 311 | pub terminal: Option<TurnStopDiagnostics>, |
| 312 | pub tools_field_present: bool, |
| 313 | pub tool_count: usize, |
| 314 | pub rendered_tool_count: usize, |
| 315 | pub omitted_tool_count: usize, |
| 316 | pub payload_json_bytes: Option<usize>, |
| 317 | pub payload_measurement_status: String, |
| 318 | /// The active-tool-catalog digest, computed by the *same* function the |
| 319 | /// request manifest uses for `active_tool_catalog_sha256`. Absent only when |
| 320 | /// the request carried no tools field at all. Covers tool name, |
| 321 | /// description, and canonical input schema — not transport-only fields — |
| 322 | /// exactly as the manifest does. |
| 323 | pub active_tool_catalog_sha256: Option<String>, |
| 324 | /// Facts nothing on this path can observe for this request. Shrinks when a |
| 325 | /// surface context supplies registry- and client-derived truth; never |
| 326 | /// empties, because the provider adapter's wire payload is never visible |
| 327 | /// here. |
| 328 | pub unavailable_for_this_request: Vec<&'static str>, |
| 329 | /// Provider receipt from the resolved model client, or `Unknown` when no |
| 330 | /// surface context was captured. |
| 331 | pub provider: ProviderAvailability, |
| 332 | /// Whether registry-derived facts were captured at all. Absent stays |
| 333 | /// distinct from an empty registry. |
| 334 | pub registry_facts_present: bool, |
| 335 | /// Size of the flattened registry, when it was captured. |
| 336 | pub registry_tool_count: Evidence<usize>, |
| 337 | /// Registered, model-visible tools this request does *not* carry. Bounded, |
| 338 | /// with an explicit omission count. |
| 339 | pub registry_only_tools: Evidence<BoundedList>, |
| 340 | pub tools: Vec<ToolProjection>, |
| 341 | } |
| 342 | |
| 343 | impl ToolInspectionSnapshot { |
| 344 | /// Wire facts only. Provenance, attribution, capabilities, approval, and |
| 345 | /// provider identity stay explicitly unknown. |
| 346 | #[must_use] |
| 347 | #[cfg(test)] |
| 348 | pub fn from_prepared_request(turn_id: &str, step: u32, tools: Option<&[Tool]>) -> Self { |
| 349 | Self::from_prepared_request_with_surface(turn_id, step, tools, None) |
| 350 | } |
| 351 | |
| 352 | /// Wire facts joined against the turn's resolved surface context. |
| 353 | /// |
| 354 | /// The context is what turns "unavailable" into truth: it is derived from |
| 355 | /// the real registry, the real MCP pool's own attribution, the engine's own |
| 356 | /// synthetic-name list, and the resolved model client. Passing `None` |
| 357 | /// reproduces the wire-only projection exactly. |
| 358 | #[must_use] |
| 359 | pub fn from_prepared_request_with_surface( |
| 360 | turn_id: &str, |
| 361 | step: u32, |
| 362 | tools: Option<&[Tool]>, |
| 363 | surface: Option<&ToolSurfaceContext>, |
| 364 | ) -> Self { |
| 365 | let tool_count = tools.map_or(0, <[Tool]>::len); |
| 366 | let projected = tools |
| 367 | .unwrap_or_default() |
| 368 | .iter() |
| 369 | .take(MAX_RENDERED_TOOLS) |
| 370 | .enumerate() |
| 371 | .map(|(index, tool)| project_tool(index, tool, surface)) |
| 372 | .collect::<Vec<_>>(); |
| 373 | let (payload_json_bytes, payload_measurement_status) = measure_payload(tools); |
| 374 | |
| 375 | let request_names = tools |
| 376 | .unwrap_or_default() |
| 377 | .iter() |
| 378 | .map(|tool| tool.name.as_str()) |
| 379 | .collect::<std::collections::BTreeSet<_>>(); |
| 380 | let registry_only_tools = surface.map_or_else( |
| 381 | || unknown("registry facts not captured for this request"), |
| 382 | |surface| { |
| 383 | let names = surface |
| 384 | .registry |
| 385 | .iter() |
| 386 | .filter(|facts| { |
| 387 | facts.model_visible && !request_names.contains(facts.name.as_str()) |
| 388 | }) |
| 389 | .map(|facts| facts.name.as_str()) |
| 390 | .collect::<Vec<_>>(); |
| 391 | let rendered = names |
| 392 | .iter() |
| 393 | .take(MAX_RENDERED_TOOLS) |
| 394 | .map(|name| bounded_chars(name, MAX_NAME_CHARS)) |
| 395 | .collect::<Vec<_>>(); |
| 396 | Evidence::Known { |
| 397 | value: BoundedList { |
| 398 | count: names.len(), |
| 399 | omitted: names.len().saturating_sub(rendered.len()), |
| 400 | rendered, |
| 401 | }, |
| 402 | } |
| 403 | }, |
| 404 | ); |
| 405 | |
| 406 | let mut unavailable_for_this_request = vec!["provider_wire_payload"]; |
| 407 | if surface.is_none() { |
| 408 | unavailable_for_this_request.extend([ |
| 409 | "provider", |
| 410 | "model", |
| 411 | "approval", |
| 412 | "provenance", |
| 413 | "capabilities", |
| 414 | ]); |
| 415 | } else if !surface.is_some_and(|surface| surface.provider.is_available()) { |
| 416 | unavailable_for_this_request.extend(["provider", "model"]); |
| 417 | } |
| 418 | |
| 419 | Self { |
| 420 | schema_version: 1, |
| 421 | capture_source: "prepared model-client request", |
| 422 | delivery_status: "unknown (capture does not prove provider delivery)", |
| 423 | turn_id: bounded_chars(turn_id, MAX_AUXILIARY_CHARS), |
| 424 | step, |
| 425 | terminal: None, |
| 426 | tools_field_present: tools.is_some(), |
| 427 | tool_count, |
| 428 | rendered_tool_count: projected.len(), |
| 429 | omitted_tool_count: tool_count.saturating_sub(projected.len()), |
| 430 | payload_json_bytes, |
| 431 | payload_measurement_status, |
| 432 | active_tool_catalog_sha256: tools |
| 433 | .map(crate::core::engine::preview::active_tool_catalog_sha256), |
| 434 | unavailable_for_this_request, |
| 435 | provider: surface.map_or(ProviderAvailability::Unknown, |surface| { |
| 436 | surface.provider.clone() |
| 437 | }), |
| 438 | registry_facts_present: surface.is_some(), |
| 439 | registry_tool_count: surface.map_or_else( |
| 440 | || unknown("registry facts not captured for this request"), |
| 441 | |surface| Evidence::Known { |
| 442 | value: surface.registry.len(), |
| 443 | }, |
| 444 | ), |
| 445 | registry_only_tools, |
| 446 | tools: projected, |
| 447 | } |
| 448 | } |
| 449 | |
| 450 | #[must_use] |
| 451 | #[cfg(test)] |
| 452 | pub fn render_text(&self) -> String { |
| 453 | crate::diagnostics_reports::render_tool_snapshot_text(&project_snapshot(self)) |
| 454 | } |
| 455 | |
| 456 | #[cfg(test)] |
| 457 | pub fn render_json(&self) -> Result<String, serde_json::Error> { |
| 458 | crate::diagnostics_reports::render_tool_snapshot_json(&project_snapshot(self)) |
| 459 | } |
| 460 | } |
| 461 | |
| 462 | fn project_tool(index: usize, tool: &Tool, surface: Option<&ToolSurfaceContext>) -> ToolProjection { |
| 463 | let facts = surface.and_then(|surface| { |
| 464 | surface |
| 465 | .registry |
| 466 | .iter() |
| 467 | .find(|facts| facts.name == tool.name) |
| 468 | }); |
| 469 | let no_surface = "surface context not captured for this request"; |
| 470 | let not_registered = "tool is not in the registry"; |
| 471 | ToolProjection { |
| 472 | ordinal: index + 1, |
| 473 | name: bounded_chars(&tool.name, MAX_NAME_CHARS), |
| 474 | tool_type: optional_bounded(tool.tool_type.as_deref()), |
| 475 | description: bounded_chars(&tool.description, MAX_DESCRIPTION_CHARS), |
| 476 | input_schema_json: bounded_json(&tool.input_schema, MAX_SCHEMA_BYTES), |
| 477 | allowed_callers: tool.allowed_callers.as_ref().map_or_else( |
| 478 | || unknown("request field absent"), |
| 479 | |values| { |
| 480 | let rendered = values |
| 481 | .iter() |
| 482 | .take(MAX_ALLOWED_CALLERS) |
| 483 | .map(|value| bounded_chars(value, MAX_ALLOWED_CALLER_CHARS)) |
| 484 | .collect::<Vec<_>>(); |
| 485 | Evidence::Known { |
| 486 | value: BoundedList { |
| 487 | count: values.len(), |
| 488 | omitted: values.len().saturating_sub(rendered.len()), |
| 489 | rendered, |
| 490 | }, |
| 491 | } |
| 492 | }, |
| 493 | ), |
| 494 | defer_loading: optional_copy(tool.defer_loading.as_ref()), |
| 495 | input_examples: tool.input_examples.as_ref().map_or_else( |
| 496 | || unknown("request field absent"), |
| 497 | |values| Evidence::Known { |
| 498 | value: CountOnly { |
| 499 | count: values.len(), |
| 500 | values: "values omitted from bounded projection", |
| 501 | }, |
| 502 | }, |
| 503 | ), |
| 504 | strict: optional_copy(tool.strict.as_ref()), |
| 505 | cache_control_type: optional_bounded( |
| 506 | tool.cache_control |
| 507 | .as_ref() |
| 508 | .map(|value| value.cache_type.as_str()), |
| 509 | ), |
| 510 | provenance: surface.map_or_else( |
| 511 | || unknown(no_surface), |
| 512 | |surface| Evidence::Known { |
| 513 | value: surface.provenance(&tool.name), |
| 514 | }, |
| 515 | ), |
| 516 | mcp_server: surface.map_or_else( |
| 517 | || unknown(no_surface), |
| 518 | |surface| { |
| 519 | surface.mcp_servers.get(&tool.name).map_or_else( |
| 520 | || unknown("the MCP pool did not attribute this tool name"), |
| 521 | |server| Evidence::Known { |
| 522 | value: bounded_chars(server, MAX_AUXILIARY_CHARS), |
| 523 | }, |
| 524 | ) |
| 525 | }, |
| 526 | ), |
| 527 | capabilities: match (surface, facts) { |
| 528 | (None, _) => unknown(no_surface), |
| 529 | (Some(_), None) => unknown(not_registered), |
| 530 | (Some(_), Some(facts)) => { |
| 531 | let rendered = facts |
| 532 | .capabilities |
| 533 | .iter() |
| 534 | .take(MAX_ALLOWED_CALLERS) |
| 535 | .map(|value| bounded_chars(value, MAX_ALLOWED_CALLER_CHARS)) |
| 536 | .collect::<Vec<_>>(); |
| 537 | Evidence::Known { |
| 538 | value: BoundedList { |
| 539 | count: facts.capabilities.len(), |
| 540 | omitted: facts.capabilities.len().saturating_sub(rendered.len()), |
| 541 | rendered, |
| 542 | }, |
| 543 | } |
| 544 | } |
| 545 | }, |
| 546 | approval: match (surface, facts) { |
| 547 | (None, _) => unknown(no_surface), |
| 548 | (Some(_), None) => unknown(not_registered), |
| 549 | (Some(_), Some(facts)) => Evidence::Known { |
| 550 | value: bounded_chars(&facts.approval, MAX_AUXILIARY_CHARS), |
| 551 | }, |
| 552 | }, |
| 553 | model_visible: match (surface, facts) { |
| 554 | (None, _) => unknown(no_surface), |
| 555 | (Some(_), None) => unknown(not_registered), |
| 556 | (Some(_), Some(facts)) => Evidence::Known { |
| 557 | value: facts.model_visible, |
| 558 | }, |
| 559 | }, |
| 560 | // Every projected tool is carried by this prepared request; the |
| 561 | // transport flag only says whether its schema rides along now. |
| 562 | visibility: match tool.defer_loading { |
| 563 | Some(true) => ToolVisibility::Deferred, |
| 564 | Some(false) => ToolVisibility::Active, |
| 565 | None => ToolVisibility::InRequest, |
| 566 | }, |
| 567 | } |
| 568 | } |
| 569 | |
| 570 | /// Byte accounting only. The digest is deliberately *not* computed here: it is |
| 571 | /// the request path's [`crate::core::engine::preview::active_tool_catalog_sha256`], |
| 572 | /// so this projection never defines a second catalog hash. |
| 573 | fn measure_payload(tools: Option<&[Tool]>) -> (Option<usize>, String) { |
| 574 | let Some(tools) = tools else { |
| 575 | return (None, "unavailable (tools field absent)".to_string()); |
| 576 | }; |
| 577 | let mut writer = BoundedWriter::new(MAX_PAYLOAD_MEASUREMENT_BYTES); |
| 578 | match serde_json::to_writer(&mut writer, tools) { |
| 579 | Ok(()) => ( |
| 580 | Some(writer.bytes.len()), |
| 581 | "exact (within 1048576-byte measurement bound)".to_string(), |
| 582 | ), |
| 583 | Err(_) if writer.exceeded => ( |
| 584 | None, |
| 585 | "unavailable (payload exceeds 1048576-byte measurement bound)".to_string(), |
| 586 | ), |
| 587 | Err(_) => (None, "unavailable (serialization failed)".to_string()), |
| 588 | } |
| 589 | } |
| 590 | |
| 591 | fn bounded_json(value: &Value, limit: usize) -> BoundedString { |
| 592 | let mut writer = BoundedWriter::new(limit); |
| 593 | let result = serde_json::to_writer(&mut writer, value); |
| 594 | BoundedString { |
| 595 | value: String::from_utf8_lossy(&writer.bytes).into_owned(), |
| 596 | truncated: result.is_err() && writer.exceeded, |
| 597 | } |
| 598 | } |
| 599 | |
| 600 | struct BoundedWriter { |
| 601 | bytes: Vec<u8>, |
| 602 | limit: usize, |
| 603 | exceeded: bool, |
| 604 | } |
| 605 | |
| 606 | impl BoundedWriter { |
| 607 | fn new(limit: usize) -> Self { |
| 608 | Self { |
| 609 | bytes: Vec::with_capacity(limit.min(8_192)), |
| 610 | limit, |
| 611 | exceeded: false, |
| 612 | } |
| 613 | } |
| 614 | } |
| 615 | |
| 616 | impl Write for BoundedWriter { |
| 617 | fn write(&mut self, buffer: &[u8]) -> io::Result<usize> { |
| 618 | let remaining = self.limit.saturating_sub(self.bytes.len()); |
| 619 | let accepted = buffer.len().min(remaining); |
| 620 | self.bytes.extend_from_slice(&buffer[..accepted]); |
| 621 | if accepted < buffer.len() { |
| 622 | self.exceeded = true; |
| 623 | return Err(io::Error::other("inspection bound exceeded")); |
| 624 | } |
| 625 | Ok(accepted) |
| 626 | } |
| 627 | |
| 628 | fn flush(&mut self) -> io::Result<()> { |
| 629 | Ok(()) |
| 630 | } |
| 631 | } |
| 632 | |
| 633 | fn bounded_chars(value: &str, limit: usize) -> BoundedString { |
| 634 | let mut chars = value.chars(); |
| 635 | let value = chars.by_ref().take(limit).collect::<String>(); |
| 636 | BoundedString { |
| 637 | value, |
| 638 | truncated: chars.next().is_some(), |
| 639 | } |
| 640 | } |
| 641 | |
| 642 | fn optional_bounded(value: Option<&str>) -> Evidence<BoundedString> { |
| 643 | value.map_or_else( |
| 644 | || unknown("request field absent"), |
| 645 | |value| Evidence::Known { |
| 646 | value: bounded_chars(value, MAX_AUXILIARY_CHARS), |
| 647 | }, |
| 648 | ) |
| 649 | } |
| 650 | |
| 651 | fn optional_copy<T: Copy>(value: Option<&T>) -> Evidence<T> { |
| 652 | value.map_or_else( |
| 653 | || unknown("request field absent"), |
| 654 | |value| Evidence::Known { value: *value }, |
| 655 | ) |
| 656 | } |
| 657 | |
| 658 | fn unknown<T>(reason: &str) -> Evidence<T> { |
| 659 | Evidence::Unknown { |
| 660 | reason: reason.to_string(), |
| 661 | } |
| 662 | } |
| 663 | |
| 664 | #[cfg(test)] |
| 665 | mod tests { |
| 666 | use super::*; |
| 667 | use serde_json::json; |
| 668 | |
| 669 | fn tool(name: &str) -> Tool { |
| 670 | Tool { |
| 671 | tool_type: Some("function".to_string()), |
| 672 | name: name.to_string(), |
| 673 | description: "Read a file".to_string(), |
| 674 | input_schema: json!({"type": "object"}), |
| 675 | allowed_callers: None, |
| 676 | defer_loading: Some(false), |
| 677 | input_examples: None, |
| 678 | strict: Some(true), |
| 679 | cache_control: None, |
| 680 | } |
| 681 | } |
| 682 | |
| 683 | #[test] |
| 684 | fn absent_field_stays_distinct_from_present_empty_array() { |
| 685 | let absent = ToolInspectionSnapshot::from_prepared_request("turn", 1, None); |
| 686 | let empty = ToolInspectionSnapshot::from_prepared_request("turn", 1, Some(&[])); |
| 687 | assert!(!absent.tools_field_present); |
| 688 | assert_eq!(absent.payload_json_bytes, None); |
| 689 | assert!(absent.active_tool_catalog_sha256.is_none()); |
| 690 | assert!(empty.tools_field_present); |
| 691 | assert_eq!(empty.payload_json_bytes, Some(2)); |
| 692 | assert!(empty.active_tool_catalog_sha256.is_some()); |
| 693 | } |
| 694 | |
| 695 | #[test] |
| 696 | fn catalog_digest_is_the_request_manifest_digest_not_a_second_definition() { |
| 697 | let tools = vec![tool("read_file"), tool("write_file")]; |
| 698 | let snapshot = ToolInspectionSnapshot::from_prepared_request("turn", 1, Some(&tools)); |
| 699 | |
| 700 | // Same prepared request, same accounting object: the value the request |
| 701 | // manifest publishes as `active_tool_catalog_sha256`. |
| 702 | assert_eq!( |
| 703 | snapshot.active_tool_catalog_sha256.as_deref(), |
| 704 | Some(crate::core::engine::preview::active_tool_catalog_sha256(&tools).as_str()), |
| 705 | ); |
| 706 | |
| 707 | // And it is a catalog digest, not an incidental byte hash: reordering |
| 708 | // the same tools changes it. |
| 709 | let reordered = vec![tools[1].clone(), tools[0].clone()]; |
| 710 | let reordered = ToolInspectionSnapshot::from_prepared_request("turn", 1, Some(&reordered)); |
| 711 | assert_ne!( |
| 712 | snapshot.active_tool_catalog_sha256, |
| 713 | reordered.active_tool_catalog_sha256 |
| 714 | ); |
| 715 | } |
| 716 | |
| 717 | #[test] |
| 718 | fn projection_preserves_known_false_and_marks_unknown() { |
| 719 | let snapshot = |
| 720 | ToolInspectionSnapshot::from_prepared_request("turn", 3, Some(&[tool("read_file")])); |
| 721 | let text = snapshot.render_text(); |
| 722 | assert!(text.contains("deferred loading: false"), "{text}"); |
| 723 | assert!(text.contains("strict: true"), "{text}"); |
| 724 | assert!(text.contains("allowed callers: unknown (request field absent)")); |
| 725 | assert!(text.contains("Delivery: unknown")); |
| 726 | assert!(text.contains("Provider-wire tool payload: unavailable")); |
| 727 | } |
| 728 | |
| 729 | fn facts(name: &str, plugin: bool, model_visible: bool) -> RegistryFacts { |
| 730 | RegistryFacts { |
| 731 | name: name.to_string(), |
| 732 | description: format!("{name} registry description"), |
| 733 | model_visible, |
| 734 | capabilities: vec!["ReadOnly".to_string()], |
| 735 | approval: "Auto".to_string(), |
| 736 | plugin, |
| 737 | } |
| 738 | } |
| 739 | |
| 740 | fn surface() -> ToolSurfaceContext { |
| 741 | ToolSurfaceContext { |
| 742 | registry: vec![ |
| 743 | facts("read_file", false, true), |
| 744 | facts("plugin_tool", true, true), |
| 745 | facts("hidden_alias", false, false), |
| 746 | facts("not_sent", false, true), |
| 747 | ], |
| 748 | mcp_servers: BTreeMap::from([( |
| 749 | "mcp_my_server_read_file".to_string(), |
| 750 | "my_server".to_string(), |
| 751 | )]), |
| 752 | synthetic_names: vec!["tool_search".to_string()], |
| 753 | provider: ProviderAvailability::Available { |
| 754 | provider: "Deepseek".to_string(), |
| 755 | model: "deepseek-chat".to_string(), |
| 756 | }, |
| 757 | } |
| 758 | } |
| 759 | |
| 760 | #[test] |
| 761 | fn surface_context_turns_provenance_and_attribution_into_truth() { |
| 762 | let tools = vec![ |
| 763 | tool("read_file"), |
| 764 | tool("plugin_tool"), |
| 765 | tool("mcp_my_server_read_file"), |
| 766 | tool("tool_search"), |
| 767 | tool("stranger"), |
| 768 | ]; |
| 769 | let surface = surface(); |
| 770 | let snapshot = ToolInspectionSnapshot::from_prepared_request_with_surface( |
| 771 | "turn", |
| 772 | 1, |
| 773 | Some(&tools), |
| 774 | Some(&surface), |
| 775 | ); |
| 776 | |
| 777 | let provenance = |name: &str| { |
| 778 | snapshot |
| 779 | .tools |
| 780 | .iter() |
| 781 | .find(|entry| entry.name.value == name) |
| 782 | .map(|entry| entry.provenance.clone()) |
| 783 | .expect("projected tool") |
| 784 | }; |
| 785 | for (name, expected) in [ |
| 786 | ("read_file", ToolProvenance::Builtin), |
| 787 | ("plugin_tool", ToolProvenance::Plugin), |
| 788 | ("mcp_my_server_read_file", ToolProvenance::Mcp), |
| 789 | ("tool_search", ToolProvenance::Synthetic), |
| 790 | // Captured context, no resolved origin: unknown is the answer. |
| 791 | ("stranger", ToolProvenance::Unknown), |
| 792 | ] { |
| 793 | assert_eq!( |
| 794 | provenance(name), |
| 795 | Evidence::Known { value: expected }, |
| 796 | "provenance for {name}" |
| 797 | ); |
| 798 | } |
| 799 | |
| 800 | let mcp = snapshot |
| 801 | .tools |
| 802 | .iter() |
| 803 | .find(|entry| entry.name.value == "mcp_my_server_read_file") |
| 804 | .expect("mcp tool"); |
| 805 | // Attribution comes from the pool, not from splitting on `_`. |
| 806 | assert_eq!( |
| 807 | mcp.mcp_server, |
| 808 | Evidence::Known { |
| 809 | value: BoundedString { |
| 810 | value: "my_server".to_string(), |
| 811 | truncated: false, |
| 812 | } |
| 813 | } |
| 814 | ); |
| 815 | |
| 816 | let read_file = snapshot |
| 817 | .tools |
| 818 | .iter() |
| 819 | .find(|entry| entry.name.value == "read_file") |
| 820 | .expect("read_file"); |
| 821 | assert!(matches!(read_file.capabilities, Evidence::Known { .. })); |
| 822 | assert_eq!( |
| 823 | read_file.approval, |
| 824 | Evidence::Known { |
| 825 | value: BoundedString { |
| 826 | value: "Auto".to_string(), |
| 827 | truncated: false, |
| 828 | } |
| 829 | } |
| 830 | ); |
| 831 | assert_eq!(read_file.model_visible, Evidence::Known { value: true }); |
| 832 | |
| 833 | // Unregistered tools stay unknown rather than being reported as "none". |
| 834 | let stranger = snapshot |
| 835 | .tools |
| 836 | .iter() |
| 837 | .find(|entry| entry.name.value == "stranger") |
| 838 | .expect("stranger"); |
| 839 | assert!(matches!(stranger.capabilities, Evidence::Unknown { .. })); |
| 840 | assert!(matches!(stranger.approval, Evidence::Unknown { .. })); |
| 841 | |
| 842 | // The unavailable set shrinks to what nothing here can observe. |
| 843 | assert_eq!( |
| 844 | snapshot.unavailable_for_this_request, |
| 845 | vec!["provider_wire_payload"] |
| 846 | ); |
| 847 | assert!(snapshot.provider.is_available()); |
| 848 | assert!(snapshot.registry_facts_present); |
| 849 | assert_eq!(snapshot.registry_tool_count, Evidence::Known { value: 4 }); |
| 850 | |
| 851 | let text = snapshot.render_text(); |
| 852 | assert!(text.contains("provenance: synthetic"), "{text}"); |
| 853 | assert!(text.contains("MCP server: \"my_server\""), "{text}"); |
| 854 | assert!(text.contains("Provider: \"Deepseek\""), "{text}"); |
| 855 | assert!( |
| 856 | text.contains("Provider-wire tool payload: unavailable"), |
| 857 | "{text}" |
| 858 | ); |
| 859 | } |
| 860 | |
| 861 | #[test] |
| 862 | fn registry_only_tools_are_counted_without_expanding_the_projection() { |
| 863 | let tools = vec![tool("read_file")]; |
| 864 | let surface = surface(); |
| 865 | let snapshot = ToolInspectionSnapshot::from_prepared_request_with_surface( |
| 866 | "turn", |
| 867 | 1, |
| 868 | Some(&tools), |
| 869 | Some(&surface), |
| 870 | ); |
| 871 | |
| 872 | // Only the request's tools are projected; the rest are counted. |
| 873 | assert_eq!(snapshot.tools.len(), 1); |
| 874 | let Evidence::Known { value } = &snapshot.registry_only_tools else { |
| 875 | panic!("registry-only tools must be known when facts were captured"); |
| 876 | }; |
| 877 | // `hidden_alias` is not model-visible, so it is not a missing tool. |
| 878 | assert_eq!(value.count, 2); |
| 879 | let rendered = value |
| 880 | .rendered |
| 881 | .iter() |
| 882 | .map(|entry| entry.value.as_str()) |
| 883 | .collect::<Vec<_>>(); |
| 884 | // Order follows the registry facts as supplied; the producer sorts. |
| 885 | assert_eq!(rendered, vec!["plugin_tool", "not_sent"]); |
| 886 | |
| 887 | // Everything projected is in the request; the request is the evidence. |
| 888 | assert!(snapshot.tools.iter().all(|entry| { |
| 889 | entry.visibility.in_request() && entry.visibility == ToolVisibility::Active |
| 890 | })); |
| 891 | } |
| 892 | |
| 893 | #[test] |
| 894 | fn absent_surface_keeps_every_registry_derived_field_unknown() { |
| 895 | let tools = vec![tool("read_file")]; |
| 896 | let snapshot = ToolInspectionSnapshot::from_prepared_request("turn", 1, Some(&tools)); |
| 897 | |
| 898 | assert_eq!(snapshot.provider, ProviderAvailability::Unknown); |
| 899 | assert!(!snapshot.registry_facts_present); |
| 900 | assert!(matches!( |
| 901 | snapshot.registry_tool_count, |
| 902 | Evidence::Unknown { .. } |
| 903 | )); |
| 904 | assert!(matches!( |
| 905 | snapshot.registry_only_tools, |
| 906 | Evidence::Unknown { .. } |
| 907 | )); |
| 908 | assert_eq!( |
| 909 | snapshot.unavailable_for_this_request, |
| 910 | vec![ |
| 911 | "provider_wire_payload", |
| 912 | "provider", |
| 913 | "model", |
| 914 | "approval", |
| 915 | "provenance", |
| 916 | "capabilities", |
| 917 | ] |
| 918 | ); |
| 919 | let entry = &snapshot.tools[0]; |
| 920 | for evidence in [ |
| 921 | matches!(entry.provenance, Evidence::Unknown { .. }), |
| 922 | matches!(entry.mcp_server, Evidence::Unknown { .. }), |
| 923 | matches!(entry.capabilities, Evidence::Unknown { .. }), |
| 924 | matches!(entry.approval, Evidence::Unknown { .. }), |
| 925 | matches!(entry.model_visible, Evidence::Unknown { .. }), |
| 926 | ] { |
| 927 | assert!(evidence); |
| 928 | } |
| 929 | // Wire facts are still exact without a surface context. |
| 930 | assert!(entry.visibility.in_request()); |
| 931 | assert!(snapshot.active_tool_catalog_sha256.is_some()); |
| 932 | } |
| 933 | |
| 934 | #[test] |
| 935 | fn provider_receipt_records_an_unresolved_client_without_borrowing_registry_truth() { |
| 936 | let tools = vec![tool("read_file")]; |
| 937 | let surface = ToolSurfaceContext { |
| 938 | registry: vec![facts("read_file", false, true)], |
| 939 | provider: ProviderAvailability::Unavailable { |
| 940 | reason: "no model client resolved for this turn".to_string(), |
| 941 | }, |
| 942 | ..ToolSurfaceContext::default() |
| 943 | }; |
| 944 | let snapshot = ToolInspectionSnapshot::from_prepared_request_with_surface( |
| 945 | "turn", |
| 946 | 1, |
| 947 | Some(&tools), |
| 948 | Some(&surface), |
| 949 | ); |
| 950 | |
| 951 | // A full registry does not make a provider available. |
| 952 | assert!(!snapshot.provider.is_available()); |
| 953 | assert_eq!(snapshot.provider.label(), "unavailable"); |
| 954 | assert!(snapshot.unavailable_for_this_request.contains(&"provider")); |
| 955 | // Registry-derived truth is unaffected by the missing client. |
| 956 | assert!( |
| 957 | !snapshot |
| 958 | .unavailable_for_this_request |
| 959 | .contains(&"provenance") |
| 960 | ); |
| 961 | assert_eq!( |
| 962 | snapshot.tools[0].provenance, |
| 963 | Evidence::Known { |
| 964 | value: ToolProvenance::Builtin |
| 965 | } |
| 966 | ); |
| 967 | } |
| 968 | |
| 969 | #[test] |
| 970 | fn capture_and_rendering_are_bounded_with_explicit_receipts() { |
| 971 | let mut tools = (0..40) |
| 972 | .map(|index| { |
| 973 | let mut value = tool(&format!("tool_{index}")); |
| 974 | value.description = "x".repeat(MAX_DESCRIPTION_CHARS + 10); |
| 975 | value.input_schema = json!({"large": "y".repeat(MAX_SCHEMA_BYTES * 600)}); |
| 976 | value |
| 977 | }) |
| 978 | .collect::<Vec<_>>(); |
| 979 | tools[0].allowed_callers = Some( |
| 980 | (0..20) |
| 981 | .map(|caller| format!("caller-{caller}-{}", "z".repeat(200))) |
| 982 | .collect(), |
| 983 | ); |
| 984 | let snapshot = ToolInspectionSnapshot::from_prepared_request( |
| 985 | &"t".repeat(MAX_AUXILIARY_CHARS + 1), |
| 986 | 1, |
| 987 | Some(&tools), |
| 988 | ); |
| 989 | assert_eq!(snapshot.rendered_tool_count, MAX_RENDERED_TOOLS); |
| 990 | assert_eq!(snapshot.omitted_tool_count, 8); |
| 991 | assert!(snapshot.turn_id.truncated); |
| 992 | assert!(snapshot.tools[0].description.truncated); |
| 993 | assert!(snapshot.tools[0].input_schema_json.truncated); |
| 994 | assert_eq!(snapshot.payload_json_bytes, None); |
| 995 | assert!(snapshot.payload_measurement_status.contains("exceeds")); |
| 996 | // The catalog digest is fixed-width, so it survives the byte bound. |
| 997 | assert!(snapshot.active_tool_catalog_sha256.is_some()); |
| 998 | let json = snapshot.render_json().expect("bounded JSON"); |
| 999 | // 160 KiB: the per-tool evidence fields (provenance, MCP server, |
| 1000 | // capabilities, approval, visibility) each carry an explicit reason |
| 1001 | // string when unresolved, which is the point — the cap moved, the |
| 1002 | // bound did not disappear. |
| 1003 | assert!( |
| 1004 | json.len() < 163_840, |
| 1005 | "projection grew to {} bytes", |
| 1006 | json.len() |
| 1007 | ); |
| 1008 | } |
| 1009 | } |
| 1010 |