| 1 | //! `retrieve_tool_result` - selective retrieval for spilled tool outputs. |
| 2 | //! |
| 3 | //! Exact tool evidence is retained under its origin session. Historical |
| 4 | //! payloads in the global `tool_outputs/` compatibility directory are readable |
| 5 | //! only when a digest-bound ownership sidecar proves they belong to the active |
| 6 | //! session. |
| 7 | |
| 8 | use std::fs; |
| 9 | use std::path::PathBuf; |
| 10 | |
| 11 | use async_trait::async_trait; |
| 12 | use serde_json::{Value, json}; |
| 13 | |
| 14 | use super::spec::{ |
| 15 | ToolCapability, ToolContext, ToolError, ToolResult, ToolSpec, optional_str, optional_u64, |
| 16 | required_str, |
| 17 | }; |
| 18 | |
| 19 | const DEFAULT_MAX_BYTES: usize = 8 * 1024; |
| 20 | const HARD_MAX_BYTES: usize = 128 * 1024; |
| 21 | const DEFAULT_LINE_COUNT: usize = 40; |
| 22 | const HARD_LINE_COUNT: usize = 500; |
| 23 | const DEFAULT_MAX_MATCHES: usize = 20; |
| 24 | const HARD_MAX_MATCHES: usize = 100; |
| 25 | const DEFAULT_CONTEXT_LINES: usize = 1; |
| 26 | const HARD_CONTEXT_LINES: usize = 5; |
| 27 | |
| 28 | /// Retrieve summaries or slices of a prior spilled tool result. |
| 29 | pub struct RetrieveToolResultTool; |
| 30 | |
| 31 | #[async_trait] |
| 32 | impl ToolSpec for RetrieveToolResultTool { |
| 33 | fn name(&self) -> &'static str { |
| 34 | "retrieve_tool_result" |
| 35 | } |
| 36 | |
| 37 | fn description(&self) -> &'static str { |
| 38 | "Inspect retained tool evidence with strict session ownership and bounds. Accepts an artifact id, validated session-relative path, or an ownership-proven legacy call/SHA reference. Unowned legacy-global evidence fails closed. Modes: metadata, summary, head, tail, lines, query, bytes. bytes returns a bounded base64 slice for exact text or binary recovery." |
| 39 | } |
| 40 | |
| 41 | fn input_schema(&self) -> Value { |
| 42 | json!({ |
| 43 | "type": "object", |
| 44 | "properties": { |
| 45 | "ref": { |
| 46 | "type": "string", |
| 47 | "description": "Session-owned artifact id (`art_<id>`) or validated artifact-relative path. Legacy call-id/SHA references work only when origin-session ownership was recorded." |
| 48 | }, |
| 49 | "mode": { |
| 50 | "type": "string", |
| 51 | "enum": ["metadata", "summary", "head", "tail", "lines", "query", "bytes"], |
| 52 | "description": "Retrieval mode. Defaults to summary." |
| 53 | }, |
| 54 | "query": { |
| 55 | "type": "string", |
| 56 | "description": "Case-insensitive substring to search for when mode=query." |
| 57 | }, |
| 58 | "lines": { |
| 59 | "type": "string", |
| 60 | "description": "Line selector for mode=lines, e.g. \"10\" or \"10-40\"." |
| 61 | }, |
| 62 | "start_line": { |
| 63 | "type": "integer", |
| 64 | "description": "1-based first line for mode=lines." |
| 65 | }, |
| 66 | "end_line": { |
| 67 | "type": "integer", |
| 68 | "description": "1-based final line for mode=lines." |
| 69 | }, |
| 70 | "line_count": { |
| 71 | "type": "integer", |
| 72 | "description": "Number of lines for head/tail modes. Default 40, hard cap 500." |
| 73 | }, |
| 74 | "max_bytes": { |
| 75 | "type": "integer", |
| 76 | "description": "Maximum bytes of excerpt text returned. Default 8192, hard cap 131072." |
| 77 | }, |
| 78 | "max_matches": { |
| 79 | "type": "integer", |
| 80 | "description": "Maximum query matches or signal lines returned. Default 20, hard cap 100." |
| 81 | }, |
| 82 | "context_lines": { |
| 83 | "type": "integer", |
| 84 | "description": "Extra lines around each query match. Default 1, hard cap 5." |
| 85 | }, |
| 86 | "generation": { |
| 87 | "type": "integer", |
| 88 | "minimum": 1, |
| 89 | "description": "Optional expected evidence generation; mismatches fail closed." |
| 90 | }, |
| 91 | "offset": { |
| 92 | "type": "integer", |
| 93 | "minimum": 0, |
| 94 | "description": "Zero-based byte offset for mode=bytes." |
| 95 | }, |
| 96 | "length": { |
| 97 | "type": "integer", |
| 98 | "minimum": 1, |
| 99 | "description": "Byte count for mode=bytes, capped at max_bytes." |
| 100 | } |
| 101 | }, |
| 102 | "required": ["ref"] |
| 103 | }) |
| 104 | } |
| 105 | |
| 106 | fn capabilities(&self) -> Vec<ToolCapability> { |
| 107 | vec![ToolCapability::ReadOnly] |
| 108 | } |
| 109 | |
| 110 | fn supports_parallel(&self) -> bool { |
| 111 | true |
| 112 | } |
| 113 | |
| 114 | async fn execute(&self, input: Value, context: &ToolContext) -> Result<ToolResult, ToolError> { |
| 115 | let reference = required_str(&input, "ref")?.trim(); |
| 116 | if reference.is_empty() { |
| 117 | return Err(ToolError::invalid_input("ref cannot be empty")); |
| 118 | } |
| 119 | |
| 120 | let mode = optional_str(&input, "mode")? |
| 121 | .unwrap_or("summary") |
| 122 | .trim() |
| 123 | .to_ascii_lowercase(); |
| 124 | let max_bytes = clamp_u64( |
| 125 | optional_u64(&input, "max_bytes", DEFAULT_MAX_BYTES as u64)?, |
| 126 | 1, |
| 127 | HARD_MAX_BYTES, |
| 128 | ); |
| 129 | let resolved = resolve_spillover_reference(reference, &context.state_namespace)?; |
| 130 | let legacy_ownership = if resolved.kind == ResolvedReferenceKind::LegacyGlobal { |
| 131 | Some(authorize_legacy_spillover( |
| 132 | &resolved.path, |
| 133 | &context.state_namespace, |
| 134 | )?) |
| 135 | } else { |
| 136 | None |
| 137 | }; |
| 138 | let bytes = fs::read(&resolved.path).map_err(|_| { |
| 139 | ToolError::execution_failed("evidence is missing or no longer retained") |
| 140 | })?; |
| 141 | if let Some(ownership) = legacy_ownership { |
| 142 | let size = u64::try_from(bytes.len()).unwrap_or(u64::MAX); |
| 143 | if ownership.size_bytes != size |
| 144 | || ownership.digest != crate::hashing::sha256_hex(&bytes) |
| 145 | { |
| 146 | return Err(ToolError::execution_failed( |
| 147 | "legacy evidence content is corrupt", |
| 148 | )); |
| 149 | } |
| 150 | } |
| 151 | let evidence = validate_evidence_if_present( |
| 152 | reference, |
| 153 | &resolved.path, |
| 154 | &bytes, |
| 155 | &context.state_namespace, |
| 156 | &input, |
| 157 | )?; |
| 158 | if mode == "metadata" { |
| 159 | return ToolResult::json(&json!({ |
| 160 | "ref": reference, |
| 161 | "available": true, |
| 162 | "total_bytes": bytes.len(), |
| 163 | "evidence": evidence, |
| 164 | })) |
| 165 | .map_err(|err| ToolError::execution_failed(err.to_string())); |
| 166 | } |
| 167 | if mode == "bytes" { |
| 168 | use base64::Engine as _; |
| 169 | let offset = input.get("offset").and_then(Value::as_u64).unwrap_or(0) as usize; |
| 170 | let requested = input |
| 171 | .get("length") |
| 172 | .and_then(Value::as_u64) |
| 173 | .unwrap_or(max_bytes as u64) as usize; |
| 174 | let end = offset |
| 175 | .saturating_add(requested.min(max_bytes)) |
| 176 | .min(bytes.len()); |
| 177 | let slice = bytes.get(offset.min(bytes.len())..end).unwrap_or_default(); |
| 178 | return ToolResult::json(&json!({ |
| 179 | "ref": reference, |
| 180 | "mode": "bytes", |
| 181 | "offset": offset, |
| 182 | "returned_bytes": slice.len(), |
| 183 | "total_bytes": bytes.len(), |
| 184 | "encoding": "base64", |
| 185 | "data": base64::engine::general_purpose::STANDARD.encode(slice), |
| 186 | })) |
| 187 | .map_err(|err| ToolError::execution_failed(err.to_string())); |
| 188 | } |
| 189 | let content = String::from_utf8(bytes).map_err(|_| { |
| 190 | ToolError::execution_failed( |
| 191 | "evidence encoding is binary; bounded text inspection is unavailable", |
| 192 | ) |
| 193 | })?; |
| 194 | |
| 195 | let lines: Vec<&str> = content.lines().collect(); |
| 196 | let payload = match mode.as_str() { |
| 197 | "summary" => build_summary_payload(reference, &content, &lines, &input, max_bytes)?, |
| 198 | "head" => build_head_tail_payload(reference, "head", &lines, &input, max_bytes)?, |
| 199 | "tail" => build_head_tail_payload(reference, "tail", &lines, &input, max_bytes)?, |
| 200 | "lines" => build_lines_payload(reference, &lines, &input, max_bytes)?, |
| 201 | "query" => build_query_payload(reference, &lines, &input, max_bytes)?, |
| 202 | other => { |
| 203 | return Err(ToolError::invalid_input(format!( |
| 204 | "unsupported mode `{other}` (expected metadata, summary, head, tail, lines, query, or bytes)" |
| 205 | ))); |
| 206 | } |
| 207 | }; |
| 208 | |
| 209 | ToolResult::json(&payload).map_err(|err| { |
| 210 | ToolError::execution_failed(format!("failed to serialize result: {err}")) |
| 211 | }) |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | fn validate_evidence_if_present( |
| 216 | reference: &str, |
| 217 | path: &std::path::Path, |
| 218 | bytes: &[u8], |
| 219 | session_id: &str, |
| 220 | input: &Value, |
| 221 | ) -> Result<Option<crate::tools::large_output_router::EvidenceArtifact>, ToolError> { |
| 222 | let handle = path |
| 223 | .file_stem() |
| 224 | .and_then(|stem| stem.to_str()) |
| 225 | .filter(|stem| stem.starts_with("art_")) |
| 226 | .or_else(|| { |
| 227 | reference |
| 228 | .trim() |
| 229 | .starts_with("art_") |
| 230 | .then(|| reference.trim()) |
| 231 | }); |
| 232 | let Some(handle) = handle else { |
| 233 | return Ok(None); |
| 234 | }; |
| 235 | let metadata = |
| 236 | match crate::tools::large_output_router::read_evidence_metadata(session_id, handle) { |
| 237 | Ok(metadata) => metadata, |
| 238 | Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ok(None), |
| 239 | Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => { |
| 240 | return Err(ToolError::permission_denied( |
| 241 | "evidence belongs to another session", |
| 242 | )); |
| 243 | } |
| 244 | Err(_) => return Err(ToolError::execution_failed("evidence metadata is corrupt")), |
| 245 | }; |
| 246 | if metadata.origin_session != session_id || metadata.handle != handle { |
| 247 | return Err(ToolError::permission_denied( |
| 248 | "evidence belongs to another session", |
| 249 | )); |
| 250 | } |
| 251 | if metadata.redacted { |
| 252 | return Err(ToolError::permission_denied("evidence has been redacted")); |
| 253 | } |
| 254 | if crate::tools::large_output_router::evidence_is_expired( |
| 255 | &metadata, |
| 256 | crate::tools::large_output_router::unix_millis_now(), |
| 257 | ) { |
| 258 | return Err(ToolError::execution_failed( |
| 259 | "evidence retention has expired", |
| 260 | )); |
| 261 | } |
| 262 | if input |
| 263 | .get("generation") |
| 264 | .and_then(Value::as_u64) |
| 265 | .is_some_and(|generation| generation != u64::from(metadata.generation)) |
| 266 | { |
| 267 | return Err(ToolError::execution_failed( |
| 268 | "evidence generation does not match", |
| 269 | )); |
| 270 | } |
| 271 | if metadata.size_bytes != u64::try_from(bytes.len()).unwrap_or(u64::MAX) |
| 272 | || metadata.digest != crate::hashing::sha256_hex(bytes) |
| 273 | { |
| 274 | return Err(ToolError::execution_failed("evidence content is corrupt")); |
| 275 | } |
| 276 | Ok(Some(metadata)) |
| 277 | } |
| 278 | |
| 279 | #[derive(Debug, Clone, Copy, PartialEq, Eq)] |
| 280 | enum ResolvedReferenceKind { |
| 281 | ActiveSession, |
| 282 | LegacyGlobal, |
| 283 | } |
| 284 | |
| 285 | #[derive(Debug, Clone, PartialEq, Eq)] |
| 286 | struct ResolvedSpilloverReference { |
| 287 | path: PathBuf, |
| 288 | kind: ResolvedReferenceKind, |
| 289 | } |
| 290 | |
| 291 | fn authorize_legacy_spillover( |
| 292 | path: &std::path::Path, |
| 293 | session_id: &str, |
| 294 | ) -> Result<crate::tools::truncate::LegacySpilloverOwnership, ToolError> { |
| 295 | if session_id.trim().is_empty() { |
| 296 | return Err(ToolError::permission_denied( |
| 297 | "legacy evidence has no verifiable session owner", |
| 298 | )); |
| 299 | } |
| 300 | let ownership = match crate::tools::truncate::read_legacy_spillover_ownership(path) { |
| 301 | Ok(ownership) => ownership, |
| 302 | Err(error) if error.kind() == std::io::ErrorKind::NotFound => { |
| 303 | return Err(ToolError::permission_denied( |
| 304 | "legacy evidence has no verifiable session owner", |
| 305 | )); |
| 306 | } |
| 307 | Err(error) if error.kind() == std::io::ErrorKind::PermissionDenied => { |
| 308 | return Err(ToolError::permission_denied( |
| 309 | "legacy evidence ownership proof is invalid", |
| 310 | )); |
| 311 | } |
| 312 | Err(_) => { |
| 313 | return Err(ToolError::execution_failed( |
| 314 | "legacy evidence ownership metadata is corrupt", |
| 315 | )); |
| 316 | } |
| 317 | }; |
| 318 | if ownership.origin_session != session_id { |
| 319 | return Err(ToolError::permission_denied( |
| 320 | "legacy evidence belongs to another session", |
| 321 | )); |
| 322 | } |
| 323 | Ok(ownership) |
| 324 | } |
| 325 | |
| 326 | /// Resolve a tool-result ref without weakening its ownership boundary. |
| 327 | /// |
| 328 | /// Current-session artifacts always win over same-named global compatibility |
| 329 | /// files. A legacy call-id, SHA, relative path, or absolute path can resolve, |
| 330 | /// but the caller must validate its ownership sidecar before reading bytes. |
| 331 | fn resolve_spillover_reference( |
| 332 | reference: &str, |
| 333 | session_id: &str, |
| 334 | ) -> Result<ResolvedSpilloverReference, ToolError> { |
| 335 | let root = crate::tools::truncate::spillover_root() |
| 336 | .ok_or_else(|| ToolError::execution_failed("retained evidence storage is unavailable"))?; |
| 337 | let root_canonical = root.canonicalize().ok(); |
| 338 | |
| 339 | // Resolve the session's `artifacts/` directory. |
| 340 | // `session_artifact_absolute_path(sid, p)` returns |
| 341 | // `~/.codewhale/sessions/<sid>/<p>` — so passing the literal |
| 342 | // `ARTIFACTS_DIR_NAME` ("artifacts") gets us the real artifacts |
| 343 | // root. An earlier draft passed `Path::new(".")` and took |
| 344 | // `.parent()`, which landed one directory too high (`<sid>` instead |
| 345 | // of `<sid>/artifacts`) and silently broke every bare `art_<id>` |
| 346 | // ref — only the legacy-spillover fallback survived. The test |
| 347 | // `resolves_art_prefix_to_legacy_spillover_id` masked it because |
| 348 | // it ONLY wrote a legacy spillover file. The new test |
| 349 | // `resolves_art_prefix_via_session_artifacts` exercises the real |
| 350 | // path. |
| 351 | let session_artifacts_root = if !session_id.is_empty() { |
| 352 | crate::artifacts::session_artifact_absolute_path( |
| 353 | session_id, |
| 354 | std::path::Path::new(crate::artifacts::ARTIFACTS_DIR_NAME), |
| 355 | ) |
| 356 | } else { |
| 357 | None |
| 358 | }; |
| 359 | let session_artifacts_root_canonical = session_artifacts_root |
| 360 | .as_ref() |
| 361 | .and_then(|p| p.canonicalize().ok()); |
| 362 | |
| 363 | let trimmed = reference.trim(); |
| 364 | let stripped = trimmed |
| 365 | .strip_prefix("tool_result:") |
| 366 | .unwrap_or(trimmed) |
| 367 | .trim(); |
| 368 | |
| 369 | let mut tried = 0_usize; |
| 370 | let try_path = |candidate: PathBuf, tried: &mut usize| -> Option<ResolvedSpilloverReference> { |
| 371 | *tried = (*tried).saturating_add(1); |
| 372 | |
| 373 | // Reject symlinks at the leaf BEFORE canonicalizing so an |
| 374 | // attacker who can write under `<sid>/artifacts/` cannot |
| 375 | // plant a symlink to `/etc/passwd` and read it back through |
| 376 | // `retrieve_tool_result`. canonicalize() would happily |
| 377 | // follow such a link and then pass the `starts_with(root)` |
| 378 | // check because of the resolved-then-compare order. Both session and |
| 379 | // compatibility roots reject leaf symlinks. Legacy |
| 380 | // compatibility files also need a separate ownership sidecar below. |
| 381 | if let Ok(meta) = std::fs::symlink_metadata(&candidate) |
| 382 | && meta.file_type().is_symlink() |
| 383 | { |
| 384 | return None; |
| 385 | } |
| 386 | |
| 387 | let canonical = candidate.canonicalize().ok()?; |
| 388 | if !canonical.is_file() { |
| 389 | return None; |
| 390 | } |
| 391 | let inside_legacy = root_canonical |
| 392 | .as_ref() |
| 393 | .is_some_and(|root| canonical.starts_with(root)); |
| 394 | let inside_session = session_artifacts_root_canonical |
| 395 | .as_ref() |
| 396 | .is_some_and(|root| canonical.starts_with(root)); |
| 397 | if inside_session { |
| 398 | Some(ResolvedSpilloverReference { |
| 399 | path: canonical, |
| 400 | kind: ResolvedReferenceKind::ActiveSession, |
| 401 | }) |
| 402 | } else if inside_legacy { |
| 403 | Some(ResolvedSpilloverReference { |
| 404 | path: canonical, |
| 405 | kind: ResolvedReferenceKind::LegacyGlobal, |
| 406 | }) |
| 407 | } else { |
| 408 | None |
| 409 | } |
| 410 | }; |
| 411 | |
| 412 | // Form 1/3: absolute path. Validate it lives under one of the allowed roots. |
| 413 | let raw_path = PathBuf::from(stripped); |
| 414 | if raw_path.is_absolute() { |
| 415 | if let Some(found) = try_path(raw_path, &mut tried) { |
| 416 | return Ok(found); |
| 417 | } |
| 418 | return Err(ToolError::permission_denied( |
| 419 | "evidence path is not owned by the active session", |
| 420 | )); |
| 421 | } |
| 422 | |
| 423 | // Session artifact paths take priority over legacy-global lookups. |
| 424 | let looks_like_path = stripped.ends_with(".txt") |
| 425 | || stripped.contains('/') |
| 426 | || (std::path::MAIN_SEPARATOR != '/' && stripped.contains(std::path::MAIN_SEPARATOR)); |
| 427 | if looks_like_path { |
| 428 | if let Some(sa_root) = session_artifacts_root.as_ref() { |
| 429 | let rel = stripped.strip_prefix("artifacts/").unwrap_or(stripped); |
| 430 | if let Some(found) = try_path(sa_root.join(rel), &mut tried) { |
| 431 | return Ok(found); |
| 432 | } |
| 433 | } |
| 434 | if let Some(found) = try_path(root.join(stripped), &mut tried) { |
| 435 | return Ok(found); |
| 436 | } |
| 437 | return Err(not_found(reference, tried)); |
| 438 | } |
| 439 | |
| 440 | if let Some(sa_root) = session_artifacts_root.as_ref() { |
| 441 | let file_name = if stripped.starts_with("art_") { |
| 442 | format!("{stripped}.txt") |
| 443 | } else { |
| 444 | format!("art_{stripped}.txt") |
| 445 | }; |
| 446 | if let Some(found) = try_path(sa_root.join(file_name), &mut tried) { |
| 447 | return Ok(found); |
| 448 | } |
| 449 | } |
| 450 | |
| 451 | // `sha:<hex>` or bare 64-hex resolves only as legacy-global evidence and |
| 452 | // therefore still requires an ownership sidecar in the caller. |
| 453 | let sha_candidate = stripped |
| 454 | .strip_prefix("sha:") |
| 455 | .or_else(|| stripped.strip_prefix("sha_")) |
| 456 | .unwrap_or(stripped) |
| 457 | .trim(); |
| 458 | if crate::tools::truncate::is_valid_sha256(&sha_candidate.to_ascii_lowercase()) |
| 459 | && let Some(p) = crate::tools::truncate::sha_spillover_path(sha_candidate) |
| 460 | && let Some(found) = try_path(p, &mut tried) |
| 461 | { |
| 462 | return Ok(found); |
| 463 | } |
| 464 | |
| 465 | // Compatibility lookup: `art_<id>` may name the historical `<id>.txt`. |
| 466 | if let Some(stripped_art) = stripped.strip_prefix("art_") |
| 467 | && let Some(p) = crate::tools::truncate::spillover_path(stripped_art) |
| 468 | && let Some(found) = try_path(p, &mut tried) |
| 469 | { |
| 470 | return Ok(found); |
| 471 | } |
| 472 | |
| 473 | if let Some(path) = crate::tools::truncate::spillover_path(stripped) |
| 474 | && let Some(found) = try_path(path, &mut tried) |
| 475 | { |
| 476 | return Ok(found); |
| 477 | } |
| 478 | |
| 479 | Err(not_found(reference, tried)) |
| 480 | } |
| 481 | |
| 482 | /// Missing evidence is distinct without revealing storage roots or generated |
| 483 | /// session identifiers to the model. |
| 484 | fn not_found(reference: &str, tried: usize) -> ToolError { |
| 485 | ToolError::execution_failed(format!( |
| 486 | "retained evidence `{reference}` was not found for the active session \ |
| 487 | ({tried} bounded candidate forms checked). Use the session-owned \ |
| 488 | `art_<id>` handle from the original receipt." |
| 489 | )) |
| 490 | } |
| 491 | |
| 492 | fn build_summary_payload( |
| 493 | reference: &str, |
| 494 | content: &str, |
| 495 | lines: &[&str], |
| 496 | input: &Value, |
| 497 | max_bytes: usize, |
| 498 | ) -> Result<Value, ToolError> { |
| 499 | let max_matches = clamp_u64( |
| 500 | optional_u64(input, "max_matches", DEFAULT_MAX_MATCHES as u64)?, |
| 501 | 1, |
| 502 | HARD_MAX_MATCHES, |
| 503 | ); |
| 504 | let signal_lines = collect_signal_lines(lines, max_matches); |
| 505 | let head_count = DEFAULT_LINE_COUNT.min(lines.len()); |
| 506 | let tail_count = DEFAULT_LINE_COUNT.min(lines.len()); |
| 507 | let head = render_numbered_lines( |
| 508 | lines |
| 509 | .iter() |
| 510 | .take(head_count) |
| 511 | .enumerate() |
| 512 | .map(|(idx, line)| (idx + 1, *line)), |
| 513 | max_bytes / 2, |
| 514 | ); |
| 515 | let tail_start = lines.len().saturating_sub(tail_count); |
| 516 | let tail = render_numbered_lines( |
| 517 | lines |
| 518 | .iter() |
| 519 | .enumerate() |
| 520 | .skip(tail_start) |
| 521 | .map(|(idx, line)| (idx + 1, *line)), |
| 522 | max_bytes / 2, |
| 523 | ); |
| 524 | |
| 525 | Ok(json!({ |
| 526 | "ref": reference, |
| 527 | "mode": "summary", |
| 528 | "total_bytes": content.len(), |
| 529 | "total_lines": lines.len(), |
| 530 | "non_empty_lines": lines.iter().filter(|line| !line.trim().is_empty()).count(), |
| 531 | "signal_lines": signal_lines, |
| 532 | "head": head, |
| 533 | "tail": tail, |
| 534 | "hint": "Use mode=head, tail, lines, or query to retrieve a narrower slice." |
| 535 | })) |
| 536 | } |
| 537 | |
| 538 | fn build_head_tail_payload( |
| 539 | reference: &str, |
| 540 | mode: &str, |
| 541 | lines: &[&str], |
| 542 | input: &Value, |
| 543 | max_bytes: usize, |
| 544 | ) -> Result<Value, ToolError> { |
| 545 | let count = clamp_u64( |
| 546 | optional_u64(input, "line_count", DEFAULT_LINE_COUNT as u64)?, |
| 547 | 1, |
| 548 | HARD_LINE_COUNT, |
| 549 | ); |
| 550 | let selected: Vec<(usize, &str)> = if mode == "head" { |
| 551 | lines |
| 552 | .iter() |
| 553 | .take(count) |
| 554 | .enumerate() |
| 555 | .map(|(idx, line)| (idx + 1, *line)) |
| 556 | .collect() |
| 557 | } else { |
| 558 | let start = lines.len().saturating_sub(count); |
| 559 | lines |
| 560 | .iter() |
| 561 | .enumerate() |
| 562 | .skip(start) |
| 563 | .map(|(idx, line)| (idx + 1, *line)) |
| 564 | .collect() |
| 565 | }; |
| 566 | let excerpt = render_numbered_lines(selected.iter().copied(), max_bytes); |
| 567 | |
| 568 | Ok(json!({ |
| 569 | "ref": reference, |
| 570 | "mode": mode, |
| 571 | "total_lines": lines.len(), |
| 572 | "line_count": count, |
| 573 | "excerpt": excerpt, |
| 574 | })) |
| 575 | } |
| 576 | |
| 577 | fn build_lines_payload( |
| 578 | reference: &str, |
| 579 | lines: &[&str], |
| 580 | input: &Value, |
| 581 | max_bytes: usize, |
| 582 | ) -> Result<Value, ToolError> { |
| 583 | let (start, end) = parse_line_selector(input)?; |
| 584 | let excerpt = if start > lines.len() { |
| 585 | String::new() |
| 586 | } else { |
| 587 | let end = end.min(lines.len()); |
| 588 | render_numbered_lines( |
| 589 | lines |
| 590 | .iter() |
| 591 | .enumerate() |
| 592 | .skip(start - 1) |
| 593 | .take(end.saturating_sub(start) + 1) |
| 594 | .map(|(idx, line)| (idx + 1, *line)), |
| 595 | max_bytes, |
| 596 | ) |
| 597 | }; |
| 598 | |
| 599 | Ok(json!({ |
| 600 | "ref": reference, |
| 601 | "mode": "lines", |
| 602 | "total_lines": lines.len(), |
| 603 | "start_line": start, |
| 604 | "end_line": end.min(lines.len()), |
| 605 | "excerpt": excerpt, |
| 606 | })) |
| 607 | } |
| 608 | |
| 609 | fn build_query_payload( |
| 610 | reference: &str, |
| 611 | lines: &[&str], |
| 612 | input: &Value, |
| 613 | max_bytes: usize, |
| 614 | ) -> Result<Value, ToolError> { |
| 615 | let query = optional_str(input, "query")? |
| 616 | .map(str::trim) |
| 617 | .filter(|q| !q.is_empty()) |
| 618 | .ok_or_else(|| ToolError::invalid_input("query is required when mode=query"))?; |
| 619 | let query_lower = query.to_lowercase(); |
| 620 | let max_matches = clamp_u64( |
| 621 | optional_u64(input, "max_matches", DEFAULT_MAX_MATCHES as u64)?, |
| 622 | 1, |
| 623 | HARD_MAX_MATCHES, |
| 624 | ); |
| 625 | let context_lines = clamp_u64( |
| 626 | optional_u64(input, "context_lines", DEFAULT_CONTEXT_LINES as u64)?, |
| 627 | 0, |
| 628 | HARD_CONTEXT_LINES, |
| 629 | ); |
| 630 | |
| 631 | let mut matched_lines = 0usize; |
| 632 | let mut results = Vec::new(); |
| 633 | for (idx, line) in lines.iter().enumerate() { |
| 634 | if !line.to_lowercase().contains(&query_lower) { |
| 635 | continue; |
| 636 | } |
| 637 | matched_lines += 1; |
| 638 | if results.len() >= max_matches { |
| 639 | continue; |
| 640 | } |
| 641 | let start = idx.saturating_sub(context_lines); |
| 642 | let end = (idx + context_lines).min(lines.len().saturating_sub(1)); |
| 643 | let excerpt = render_numbered_lines( |
| 644 | lines |
| 645 | .iter() |
| 646 | .enumerate() |
| 647 | .skip(start) |
| 648 | .take(end.saturating_sub(start) + 1) |
| 649 | .map(|(line_idx, text)| (line_idx + 1, *text)), |
| 650 | max_bytes / max_matches.max(1), |
| 651 | ); |
| 652 | results.push(json!({ |
| 653 | "line": idx + 1, |
| 654 | "excerpt": excerpt, |
| 655 | })); |
| 656 | } |
| 657 | |
| 658 | Ok(json!({ |
| 659 | "ref": reference, |
| 660 | "mode": "query", |
| 661 | "query": query, |
| 662 | "total_lines": lines.len(), |
| 663 | "matched_lines": matched_lines, |
| 664 | "matches_returned": results.len(), |
| 665 | "results": results, |
| 666 | })) |
| 667 | } |
| 668 | |
| 669 | fn parse_line_selector(input: &Value) -> Result<(usize, usize), ToolError> { |
| 670 | let explicit_start = input.get("start_line").and_then(Value::as_u64); |
| 671 | let explicit_end = input.get("end_line").and_then(Value::as_u64); |
| 672 | if explicit_start.is_some() || explicit_end.is_some() { |
| 673 | let start = explicit_start.ok_or_else(|| { |
| 674 | ToolError::invalid_input("start_line is required when end_line is supplied") |
| 675 | })?; |
| 676 | let end = explicit_end.unwrap_or(start); |
| 677 | return validate_line_range(start as usize, end as usize); |
| 678 | } |
| 679 | |
| 680 | let spec = optional_str(input, "lines")? |
| 681 | .map(str::trim) |
| 682 | .filter(|s| !s.is_empty()) |
| 683 | .ok_or_else(|| { |
| 684 | ToolError::invalid_input( |
| 685 | "mode=lines requires `lines` (for example \"10-40\") or start_line/end_line", |
| 686 | ) |
| 687 | })?; |
| 688 | |
| 689 | if let Some((start, end)) = spec.split_once('-') { |
| 690 | let start = parse_positive_line(start.trim(), "lines start")?; |
| 691 | let end = parse_positive_line(end.trim(), "lines end")?; |
| 692 | validate_line_range(start, end) |
| 693 | } else { |
| 694 | let line = parse_positive_line(spec, "lines")?; |
| 695 | validate_line_range(line, line) |
| 696 | } |
| 697 | } |
| 698 | |
| 699 | fn validate_line_range(start: usize, end: usize) -> Result<(usize, usize), ToolError> { |
| 700 | if start == 0 || end == 0 { |
| 701 | return Err(ToolError::invalid_input("line numbers are 1-based")); |
| 702 | } |
| 703 | if end < start { |
| 704 | return Err(ToolError::invalid_input( |
| 705 | "end_line must be greater than or equal to start_line", |
| 706 | )); |
| 707 | } |
| 708 | Ok((start, end)) |
| 709 | } |
| 710 | |
| 711 | fn parse_positive_line(raw: &str, field: &str) -> Result<usize, ToolError> { |
| 712 | raw.parse::<usize>().map_err(|_| { |
| 713 | ToolError::invalid_input(format!("{field} must be a positive integer line number")) |
| 714 | }) |
| 715 | } |
| 716 | |
| 717 | fn collect_signal_lines(lines: &[&str], max_matches: usize) -> Vec<Value> { |
| 718 | let mut out = Vec::new(); |
| 719 | for (idx, line) in lines.iter().enumerate() { |
| 720 | if !is_signal_line(line) { |
| 721 | continue; |
| 722 | } |
| 723 | out.push(json!({ |
| 724 | "line": idx + 1, |
| 725 | "text": truncate_line(line.trim(), 300), |
| 726 | })); |
| 727 | if out.len() >= max_matches { |
| 728 | break; |
| 729 | } |
| 730 | } |
| 731 | out |
| 732 | } |
| 733 | |
| 734 | fn is_signal_line(line: &str) -> bool { |
| 735 | let lower = line.to_lowercase(); |
| 736 | [ |
| 737 | "error", |
| 738 | "failed", |
| 739 | "failure", |
| 740 | "panic", |
| 741 | "warning", |
| 742 | "exception", |
| 743 | "traceback", |
| 744 | "assertion", |
| 745 | "exit code", |
| 746 | "test result", |
| 747 | "thread '", |
| 748 | ] |
| 749 | .iter() |
| 750 | .any(|needle| lower.contains(needle)) |
| 751 | } |
| 752 | |
| 753 | fn render_numbered_lines<'a>( |
| 754 | lines: impl IntoIterator<Item = (usize, &'a str)>, |
| 755 | max_bytes: usize, |
| 756 | ) -> String { |
| 757 | let mut rendered = String::new(); |
| 758 | for (line_no, line) in lines { |
| 759 | rendered.push_str(&format!("{line_no}: {line}\n")); |
| 760 | if rendered.len() > max_bytes { |
| 761 | break; |
| 762 | } |
| 763 | } |
| 764 | truncate_text(&rendered, max_bytes) |
| 765 | } |
| 766 | |
| 767 | fn truncate_text(text: &str, max_bytes: usize) -> String { |
| 768 | if text.len() <= max_bytes { |
| 769 | return text.trim_end_matches('\n').to_string(); |
| 770 | } |
| 771 | let note = "\n[truncated to max_bytes]"; |
| 772 | let budget = max_bytes.saturating_sub(note.len()).max(1); |
| 773 | let cut = (0..=budget) |
| 774 | .rev() |
| 775 | .find(|idx| text.is_char_boundary(*idx)) |
| 776 | .unwrap_or(0); |
| 777 | format!("{}{}", text[..cut].trim_end_matches('\n'), note) |
| 778 | } |
| 779 | |
| 780 | fn truncate_line(line: &str, max_chars: usize) -> String { |
| 781 | if line.chars().count() <= max_chars { |
| 782 | return line.to_string(); |
| 783 | } |
| 784 | let mut out: String = line.chars().take(max_chars.saturating_sub(3)).collect(); |
| 785 | out.push_str("..."); |
| 786 | out |
| 787 | } |
| 788 | |
| 789 | fn clamp_u64(value: u64, min: usize, max: usize) -> usize { |
| 790 | (value as usize).clamp(min, max) |
| 791 | } |
| 792 | |
| 793 | #[cfg(test)] |
| 794 | mod tests { |
| 795 | use super::*; |
| 796 | use std::sync::MutexGuard; |
| 797 | use tempfile::tempdir; |
| 798 | |
| 799 | struct SpilloverRootGuard { |
| 800 | prior: Option<PathBuf>, |
| 801 | } |
| 802 | |
| 803 | impl Drop for SpilloverRootGuard { |
| 804 | fn drop(&mut self) { |
| 805 | crate::tools::truncate::set_test_spillover_root(self.prior.take()); |
| 806 | } |
| 807 | } |
| 808 | |
| 809 | fn set_spillover_root(path: PathBuf) -> SpilloverRootGuard { |
| 810 | let prior = crate::tools::truncate::set_test_spillover_root(Some(path)); |
| 811 | SpilloverRootGuard { prior } |
| 812 | } |
| 813 | |
| 814 | fn context() -> ToolContext { |
| 815 | let tmp = tempdir().unwrap(); |
| 816 | ToolContext::new(tmp.path()) |
| 817 | } |
| 818 | |
| 819 | fn test_lock() -> MutexGuard<'static, ()> { |
| 820 | crate::tools::truncate::TEST_SPILLOVER_GUARD |
| 821 | .lock() |
| 822 | .unwrap_or_else(|err| err.into_inner()) |
| 823 | } |
| 824 | |
| 825 | fn execute_tool(input: Value) -> Result<ToolResult, ToolError> { |
| 826 | let runtime = tokio::runtime::Builder::new_current_thread() |
| 827 | .enable_all() |
| 828 | .build() |
| 829 | .unwrap(); |
| 830 | runtime.block_on(RetrieveToolResultTool.execute(input, &context())) |
| 831 | } |
| 832 | |
| 833 | fn execute_tool_in_session(input: Value, session_id: &str) -> Result<ToolResult, ToolError> { |
| 834 | let runtime = tokio::runtime::Builder::new_current_thread() |
| 835 | .enable_all() |
| 836 | .build() |
| 837 | .unwrap(); |
| 838 | let mut context = context(); |
| 839 | context.state_namespace = session_id.to_string(); |
| 840 | runtime.block_on(RetrieveToolResultTool.execute(input, &context)) |
| 841 | } |
| 842 | |
| 843 | fn publish_test_evidence( |
| 844 | session_id: &str, |
| 845 | handle: &str, |
| 846 | bytes: &[u8], |
| 847 | expired: bool, |
| 848 | ) -> crate::tools::large_output_router::EvidenceArtifact { |
| 849 | let relative = crate::artifacts::session_artifact_relative_path(handle); |
| 850 | crate::artifacts::write_session_relative_immutable(session_id, &relative, bytes).unwrap(); |
| 851 | let now = crate::tools::large_output_router::unix_millis_now(); |
| 852 | let artifact = crate::tools::large_output_router::EvidenceArtifact { |
| 853 | handle: handle.to_string(), |
| 854 | digest: crate::hashing::sha256_hex(bytes), |
| 855 | size_bytes: bytes.len() as u64, |
| 856 | content_type: "application/octet-stream".to_string(), |
| 857 | tool_name: "exec_shell".to_string(), |
| 858 | call_id: handle.trim_start_matches("art_").to_string(), |
| 859 | origin_session: session_id.to_string(), |
| 860 | generation: 1, |
| 861 | redacted: false, |
| 862 | encoding: "binary".to_string(), |
| 863 | retention_state: if expired { |
| 864 | crate::tools::large_output_router::EvidenceRetentionState::Expired |
| 865 | } else { |
| 866 | crate::tools::large_output_router::EvidenceRetentionState::Live |
| 867 | }, |
| 868 | created_at_unix_ms: now, |
| 869 | retain_until_unix_ms: now.saturating_add(60_000), |
| 870 | storage_path: relative, |
| 871 | }; |
| 872 | crate::tools::large_output_router::publish_evidence_metadata(session_id, &artifact) |
| 873 | .unwrap(); |
| 874 | artifact |
| 875 | } |
| 876 | |
| 877 | fn write_owned_legacy(id: &str, content: &str, session_id: &str) -> PathBuf { |
| 878 | let path = crate::tools::truncate::write_spillover(id, content).unwrap(); |
| 879 | crate::tools::truncate::publish_legacy_spillover_ownership( |
| 880 | &path, |
| 881 | session_id, |
| 882 | content.as_bytes(), |
| 883 | ) |
| 884 | .unwrap(); |
| 885 | path |
| 886 | } |
| 887 | |
| 888 | fn write_owned_sha(content: &str, session_id: &str) -> (String, PathBuf) { |
| 889 | let sha = crate::hashing::sha256_hex(content.as_bytes()); |
| 890 | let path = crate::tools::truncate::write_sha_spillover(&sha, content).unwrap(); |
| 891 | crate::tools::truncate::publish_legacy_spillover_ownership( |
| 892 | &path, |
| 893 | session_id, |
| 894 | content.as_bytes(), |
| 895 | ) |
| 896 | .unwrap(); |
| 897 | (sha, path) |
| 898 | } |
| 899 | |
| 900 | #[test] |
| 901 | fn summary_reads_spillover_by_tool_call_id() { |
| 902 | let _lock = test_lock(); |
| 903 | let tmp = tempdir().unwrap(); |
| 904 | let _guard = set_spillover_root(tmp.path().join("tool_outputs")); |
| 905 | let session_id = "session-legacy-summary"; |
| 906 | write_owned_legacy( |
| 907 | "call-abc", |
| 908 | "checking crate\nerror[E0425]: missing value\nwarning: unused import\nfinished", |
| 909 | session_id, |
| 910 | ); |
| 911 | |
| 912 | let result = execute_tool_in_session(json!({"ref": "call-abc"}), session_id).unwrap(); |
| 913 | |
| 914 | assert!(result.success); |
| 915 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 916 | assert_eq!(body["mode"], "summary"); |
| 917 | assert!(body["signal_lines"].to_string().contains("error[E0425]")); |
| 918 | assert!(body["signal_lines"].to_string().contains("warning")); |
| 919 | } |
| 920 | |
| 921 | #[test] |
| 922 | fn adaptive_evidence_binary_bytes_are_exact_and_bounded() { |
| 923 | let _spill = test_lock(); |
| 924 | let _artifact = crate::artifacts::TEST_ARTIFACT_SESSIONS_GUARD |
| 925 | .lock() |
| 926 | .unwrap_or_else(|err| err.into_inner()); |
| 927 | let tmp = tempdir().unwrap(); |
| 928 | let _root = set_spillover_root(tmp.path().join("tool_outputs")); |
| 929 | let prior = |
| 930 | crate::artifacts::set_test_artifact_sessions_root(Some(tmp.path().join("sessions"))); |
| 931 | struct Restore(Option<PathBuf>); |
| 932 | impl Drop for Restore { |
| 933 | fn drop(&mut self) { |
| 934 | crate::artifacts::set_test_artifact_sessions_root(self.0.take()); |
| 935 | } |
| 936 | } |
| 937 | let _restore = Restore(prior); |
| 938 | let bytes = b"\0\xffbinary\nDEEP_SENTINEL\x80tail"; |
| 939 | publish_test_evidence("session-a", "art_call-binary", bytes, false); |
| 940 | |
| 941 | let result = execute_tool_in_session( |
| 942 | json!({"ref": "art_call-binary", "mode": "bytes", "offset": 0, "length": 1024}), |
| 943 | "session-a", |
| 944 | ) |
| 945 | .unwrap(); |
| 946 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 947 | use base64::Engine as _; |
| 948 | let decoded = base64::engine::general_purpose::STANDARD |
| 949 | .decode(body["data"].as_str().unwrap()) |
| 950 | .unwrap(); |
| 951 | assert_eq!(decoded, bytes); |
| 952 | assert_eq!(body["total_bytes"], bytes.len()); |
| 953 | } |
| 954 | |
| 955 | #[test] |
| 956 | fn adaptive_evidence_retrieves_after_restart_without_memory_state() { |
| 957 | let _spill = test_lock(); |
| 958 | let _artifact = crate::artifacts::TEST_ARTIFACT_SESSIONS_GUARD |
| 959 | .lock() |
| 960 | .unwrap_or_else(|err| err.into_inner()); |
| 961 | let tmp = tempdir().unwrap(); |
| 962 | let _root = set_spillover_root(tmp.path().join("tool_outputs")); |
| 963 | let prior = |
| 964 | crate::artifacts::set_test_artifact_sessions_root(Some(tmp.path().join("sessions"))); |
| 965 | struct Restore(Option<PathBuf>); |
| 966 | impl Drop for Restore { |
| 967 | fn drop(&mut self) { |
| 968 | crate::artifacts::set_test_artifact_sessions_root(self.0.take()); |
| 969 | } |
| 970 | } |
| 971 | let _restore = Restore(prior); |
| 972 | |
| 973 | let bytes = b"restart-proof\nDEEP_RESTART_SENTINEL\nend"; |
| 974 | publish_test_evidence("session-restart", "art_call-restart", bytes, false); |
| 975 | |
| 976 | // Construct two independent contexts to model process teardown and |
| 977 | // resume. Retrieval must depend only on the sealed session artifact |
| 978 | // and metadata, never an in-memory routing table from publication. |
| 979 | let first = execute_tool_in_session( |
| 980 | json!({"ref": "art_call-restart", "mode": "metadata"}), |
| 981 | "session-restart", |
| 982 | ) |
| 983 | .unwrap(); |
| 984 | drop(first); |
| 985 | let resumed = execute_tool_in_session( |
| 986 | json!({"ref": "art_call-restart", "mode": "bytes", "length": 4096}), |
| 987 | "session-restart", |
| 988 | ) |
| 989 | .unwrap(); |
| 990 | let body: Value = serde_json::from_str(&resumed.content).unwrap(); |
| 991 | use base64::Engine as _; |
| 992 | let decoded = base64::engine::general_purpose::STANDARD |
| 993 | .decode(body["data"].as_str().unwrap()) |
| 994 | .unwrap(); |
| 995 | assert_eq!(decoded, bytes); |
| 996 | assert_eq!(body["total_bytes"], bytes.len()); |
| 997 | } |
| 998 | |
| 999 | #[test] |
| 1000 | fn adaptive_evidence_distinguishes_corrupt_expired_and_generation_mismatch() { |
| 1001 | let _spill = test_lock(); |
| 1002 | let _artifact = crate::artifacts::TEST_ARTIFACT_SESSIONS_GUARD |
| 1003 | .lock() |
| 1004 | .unwrap_or_else(|err| err.into_inner()); |
| 1005 | let tmp = tempdir().unwrap(); |
| 1006 | let _root = set_spillover_root(tmp.path().join("tool_outputs")); |
| 1007 | let prior = |
| 1008 | crate::artifacts::set_test_artifact_sessions_root(Some(tmp.path().join("sessions"))); |
| 1009 | struct Restore(Option<PathBuf>); |
| 1010 | impl Drop for Restore { |
| 1011 | fn drop(&mut self) { |
| 1012 | crate::artifacts::set_test_artifact_sessions_root(self.0.take()); |
| 1013 | } |
| 1014 | } |
| 1015 | let _restore = Restore(prior); |
| 1016 | |
| 1017 | publish_test_evidence("session-a", "art_call-expired", b"expired", true); |
| 1018 | let expired = execute_tool_in_session(json!({"ref": "art_call-expired"}), "session-a") |
| 1019 | .unwrap_err() |
| 1020 | .to_string(); |
| 1021 | assert!(expired.contains("expired"), "{expired}"); |
| 1022 | |
| 1023 | let artifact = publish_test_evidence("session-a", "art_call-corrupt", b"original", false); |
| 1024 | let absolute = |
| 1025 | crate::artifacts::session_artifact_absolute_path("session-a", &artifact.storage_path) |
| 1026 | .unwrap(); |
| 1027 | std::fs::write(absolute, b"changed").unwrap(); |
| 1028 | let corrupt = execute_tool_in_session(json!({"ref": "art_call-corrupt"}), "session-a") |
| 1029 | .unwrap_err() |
| 1030 | .to_string(); |
| 1031 | assert!(corrupt.contains("corrupt"), "{corrupt}"); |
| 1032 | |
| 1033 | publish_test_evidence("session-a", "art_call-generation", b"stable", false); |
| 1034 | let mismatch = execute_tool_in_session( |
| 1035 | json!({"ref": "art_call-generation", "generation": 2}), |
| 1036 | "session-a", |
| 1037 | ) |
| 1038 | .unwrap_err() |
| 1039 | .to_string(); |
| 1040 | assert!(mismatch.contains("generation"), "{mismatch}"); |
| 1041 | } |
| 1042 | |
| 1043 | #[test] |
| 1044 | fn query_returns_matching_line_with_context() { |
| 1045 | let _lock = test_lock(); |
| 1046 | let tmp = tempdir().unwrap(); |
| 1047 | let _guard = set_spillover_root(tmp.path().join("tool_outputs")); |
| 1048 | let session_id = "session-legacy-query"; |
| 1049 | write_owned_legacy( |
| 1050 | "call-query", |
| 1051 | "one\ntwo before\nneedle here\nafter\nlast", |
| 1052 | session_id, |
| 1053 | ); |
| 1054 | |
| 1055 | let result = execute_tool_in_session( |
| 1056 | json!({ |
| 1057 | "ref": "tool_result:call-query", |
| 1058 | "mode": "query", |
| 1059 | "query": "needle", |
| 1060 | "context_lines": 1 |
| 1061 | }), |
| 1062 | session_id, |
| 1063 | ) |
| 1064 | .unwrap(); |
| 1065 | |
| 1066 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 1067 | assert_eq!(body["matched_lines"], 1); |
| 1068 | let rendered = body["results"].to_string(); |
| 1069 | assert!(rendered.contains("2: two before")); |
| 1070 | assert!(rendered.contains("3: needle here")); |
| 1071 | assert!(rendered.contains("4: after")); |
| 1072 | } |
| 1073 | |
| 1074 | #[test] |
| 1075 | fn lines_mode_accepts_filename_inside_spillover_root() { |
| 1076 | let _lock = test_lock(); |
| 1077 | let tmp = tempdir().unwrap(); |
| 1078 | let root = tmp.path().join("tool_outputs"); |
| 1079 | let _guard = set_spillover_root(root.clone()); |
| 1080 | let session_id = "session-legacy-lines"; |
| 1081 | write_owned_legacy("call-lines", "a\nb\nc\nd", session_id); |
| 1082 | |
| 1083 | let result = execute_tool_in_session( |
| 1084 | json!({ |
| 1085 | "ref": "call-lines.txt", |
| 1086 | "mode": "lines", |
| 1087 | "lines": "2-3" |
| 1088 | }), |
| 1089 | session_id, |
| 1090 | ) |
| 1091 | .unwrap(); |
| 1092 | |
| 1093 | let body: Value = serde_json::from_str(&result.content).unwrap(); |
| 1094 | let excerpt = body["excerpt"].as_str().unwrap(); |
| 1095 | assert!(excerpt.contains("2: b")); |
| 1096 | assert!(excerpt.contains("3: c")); |
| 1097 | assert!(!excerpt.contains("1: a")); |
| 1098 | assert!(!excerpt.contains("4: d")); |
| 1099 | } |
| 1100 | |
| 1101 | #[test] |
| 1102 | fn rejects_path_outside_spillover_root() { |
| 1103 | let _lock = test_lock(); |
| 1104 | let tmp = tempdir().unwrap(); |
| 1105 | let root = tmp.path().join("tool_outputs"); |
| 1106 | fs::create_dir_all(&root).unwrap(); |
| 1107 | let outside = tmp.path().join("outside.txt"); |
| 1108 | fs::write(&outside, "secret").unwrap(); |
| 1109 | let _guard = set_spillover_root(root); |
| 1110 | |
| 1111 | let err = execute_tool(json!({"ref": outside.display().to_string()})).unwrap_err(); |
| 1112 | |
| 1113 | // Unauthorized is distinct but non-leaking: no outside path detail is |
| 1114 | // echoed beyond the caller-supplied ref. |
| 1115 | let msg = err.to_string(); |
| 1116 | assert!( |
| 1117 | msg.contains("authorize") && msg.contains("active session"), |
| 1118 | "expected non-leaking authorization diagnostic, got: {msg}" |
| 1119 | ); |
| 1120 | } |
| 1121 | |
| 1122 | #[test] |
| 1123 | fn resolves_sha_reference_from_wire_dedup() { |
| 1124 | // A SHA-keyed lookup — emulates what happens when the model |
| 1125 | // sees a `<TOOL_RESULT_REF sha="..." />` block and passes the |
| 1126 | // SHA to retrieve_tool_result. |
| 1127 | let _lock = test_lock(); |
| 1128 | let tmp = tempdir().unwrap(); |
| 1129 | let _guard = set_spillover_root(tmp.path().join("tool_outputs")); |
| 1130 | let body = "checking crate ... error[E0425]: cannot find value\n".repeat(80); |
| 1131 | let session_id = "session-legacy-sha"; |
| 1132 | let (sha, _) = write_owned_sha(&body, session_id); |
| 1133 | |
| 1134 | // Form: `sha:<hex>` |
| 1135 | let result = |
| 1136 | execute_tool_in_session(json!({"ref": format!("sha:{sha}")}), session_id).unwrap(); |
| 1137 | assert!(result.success, "sha:<hex> form should resolve"); |
| 1138 | |
| 1139 | // Form: bare 64-hex |
| 1140 | let result = execute_tool_in_session(json!({"ref": &sha}), session_id).unwrap(); |
| 1141 | assert!(result.success, "bare 64-hex form should resolve"); |
| 1142 | } |
| 1143 | |
| 1144 | #[test] |
| 1145 | fn resolves_art_prefix_to_legacy_spillover_id() { |
| 1146 | // The model commonly sees `id: art_call_xyz` in artifact |
| 1147 | // ref blocks. retrieve_tool_result should strip the `art_` |
| 1148 | // prefix and find the legacy `<id>.txt` file if no |
| 1149 | // session-artifact equivalent exists. |
| 1150 | let _lock = test_lock(); |
| 1151 | let tmp = tempdir().unwrap(); |
| 1152 | let _guard = set_spillover_root(tmp.path().join("tool_outputs")); |
| 1153 | let session_id = "session-legacy-art-prefix"; |
| 1154 | write_owned_legacy("call_xyz", "line1\nline2\nline3", session_id); |
| 1155 | |
| 1156 | let result = execute_tool_in_session(json!({"ref": "art_call_xyz"}), session_id).unwrap(); |
| 1157 | assert!(result.success, "art_ prefix should resolve to legacy id"); |
| 1158 | } |
| 1159 | |
| 1160 | #[test] |
| 1161 | fn unowned_and_foreign_legacy_spillovers_fail_closed_without_leaking_content() { |
| 1162 | let _lock = test_lock(); |
| 1163 | let tmp = tempdir().unwrap(); |
| 1164 | let _guard = set_spillover_root(tmp.path().join("tool_outputs")); |
| 1165 | let sentinel = "SESSION_A_PRIVATE_SENTINEL"; |
| 1166 | |
| 1167 | crate::tools::truncate::write_spillover("call-unowned", sentinel).unwrap(); |
| 1168 | let unowned = |
| 1169 | execute_tool_in_session(json!({"ref": "call-unowned", "mode": "bytes"}), "session-b") |
| 1170 | .unwrap_err() |
| 1171 | .to_string(); |
| 1172 | assert!(unowned.contains("no verifiable session owner"), "{unowned}"); |
| 1173 | assert!(!unowned.contains(sentinel), "{unowned}"); |
| 1174 | assert!( |
| 1175 | !unowned.contains(tmp.path().to_string_lossy().as_ref()), |
| 1176 | "{unowned}" |
| 1177 | ); |
| 1178 | |
| 1179 | write_owned_legacy("call-foreign", sentinel, "session-a"); |
| 1180 | let foreign = |
| 1181 | execute_tool_in_session(json!({"ref": "call-foreign", "mode": "bytes"}), "session-b") |
| 1182 | .unwrap_err() |
| 1183 | .to_string(); |
| 1184 | assert!(foreign.contains("another session"), "{foreign}"); |
| 1185 | assert!(!foreign.contains(sentinel), "{foreign}"); |
| 1186 | assert!( |
| 1187 | !foreign.contains(tmp.path().to_string_lossy().as_ref()), |
| 1188 | "{foreign}" |
| 1189 | ); |
| 1190 | } |
| 1191 | |
| 1192 | #[test] |
| 1193 | fn owned_legacy_digest_mismatch_is_distinct_from_unauthorized() { |
| 1194 | let _lock = test_lock(); |
| 1195 | let tmp = tempdir().unwrap(); |
| 1196 | let _guard = set_spillover_root(tmp.path().join("tool_outputs")); |
| 1197 | let path = write_owned_legacy("call-corrupt-owned", "original", "session-a"); |
| 1198 | std::fs::write(path, "changed").unwrap(); |
| 1199 | |
| 1200 | let error = execute_tool_in_session(json!({"ref": "call-corrupt-owned"}), "session-a") |
| 1201 | .unwrap_err() |
| 1202 | .to_string(); |
| 1203 | assert!(error.contains("content is corrupt"), "{error}"); |
| 1204 | assert!(!error.contains("another session"), "{error}"); |
| 1205 | } |
| 1206 | |
| 1207 | #[test] |
| 1208 | fn not_found_error_lists_tried_candidates_and_accepted_forms() { |
| 1209 | let _lock = test_lock(); |
| 1210 | let tmp = tempdir().unwrap(); |
| 1211 | let _guard = set_spillover_root(tmp.path().join("tool_outputs")); |
| 1212 | fs::create_dir_all(tmp.path().join("tool_outputs")).unwrap(); |
| 1213 | |
| 1214 | let err = execute_tool(json!({"ref": "definitely_missing_id"})).unwrap_err(); |
| 1215 | let msg = err.to_string(); |
| 1216 | assert!(msg.contains("not found"), "got: {msg}"); |
| 1217 | assert!(msg.contains("active session"), "got: {msg}"); |
| 1218 | assert!(msg.contains("art_<id>"), "got: {msg}"); |
| 1219 | assert!(!msg.contains("tool_outputs"), "storage root leaked: {msg}"); |
| 1220 | assert!( |
| 1221 | !msg.contains(tmp.path().to_string_lossy().as_ref()), |
| 1222 | "path leaked: {msg}" |
| 1223 | ); |
| 1224 | } |
| 1225 | |
| 1226 | #[test] |
| 1227 | fn resolves_art_prefix_via_session_artifacts() { |
| 1228 | let _lock = test_lock(); |
| 1229 | let tmp = tempdir().unwrap(); |
| 1230 | let _spill_guard = set_spillover_root(tmp.path().join("tool_outputs")); |
| 1231 | let _art_guard = { |
| 1232 | let prior = crate::artifacts::set_test_artifact_sessions_root(Some( |
| 1233 | tmp.path().join("sessions"), |
| 1234 | )); |
| 1235 | scopeguard_for_test(prior) |
| 1236 | }; |
| 1237 | let session_id = "session-abc"; |
| 1238 | let body = "this is the canonical session artifact body, not a legacy file"; |
| 1239 | crate::artifacts::write_session_artifact(session_id, "art_call_real", body).unwrap(); |
| 1240 | |
| 1241 | let runtime = tokio::runtime::Builder::new_current_thread() |
| 1242 | .enable_all() |
| 1243 | .build() |
| 1244 | .unwrap(); |
| 1245 | let workspace_tmp = tempdir().unwrap(); |
| 1246 | let ctx = ToolContext::new(workspace_tmp.path()).with_state_namespace(session_id); |
| 1247 | let result = runtime |
| 1248 | .block_on(RetrieveToolResultTool.execute(json!({"ref": "art_call_real"}), &ctx)) |
| 1249 | .expect("art_<id> should resolve via session artifacts"); |
| 1250 | assert!(result.success); |
| 1251 | let payload: Value = serde_json::from_str(&result.content).unwrap(); |
| 1252 | assert!( |
| 1253 | payload |
| 1254 | .to_string() |
| 1255 | .contains("canonical session artifact body"), |
| 1256 | "summary should pull from session artifact, got: {payload}" |
| 1257 | ); |
| 1258 | } |
| 1259 | |
| 1260 | #[cfg(unix)] |
| 1261 | #[test] |
| 1262 | fn rejects_symlink_inside_session_artifacts() { |
| 1263 | let _lock = test_lock(); |
| 1264 | let tmp = tempdir().unwrap(); |
| 1265 | let _spill_guard = set_spillover_root(tmp.path().join("tool_outputs")); |
| 1266 | let _art_guard = { |
| 1267 | let prior = crate::artifacts::set_test_artifact_sessions_root(Some( |
| 1268 | tmp.path().join("sessions"), |
| 1269 | )); |
| 1270 | scopeguard_for_test(prior) |
| 1271 | }; |
| 1272 | let session_id = "session-xyz"; |
| 1273 | // Plant a sensitive file outside the artifact dir. |
| 1274 | let secret = tmp.path().join("secret.txt"); |
| 1275 | fs::write(&secret, "do not leak").unwrap(); |
| 1276 | // Create the artifact dir, then drop a symlink inside it |
| 1277 | // pointing at the secret. |
| 1278 | let art_dir = tmp |
| 1279 | .path() |
| 1280 | .join("sessions") |
| 1281 | .join(session_id) |
| 1282 | .join("artifacts"); |
| 1283 | fs::create_dir_all(&art_dir).unwrap(); |
| 1284 | std::os::unix::fs::symlink(&secret, art_dir.join("art_evil.txt")).unwrap(); |
| 1285 | |
| 1286 | let runtime = tokio::runtime::Builder::new_current_thread() |
| 1287 | .enable_all() |
| 1288 | .build() |
| 1289 | .unwrap(); |
| 1290 | let workspace_tmp = tempdir().unwrap(); |
| 1291 | let ctx = ToolContext::new(workspace_tmp.path()).with_state_namespace(session_id); |
| 1292 | let result = |
| 1293 | runtime.block_on(RetrieveToolResultTool.execute(json!({"ref": "art_evil"}), &ctx)); |
| 1294 | let err = result.expect_err("symlink artifact must not resolve"); |
| 1295 | assert!( |
| 1296 | err.to_string().contains("not found"), |
| 1297 | "expected `not found`, got: {err}" |
| 1298 | ); |
| 1299 | } |
| 1300 | |
| 1301 | struct ArtifactRootGuard { |
| 1302 | prior: Option<PathBuf>, |
| 1303 | } |
| 1304 | impl Drop for ArtifactRootGuard { |
| 1305 | fn drop(&mut self) { |
| 1306 | crate::artifacts::set_test_artifact_sessions_root(self.prior.take()); |
| 1307 | } |
| 1308 | } |
| 1309 | fn scopeguard_for_test(prior: Option<PathBuf>) -> ArtifactRootGuard { |
| 1310 | ArtifactRootGuard { prior } |
| 1311 | } |
| 1312 | } |
| 1313 |