| 1 | //! Tool-output spillover writer (#422). |
| 2 | //! |
| 3 | //! When a tool produces output that's too large to land in the model's |
| 4 | //! context budget, we want two things at once: |
| 5 | //! |
| 6 | //! 1. The transcript / tool-cell renders a bounded preview so the UI |
| 7 | //! stays scannable. |
| 8 | //! 2. The full router input is preserved under its origin session so bounded |
| 9 | //! retrieval and the raw-detail pager can inspect it without leaking a |
| 10 | //! process-global filesystem path. |
| 11 | //! |
| 12 | //! The default adaptive path writes immutable artifacts under |
| 13 | //! `~/.codewhale/sessions/<session>/artifacts/`. The historical |
| 14 | //! `~/.codewhale/tool_outputs/<sanitised-id>.txt` directory remains only for |
| 15 | //! classic-routing compatibility, protected by a digest-bound origin sidecar. |
| 16 | //! |
| 17 | //! Boot prune drops files whose mtime is older than [`SPILLOVER_MAX_AGE`] |
| 18 | //! (7 days). Prune failures are logged and never fatal — the user |
| 19 | //! shouldn't see startup wedge because of a stale tool-output file. |
| 20 | //! |
| 21 | //! ## Live callers |
| 22 | //! |
| 23 | //! * [`apply_spillover`] — invoked from the engine's tool-execution |
| 24 | //! path (`turn_loop.rs`) so any successful tool result over |
| 25 | //! [`SPILLOVER_THRESHOLD_BYTES`] spills to disk and the model |
| 26 | //! receives a bounded plain preview: a [`SPILLOVER_HEAD_BYTES`] head, |
| 27 | //! a short retained tail, and an honest footer naming the on-disk |
| 28 | //! path of the full output plus a one-line recovery instruction. |
| 29 | //! * Boot prune in `main.rs` deletes files older than |
| 30 | //! [`SPILLOVER_MAX_AGE`]. |
| 31 | //! |
| 32 | //! UI-side rendering is owned by `tui/history.rs::render_spillover_annotation`; |
| 33 | //! it exposes a calm expand affordance and the tool-details shortcut opens the |
| 34 | //! full retained output. |
| 35 | |
| 36 | use std::fs; |
| 37 | use std::io; |
| 38 | use std::path::{Path, PathBuf}; |
| 39 | use std::time::{Duration, SystemTime}; |
| 40 | |
| 41 | use crate::tools::spec::ToolResult; |
| 42 | |
| 43 | /// Name of the spillover directory under the CodeWhale home. |
| 44 | pub const SPILLOVER_DIR_NAME: &str = "tool_outputs"; |
| 45 | |
| 46 | const LEGACY_SPILLOVER_OWNER_SCHEMA_VERSION: u32 = 1; |
| 47 | |
| 48 | /// Session proof for compatibility payloads kept in the historical global |
| 49 | /// `tool_outputs/` directory. |
| 50 | /// |
| 51 | /// The payload remains in its legacy location so classic-routing rollback and |
| 52 | /// existing detail pagers keep working, but model retrieval is authorized only |
| 53 | /// when this sidecar names the active origin session and still matches the |
| 54 | /// immutable bytes being returned. |
| 55 | #[derive(Debug, Clone, serde::Serialize, serde::Deserialize, PartialEq, Eq)] |
| 56 | pub(crate) struct LegacySpilloverOwnership { |
| 57 | pub schema_version: u32, |
| 58 | pub origin_session: String, |
| 59 | pub digest: String, |
| 60 | pub size_bytes: u64, |
| 61 | } |
| 62 | |
| 63 | /// Default threshold above which a tool result is a candidate for |
| 64 | /// spillover. Mirrors the `MAX_MEMORY_SIZE` ceiling we use elsewhere |
| 65 | /// for "too large to inline" so the rules feel consistent. Wired |
| 66 | /// callers can pass a different value if a tool family has different |
| 67 | /// economics. |
| 68 | pub const SPILLOVER_THRESHOLD_BYTES: usize = 100 * 1024; // 100 KiB |
| 69 | |
| 70 | /// Default boot-prune age. Older spillover files are deleted on |
| 71 | /// startup to keep `~/.codewhale/tool_outputs/` from growing without |
| 72 | /// bound. Mirrors the workspace-snapshot 7-day default. |
| 73 | pub const SPILLOVER_MAX_AGE: Duration = Duration::from_secs(7 * 24 * 60 * 60); |
| 74 | |
| 75 | #[cfg(test)] |
| 76 | static TEST_SPILLOVER_ROOT: std::sync::Mutex<Option<PathBuf>> = std::sync::Mutex::new(None); |
| 77 | |
| 78 | #[cfg(test)] |
| 79 | pub(crate) static TEST_SPILLOVER_GUARD: std::sync::Mutex<()> = std::sync::Mutex::new(()); |
| 80 | |
| 81 | /// Resolve `~/.codewhale/tool_outputs/`. Returns `None` if the home |
| 82 | /// directory can't be determined (CI containers occasionally hit |
| 83 | /// this). Callers should treat `None` as "spillover unavailable" and |
| 84 | /// degrade gracefully rather than fail the tool call. |
| 85 | #[must_use] |
| 86 | pub fn spillover_root() -> Option<PathBuf> { |
| 87 | #[cfg(test)] |
| 88 | if let Some(root) = TEST_SPILLOVER_ROOT |
| 89 | .lock() |
| 90 | .unwrap_or_else(|err| err.into_inner()) |
| 91 | .clone() |
| 92 | { |
| 93 | return Some(root); |
| 94 | } |
| 95 | |
| 96 | let home = crate::config::effective_home_dir()?; |
| 97 | let primary = home.join(".codewhale").join(SPILLOVER_DIR_NAME); |
| 98 | let legacy = home.join(".deepseek").join(SPILLOVER_DIR_NAME); |
| 99 | if primary.exists() || !legacy.exists() { |
| 100 | return Some(primary); |
| 101 | } |
| 102 | Some(legacy) |
| 103 | } |
| 104 | |
| 105 | /// Override the spillover root for tests without mutating `$HOME`. |
| 106 | #[cfg(test)] |
| 107 | pub(crate) fn set_test_spillover_root(root: Option<PathBuf>) -> Option<PathBuf> { |
| 108 | let mut guard = TEST_SPILLOVER_ROOT |
| 109 | .lock() |
| 110 | .unwrap_or_else(|err| err.into_inner()); |
| 111 | std::mem::replace(&mut *guard, root) |
| 112 | } |
| 113 | |
| 114 | /// Resolve the spillover-file path for a tool call id. Sanitises the |
| 115 | /// id so that a hostile value can't escape the storage directory. |
| 116 | /// Returns `None` for empty / fully-invalid ids; the caller should |
| 117 | /// treat that as "spillover unavailable" and skip the write. |
| 118 | #[must_use] |
| 119 | pub fn spillover_path(id: &str) -> Option<PathBuf> { |
| 120 | let sanitised = sanitise_id(id)?; |
| 121 | Some(spillover_root()?.join(format!("{sanitised}.txt"))) |
| 122 | } |
| 123 | |
| 124 | #[must_use] |
| 125 | pub(crate) fn legacy_spillover_ownership_path(payload_path: &Path) -> PathBuf { |
| 126 | payload_path.with_extension("owner.json") |
| 127 | } |
| 128 | |
| 129 | /// Publish the proof needed to retrieve a legacy-global spillover safely. |
| 130 | /// |
| 131 | /// Payload publication happens first. If this atomic sidecar write fails, the |
| 132 | /// payload is deliberately left unowned and therefore inaccessible through |
| 133 | /// `retrieve_tool_result`; callers must not advertise a retrieval hint. |
| 134 | pub(crate) fn publish_legacy_spillover_ownership( |
| 135 | payload_path: &Path, |
| 136 | session_id: &str, |
| 137 | bytes: &[u8], |
| 138 | ) -> io::Result<PathBuf> { |
| 139 | if session_id.trim().is_empty() { |
| 140 | return Err(io::Error::new( |
| 141 | io::ErrorKind::InvalidInput, |
| 142 | "legacy spillover ownership requires a session id", |
| 143 | )); |
| 144 | } |
| 145 | let ownership = LegacySpilloverOwnership { |
| 146 | schema_version: LEGACY_SPILLOVER_OWNER_SCHEMA_VERSION, |
| 147 | origin_session: session_id.to_string(), |
| 148 | digest: crate::hashing::sha256_hex(bytes), |
| 149 | size_bytes: bytes.len().try_into().unwrap_or(u64::MAX), |
| 150 | }; |
| 151 | let sidecar = legacy_spillover_ownership_path(payload_path); |
| 152 | let encoded = serde_json::to_vec_pretty(&ownership) |
| 153 | .map_err(|error| io::Error::new(io::ErrorKind::InvalidData, error))?; |
| 154 | crate::utils::write_atomic(&sidecar, &encoded)?; |
| 155 | Ok(sidecar) |
| 156 | } |
| 157 | |
| 158 | pub(crate) fn read_legacy_spillover_ownership( |
| 159 | payload_path: &Path, |
| 160 | ) -> io::Result<LegacySpilloverOwnership> { |
| 161 | let sidecar = legacy_spillover_ownership_path(payload_path); |
| 162 | if std::fs::symlink_metadata(&sidecar)? |
| 163 | .file_type() |
| 164 | .is_symlink() |
| 165 | { |
| 166 | return Err(io::Error::new( |
| 167 | io::ErrorKind::PermissionDenied, |
| 168 | "legacy spillover ownership sidecar must not be a symlink", |
| 169 | )); |
| 170 | } |
| 171 | let ownership = serde_json::from_slice::<LegacySpilloverOwnership>(&std::fs::read(sidecar)?) |
| 172 | .map_err(|error| io::Error::new(io::ErrorKind::InvalidData, error))?; |
| 173 | if ownership.schema_version != LEGACY_SPILLOVER_OWNER_SCHEMA_VERSION { |
| 174 | return Err(io::Error::new( |
| 175 | io::ErrorKind::InvalidData, |
| 176 | "unsupported legacy spillover ownership schema", |
| 177 | )); |
| 178 | } |
| 179 | Ok(ownership) |
| 180 | } |
| 181 | |
| 182 | /// Resolve the spillover-file path for a SHA256 content hash. Separate |
| 183 | /// namespace (`sha_<hex>.txt`) from the tool-call-id files so legacy |
| 184 | /// SHA-addressed evidence can be recognized without colliding with |
| 185 | /// tool-call references. Retrieval still requires matching ownership |
| 186 | /// metadata. `sha` must be the raw 64-char lowercase hex digest — |
| 187 | /// case-insensitive matching is done by the caller. |
| 188 | #[must_use] |
| 189 | pub fn sha_spillover_path(sha: &str) -> Option<PathBuf> { |
| 190 | let sha = sha.trim().to_ascii_lowercase(); |
| 191 | if !is_valid_sha256(&sha) { |
| 192 | return None; |
| 193 | } |
| 194 | Some(spillover_root()?.join(format!("sha_{sha}.txt"))) |
| 195 | } |
| 196 | |
| 197 | /// True when `s` is a 64-character lowercase ASCII hex string. Used |
| 198 | /// to detect bare SHA refs the model might pass to retrieval and to |
| 199 | /// validate input to [`sha_spillover_path`]. |
| 200 | #[must_use] |
| 201 | pub fn is_valid_sha256(s: &str) -> bool { |
| 202 | s.len() == 64 |
| 203 | && s.chars() |
| 204 | .all(|c| c.is_ascii_hexdigit() && !c.is_ascii_uppercase()) |
| 205 | } |
| 206 | |
| 207 | /// Write a legacy SHA-addressed spillover fixture for ownership tests. |
| 208 | #[cfg(test)] |
| 209 | pub fn write_sha_spillover(sha: &str, content: &str) -> io::Result<PathBuf> { |
| 210 | let path = sha_spillover_path(sha).ok_or_else(|| { |
| 211 | io::Error::new( |
| 212 | io::ErrorKind::InvalidInput, |
| 213 | "sha must be a 64-char lowercase hex digest", |
| 214 | ) |
| 215 | })?; |
| 216 | if path.exists() { |
| 217 | return Ok(path); |
| 218 | } |
| 219 | if let Some(parent) = path.parent() { |
| 220 | fs::create_dir_all(parent)?; |
| 221 | } |
| 222 | crate::utils::write_atomic(&path, content.as_bytes())?; |
| 223 | Ok(path) |
| 224 | } |
| 225 | |
| 226 | /// Write `content` to the spillover file for `id`. Creates the |
| 227 | /// parent directory if needed. Returns the resolved path on success. |
| 228 | /// |
| 229 | /// Atomic via `write` + filesystem rename guarantees from the |
| 230 | /// underlying OS — the file is created at a temp name first and |
| 231 | /// then renamed into place. Failures bubble up as `io::Error` so the |
| 232 | /// caller can decide whether to surface them. |
| 233 | pub fn write_spillover(id: &str, content: &str) -> io::Result<PathBuf> { |
| 234 | let path = spillover_path(id).ok_or_else(|| { |
| 235 | io::Error::new( |
| 236 | io::ErrorKind::InvalidInput, |
| 237 | "could not resolve spillover path (empty/invalid id or missing home directory)", |
| 238 | ) |
| 239 | })?; |
| 240 | if let Some(parent) = path.parent() { |
| 241 | fs::create_dir_all(parent)?; |
| 242 | } |
| 243 | crate::utils::write_atomic(&path, content.as_bytes())?; |
| 244 | Ok(path) |
| 245 | } |
| 246 | |
| 247 | /// Drop spillover files older than `max_age`. Returns the number of |
| 248 | /// files removed. Non-fatal: directory-missing returns 0; per-file |
| 249 | /// errors are logged and skipped. Mirrors |
| 250 | /// [`crate::session_manager::prune_workspace_snapshots`]. |
| 251 | pub fn prune_older_than(max_age: Duration) -> io::Result<usize> { |
| 252 | let Some(root) = spillover_root() else { |
| 253 | return Ok(0); |
| 254 | }; |
| 255 | if !root.exists() { |
| 256 | return Ok(0); |
| 257 | } |
| 258 | let cutoff = SystemTime::now() |
| 259 | .checked_sub(max_age) |
| 260 | .unwrap_or(SystemTime::UNIX_EPOCH); |
| 261 | let mut pruned = 0usize; |
| 262 | for entry in fs::read_dir(&root)? { |
| 263 | let entry = match entry { |
| 264 | Ok(e) => e, |
| 265 | Err(err) => { |
| 266 | tracing::warn!(target: "spillover", ?err, "skipping unreadable dir entry"); |
| 267 | continue; |
| 268 | } |
| 269 | }; |
| 270 | let path = entry.path(); |
| 271 | if !path.is_file() { |
| 272 | continue; |
| 273 | } |
| 274 | let modified = match entry.metadata().and_then(|m| m.modified()) { |
| 275 | Ok(t) => t, |
| 276 | Err(err) => { |
| 277 | tracing::warn!(target: "spillover", ?err, ?path, "skipping unreadable mtime"); |
| 278 | continue; |
| 279 | } |
| 280 | }; |
| 281 | if modified < cutoff { |
| 282 | if let Err(err) = fs::remove_file(&path) { |
| 283 | tracing::warn!(target: "spillover", ?err, ?path, "spillover prune skipped a file"); |
| 284 | continue; |
| 285 | } |
| 286 | pruned += 1; |
| 287 | } |
| 288 | } |
| 289 | Ok(pruned) |
| 290 | } |
| 291 | |
| 292 | /// Convenience for the common "too long? spill it." pattern. If |
| 293 | /// `content` is at or below `threshold` bytes, returns `None` and the |
| 294 | /// caller keeps the inline content. Above the threshold, writes the |
| 295 | /// full content to the spillover file and returns |
| 296 | /// `Some((head, path))` where `head` is the leading slice the caller |
| 297 | /// can show inline. The trailing tail isn't returned — `path` is the |
| 298 | /// canonical reference. |
| 299 | /// |
| 300 | /// `head_bytes` controls how much inline content the caller wants to |
| 301 | /// keep. Pass `threshold` for "preserve as much as fits inline" or |
| 302 | /// a smaller value (e.g. `4 * 1024`) for "show a peek". |
| 303 | pub fn maybe_spillover( |
| 304 | id: &str, |
| 305 | content: &str, |
| 306 | threshold: usize, |
| 307 | head_bytes: usize, |
| 308 | ) -> io::Result<Option<(String, PathBuf)>> { |
| 309 | if content.len() <= threshold { |
| 310 | return Ok(None); |
| 311 | } |
| 312 | let path = write_spillover(id, content)?; |
| 313 | // Don't slice mid-utf8: walk back to a char boundary if needed. |
| 314 | let cut = head_bytes.min(content.len()); |
| 315 | let cut = (0..=cut) |
| 316 | .rev() |
| 317 | .find(|&i| content.is_char_boundary(i)) |
| 318 | .unwrap_or(0); |
| 319 | Ok(Some((content[..cut].to_string(), path))) |
| 320 | } |
| 321 | |
| 322 | /// Inline head retained when [`apply_spillover`] truncates a tool |
| 323 | /// result. 32 KiB is large enough for the model to keep meaningful |
| 324 | /// context (a long stack trace, a `git diff` head, a directory |
| 325 | /// listing of typical depth) without consuming the lion's share of |
| 326 | /// the per-turn context budget. The full output is preserved |
| 327 | /// internally and opens in the tool details view. |
| 328 | pub const SPILLOVER_HEAD_BYTES: usize = 32 * 1024; |
| 329 | /// Inline tail retained alongside the head so compiler summaries and final |
| 330 | /// test failures are not systematically hidden by truncation. |
| 331 | pub const SPILLOVER_TAIL_BYTES: usize = 8 * 1024; |
| 332 | |
| 333 | /// Inline head/tail budgets for the adaptive evidence bands. Hybrid results |
| 334 | /// keep a generous 32 KiB head + 8 KiB tail so mid-size outputs stay mostly |
| 335 | /// readable; handle-only results keep a 16 KiB head + 4 KiB tail. The head |
| 336 | /// and tail windows never overlap ([`head_tail_windows`]). |
| 337 | const HYBRID_HEAD_BYTES: usize = 32 * 1024; |
| 338 | const HYBRID_TAIL_BYTES: usize = 8 * 1024; |
| 339 | const HANDLE_ONLY_HEAD_BYTES: usize = 16 * 1024; |
| 340 | const HANDLE_ONLY_TAIL_BYTES: usize = 4 * 1024; |
| 341 | |
| 342 | /// Phrase used only by the TUI expand affordance and the UI-side detection of |
| 343 | /// historical truncated previews. Never emitted into model-facing content: |
| 344 | /// the model cannot open the tool details view, so the model-facing footer |
| 345 | /// carries the artifact path and a recovery instruction instead. |
| 346 | pub const SPILLOVER_PREVIEW_HINT: &str = "view full output in the tool details view"; |
| 347 | |
| 348 | /// Sentinel phrase the TUI matches on to recognise a current-format truncated |
| 349 | /// preview. It must stay a literal substring of every footer variant. |
| 350 | pub const SPILLOVER_RECOVERY_HINT: &str = "omitted range recovery:"; |
| 351 | |
| 352 | /// Model-facing recovery instruction for a truncated tool result. |
| 353 | /// |
| 354 | /// The previous text — "read it back with the read_file tool or with sed line |
| 355 | /// ranges" — named `read_file`, which is not model-visible at all (only `File` |
| 356 | /// is), and otherwise leaned on reaching the artifact by path. Reaching it by |
| 357 | /// path is *conditional*: `ToolContext::resolve_path` short-circuits under |
| 358 | /// trust mode, so `File action="read"` on an artifact under |
| 359 | /// `~/.codewhale/sessions/` succeeds in a trusted/auto session and is refused |
| 360 | /// as a path escape otherwise — and even when it succeeds it pages the file |
| 361 | /// rather than seeking the omitted range. Meanwhile `retrieve_tool_result` — |
| 362 | /// model-visible, purpose-built, unconditional, and already named correctly by |
| 363 | /// the web overflow path in `tools/web/overflow.rs` — went unmentioned. |
| 364 | /// `tests/adaptive_evidence_acceptance.rs` proves end to end that a model |
| 365 | /// handed one of these receipts can take the named ref and get the omitted |
| 366 | /// bytes back. |
| 367 | /// |
| 368 | /// The distinction that matters is retrievability, not tidiness. An adaptive |
| 369 | /// session artifact carries an `art_<id>` the retrieval tool resolves, so name |
| 370 | /// it. A legacy global spillover is authorized by an ownership sidecar whose |
| 371 | /// write is allowed to fail (see [`publish_legacy_spillover_ownership`]), so |
| 372 | /// promising retrieval there would just be a fourth dead route; say plainly |
| 373 | /// that there is no tool call for it and name what does work instead. |
| 374 | fn spillover_recovery_instruction(retrieval_ref: Option<&str>) -> String { |
| 375 | match retrieval_ref { |
| 376 | Some(reference) => format!( |
| 377 | "{SPILLOVER_RECOVERY_HINT} call retrieve_tool_result with ref=\"{reference}\" \ |
| 378 | (mode=\"tail\" for the end, mode=\"lines\" with lines=\"120-160\" for a range, \ |
| 379 | mode=\"query\" with query=\"…\" to search it)" |
| 380 | ), |
| 381 | None => format!( |
| 382 | "{SPILLOVER_RECOVERY_HINT} no tool call reaches this copy — re-run the command \ |
| 383 | with narrower output (a tighter filter, or head/tail) if you need the rest" |
| 384 | ), |
| 385 | } |
| 386 | } |
| 387 | |
| 388 | /// Model-facing footer for a truncated tool result. Names how much was |
| 389 | /// omitted (bytes and lines), where the complete output lives on disk, and |
| 390 | /// how the model can read the omitted range back. |
| 391 | fn spillover_preview_footer( |
| 392 | omitted_bytes: usize, |
| 393 | omitted_lines: usize, |
| 394 | recovery_path: &str, |
| 395 | retrieval_ref: Option<&str>, |
| 396 | ) -> String { |
| 397 | format!( |
| 398 | "… {} of output omitted ({omitted_lines} lines) — full output at {recovery_path}; {}", |
| 399 | crate::artifacts::format_byte_size(omitted_bytes.try_into().unwrap_or(u64::MAX)), |
| 400 | spillover_recovery_instruction(retrieval_ref) |
| 401 | ) |
| 402 | } |
| 403 | |
| 404 | /// Split `content` into a head of at most `head_bytes` and a tail of at most |
| 405 | /// `tail_bytes` that never overlap: the tail window always starts at or after |
| 406 | /// the head window ends, so no byte of the output appears twice and the |
| 407 | /// omitted count is exact. |
| 408 | fn head_tail_windows(content: &str, head_bytes: usize, tail_bytes: usize) -> (&str, &str) { |
| 409 | let head_end = (0..=head_bytes.min(content.len())) |
| 410 | .rev() |
| 411 | .find(|&index| content.is_char_boundary(index)) |
| 412 | .unwrap_or(0); |
| 413 | let tail_floor = content.len().saturating_sub(tail_bytes).max(head_end); |
| 414 | let tail_start = (tail_floor..=content.len()) |
| 415 | .find(|&index| content.is_char_boundary(index)) |
| 416 | .unwrap_or(content.len()); |
| 417 | (&content[..head_end], &content[tail_start..]) |
| 418 | } |
| 419 | |
| 420 | /// Build the model-facing preview for a truncated tool result: the head, an |
| 421 | /// honest footer naming how much was omitted and where the full output can be |
| 422 | /// read back, and a short retained tail. When the head and tail windows cover |
| 423 | /// the whole output (nothing was actually omitted), the content is returned |
| 424 | /// unchanged — the preview never claims a truncation that did not happen. |
| 425 | fn truncated_preview( |
| 426 | head: &str, |
| 427 | tail: &str, |
| 428 | original: &str, |
| 429 | recovery_path: &str, |
| 430 | retrieval_ref: Option<&str>, |
| 431 | ) -> String { |
| 432 | let omitted = original.len().saturating_sub(head.len() + tail.len()); |
| 433 | if omitted == 0 { |
| 434 | return original.to_string(); |
| 435 | } |
| 436 | let omitted_lines = original[head.len()..original.len() - tail.len()] |
| 437 | .lines() |
| 438 | .count(); |
| 439 | format!( |
| 440 | "{head}\n\n{}\n\n…\n{tail}", |
| 441 | spillover_preview_footer(omitted, omitted_lines, recovery_path, retrieval_ref) |
| 442 | ) |
| 443 | } |
| 444 | |
| 445 | /// Apply spillover to a tool result in place. If the result's |
| 446 | /// content exceeds [`SPILLOVER_THRESHOLD_BYTES`], writes the full |
| 447 | /// content to a sibling file under `~/.codewhale/tool_outputs/`, |
| 448 | /// replaces `result.content` with a [`SPILLOVER_HEAD_BYTES`] head |
| 449 | /// plus a footer naming the spillover path and how to read the |
| 450 | /// omitted range back, and stamps `metadata.spillover_path` so the |
| 451 | /// UI can render its expand annotation. |
| 452 | /// |
| 453 | /// Returns the spillover path on success, `None` if no spillover |
| 454 | /// happened (content small enough, error result, write failure). |
| 455 | /// Failures are logged but never bubble up — a tool that produced a |
| 456 | /// result shouldn't be marked failed because the spillover writer |
| 457 | /// couldn't reach disk; we degrade to no-op and the model gets the |
| 458 | /// original (large) content. |
| 459 | /// |
| 460 | /// Error results (`success == false`) are skipped: error messages |
| 461 | /// are typically short, and turning them into a truncated preview |
| 462 | /// would just hide the error from the model's reasoning. |
| 463 | #[cfg_attr(not(test), expect(dead_code))] |
| 464 | pub fn apply_spillover(result: &mut ToolResult, tool_id: &str) -> Option<PathBuf> { |
| 465 | apply_spillover_inner(result, tool_id, None, false) |
| 466 | } |
| 467 | |
| 468 | /// Apply spillover and publish session-scoped exact evidence. |
| 469 | /// |
| 470 | /// The default (classic) path writes the full bytes under the origin session |
| 471 | /// and replaces oversized content with a bounded head/tail preview whose |
| 472 | /// footer names the artifact path and how to read the omitted range back. |
| 473 | /// The adaptive evidence lane is reachable only through the explicit |
| 474 | /// `CODEWHALE_ADAPTIVE_OUTPUT_ROUTING` opt-in. |
| 475 | pub fn apply_spillover_with_artifact( |
| 476 | result: &mut ToolResult, |
| 477 | tool_id: &str, |
| 478 | tool_name: &str, |
| 479 | session_id: &str, |
| 480 | ) -> Option<PathBuf> { |
| 481 | apply_spillover_inner( |
| 482 | result, |
| 483 | tool_id, |
| 484 | Some(ArtifactSpilloverContext { |
| 485 | tool_name, |
| 486 | session_id, |
| 487 | }), |
| 488 | false, |
| 489 | ) |
| 490 | } |
| 491 | |
| 492 | /// [`apply_spillover_with_artifact`] for callers whose error payloads are |
| 493 | /// routinely as large as their successes — sub-agent tool output is often a |
| 494 | /// full build log, so the root loop's pass-errors-through rationale does not |
| 495 | /// hold there. |
| 496 | pub(crate) fn apply_spillover_with_artifact_including_errors( |
| 497 | result: &mut ToolResult, |
| 498 | tool_id: &str, |
| 499 | tool_name: &str, |
| 500 | session_id: &str, |
| 501 | ) -> Option<PathBuf> { |
| 502 | apply_spillover_inner( |
| 503 | result, |
| 504 | tool_id, |
| 505 | Some(ArtifactSpilloverContext { |
| 506 | tool_name, |
| 507 | session_id, |
| 508 | }), |
| 509 | true, |
| 510 | ) |
| 511 | } |
| 512 | |
| 513 | #[derive(Clone, Copy)] |
| 514 | struct ArtifactSpilloverContext<'a> { |
| 515 | tool_name: &'a str, |
| 516 | session_id: &'a str, |
| 517 | } |
| 518 | |
| 519 | fn apply_spillover_inner( |
| 520 | result: &mut ToolResult, |
| 521 | tool_id: &str, |
| 522 | artifact_context: Option<ArtifactSpilloverContext<'_>>, |
| 523 | bound_errors: bool, |
| 524 | ) -> Option<PathBuf> { |
| 525 | if crate::tools::large_output_router::adaptive_output_routing_enabled() |
| 526 | && let Some(context) = artifact_context |
| 527 | { |
| 528 | return apply_adaptive_evidence_inner(result, tool_id, context); |
| 529 | } |
| 530 | if !result.success && !bound_errors { |
| 531 | return None; |
| 532 | } |
| 533 | if result.content.len() <= SPILLOVER_THRESHOLD_BYTES { |
| 534 | return None; |
| 535 | } |
| 536 | let original_content = result.content.clone(); |
| 537 | let outcome = match maybe_spillover( |
| 538 | tool_id, |
| 539 | &original_content, |
| 540 | SPILLOVER_THRESHOLD_BYTES, |
| 541 | SPILLOVER_HEAD_BYTES, |
| 542 | ) { |
| 543 | Ok(Some(pair)) => pair, |
| 544 | Ok(None) => return None, |
| 545 | Err(err) => { |
| 546 | tracing::warn!( |
| 547 | target: "spillover", |
| 548 | ?err, |
| 549 | tool_id, |
| 550 | "spillover write failed; passing original content through" |
| 551 | ); |
| 552 | return None; |
| 553 | } |
| 554 | }; |
| 555 | let (_head, path) = outcome; |
| 556 | let (head, tail) = head_tail_windows( |
| 557 | &original_content, |
| 558 | SPILLOVER_HEAD_BYTES, |
| 559 | SPILLOVER_TAIL_BYTES, |
| 560 | ); |
| 561 | let digest = crate::hashing::sha256_hex(original_content.as_bytes()); |
| 562 | let path_str = path.display().to_string(); |
| 563 | |
| 564 | // Keep publishing the legacy ownership proof even though the model-facing |
| 565 | // footer no longer mentions retrieval: the tool-details pager authorizes |
| 566 | // legacy spillover reads through this sidecar. |
| 567 | if let Some(context) = artifact_context |
| 568 | && let Err(err) = publish_legacy_spillover_ownership( |
| 569 | &path, |
| 570 | context.session_id, |
| 571 | original_content.as_bytes(), |
| 572 | ) |
| 573 | { |
| 574 | tracing::warn!( |
| 575 | target: "spillover", |
| 576 | ?err, |
| 577 | tool_id, |
| 578 | "legacy spillover ownership publication failed" |
| 579 | ); |
| 580 | } |
| 581 | |
| 582 | let mut artifact_path = None; |
| 583 | if let Some(context) = artifact_context { |
| 584 | let artifact_id = crate::artifacts::artifact_id_for_tool_call(tool_id); |
| 585 | match crate::artifacts::write_session_artifact( |
| 586 | context.session_id, |
| 587 | &artifact_id, |
| 588 | &original_content, |
| 589 | ) { |
| 590 | Ok((absolute_path, relative_path)) => { |
| 591 | let record = crate::artifacts::record_tool_output_artifact( |
| 592 | context.session_id, |
| 593 | tool_id, |
| 594 | context.tool_name, |
| 595 | relative_path.clone(), |
| 596 | &original_content, |
| 597 | ); |
| 598 | result.content = truncated_preview( |
| 599 | head, |
| 600 | tail, |
| 601 | &original_content, |
| 602 | &crate::artifacts::format_artifact_relative_path(&absolute_path), |
| 603 | Some(artifact_id.as_str()), |
| 604 | ); |
| 605 | artifact_path = Some((absolute_path, relative_path, record)); |
| 606 | } |
| 607 | Err(err) => { |
| 608 | tracing::warn!( |
| 609 | target: "spillover", |
| 610 | ?err, |
| 611 | tool_id, |
| 612 | "session artifact write failed; falling back to legacy spillover footer" |
| 613 | ); |
| 614 | } |
| 615 | } |
| 616 | } |
| 617 | |
| 618 | if artifact_path.is_none() { |
| 619 | // Legacy fallback: no session artifact was written, so there is no |
| 620 | // `art_<id>` ref to hand over — only the on-disk path. |
| 621 | result.content = truncated_preview(head, tail, &original_content, &path_str, None); |
| 622 | } |
| 623 | |
| 624 | let metadata = result.metadata.get_or_insert_with(|| serde_json::json!({})); |
| 625 | if let Some(obj) = metadata.as_object_mut() { |
| 626 | if let Some((absolute_path, relative_path, record)) = artifact_path.as_ref() { |
| 627 | obj.insert( |
| 628 | "spillover_path".into(), |
| 629 | serde_json::Value::String(absolute_path.display().to_string()), |
| 630 | ); |
| 631 | obj.insert( |
| 632 | "legacy_spillover_path".into(), |
| 633 | serde_json::Value::String(path_str), |
| 634 | ); |
| 635 | obj.insert( |
| 636 | "artifact_id".into(), |
| 637 | serde_json::Value::String(record.id.clone()), |
| 638 | ); |
| 639 | obj.insert( |
| 640 | "artifact_session_id".into(), |
| 641 | serde_json::Value::String(record.session_id.clone()), |
| 642 | ); |
| 643 | obj.insert( |
| 644 | "artifact_relative_path".into(), |
| 645 | serde_json::Value::String(crate::artifacts::format_artifact_relative_path( |
| 646 | relative_path, |
| 647 | )), |
| 648 | ); |
| 649 | obj.insert( |
| 650 | "artifact_path".into(), |
| 651 | serde_json::Value::String(absolute_path.display().to_string()), |
| 652 | ); |
| 653 | obj.insert( |
| 654 | "artifact_byte_size".into(), |
| 655 | serde_json::Value::Number(serde_json::Number::from(record.byte_size)), |
| 656 | ); |
| 657 | obj.insert( |
| 658 | "artifact_preview".into(), |
| 659 | serde_json::Value::String(record.preview.clone()), |
| 660 | ); |
| 661 | } else { |
| 662 | obj.insert("spillover_path".into(), serde_json::Value::String(path_str)); |
| 663 | } |
| 664 | } else { |
| 665 | // Pre-existing metadata that wasn't a JSON object (rare, |
| 666 | // possibly an array). Replace with an object so we can |
| 667 | // attach our key without losing prior data — wrap it under |
| 668 | // a `_prior` field so callers that introspect can recover. |
| 669 | let prior = std::mem::replace(metadata, serde_json::json!({})); |
| 670 | if let Some(obj) = metadata.as_object_mut() { |
| 671 | obj.insert("_prior".into(), prior); |
| 672 | if let Some((absolute_path, relative_path, record)) = artifact_path.as_ref() { |
| 673 | obj.insert( |
| 674 | "spillover_path".into(), |
| 675 | serde_json::Value::String(absolute_path.display().to_string()), |
| 676 | ); |
| 677 | obj.insert( |
| 678 | "legacy_spillover_path".into(), |
| 679 | serde_json::Value::String(path.display().to_string()), |
| 680 | ); |
| 681 | obj.insert( |
| 682 | "artifact_id".into(), |
| 683 | serde_json::Value::String(record.id.clone()), |
| 684 | ); |
| 685 | obj.insert( |
| 686 | "artifact_session_id".into(), |
| 687 | serde_json::Value::String(record.session_id.clone()), |
| 688 | ); |
| 689 | obj.insert( |
| 690 | "artifact_relative_path".into(), |
| 691 | serde_json::Value::String(crate::artifacts::format_artifact_relative_path( |
| 692 | relative_path, |
| 693 | )), |
| 694 | ); |
| 695 | obj.insert( |
| 696 | "artifact_path".into(), |
| 697 | serde_json::Value::String(absolute_path.display().to_string()), |
| 698 | ); |
| 699 | obj.insert( |
| 700 | "artifact_byte_size".into(), |
| 701 | serde_json::Value::Number(serde_json::Number::from(record.byte_size)), |
| 702 | ); |
| 703 | obj.insert( |
| 704 | "artifact_preview".into(), |
| 705 | serde_json::Value::String(record.preview.clone()), |
| 706 | ); |
| 707 | } else { |
| 708 | obj.insert( |
| 709 | "spillover_path".into(), |
| 710 | serde_json::Value::String(path.display().to_string()), |
| 711 | ); |
| 712 | } |
| 713 | } |
| 714 | } |
| 715 | if let Some(obj) = result |
| 716 | .metadata |
| 717 | .as_mut() |
| 718 | .and_then(serde_json::Value::as_object_mut) |
| 719 | { |
| 720 | obj.insert("truncated".into(), serde_json::Value::Bool(true)); |
| 721 | obj.insert( |
| 722 | "content_digest".into(), |
| 723 | serde_json::Value::String(format!("sha256:{digest}")), |
| 724 | ); |
| 725 | obj.insert( |
| 726 | "original_byte_count".into(), |
| 727 | serde_json::Value::Number(serde_json::Number::from(original_content.len() as u64)), |
| 728 | ); |
| 729 | obj.insert( |
| 730 | "retained_head_bytes".into(), |
| 731 | serde_json::Value::Number(serde_json::Number::from(head.len() as u64)), |
| 732 | ); |
| 733 | obj.insert( |
| 734 | "retained_tail_bytes".into(), |
| 735 | serde_json::Value::Number(serde_json::Number::from(tail.len() as u64)), |
| 736 | ); |
| 737 | } |
| 738 | artifact_path |
| 739 | .map(|(absolute_path, _, _)| absolute_path) |
| 740 | .or(Some(path)) |
| 741 | } |
| 742 | |
| 743 | fn apply_adaptive_evidence_inner( |
| 744 | result: &mut ToolResult, |
| 745 | tool_id: &str, |
| 746 | context: ArtifactSpilloverContext<'_>, |
| 747 | ) -> Option<PathBuf> { |
| 748 | use crate::tools::large_output_router::{ |
| 749 | DEFAULT_LARGE_OUTPUT_THRESHOLD_TOKENS, EVIDENCE_RETENTION_SECS, EvidenceArtifact, |
| 750 | EvidenceRetentionState, EvidenceRouting, estimate_tokens, publish_evidence_metadata, |
| 751 | unix_millis_now, |
| 752 | }; |
| 753 | |
| 754 | let estimated_tokens = estimate_tokens(&result.content); |
| 755 | let threshold = result |
| 756 | .metadata |
| 757 | .as_ref() |
| 758 | .and_then(|metadata| metadata.get("evidence_threshold_tokens")) |
| 759 | .and_then(serde_json::Value::as_u64) |
| 760 | .and_then(|value| usize::try_from(value).ok()) |
| 761 | .unwrap_or(DEFAULT_LARGE_OUTPUT_THRESHOLD_TOKENS); |
| 762 | let routing = result |
| 763 | .metadata |
| 764 | .as_ref() |
| 765 | .and_then(|metadata| metadata.get("evidence_routing")) |
| 766 | .cloned() |
| 767 | .and_then(|value| serde_json::from_value::<EvidenceRouting>(value).ok()) |
| 768 | .unwrap_or_else(|| EvidenceRouting::from_token_estimate(estimated_tokens, threshold)); |
| 769 | if routing == EvidenceRouting::Inline { |
| 770 | return None; |
| 771 | } |
| 772 | |
| 773 | let original = result.content.clone(); |
| 774 | let (head_bytes, tail_bytes) = if routing == EvidenceRouting::Hybrid { |
| 775 | (HYBRID_HEAD_BYTES, HYBRID_TAIL_BYTES) |
| 776 | } else { |
| 777 | (HANDLE_ONLY_HEAD_BYTES, HANDLE_ONLY_TAIL_BYTES) |
| 778 | }; |
| 779 | let (head, tail) = head_tail_windows(&original, head_bytes, tail_bytes); |
| 780 | let omitted = original.len().saturating_sub(head.len() + tail.len()); |
| 781 | if omitted == 0 { |
| 782 | // The whole output fits inside the preview budget: there is nothing |
| 783 | // to recover, so publishing an artifact and claiming a truncation |
| 784 | // would both be dishonest. Pass the content through unchanged. |
| 785 | return None; |
| 786 | } |
| 787 | let head_len = head.len(); |
| 788 | let tail_len = tail.len(); |
| 789 | |
| 790 | let artifact_id = crate::artifacts::artifact_id_for_tool_call(tool_id); |
| 791 | let relative_path = crate::artifacts::session_artifact_relative_path(&artifact_id); |
| 792 | let digest = crate::hashing::sha256_hex(original.as_bytes()); |
| 793 | let now_ms = unix_millis_now(); |
| 794 | let proposed_artifact = EvidenceArtifact { |
| 795 | handle: artifact_id.clone(), |
| 796 | digest: digest.clone(), |
| 797 | size_bytes: original.len().try_into().unwrap_or(u64::MAX), |
| 798 | content_type: if serde_json::from_str::<serde_json::Value>(&original).is_ok() { |
| 799 | "application/json".to_string() |
| 800 | } else { |
| 801 | "text/plain".to_string() |
| 802 | }, |
| 803 | tool_name: context.tool_name.to_string(), |
| 804 | call_id: tool_id.to_string(), |
| 805 | origin_session: context.session_id.to_string(), |
| 806 | generation: 1, |
| 807 | redacted: false, |
| 808 | encoding: "utf-8".to_string(), |
| 809 | retention_state: EvidenceRetentionState::Live, |
| 810 | created_at_unix_ms: now_ms, |
| 811 | retain_until_unix_ms: now_ms.saturating_add(EVIDENCE_RETENTION_SECS * 1_000), |
| 812 | storage_path: relative_path.clone(), |
| 813 | }; |
| 814 | let artifact = match crate::tools::large_output_router::read_evidence_metadata( |
| 815 | context.session_id, |
| 816 | &artifact_id, |
| 817 | ) { |
| 818 | Ok(existing) |
| 819 | if existing.digest == proposed_artifact.digest |
| 820 | && existing.size_bytes == proposed_artifact.size_bytes |
| 821 | && existing.call_id == proposed_artifact.call_id |
| 822 | && existing.origin_session == proposed_artifact.origin_session => |
| 823 | { |
| 824 | existing |
| 825 | } |
| 826 | Ok(_) => { |
| 827 | tracing::warn!(target: "evidence", tool_id, "adaptive evidence replay conflicts with immutable metadata"); |
| 828 | return None; |
| 829 | } |
| 830 | Err(err) if err.kind() == std::io::ErrorKind::NotFound => { |
| 831 | if let Err(err) = publish_evidence_metadata(context.session_id, &proposed_artifact) { |
| 832 | tracing::warn!(target: "evidence", ?err, tool_id, "adaptive evidence metadata publication failed"); |
| 833 | return None; |
| 834 | } |
| 835 | proposed_artifact |
| 836 | } |
| 837 | Err(err) => { |
| 838 | tracing::warn!(target: "evidence", ?err, tool_id, "adaptive evidence metadata validation failed"); |
| 839 | return None; |
| 840 | } |
| 841 | }; |
| 842 | |
| 843 | // Seal the ownership/integrity record before publishing predictable |
| 844 | // `art_<call>.txt` bytes. If metadata publication fails, no payload exists |
| 845 | // for a guessed handle to retrieve without the generation, redaction, |
| 846 | // retention, size, and digest checks above. A metadata-only interruption |
| 847 | // is safe: the handle is never advertised and a retry can idempotently |
| 848 | // publish the matching bytes. |
| 849 | let (absolute_path, relative_path) = match crate::artifacts::write_session_artifact_immutable( |
| 850 | context.session_id, |
| 851 | &artifact_id, |
| 852 | original.as_bytes(), |
| 853 | ) { |
| 854 | Ok(paths) => paths, |
| 855 | Err(err) => { |
| 856 | tracing::warn!(target: "evidence", ?err, tool_id, "adaptive evidence content publication failed"); |
| 857 | return None; |
| 858 | } |
| 859 | }; |
| 860 | |
| 861 | let record = crate::artifacts::record_tool_output_artifact( |
| 862 | context.session_id, |
| 863 | tool_id, |
| 864 | context.tool_name, |
| 865 | relative_path.clone(), |
| 866 | &original, |
| 867 | ); |
| 868 | result.content = truncated_preview( |
| 869 | head, |
| 870 | tail, |
| 871 | &original, |
| 872 | &crate::artifacts::format_artifact_relative_path(&absolute_path), |
| 873 | Some(artifact_id.as_str()), |
| 874 | ); |
| 875 | let metadata = result.metadata.get_or_insert_with(|| serde_json::json!({})); |
| 876 | if let Some(object) = metadata.as_object_mut() { |
| 877 | object.insert( |
| 878 | "spillover_path".into(), |
| 879 | absolute_path.display().to_string().into(), |
| 880 | ); |
| 881 | object.insert("artifact_id".into(), artifact_id.into()); |
| 882 | object.insert("artifact_session_id".into(), context.session_id.into()); |
| 883 | object.insert( |
| 884 | "artifact_relative_path".into(), |
| 885 | crate::artifacts::format_artifact_relative_path(&relative_path).into(), |
| 886 | ); |
| 887 | object.insert("artifact_byte_size".into(), artifact.size_bytes.into()); |
| 888 | object.insert("artifact_digest".into(), digest.into()); |
| 889 | object.insert("artifact_generation".into(), artifact.generation.into()); |
| 890 | object.insert("artifact_encoding".into(), artifact.encoding.into()); |
| 891 | object.insert("artifact_retention_state".into(), "live".into()); |
| 892 | object.insert("evidence_available".into(), true.into()); |
| 893 | object.insert("truncated".into(), true.into()); |
| 894 | object.insert("original_byte_count".into(), artifact.size_bytes.into()); |
| 895 | object.insert("retained_head_bytes".into(), head_len.into()); |
| 896 | object.insert("retained_tail_bytes".into(), tail_len.into()); |
| 897 | object.insert( |
| 898 | "artifact_preview".into(), |
| 899 | original.chars().take(200).collect::<String>().into(), |
| 900 | ); |
| 901 | object.insert( |
| 902 | "artifact_record".into(), |
| 903 | serde_json::to_value(record).unwrap_or(serde_json::Value::Null), |
| 904 | ); |
| 905 | } |
| 906 | Some(absolute_path) |
| 907 | } |
| 908 | |
| 909 | /// Sanitise a tool call id for use as a filename. Keeps ASCII |
| 910 | /// alphanumerics, `-`, and `_`; rejects `.` to keep `..` traversal |
| 911 | /// out, rejects empty results. Returns `None` if the input contains |
| 912 | /// no acceptable characters. |
| 913 | fn sanitise_id(id: &str) -> Option<String> { |
| 914 | let cleaned: String = id |
| 915 | .chars() |
| 916 | .filter(|c| c.is_ascii_alphanumeric() || *c == '-' || *c == '_') |
| 917 | .collect(); |
| 918 | if cleaned.is_empty() { |
| 919 | None |
| 920 | } else { |
| 921 | Some(cleaned) |
| 922 | } |
| 923 | } |
| 924 | |
| 925 | /// Override the storage roots for tests so they don't pollute the |
| 926 | /// user's real `~/.codewhale/` directory. This uses explicit test hooks instead |
| 927 | /// of `$HOME` because Windows home-dir resolution can ignore environment |
| 928 | /// overrides and return the runner profile directory. |
| 929 | #[cfg(test)] |
| 930 | fn with_test_home<F, R>(home: &Path, f: F) -> R |
| 931 | where |
| 932 | F: FnOnce() -> R, |
| 933 | { |
| 934 | let _artifact_guard = crate::artifacts::TEST_ARTIFACT_SESSIONS_GUARD |
| 935 | .lock() |
| 936 | .unwrap_or_else(|err| err.into_inner()); |
| 937 | |
| 938 | struct StorageRootOverride { |
| 939 | prior_spillover: Option<PathBuf>, |
| 940 | prior_artifacts: Option<PathBuf>, |
| 941 | } |
| 942 | |
| 943 | impl Drop for StorageRootOverride { |
| 944 | fn drop(&mut self) { |
| 945 | set_test_spillover_root(self.prior_spillover.take()); |
| 946 | crate::artifacts::set_test_artifact_sessions_root(self.prior_artifacts.take()); |
| 947 | } |
| 948 | } |
| 949 | |
| 950 | // Tests in this module serialize spillover through `TEST_GUARD`; the |
| 951 | // artifact guard above protects the session-artifact root shared with |
| 952 | // artifacts.rs tests. |
| 953 | let prior_spillover = |
| 954 | set_test_spillover_root(Some(home.join(".codewhale").join(SPILLOVER_DIR_NAME))); |
| 955 | let prior_artifacts = crate::artifacts::set_test_artifact_sessions_root(Some( |
| 956 | home.join(".codewhale").join("sessions"), |
| 957 | )); |
| 958 | let _restore = StorageRootOverride { |
| 959 | prior_spillover, |
| 960 | prior_artifacts, |
| 961 | }; |
| 962 | f() |
| 963 | } |
| 964 | |
| 965 | #[cfg(test)] |
| 966 | mod tests { |
| 967 | use super::*; |
| 968 | use tempfile::tempdir; |
| 969 | |
| 970 | /// Tests in this module serialize through this guard because they mutate |
| 971 | /// process-global test storage roots. Without it, cargo's parallel runner |
| 972 | /// would observe interleaved overrides. |
| 973 | fn setup() -> std::sync::MutexGuard<'static, ()> { |
| 974 | super::TEST_SPILLOVER_GUARD |
| 975 | .lock() |
| 976 | .unwrap_or_else(|e| e.into_inner()) |
| 977 | } |
| 978 | |
| 979 | /// Run the adaptive evidence lane directly. These cases exercise it |
| 980 | /// without the `CODEWHALE_ADAPTIVE_OUTPUT_ROUTING` process opt-in so |
| 981 | /// parallel tests keep a deterministic routing decision. |
| 982 | fn adaptive_spillover( |
| 983 | result: &mut ToolResult, |
| 984 | tool_id: &str, |
| 985 | tool_name: &str, |
| 986 | session_id: &str, |
| 987 | ) -> Option<PathBuf> { |
| 988 | apply_adaptive_evidence_inner( |
| 989 | result, |
| 990 | tool_id, |
| 991 | ArtifactSpilloverContext { |
| 992 | tool_name, |
| 993 | session_id, |
| 994 | }, |
| 995 | ) |
| 996 | } |
| 997 | |
| 998 | /// The old hint named `read_file`, which is not registered for the model |
| 999 | /// at all, and otherwise pointed at routes that only reach the artifact |
| 1000 | /// under trust mode (see [`spillover_recovery_instruction`]), while never |
| 1001 | /// naming `retrieve_tool_result` — model-visible, unconditional, and built |
| 1002 | /// for exactly this. |
| 1003 | #[test] |
| 1004 | fn truncation_footer_names_a_recovery_route_that_works() { |
| 1005 | let footer = spillover_preview_footer( |
| 1006 | 4096, |
| 1007 | 120, |
| 1008 | "/tmp/artifacts/art_call-1.txt", |
| 1009 | Some("art_call-1"), |
| 1010 | ); |
| 1011 | |
| 1012 | assert!(footer.contains("retrieve_tool_result"), "{footer}"); |
| 1013 | assert!(footer.contains("ref=\"art_call-1\""), "{footer}"); |
| 1014 | assert!(!footer.contains("read_file"), "{footer}"); |
| 1015 | assert!(!footer.contains("sed"), "{footer}"); |
| 1016 | } |
| 1017 | |
| 1018 | /// The legacy global copy has no guaranteed authorization sidecar, so the |
| 1019 | /// footer must not invent a fourth dead route — but it still must not name |
| 1020 | /// the three it used to. |
| 1021 | #[test] |
| 1022 | fn truncation_footer_without_an_artifact_promises_nothing_it_cannot_deliver() { |
| 1023 | let footer = spillover_preview_footer(4096, 120, "/tmp/tool_outputs/call-1.txt", None); |
| 1024 | |
| 1025 | assert!( |
| 1026 | footer.contains("no tool call reaches this copy"), |
| 1027 | "{footer}" |
| 1028 | ); |
| 1029 | assert!(!footer.contains("retrieve_tool_result"), "{footer}"); |
| 1030 | assert!(!footer.contains("read_file"), "{footer}"); |
| 1031 | assert!(!footer.contains("sed"), "{footer}"); |
| 1032 | } |
| 1033 | |
| 1034 | /// The TUI keys its "this preview was truncated" detection off the shared |
| 1035 | /// constant, so it has to stay a literal substring of every variant. |
| 1036 | #[test] |
| 1037 | fn every_footer_variant_carries_the_ui_detection_marker() { |
| 1038 | for reference in [Some("art_call-1"), None] { |
| 1039 | let footer = spillover_preview_footer(4096, 120, "/tmp/x.txt", reference); |
| 1040 | assert!(footer.contains(SPILLOVER_RECOVERY_HINT), "{footer}"); |
| 1041 | } |
| 1042 | } |
| 1043 | |
| 1044 | #[test] |
| 1045 | fn with_test_home_overrides_storage_roots_without_home_resolution() { |
| 1046 | let _g = setup(); |
| 1047 | let tmp = tempdir().unwrap(); |
| 1048 | |
| 1049 | with_test_home(tmp.path(), || { |
| 1050 | assert_eq!( |
| 1051 | spillover_root().as_deref(), |
| 1052 | Some(tmp.path().join(".codewhale").join("tool_outputs").as_path()) |
| 1053 | ); |
| 1054 | assert_eq!( |
| 1055 | crate::artifacts::session_artifact_absolute_path( |
| 1056 | "session-123", |
| 1057 | &PathBuf::from("artifacts").join("art_call-big.txt") |
| 1058 | ) |
| 1059 | .as_deref(), |
| 1060 | Some( |
| 1061 | tmp.path() |
| 1062 | .join(".codewhale") |
| 1063 | .join("sessions") |
| 1064 | .join("session-123") |
| 1065 | .join("artifacts") |
| 1066 | .join("art_call-big.txt") |
| 1067 | .as_path() |
| 1068 | ) |
| 1069 | ); |
| 1070 | }); |
| 1071 | } |
| 1072 | |
| 1073 | #[test] |
| 1074 | fn sanitise_id_keeps_safe_chars_and_drops_dangerous() { |
| 1075 | assert_eq!(super::sanitise_id("abc-123_x"), Some("abc-123_x".into())); |
| 1076 | // `.` is dropped to keep `..` out of the path. |
| 1077 | assert_eq!(super::sanitise_id("../etc"), Some("etc".into())); |
| 1078 | assert_eq!(super::sanitise_id("/etc/passwd"), Some("etcpasswd".into())); |
| 1079 | // Empty-after-sanitise → None. |
| 1080 | assert!(super::sanitise_id("...").is_none()); |
| 1081 | assert!(super::sanitise_id("").is_none()); |
| 1082 | } |
| 1083 | |
| 1084 | #[test] |
| 1085 | fn write_spillover_creates_directory_and_writes_file() { |
| 1086 | let _g = setup(); |
| 1087 | let tmp = tempdir().unwrap(); |
| 1088 | with_test_home(tmp.path(), || { |
| 1089 | let path = write_spillover("call-abc", "hello world").expect("write"); |
| 1090 | assert!(path.exists(), "{path:?} missing"); |
| 1091 | let body = fs::read_to_string(&path).unwrap(); |
| 1092 | assert_eq!(body, "hello world"); |
| 1093 | // Directory landed under `<HOME>/.codewhale/tool_outputs/`. |
| 1094 | // Compare components instead of a substring on `to_string_lossy` |
| 1095 | // — Windows uses `\` as the separator so a `/` substring match |
| 1096 | // would falsely fail there. |
| 1097 | let components: Vec<&str> = path |
| 1098 | .components() |
| 1099 | .filter_map(|c| c.as_os_str().to_str()) |
| 1100 | .collect(); |
| 1101 | assert!( |
| 1102 | components.contains(&".codewhale") && components.contains(&"tool_outputs"), |
| 1103 | "spillover path missing expected `.codewhale/tool_outputs/...` segments: {path:?}" |
| 1104 | ); |
| 1105 | }); |
| 1106 | } |
| 1107 | |
| 1108 | #[test] |
| 1109 | fn write_spillover_rejects_empty_id() { |
| 1110 | let _g = setup(); |
| 1111 | let tmp = tempdir().unwrap(); |
| 1112 | with_test_home(tmp.path(), || { |
| 1113 | let err = write_spillover("...", "x").unwrap_err(); |
| 1114 | assert_eq!(err.kind(), io::ErrorKind::InvalidInput); |
| 1115 | }); |
| 1116 | } |
| 1117 | |
| 1118 | #[test] |
| 1119 | fn maybe_spillover_returns_none_below_threshold() { |
| 1120 | let _g = setup(); |
| 1121 | let tmp = tempdir().unwrap(); |
| 1122 | with_test_home(tmp.path(), || { |
| 1123 | let out = maybe_spillover("call-1", "tiny content", 100 * 1024, 4 * 1024).expect("ok"); |
| 1124 | assert!(out.is_none()); |
| 1125 | }); |
| 1126 | } |
| 1127 | |
| 1128 | #[test] |
| 1129 | fn maybe_spillover_writes_and_returns_head_above_threshold() { |
| 1130 | let _g = setup(); |
| 1131 | let tmp = tempdir().unwrap(); |
| 1132 | with_test_home(tmp.path(), || { |
| 1133 | // Content larger than the threshold. |
| 1134 | let big = "A".repeat(2_000); |
| 1135 | let (head, path) = maybe_spillover("call-2", &big, 1_000, 256) |
| 1136 | .expect("ok") |
| 1137 | .expect("should have spilled"); |
| 1138 | // Head is bounded. |
| 1139 | assert_eq!(head.len(), 256); |
| 1140 | // Full content on disk. |
| 1141 | let body = fs::read_to_string(&path).unwrap(); |
| 1142 | assert_eq!(body.len(), 2_000); |
| 1143 | }); |
| 1144 | } |
| 1145 | |
| 1146 | #[test] |
| 1147 | fn maybe_spillover_does_not_split_inside_a_codepoint() { |
| 1148 | let _g = setup(); |
| 1149 | let tmp = tempdir().unwrap(); |
| 1150 | with_test_home(tmp.path(), || { |
| 1151 | // 4 byte chars; ask for 3 bytes of head → walks back to |
| 1152 | // the previous char boundary (0). |
| 1153 | let s = "🐳🐳🐳🐳"; // 4 × 4-byte codepoints |
| 1154 | assert_eq!(s.len(), 16); |
| 1155 | let (head, _) = maybe_spillover("call-3", s, 1, 3) |
| 1156 | .expect("ok") |
| 1157 | .expect("spilled"); |
| 1158 | // 3 isn't a char boundary in this string; walk back → 0. |
| 1159 | assert_eq!(head, ""); |
| 1160 | // Asking for 4 bytes lands on the first char boundary. |
| 1161 | let (head, _) = maybe_spillover("call-3b", s, 1, 4) |
| 1162 | .expect("ok") |
| 1163 | .expect("spilled"); |
| 1164 | assert_eq!(head, "🐳"); |
| 1165 | }); |
| 1166 | } |
| 1167 | |
| 1168 | #[test] |
| 1169 | fn prune_older_than_handles_missing_root() { |
| 1170 | let _g = setup(); |
| 1171 | let tmp = tempdir().unwrap(); |
| 1172 | with_test_home(tmp.path(), || { |
| 1173 | // Nothing has ever written; root doesn't exist; that's fine. |
| 1174 | let count = prune_older_than(SPILLOVER_MAX_AGE).expect("ok"); |
| 1175 | assert_eq!(count, 0); |
| 1176 | }); |
| 1177 | } |
| 1178 | |
| 1179 | // The mtime backdate uses utimensat (Unix-only). On Windows the |
| 1180 | // filetime_set_modified helper is a no-op, so the prune wouldn't see |
| 1181 | // any stale files. Gate the whole test on `cfg(unix)` instead of |
| 1182 | // testing a no-op path that can't fail meaningfully. |
| 1183 | #[test] |
| 1184 | #[cfg(unix)] |
| 1185 | fn prune_older_than_keeps_fresh_files_drops_stale_ones() { |
| 1186 | let _g = setup(); |
| 1187 | let tmp = tempdir().unwrap(); |
| 1188 | with_test_home(tmp.path(), || { |
| 1189 | let fresh = write_spillover("fresh", "x").unwrap(); |
| 1190 | let stale = write_spillover("stale", "y").unwrap(); |
| 1191 | |
| 1192 | // Backdate `stale` to 30 days ago. |
| 1193 | let thirty_days = SystemTime::now() - Duration::from_secs(30 * 24 * 60 * 60); |
| 1194 | filetime_set_modified(&stale, thirty_days); |
| 1195 | |
| 1196 | let pruned = prune_older_than(SPILLOVER_MAX_AGE).unwrap(); |
| 1197 | assert_eq!(pruned, 1); |
| 1198 | assert!(fresh.exists()); |
| 1199 | assert!(!stale.exists()); |
| 1200 | }); |
| 1201 | } |
| 1202 | |
| 1203 | /// Set the mtime on a file. The workspace doesn't pull the |
| 1204 | /// `filetime` crate, so we reach for `utimensat` directly on |
| 1205 | /// Unix. Windows is a no-op — the prune semantics are the same |
| 1206 | /// and the per-cycle stress test lives on the Unix path. |
| 1207 | #[cfg(unix)] |
| 1208 | fn filetime_set_modified(path: &Path, when: SystemTime) { |
| 1209 | let secs = when |
| 1210 | .duration_since(SystemTime::UNIX_EPOCH) |
| 1211 | .unwrap_or_default() |
| 1212 | .as_secs() as libc::time_t; |
| 1213 | let times = [ |
| 1214 | libc::timespec { |
| 1215 | tv_sec: secs, |
| 1216 | tv_nsec: 0, |
| 1217 | }, |
| 1218 | libc::timespec { |
| 1219 | tv_sec: secs, |
| 1220 | tv_nsec: 0, |
| 1221 | }, |
| 1222 | ]; |
| 1223 | let path_c = std::ffi::CString::new(path.as_os_str().as_encoded_bytes()).unwrap(); |
| 1224 | // SAFETY: path_c is a valid CString; times is a 2-element array |
| 1225 | // matching utimensat's signature. |
| 1226 | let rc = unsafe { libc::utimensat(libc::AT_FDCWD, path_c.as_ptr(), times.as_ptr(), 0) }; |
| 1227 | assert_eq!( |
| 1228 | rc, |
| 1229 | 0, |
| 1230 | "utimensat failed: {}", |
| 1231 | std::io::Error::last_os_error() |
| 1232 | ); |
| 1233 | } |
| 1234 | |
| 1235 | // Windows stub removed in v0.8.8 — the only caller of |
| 1236 | // `filetime_set_modified` is `prune_older_than_keeps_fresh_files_drops_stale_ones`, |
| 1237 | // which is now `#[cfg(unix)]` because mtime backdating requires |
| 1238 | // `utimensat` and a Windows no-op stub can't make the assertion pass |
| 1239 | // anyway. Keeping the stub triggered `-D dead-code` on Windows builds |
| 1240 | // (the prune test was the only caller) and broke `Test (windows-latest)`. |
| 1241 | |
| 1242 | #[test] |
| 1243 | fn apply_spillover_is_noop_below_threshold() { |
| 1244 | let _g = setup(); |
| 1245 | let tmp = tempdir().unwrap(); |
| 1246 | with_test_home(tmp.path(), || { |
| 1247 | let mut result = ToolResult::success("small payload"); |
| 1248 | let path = apply_spillover(&mut result, "call-small"); |
| 1249 | assert!(path.is_none()); |
| 1250 | assert_eq!(result.content, "small payload"); |
| 1251 | assert!(result.metadata.is_none()); |
| 1252 | }); |
| 1253 | } |
| 1254 | |
| 1255 | #[test] |
| 1256 | fn apply_spillover_is_noop_for_error_results() { |
| 1257 | let _g = setup(); |
| 1258 | let tmp = tempdir().unwrap(); |
| 1259 | with_test_home(tmp.path(), || { |
| 1260 | // Even very large error messages are passed through — |
| 1261 | // truncating an error would hide it from the model. |
| 1262 | let big_err = "boom\n".repeat(50_000); |
| 1263 | let mut result = ToolResult::error(big_err.clone()); |
| 1264 | let path = apply_spillover(&mut result, "call-err"); |
| 1265 | assert!(path.is_none()); |
| 1266 | assert_eq!(result.content, big_err); |
| 1267 | }); |
| 1268 | } |
| 1269 | |
| 1270 | #[test] |
| 1271 | fn apply_spillover_truncates_and_stamps_metadata_above_threshold() { |
| 1272 | let _g = setup(); |
| 1273 | let tmp = tempdir().unwrap(); |
| 1274 | with_test_home(tmp.path(), || { |
| 1275 | // 200 KiB body — well above the 100 KiB threshold. |
| 1276 | let big = "X".repeat(200 * 1024); |
| 1277 | let mut result = ToolResult::success(big.clone()); |
| 1278 | let path = apply_spillover(&mut result, "call-big").expect("should spill"); |
| 1279 | |
| 1280 | // Inline content shrunk to head + honest preview footer. |
| 1281 | assert!(result.content.len() < big.len()); |
| 1282 | assert!( |
| 1283 | !result.content.contains(SPILLOVER_PREVIEW_HINT), |
| 1284 | "the tool-details phrase is a UI affordance, not model-facing" |
| 1285 | ); |
| 1286 | assert!( |
| 1287 | result.content.contains("of output omitted"), |
| 1288 | "footer missing: {}", |
| 1289 | &result.content[result.content.len().saturating_sub(200)..] |
| 1290 | ); |
| 1291 | // The footer tells the model where the full output lives and how |
| 1292 | // to read the omitted range back. |
| 1293 | assert!(result.content.contains("full output at")); |
| 1294 | assert!(result.content.contains(&path.display().to_string())); |
| 1295 | assert!(result.content.contains(SPILLOVER_RECOVERY_HINT)); |
| 1296 | assert!( |
| 1297 | !result.content.contains("retrieve_tool_result"), |
| 1298 | "legacy spillover ownership can fail to publish; promising \ |
| 1299 | retrieval here would be another dead route" |
| 1300 | ); |
| 1301 | assert!(!result.content.contains("read_file")); |
| 1302 | assert!(!result.content.contains("sed")); |
| 1303 | |
| 1304 | // Full bytes are on disk at the returned path. |
| 1305 | assert!(path.exists(), "spillover file missing: {path:?}"); |
| 1306 | let body = fs::read_to_string(&path).unwrap(); |
| 1307 | assert_eq!(body.len(), 200 * 1024); |
| 1308 | |
| 1309 | // metadata.spillover_path stamped for the UI to find. |
| 1310 | let metadata = result.metadata.expect("metadata stamped"); |
| 1311 | let stamped = metadata |
| 1312 | .get("spillover_path") |
| 1313 | .and_then(serde_json::Value::as_str) |
| 1314 | .expect("spillover_path key present"); |
| 1315 | assert_eq!(stamped, path.display().to_string()); |
| 1316 | assert_eq!(metadata["truncated"], true); |
| 1317 | assert_eq!(metadata["original_byte_count"], 200 * 1024); |
| 1318 | assert_eq!(metadata["retained_head_bytes"], SPILLOVER_HEAD_BYTES); |
| 1319 | assert_eq!(metadata["retained_tail_bytes"], SPILLOVER_TAIL_BYTES); |
| 1320 | assert!( |
| 1321 | metadata["content_digest"] |
| 1322 | .as_str() |
| 1323 | .is_some_and(|digest| digest.starts_with("sha256:")) |
| 1324 | ); |
| 1325 | }); |
| 1326 | } |
| 1327 | |
| 1328 | #[test] |
| 1329 | fn apply_spillover_with_artifact_writes_session_file_and_plain_preview() { |
| 1330 | let _g = setup(); |
| 1331 | let tmp = tempdir().unwrap(); |
| 1332 | with_test_home(tmp.path(), || { |
| 1333 | let big = "checking crate ... error[E0425]: cannot find value\n".repeat(4_000); |
| 1334 | let mut result = ToolResult::success(big.clone()); |
| 1335 | let path = adaptive_spillover(&mut result, "call-big", "exec_shell", "session-123") |
| 1336 | .expect("should spill"); |
| 1337 | |
| 1338 | let session_artifact = tmp |
| 1339 | .path() |
| 1340 | .join(".codewhale") |
| 1341 | .join("sessions") |
| 1342 | .join("session-123") |
| 1343 | .join("artifacts") |
| 1344 | .join("art_call-big.txt"); |
| 1345 | assert_eq!(path, session_artifact); |
| 1346 | assert_eq!(fs::read_to_string(&session_artifact).unwrap(), big); |
| 1347 | assert!( |
| 1348 | !tmp.path() |
| 1349 | .join(".codewhale/tool_outputs/call-big.txt") |
| 1350 | .exists(), |
| 1351 | "adaptive evidence stores one exact origin-session copy" |
| 1352 | ); |
| 1353 | // The model sees a bounded preview with an honest footer: the |
| 1354 | // artifact path plus the retrieval call that actually resolves it. |
| 1355 | assert!(!result.content.contains(SPILLOVER_PREVIEW_HINT)); |
| 1356 | assert!(result.content.contains("\n…\n")); |
| 1357 | assert!(result.content.contains("of output omitted")); |
| 1358 | assert!(result.content.contains("full output at")); |
| 1359 | assert!(result.content.contains(SPILLOVER_RECOVERY_HINT)); |
| 1360 | assert!( |
| 1361 | result.content.contains("art_call-big.txt"), |
| 1362 | "footer must name the artifact path so the model can recover the output" |
| 1363 | ); |
| 1364 | assert!(!result.content.contains("Exact evidence retained")); |
| 1365 | assert!( |
| 1366 | result.content.contains("retrieve_tool_result"), |
| 1367 | "a session artifact is retrievable; the footer must say so: {}", |
| 1368 | result.content |
| 1369 | ); |
| 1370 | assert!( |
| 1371 | result.content.contains("ref=\"art_call-big\""), |
| 1372 | "the footer must hand over a ref that resolves: {}", |
| 1373 | result.content |
| 1374 | ); |
| 1375 | assert!( |
| 1376 | session_artifact |
| 1377 | .with_file_name("art_call-big.evidence.json") |
| 1378 | .exists() |
| 1379 | ); |
| 1380 | |
| 1381 | let metadata = result.metadata.expect("metadata stamped"); |
| 1382 | assert_eq!( |
| 1383 | metadata |
| 1384 | .get("artifact_id") |
| 1385 | .and_then(serde_json::Value::as_str), |
| 1386 | Some("art_call-big") |
| 1387 | ); |
| 1388 | assert_eq!( |
| 1389 | metadata |
| 1390 | .get("artifact_relative_path") |
| 1391 | .and_then(serde_json::Value::as_str), |
| 1392 | Some("artifacts/art_call-big.txt") |
| 1393 | ); |
| 1394 | assert_eq!( |
| 1395 | metadata |
| 1396 | .get("artifact_session_id") |
| 1397 | .and_then(serde_json::Value::as_str), |
| 1398 | Some("session-123") |
| 1399 | ); |
| 1400 | assert_eq!(metadata["original_byte_count"], big.len()); |
| 1401 | assert!(metadata["retained_head_bytes"].as_u64().unwrap_or(0) <= 16 * 1024); |
| 1402 | assert!(metadata["retained_tail_bytes"].as_u64().unwrap_or(0) <= 4 * 1024); |
| 1403 | }); |
| 1404 | } |
| 1405 | |
| 1406 | #[test] |
| 1407 | fn adaptive_evidence_keeps_success_and_failure_exact_distinct_and_out_of_context() { |
| 1408 | let _g = setup(); |
| 1409 | let tmp = tempdir().unwrap(); |
| 1410 | with_test_home(tmp.path(), || { |
| 1411 | let sentinel = "DEEP_RAW_SENTINEL"; |
| 1412 | // Payloads must exceed the 32_768-token (≈96 KiB) handle-only |
| 1413 | // threshold so adaptive routing actually spills them. |
| 1414 | let success_raw = format!( |
| 1415 | "{}{}{}", |
| 1416 | "head\n".repeat(30_000), |
| 1417 | sentinel, |
| 1418 | "tail\n".repeat(30_000) |
| 1419 | ); |
| 1420 | let failure_raw = format!("{}{}", "failure\n".repeat(30_000), "FAILURE_END"); |
| 1421 | let mut success = ToolResult::success(success_raw.clone()); |
| 1422 | let mut failure = ToolResult::error(failure_raw.clone()); |
| 1423 | |
| 1424 | let success_path = |
| 1425 | adaptive_spillover(&mut success, "call-success", "exec_shell", "session-a") |
| 1426 | .expect("success evidence"); |
| 1427 | let failure_path = |
| 1428 | adaptive_spillover(&mut failure, "call-failure", "mcp_fixture", "session-a") |
| 1429 | .expect("failure evidence"); |
| 1430 | |
| 1431 | assert_ne!(success_path, failure_path); |
| 1432 | assert_eq!( |
| 1433 | std::fs::read(&success_path).unwrap(), |
| 1434 | success_raw.as_bytes() |
| 1435 | ); |
| 1436 | assert_eq!( |
| 1437 | std::fs::read(&failure_path).unwrap(), |
| 1438 | failure_raw.as_bytes() |
| 1439 | ); |
| 1440 | assert!(!success.content.contains(sentinel)); |
| 1441 | // Handle-only preview: 16 KiB head + 4 KiB tail + footer. |
| 1442 | assert!(success.content.len() < 21 * 1024); |
| 1443 | let success_meta = success.metadata.as_ref().unwrap(); |
| 1444 | let failure_meta = failure.metadata.as_ref().unwrap(); |
| 1445 | assert_ne!( |
| 1446 | success_meta["artifact_digest"], |
| 1447 | failure_meta["artifact_digest"] |
| 1448 | ); |
| 1449 | assert_eq!(success_meta["artifact_session_id"], "session-a"); |
| 1450 | assert_eq!(failure_meta["artifact_session_id"], "session-a"); |
| 1451 | |
| 1452 | let mut replay = ToolResult::success(success_raw); |
| 1453 | let replay_path = |
| 1454 | adaptive_spillover(&mut replay, "call-success", "exec_shell", "session-a") |
| 1455 | .expect("idempotent replay"); |
| 1456 | assert_eq!(replay_path, success_path); |
| 1457 | }); |
| 1458 | } |
| 1459 | |
| 1460 | #[test] |
| 1461 | fn adaptive_evidence_publication_failure_emits_no_handle_or_details_hint() { |
| 1462 | let _g = setup(); |
| 1463 | let tmp = tempdir().unwrap(); |
| 1464 | with_test_home(tmp.path(), || { |
| 1465 | let session_dir = tmp |
| 1466 | .path() |
| 1467 | .join(".codewhale") |
| 1468 | .join("sessions") |
| 1469 | .join("session-blocked"); |
| 1470 | std::fs::create_dir_all(&session_dir).unwrap(); |
| 1471 | std::fs::write(session_dir.join("artifacts"), b"block artifact directory").unwrap(); |
| 1472 | |
| 1473 | let raw = format!( |
| 1474 | "{}{}{}", |
| 1475 | "publication failure head\n".repeat(1_500), |
| 1476 | "DEEP_FAILURE_SENTINEL", |
| 1477 | "publication failure tail\n".repeat(1_500), |
| 1478 | ); |
| 1479 | let mut result = ToolResult::error(raw.clone()); |
| 1480 | let path = adaptive_spillover( |
| 1481 | &mut result, |
| 1482 | "call-failed-publish", |
| 1483 | "mcp_fixture", |
| 1484 | "session-blocked", |
| 1485 | ); |
| 1486 | |
| 1487 | assert!(path.is_none()); |
| 1488 | assert_eq!(result.content, raw); |
| 1489 | assert!(!result.content.contains(SPILLOVER_PREVIEW_HINT)); |
| 1490 | assert!(!result.content.contains("retrieve_tool_result")); |
| 1491 | assert!( |
| 1492 | result |
| 1493 | .metadata |
| 1494 | .as_ref() |
| 1495 | .and_then(|metadata| metadata.get("evidence_available")) |
| 1496 | .is_none() |
| 1497 | ); |
| 1498 | assert!( |
| 1499 | !session_dir |
| 1500 | .join("artifacts/art_call-failed-publish.txt") |
| 1501 | .exists() |
| 1502 | ); |
| 1503 | }); |
| 1504 | } |
| 1505 | |
| 1506 | #[test] |
| 1507 | fn adaptive_evidence_metadata_atomic_failure_leaves_payload_unadvertised() { |
| 1508 | let _g = setup(); |
| 1509 | let tmp = tempdir().unwrap(); |
| 1510 | with_test_home(tmp.path(), || { |
| 1511 | let artifact_dir = tmp |
| 1512 | .path() |
| 1513 | .join(".codewhale") |
| 1514 | .join("sessions") |
| 1515 | .join("session-metadata-blocked") |
| 1516 | .join("artifacts"); |
| 1517 | std::fs::create_dir_all(artifact_dir.join("art_call-failed-metadata.evidence.json")) |
| 1518 | .unwrap(); |
| 1519 | |
| 1520 | let raw = format!( |
| 1521 | "{}{}{}", |
| 1522 | "metadata failure head\n".repeat(1_500), |
| 1523 | "DEEP_METADATA_FAILURE_SENTINEL", |
| 1524 | "metadata failure tail\n".repeat(1_500), |
| 1525 | ); |
| 1526 | let mut result = ToolResult::success(raw.clone()); |
| 1527 | let path = adaptive_spillover( |
| 1528 | &mut result, |
| 1529 | "call-failed-metadata", |
| 1530 | "exec_shell", |
| 1531 | "session-metadata-blocked", |
| 1532 | ); |
| 1533 | |
| 1534 | assert!(path.is_none()); |
| 1535 | assert_eq!(result.content, raw); |
| 1536 | assert!(!result.content.contains(SPILLOVER_PREVIEW_HINT)); |
| 1537 | assert!(!result.content.contains("retrieve_tool_result")); |
| 1538 | assert!( |
| 1539 | !artifact_dir.join("art_call-failed-metadata.txt").exists(), |
| 1540 | "metadata failure must leave no payload behind a guessable handle" |
| 1541 | ); |
| 1542 | }); |
| 1543 | } |
| 1544 | |
| 1545 | #[test] |
| 1546 | fn registry_results_spill_to_artifacts_like_every_other_tool() { |
| 1547 | // The Registry bypass is gone: an oversized payload (which today can |
| 1548 | // only be a bug, since the tool caps model-visible matches at eight) |
| 1549 | // takes the same artifact path as any other tool result. |
| 1550 | let _g = setup(); |
| 1551 | let tmp = tempdir().unwrap(); |
| 1552 | with_test_home(tmp.path(), || { |
| 1553 | let original = "registry-entry\n".repeat(10_000); |
| 1554 | assert!(original.len() > SPILLOVER_THRESHOLD_BYTES); |
| 1555 | let mut result = ToolResult::success(original); |
| 1556 | |
| 1557 | let path = apply_spillover_with_artifact( |
| 1558 | &mut result, |
| 1559 | "call-registry", |
| 1560 | "registry_sync", |
| 1561 | "session-registry", |
| 1562 | ); |
| 1563 | |
| 1564 | assert!(path.is_some(), "oversized registry payload must spill"); |
| 1565 | }); |
| 1566 | } |
| 1567 | |
| 1568 | #[test] |
| 1569 | fn apply_spillover_preserves_existing_metadata() { |
| 1570 | let _g = setup(); |
| 1571 | let tmp = tempdir().unwrap(); |
| 1572 | with_test_home(tmp.path(), || { |
| 1573 | let big = "Y".repeat(200 * 1024); |
| 1574 | let mut result = ToolResult::success(big) |
| 1575 | .with_metadata(serde_json::json!({"prior_key": "prior_value"})); |
| 1576 | let path = apply_spillover(&mut result, "call-meta").expect("should spill"); |
| 1577 | |
| 1578 | let metadata = result.metadata.expect("metadata present"); |
| 1579 | // Prior keys survive. |
| 1580 | assert_eq!( |
| 1581 | metadata |
| 1582 | .get("prior_key") |
| 1583 | .and_then(serde_json::Value::as_str), |
| 1584 | Some("prior_value") |
| 1585 | ); |
| 1586 | // New key added alongside. |
| 1587 | assert_eq!( |
| 1588 | metadata |
| 1589 | .get("spillover_path") |
| 1590 | .and_then(serde_json::Value::as_str), |
| 1591 | Some(path.display().to_string().as_str()) |
| 1592 | ); |
| 1593 | }); |
| 1594 | } |
| 1595 | |
| 1596 | #[test] |
| 1597 | fn apply_spillover_wraps_non_object_metadata_under_prior_key() { |
| 1598 | // Defends against a tool whose `metadata` is something |
| 1599 | // other than a JSON object (rare — most use the `json!({})` |
| 1600 | // pattern — but legal per `serde_json::Value`). The |
| 1601 | // spillover writer must add `spillover_path` without losing |
| 1602 | // the prior payload. |
| 1603 | let _g = setup(); |
| 1604 | let tmp = tempdir().unwrap(); |
| 1605 | with_test_home(tmp.path(), || { |
| 1606 | let big = "Z".repeat(200 * 1024); |
| 1607 | let mut result = ToolResult::success(big).with_metadata(serde_json::json!([ |
| 1608 | "unexpected", |
| 1609 | "array", |
| 1610 | "payload" |
| 1611 | ])); |
| 1612 | let path = apply_spillover(&mut result, "call-arr").expect("should spill"); |
| 1613 | |
| 1614 | let metadata = result.metadata.expect("metadata stamped"); |
| 1615 | // Prior payload re-homed under `_prior`. |
| 1616 | let prior = metadata.get("_prior").expect("_prior wrap key present"); |
| 1617 | assert_eq!( |
| 1618 | prior, |
| 1619 | &serde_json::json!(["unexpected", "array", "payload"]), |
| 1620 | "prior array should round-trip under _prior" |
| 1621 | ); |
| 1622 | // New key alongside. |
| 1623 | assert_eq!( |
| 1624 | metadata |
| 1625 | .get("spillover_path") |
| 1626 | .and_then(serde_json::Value::as_str), |
| 1627 | Some(path.display().to_string().as_str()) |
| 1628 | ); |
| 1629 | }); |
| 1630 | } |
| 1631 | |
| 1632 | // ── Honest-truncation regressions (v0.9.4) ───────────────────────────── |
| 1633 | |
| 1634 | #[test] |
| 1635 | fn truncated_preview_returns_content_unchanged_when_nothing_omitted() { |
| 1636 | let original = "line one\nline two\nline three\n"; |
| 1637 | let preview = truncated_preview(original, "", original, "/tmp/artifact.txt", None); |
| 1638 | assert_eq!(preview, original); |
| 1639 | assert!( |
| 1640 | !preview.contains("of output omitted"), |
| 1641 | "must never claim a truncation that did not happen" |
| 1642 | ); |
| 1643 | } |
| 1644 | |
| 1645 | #[test] |
| 1646 | fn head_tail_windows_never_overlap() { |
| 1647 | // Content smaller than head + tail budgets: the tail window shrinks |
| 1648 | // so it starts exactly where the head ends — no byte appears twice. |
| 1649 | let content = "x".repeat(10_000); |
| 1650 | let (head, tail) = head_tail_windows(&content, 8 * 1024, 4 * 1024); |
| 1651 | assert_eq!(head.len(), 8 * 1024); |
| 1652 | assert_eq!(tail.len(), 10_000 - 8 * 1024); |
| 1653 | assert!(head.len() + tail.len() <= content.len()); |
| 1654 | |
| 1655 | // Content larger than both budgets: full windows, exact omission. |
| 1656 | let big = "y".repeat(100_000); |
| 1657 | let (head, tail) = head_tail_windows(&big, 32 * 1024, 8 * 1024); |
| 1658 | assert_eq!(head.len(), 32 * 1024); |
| 1659 | assert_eq!(tail.len(), 8 * 1024); |
| 1660 | |
| 1661 | // UTF-8 codepoints are never split at either window edge. |
| 1662 | let emoji = "🐳".repeat(5_000); // 20_000 bytes, 4 per codepoint |
| 1663 | let (head, tail) = head_tail_windows(&emoji, 8 * 1024 + 1, 4 * 1024 + 2); |
| 1664 | assert!(emoji.is_char_boundary(head.len())); |
| 1665 | assert!(emoji.is_char_boundary(emoji.len() - tail.len())); |
| 1666 | assert!(head.len() + tail.len() <= emoji.len()); |
| 1667 | } |
| 1668 | |
| 1669 | #[test] |
| 1670 | fn adaptive_evidence_passes_through_when_preview_budget_covers_output() { |
| 1671 | let _g = setup(); |
| 1672 | let tmp = tempdir().unwrap(); |
| 1673 | with_test_home(tmp.path(), || { |
| 1674 | // 30_000 bytes → 10_000 estimated tokens → Hybrid band under the |
| 1675 | // 32_768-token default, but the 32 KiB + 8 KiB preview budget |
| 1676 | // covers the whole output, so nothing is actually omitted. |
| 1677 | let raw = "mid\n".repeat(7_500); |
| 1678 | assert_eq!(raw.len(), 30_000); |
| 1679 | let mut result = ToolResult::success(raw.clone()); |
| 1680 | let path = |
| 1681 | adaptive_spillover(&mut result, "call-covered", "exec_shell", "session-covered"); |
| 1682 | assert!(path.is_none(), "no artifact when nothing is omitted"); |
| 1683 | assert_eq!(result.content, raw); |
| 1684 | assert!(!result.content.contains("of output omitted")); |
| 1685 | assert!( |
| 1686 | !tmp.path() |
| 1687 | .join(".codewhale/sessions/session-covered/artifacts/art_call-covered.txt") |
| 1688 | .exists() |
| 1689 | ); |
| 1690 | }); |
| 1691 | } |
| 1692 | |
| 1693 | #[test] |
| 1694 | fn adaptive_evidence_footer_names_artifact_path_and_recovery() { |
| 1695 | let _g = setup(); |
| 1696 | let tmp = tempdir().unwrap(); |
| 1697 | with_test_home(tmp.path(), || { |
| 1698 | // 120_000 bytes → 40_000 estimated tokens → handle-only band. |
| 1699 | let raw = "entry\n".repeat(20_000); |
| 1700 | assert_eq!(raw.len(), 120_000); |
| 1701 | let mut result = ToolResult::success(raw); |
| 1702 | let path = |
| 1703 | adaptive_spillover(&mut result, "call-honest", "exec_shell", "session-honest") |
| 1704 | .expect("should spill"); |
| 1705 | |
| 1706 | // Footer: omitted size + line count, artifact path, recovery line. |
| 1707 | assert!(result.content.contains("of output omitted (")); |
| 1708 | assert!(result.content.contains(" lines)")); |
| 1709 | assert!(result.content.contains("full output at")); |
| 1710 | assert!( |
| 1711 | result |
| 1712 | .content |
| 1713 | .contains(&crate::artifacts::format_artifact_relative_path(&path)) |
| 1714 | ); |
| 1715 | assert!(result.content.contains(SPILLOVER_RECOVERY_HINT)); |
| 1716 | assert!(!result.content.contains(SPILLOVER_PREVIEW_HINT)); |
| 1717 | |
| 1718 | // Head and tail do not overlap: 16 KiB + 4 KiB handle-only |
| 1719 | // windows over a 120_000-byte output. |
| 1720 | let metadata = result.metadata.expect("metadata stamped"); |
| 1721 | assert_eq!(metadata["retained_head_bytes"], 16 * 1024); |
| 1722 | assert_eq!(metadata["retained_tail_bytes"], 4 * 1024); |
| 1723 | }); |
| 1724 | } |
| 1725 | |
| 1726 | #[test] |
| 1727 | fn apply_spillover_with_artifact_defaults_to_classic_head_tail_spillover() { |
| 1728 | // 120_000 bytes exceeds the 100 KiB classic threshold. Without the |
| 1729 | // `CODEWHALE_ADAPTIVE_OUTPUT_ROUTING` opt-in the public entry point |
| 1730 | // keeps a 32 KiB head + 8 KiB tail and writes the session artifact — |
| 1731 | // the adaptive handle-only windows and evidence metadata are opt-in. |
| 1732 | let _g = setup(); |
| 1733 | let tmp = tempdir().unwrap(); |
| 1734 | with_test_home(tmp.path(), || { |
| 1735 | let raw = "entry\n".repeat(20_000); |
| 1736 | assert_eq!(raw.len(), 120_000); |
| 1737 | let mut result = ToolResult::success(raw.clone()); |
| 1738 | let path = apply_spillover_with_artifact( |
| 1739 | &mut result, |
| 1740 | "call-classic", |
| 1741 | "exec_shell", |
| 1742 | "session-classic", |
| 1743 | ) |
| 1744 | .expect("should spill"); |
| 1745 | |
| 1746 | let metadata = result.metadata.expect("metadata stamped"); |
| 1747 | assert_eq!(metadata["retained_head_bytes"], 32 * 1024); |
| 1748 | assert_eq!(metadata["retained_tail_bytes"], 8 * 1024); |
| 1749 | assert!(result.content.contains(SPILLOVER_RECOVERY_HINT)); |
| 1750 | assert!( |
| 1751 | !tmp.path() |
| 1752 | .join( |
| 1753 | ".codewhale/sessions/session-classic/artifacts/art_call-classic.evidence.json" |
| 1754 | ) |
| 1755 | .exists(), |
| 1756 | "classic lane publishes no adaptive evidence metadata" |
| 1757 | ); |
| 1758 | assert_eq!(std::fs::read_to_string(&path).unwrap(), raw); |
| 1759 | }); |
| 1760 | } |
| 1761 | } |
| 1762 |