| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "bytes" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "fmt" |
| 8 | "os" |
| 9 | "path/filepath" |
| 10 | "sort" |
| 11 | "strings" |
| 12 | "time" |
| 13 | |
| 14 | "reasonix/internal/fileutil" |
| 15 | "reasonix/internal/provider" |
| 16 | "reasonix/internal/store" |
| 17 | ) |
| 18 | |
| 19 | // Recovery-branch garbage collection. Conflict recovery forks a copy of the |
| 20 | // in-memory transcript whenever a save conflicts (#5993); the triggers are |
| 21 | // fixed, but every fork that ever happened still sits in the session list |
| 22 | // until the user trashes it by hand. Most of them preserve nothing: the |
| 23 | // original session went on to contain everything the fork saved. Those — and |
| 24 | // only those — are safe to reclaim automatically. |
| 25 | |
| 26 | // RecoveryGCGracePeriod is how long a reclaimable recovery branch must sit |
| 27 | // idle before the periodic GC may collect it. A fresh fork is part of an |
| 28 | // active conflict flow — the user may be comparing it against the original. |
| 29 | const RecoveryGCGracePeriod = 24 * time.Hour |
| 30 | |
| 31 | // RecoveryGCStartupGracePeriod is used for the first post-restore sweep so |
| 32 | // upgrade storms of covered copies are cleared within minutes rather than a |
| 33 | // full day, while still protecting a conflict the user is actively inspecting. |
| 34 | const RecoveryGCStartupGracePeriod = 15 * time.Minute |
| 35 | |
| 36 | const ( |
| 37 | recoveryTrashDir = ".trash" |
| 38 | recoveryTrashMetaFile = ".trash-meta.json" |
| 39 | recoveryTrashOperationPrefix = "recovery-trash:" |
| 40 | recoveryTrashPendingFile = ".recovery-trash-pending.json" |
| 41 | recoveryTrashStagingPrefix = ".recovery-trash-staging-" |
| 42 | ) |
| 43 | |
| 44 | // ErrRecoveryBranchNotCovered means the branch cannot currently be proven |
| 45 | // redundant with its parent. Destructive callers must preserve it. |
| 46 | var ErrRecoveryBranchNotCovered = errors.New("recovery branch is not covered by its parent") |
| 47 | |
| 48 | // ErrRecoveryBranchNotIdle means the branch has not yet passed the safety |
| 49 | // grace period. It remains visible and may be retried by a later GC pass. |
| 50 | var ErrRecoveryBranchNotIdle = errors.New("recovery branch is still inside its safety grace period") |
| 51 | |
| 52 | type recoveryTrashMeta struct { |
| 53 | Key string `json:"key"` |
| 54 | DeletedAt int64 `json:"deletedAt"` |
| 55 | } |
| 56 | |
| 57 | type recoveryTrashPendingMeta struct { |
| 58 | Key string `json:"key"` |
| 59 | } |
| 60 | |
| 61 | // SessionLeaseHeld reports whether ANY live runtime — this process included — |
| 62 | // holds the session's write lease. SessionLeaseHeldByOtherRuntime deliberately |
| 63 | // answers false for the current process; GC needs the stricter question, since |
| 64 | // a branch open in one of our own tabs is just as much in use. |
| 65 | func SessionLeaseHeld(path string) bool { |
| 66 | if strings.TrimSpace(path) == "" { |
| 67 | return false |
| 68 | } |
| 69 | if _, ok := sessionLeaseOwners.Load(CanonicalSessionPath(path)); ok { |
| 70 | return true |
| 71 | } |
| 72 | return SessionLeaseHeldByOtherRuntime(path) |
| 73 | } |
| 74 | |
| 75 | // RecoveryBranchCoveredByParent reports whether a conflict-recovery branch |
| 76 | // preserves no content that is absent from its parent. It deliberately reads |
| 77 | // both transcripts instead of trusting listing sidecars: stale metadata must |
| 78 | // never authorize hiding, migration skipping, bulk trash, or permanent purge. |
| 79 | // Missing/corrupt metadata, a changed branch, or a missing/diverged parent are |
| 80 | // all treated conservatively as not covered. |
| 81 | func RecoveryBranchCoveredByParent(path, parentDir string) bool { |
| 82 | meta, ok, err := LoadBranchMeta(path) |
| 83 | if err != nil || !ok || !meta.Recovered || strings.TrimSpace(meta.RecoveryDigest) == "" { |
| 84 | return false |
| 85 | } |
| 86 | return recoveryBranchCoveredByParent(path, parentDir, meta) |
| 87 | } |
| 88 | |
| 89 | // SessionContentCovers reports whether covering contains the complete message |
| 90 | // history stored in covered. It is intentionally conservative for catalog |
| 91 | // canonical promotion: repaired or damaged loads cannot authorize a redirect. |
| 92 | func SessionContentCovers(coveringPath, coveredPath string) bool { |
| 93 | covering, ok := LoadSessionContentSnapshot(coveringPath) |
| 94 | if !ok { |
| 95 | return false |
| 96 | } |
| 97 | covered, ok := LoadSessionContentSnapshot(coveredPath) |
| 98 | return ok && covering.Covers(covered) |
| 99 | } |
| 100 | |
| 101 | // RecoveryPreferenceCurrent proves that the branch still has the exact content |
| 102 | // the user selected. Continued or externally edited branches fall back to an |
| 103 | // unresolved lineage until the user chooses again. |
| 104 | func RecoveryPreferenceCurrent(path string, meta BranchMeta) bool { |
| 105 | if !meta.RecoveryPreferred || strings.TrimSpace(meta.RecoveryPreferredDigest) == "" { |
| 106 | return false |
| 107 | } |
| 108 | session, err := LoadSession(path) |
| 109 | if err != nil || session == nil || session.normalizedDirty || session.eventLogDamaged { |
| 110 | return false |
| 111 | } |
| 112 | digest, err := digestSessionMessages(session.Snapshot()) |
| 113 | return err == nil && digestString(digest) == strings.TrimSpace(meta.RecoveryPreferredDigest) |
| 114 | } |
| 115 | |
| 116 | // SessionContentSnapshot is an immutable, validated transcript projection used |
| 117 | // by the catalog. Its messages stay private so callers cannot accidentally use |
| 118 | // a classification read as a mutable Session. |
| 119 | type SessionContentSnapshot struct { |
| 120 | messages []provider.Message |
| 121 | digest string |
| 122 | } |
| 123 | |
| 124 | // LoadSessionContentSnapshot loads one transcript once for lineage analysis. |
| 125 | // Dirty normalization and damaged event logs fail closed: neither may prove a |
| 126 | // canonical branch or authorize cleanup. |
| 127 | func LoadSessionContentSnapshot(path string) (SessionContentSnapshot, bool) { |
| 128 | session, err := LoadSession(path) |
| 129 | if err != nil || session == nil || session.normalizedDirty || session.eventLogDamaged { |
| 130 | return SessionContentSnapshot{}, false |
| 131 | } |
| 132 | messages := session.Snapshot() |
| 133 | digest, err := digestSessionMessages(messages) |
| 134 | if err != nil { |
| 135 | return SessionContentSnapshot{}, false |
| 136 | } |
| 137 | return SessionContentSnapshot{messages: messages, digest: digestString(digest)}, true |
| 138 | } |
| 139 | |
| 140 | // Len is used only to discard candidates that cannot cover the longest member. |
| 141 | func (s SessionContentSnapshot) Len() int { return len(s.messages) } |
| 142 | |
| 143 | // MatchesDigest binds catalog lineage decisions to the recovery ledger. |
| 144 | func (s SessionContentSnapshot) MatchesDigest(digest string) bool { |
| 145 | return strings.TrimSpace(digest) != "" && s.digest == strings.TrimSpace(digest) |
| 146 | } |
| 147 | |
| 148 | // Covers reports whether s contains all content in covered as a compatible |
| 149 | // prefix. Both snapshots have already passed the conservative load checks. |
| 150 | func (s SessionContentSnapshot) Covers(covered SessionContentSnapshot) bool { |
| 151 | return messagesHavePrefix(s.messages, covered.messages) || |
| 152 | messagesHavePrefixWithCompatibleSystem(s.messages, covered.messages) |
| 153 | } |
| 154 | |
| 155 | // TryAcquireRecoveryParentGuard verifies that a recovery branch is covered by |
| 156 | // its parent while holding the parent's save and lease locks. The caller must |
| 157 | // keep the returned guard until permanent deletion finishes, then Release it. |
| 158 | // If the parent is open or being rewritten, acquisition fails without waiting |
| 159 | // so bulk cleanup preserves the branch and can be retried later. |
| 160 | func TryAcquireRecoveryParentGuard(path, parentDir string) (*SessionRemovalGuard, error) { |
| 161 | meta, ok, err := LoadBranchMeta(path) |
| 162 | if err != nil || !ok || !meta.Recovered || strings.TrimSpace(meta.RecoveryDigest) == "" { |
| 163 | return nil, ErrRecoveryBranchNotCovered |
| 164 | } |
| 165 | parentID := strings.TrimSpace(meta.ParentID) |
| 166 | if parentID == "" { |
| 167 | return nil, ErrRecoveryBranchNotCovered |
| 168 | } |
| 169 | parentDir = strings.TrimSpace(parentDir) |
| 170 | if parentDir == "" { |
| 171 | parentDir = filepath.Dir(path) |
| 172 | } |
| 173 | parentPath := filepath.Join(parentDir, parentID+".jsonl") |
| 174 | if parentPath == path || !IsVisibleSession(parentPath) { |
| 175 | return nil, ErrRecoveryBranchNotCovered |
| 176 | } |
| 177 | guard, err := TryAcquireSessionRemovalGuard(parentPath) |
| 178 | if err != nil { |
| 179 | return nil, err |
| 180 | } |
| 181 | if !recoveryBranchCoveredByParent(path, parentDir, meta) { |
| 182 | guard.Release() |
| 183 | return nil, ErrRecoveryBranchNotCovered |
| 184 | } |
| 185 | return guard, nil |
| 186 | } |
| 187 | |
| 188 | func recoveryBranchCoveredByParent(path, parentDir string, meta BranchMeta) bool { |
| 189 | parentID := strings.TrimSpace(meta.ParentID) |
| 190 | if parentID == "" { |
| 191 | return false |
| 192 | } |
| 193 | branch, err := LoadSession(path) |
| 194 | if err != nil || branch == nil { |
| 195 | return false |
| 196 | } |
| 197 | branchMsgs := branch.Snapshot() |
| 198 | branchDigest, err := digestSessionMessages(branchMsgs) |
| 199 | if err != nil || digestString(branchDigest) != strings.TrimSpace(meta.RecoveryDigest) { |
| 200 | // Continued on (or undigestable): this is someone's conversation now. |
| 201 | return false |
| 202 | } |
| 203 | parentDir = strings.TrimSpace(parentDir) |
| 204 | if parentDir == "" { |
| 205 | parentDir = filepath.Dir(path) |
| 206 | } |
| 207 | parentPath := filepath.Join(parentDir, parentID+".jsonl") |
| 208 | if parentPath == path || !IsVisibleSession(parentPath) { |
| 209 | return false |
| 210 | } |
| 211 | parent, err := LoadSession(parentPath) |
| 212 | if err != nil || parent == nil { |
| 213 | return false |
| 214 | } |
| 215 | parentMsgs := parent.Snapshot() |
| 216 | parentDigest, err := digestSessionMessages(parentMsgs) |
| 217 | if err != nil { |
| 218 | return false |
| 219 | } |
| 220 | return bytes.Equal(parentDigest[:], branchDigest[:]) || |
| 221 | messagesHavePrefix(parentMsgs, branchMsgs) || |
| 222 | messagesHavePrefixWithCompatibleSystem(parentMsgs, branchMsgs) |
| 223 | } |
| 224 | |
| 225 | // ReclaimableRecoveryBranches scans dir for conflict-recovery branches that |
| 226 | // are safe to dispose of. Every condition must hold — when in doubt the branch |
| 227 | // stays, because a recovery branch exists precisely to prevent data loss: |
| 228 | // |
| 229 | // 1. The branch meta says Recovered and records the fork digest. |
| 230 | // 2. The transcript still matches that fork digest: the branch was never |
| 231 | // continued on. A single follow-up turn disqualifies it permanently. |
| 232 | // 3. The parent transcript (meta.ParentID, same directory) exists and covers |
| 233 | // the branch content — equal digest, or the branch is a strict prefix |
| 234 | // (allowing a compatible leading-system swap). These are the same checks |
| 235 | // SaveRecoveryBranch uses to declare a recovery not needed in the first |
| 236 | // place, so "covered" here means the fork preserves nothing unique. |
| 237 | // 4. No live runtime holds the branch's session lease. |
| 238 | // 5. The branch has been idle for at least grace. |
| 239 | // |
| 240 | // It returns candidate paths only; disposal (trash, delete) is caller policy. |
| 241 | func ReclaimableRecoveryBranches(dir string, now time.Time, grace time.Duration) ([]string, error) { |
| 242 | dir = strings.TrimSpace(dir) |
| 243 | if dir == "" { |
| 244 | return nil, nil |
| 245 | } |
| 246 | entries, err := os.ReadDir(dir) |
| 247 | if err != nil { |
| 248 | if os.IsNotExist(err) { |
| 249 | return nil, nil |
| 250 | } |
| 251 | return nil, err |
| 252 | } |
| 253 | var out []string |
| 254 | for _, e := range entries { |
| 255 | if e.IsDir() || !strings.HasSuffix(e.Name(), ".jsonl") || strings.HasSuffix(e.Name(), ".events.jsonl") { |
| 256 | continue |
| 257 | } |
| 258 | path := filepath.Join(dir, e.Name()) |
| 259 | if !IsVisibleSession(path) { |
| 260 | continue |
| 261 | } |
| 262 | meta, ok, err := LoadBranchMeta(path) |
| 263 | if err != nil || !ok || !meta.Recovered || strings.TrimSpace(meta.RecoveryDigest) == "" { |
| 264 | continue |
| 265 | } |
| 266 | if strings.TrimSpace(meta.ParentID) == "" { |
| 267 | continue |
| 268 | } |
| 269 | if !recoveryBranchIdle(path, meta, now, grace) { |
| 270 | continue |
| 271 | } |
| 272 | if SessionLeaseHeld(path) { |
| 273 | continue |
| 274 | } |
| 275 | if !recoveryBranchCoveredByParent(path, dir, meta) { |
| 276 | continue |
| 277 | } |
| 278 | out = append(out, path) |
| 279 | } |
| 280 | return out, nil |
| 281 | } |
| 282 | |
| 283 | // TrashCoveredRecoveryBranch moves a redundant recovery branch into the same |
| 284 | // recoverable .trash layout used by Desktop. This is the explicit/manual cleanup |
| 285 | // path, so it does not require the background GC idle grace period. Parent |
| 286 | // coverage is rechecked while both parent and branch removal guards are held. |
| 287 | func TrashCoveredRecoveryBranch(path, parentDir string) error { |
| 288 | return trashCoveredRecoveryBranch(path, parentDir, false) |
| 289 | } |
| 290 | |
| 291 | // TrashRecoveryBranchCoveredBy moves path to recoverable trash when canonical |
| 292 | // contains its complete transcript. Unlike TrashCoveredRecoveryBranch this is |
| 293 | // lineage-aware: legacy recovery storms form chains, so an adopted leaf may |
| 294 | // cover an ancestor even though that ancestor's immediate parent does not cover |
| 295 | // it. Both transcripts are held behind removal guards while coverage is proved. |
| 296 | func TrashRecoveryBranchCoveredBy(path, canonicalPath, parentDir string) error { |
| 297 | path = filepath.Clean(strings.TrimSpace(path)) |
| 298 | canonicalPath = filepath.Clean(strings.TrimSpace(canonicalPath)) |
| 299 | parentDir = filepath.Clean(strings.TrimSpace(parentDir)) |
| 300 | if path == "." || canonicalPath == "." || parentDir == "." || path == canonicalPath || |
| 301 | filepath.Dir(path) != parentDir || filepath.Dir(canonicalPath) != parentDir { |
| 302 | return ErrRecoveryBranchNotCovered |
| 303 | } |
| 304 | meta, ok, err := LoadBranchMeta(path) |
| 305 | if err != nil || !ok || !meta.Recovered { |
| 306 | return ErrRecoveryBranchNotCovered |
| 307 | } |
| 308 | paths := []string{path, canonicalPath} |
| 309 | sort.Strings(paths) |
| 310 | guards := make(map[string]*SessionRemovalGuard, len(paths)) |
| 311 | for _, guardedPath := range paths { |
| 312 | guard, guardErr := TryAcquireSessionRemovalGuard(guardedPath) |
| 313 | if guardErr != nil { |
| 314 | for _, held := range guards { |
| 315 | held.Release() |
| 316 | } |
| 317 | return guardErr |
| 318 | } |
| 319 | guards[guardedPath] = guard |
| 320 | } |
| 321 | defer func() { |
| 322 | for _, guard := range guards { |
| 323 | guard.Release() |
| 324 | } |
| 325 | }() |
| 326 | if !SessionContentCovers(canonicalPath, path) { |
| 327 | return ErrRecoveryBranchNotCovered |
| 328 | } |
| 329 | key := filepath.Base(path) |
| 330 | if !validRecoveryTrashKey(key) { |
| 331 | return fmt.Errorf("invalid recovery session path") |
| 332 | } |
| 333 | stageDir, err := reserveRecoveryTrashStage(parentDir) |
| 334 | if err != nil { |
| 335 | return err |
| 336 | } |
| 337 | if err := prepareRecoveryTrashStage(path, key, stageDir); err != nil { |
| 338 | return err |
| 339 | } |
| 340 | return finishRecoveryTrashStage(parentDir, path, key, stageDir, guards[path]) |
| 341 | } |
| 342 | |
| 343 | // ReparentRecoveryCanonical shortens a proved recovery chain to root -> |
| 344 | // canonical without rewriting either transcript. This preserves discovery for |
| 345 | // older versions after covered intermediate branches are moved to trash. |
| 346 | func ReparentRecoveryCanonical(canonicalPath, rootID, parentDir string) error { |
| 347 | canonicalPath = filepath.Clean(strings.TrimSpace(canonicalPath)) |
| 348 | parentDir = filepath.Clean(strings.TrimSpace(parentDir)) |
| 349 | rootID = strings.TrimSpace(rootID) |
| 350 | rootPath := filepath.Join(parentDir, rootID+".jsonl") |
| 351 | if canonicalPath == "." || parentDir == "." || rootID == "" || filepath.Base(rootID) != rootID || |
| 352 | filepath.Dir(canonicalPath) != parentDir || canonicalPath == rootPath { |
| 353 | return ErrRecoveryBranchNotCovered |
| 354 | } |
| 355 | paths := []string{canonicalPath, rootPath} |
| 356 | sort.Strings(paths) |
| 357 | guards := make([]*SessionRemovalGuard, 0, len(paths)) |
| 358 | for _, path := range paths { |
| 359 | guard, err := TryAcquireSessionRemovalGuard(path) |
| 360 | if err != nil { |
| 361 | for _, held := range guards { |
| 362 | held.Release() |
| 363 | } |
| 364 | return err |
| 365 | } |
| 366 | guards = append(guards, guard) |
| 367 | } |
| 368 | defer func() { |
| 369 | for _, guard := range guards { |
| 370 | guard.Release() |
| 371 | } |
| 372 | }() |
| 373 | if !SessionContentCovers(canonicalPath, rootPath) { |
| 374 | return ErrRecoveryBranchNotCovered |
| 375 | } |
| 376 | return UpdateBranchMeta(canonicalPath, false, func(meta *BranchMeta) error { |
| 377 | if !meta.Recovered { |
| 378 | return ErrRecoveryBranchNotCovered |
| 379 | } |
| 380 | meta.ParentID = rootID |
| 381 | meta.RecoveryDepth = 1 |
| 382 | return nil |
| 383 | }) |
| 384 | } |
| 385 | |
| 386 | // TrashReclaimableRecoveryBranch is the background-GC variant. In addition to |
| 387 | // the same atomic coverage proof, it requires the branch to remain idle for the |
| 388 | // full grace period. |
| 389 | func TrashReclaimableRecoveryBranch(path, parentDir string) error { |
| 390 | return trashCoveredRecoveryBranch(path, parentDir, true) |
| 391 | } |
| 392 | |
| 393 | func trashCoveredRecoveryBranch(path, parentDir string, requireIdle bool) error { |
| 394 | path = filepath.Clean(strings.TrimSpace(path)) |
| 395 | parentDir = filepath.Clean(strings.TrimSpace(parentDir)) |
| 396 | if path == "." || parentDir == "." || filepath.Dir(path) != parentDir { |
| 397 | return fmt.Errorf("recovery branch must be a direct child of its session directory") |
| 398 | } |
| 399 | key := filepath.Base(path) |
| 400 | if !strings.HasSuffix(key, ".jsonl") || strings.HasSuffix(key, ".events.jsonl") { |
| 401 | return fmt.Errorf("invalid recovery session path") |
| 402 | } |
| 403 | |
| 404 | parentGuard, err := TryAcquireRecoveryParentGuard(path, parentDir) |
| 405 | if err != nil { |
| 406 | return err |
| 407 | } |
| 408 | defer parentGuard.Release() |
| 409 | |
| 410 | branchGuard, err := TryAcquireSessionRemovalGuard(path) |
| 411 | if err != nil { |
| 412 | return err |
| 413 | } |
| 414 | defer branchGuard.Release() |
| 415 | if requireIdle { |
| 416 | meta, ok, err := LoadBranchMeta(path) |
| 417 | if err != nil || !ok || !recoveryBranchIdle(path, meta, time.Now(), RecoveryGCGracePeriod) { |
| 418 | return ErrRecoveryBranchNotIdle |
| 419 | } |
| 420 | } |
| 421 | if !RecoveryBranchCoveredByParent(path, parentDir) { |
| 422 | return ErrRecoveryBranchNotCovered |
| 423 | } |
| 424 | |
| 425 | stageDir, err := reserveRecoveryTrashStage(parentDir) |
| 426 | if err != nil { |
| 427 | return err |
| 428 | } |
| 429 | // Keep the move invisible until every artifact is staged. Older Reasonix |
| 430 | // versions ignore the non-session staging directory, while new versions can |
| 431 | // finish it from the durable in-directory marker after a crash. Publishing is |
| 432 | // one same-filesystem rename, so Desktop can never restore or purge a split |
| 433 | // transcript/sidecar set. |
| 434 | if err := prepareRecoveryTrashStage(path, key, stageDir); err != nil { |
| 435 | return err |
| 436 | } |
| 437 | return finishRecoveryTrashStage(parentDir, path, key, stageDir, branchGuard) |
| 438 | } |
| 439 | |
| 440 | func recoveryBranchIdle(path string, meta BranchMeta, now time.Time, grace time.Duration) bool { |
| 441 | idleSince := meta.UpdatedAt |
| 442 | if idleSince.IsZero() { |
| 443 | info, err := os.Stat(path) |
| 444 | if err != nil { |
| 445 | return false |
| 446 | } |
| 447 | idleSince = info.ModTime() |
| 448 | } |
| 449 | return now.Sub(idleSince) >= grace |
| 450 | } |
| 451 | |
| 452 | // reconcileRecoveryTrashPending completes an interrupted move left by the |
| 453 | // pre-staging recovery-trash protocol. Keep this compatibility path so users |
| 454 | // upgrading from an intermediate build do not strand its typed marker. |
| 455 | func reconcileRecoveryTrashPending(item CleanupPendingInfo) (bool, error) { |
| 456 | operation := strings.TrimSpace(item.Meta.Operation) |
| 457 | if !strings.HasPrefix(operation, recoveryTrashOperationPrefix) { |
| 458 | return false, nil |
| 459 | } |
| 460 | itemName := strings.TrimPrefix(operation, recoveryTrashOperationPrefix) |
| 461 | if itemName == "" || filepath.Base(itemName) != itemName || itemName == "." || itemName == ".." { |
| 462 | return true, fmt.Errorf("invalid recovery trash target") |
| 463 | } |
| 464 | path := filepath.Clean(item.SessionPath) |
| 465 | dir := filepath.Dir(path) |
| 466 | key := filepath.Base(path) |
| 467 | guard, err := TryAcquireSessionRemovalGuard(path) |
| 468 | if err != nil { |
| 469 | return true, err |
| 470 | } |
| 471 | defer guard.Release() |
| 472 | return true, finishRecoveryTrashMove(dir, path, key, filepath.Join(dir, recoveryTrashDir, itemName), guard) |
| 473 | } |
| 474 | |
| 475 | // reconcileRecoveryTrashStages completes the atomic staging protocol used by |
| 476 | // new runtimes. A staging directory is intentionally not a valid Desktop trash |
| 477 | // item: it has neither a session-shaped directory name nor .trash-meta.json. |
| 478 | // Once complete, the whole directory is renamed into place atomically. |
| 479 | func reconcileRecoveryTrashStages(dir string) error { |
| 480 | dir = strings.TrimSpace(dir) |
| 481 | if dir == "" { |
| 482 | return nil |
| 483 | } |
| 484 | root := filepath.Join(dir, recoveryTrashDir) |
| 485 | entries, err := os.ReadDir(root) |
| 486 | if err != nil { |
| 487 | if os.IsNotExist(err) { |
| 488 | return nil |
| 489 | } |
| 490 | return err |
| 491 | } |
| 492 | var errs []error |
| 493 | for _, entry := range entries { |
| 494 | if !entry.IsDir() { |
| 495 | continue |
| 496 | } |
| 497 | itemDir := filepath.Join(root, entry.Name()) |
| 498 | if _, err := os.Stat(filepath.Join(itemDir, recoveryTrashPendingFile)); err != nil { |
| 499 | if os.IsNotExist(err) { |
| 500 | continue |
| 501 | } |
| 502 | errs = append(errs, fmt.Errorf("inspect recovery trash stage %s: %w", itemDir, err)) |
| 503 | continue |
| 504 | } |
| 505 | if err := reconcileRecoveryTrashStage(dir, itemDir); err != nil { |
| 506 | errs = append(errs, fmt.Errorf("reconcile recovery trash stage %s: %w", itemDir, err)) |
| 507 | } |
| 508 | } |
| 509 | return errors.Join(errs...) |
| 510 | } |
| 511 | |
| 512 | func reconcileRecoveryTrashStage(dir, itemDir string) error { |
| 513 | pending, err := readRecoveryTrashPending(itemDir) |
| 514 | if err != nil { |
| 515 | return err |
| 516 | } |
| 517 | key := strings.TrimSpace(pending.Key) |
| 518 | if !validRecoveryTrashKey(key) { |
| 519 | return fmt.Errorf("invalid recovery trash key") |
| 520 | } |
| 521 | stagedPath := filepath.Join(itemDir, key) |
| 522 | staged, err := regularRecoveryTrashPath(stagedPath) |
| 523 | if err != nil { |
| 524 | return err |
| 525 | } |
| 526 | |
| 527 | // A non-staging name means the atomic directory rename already succeeded; |
| 528 | // only visibility metadata/final marker cleanup may remain. |
| 529 | if !strings.HasPrefix(filepath.Base(itemDir), recoveryTrashStagingPrefix) { |
| 530 | if !staged { |
| 531 | return fmt.Errorf("published recovery trash item is missing transcript") |
| 532 | } |
| 533 | if _, err := os.Stat(filepath.Join(itemDir, recoveryTrashMetaFile)); err == nil { |
| 534 | return clearRecoveryTrashPending(itemDir) |
| 535 | } else if !os.IsNotExist(err) { |
| 536 | return err |
| 537 | } |
| 538 | if err := writeRecoveryTrashMetaExisting(itemDir, key); err != nil { |
| 539 | if os.IsNotExist(err) { |
| 540 | return nil // restored or purged after complete publication |
| 541 | } |
| 542 | return err |
| 543 | } |
| 544 | return clearRecoveryTrashPending(itemDir) |
| 545 | } |
| 546 | |
| 547 | livePath := filepath.Join(dir, key) |
| 548 | live, err := regularRecoveryTrashPath(livePath) |
| 549 | if err != nil { |
| 550 | return err |
| 551 | } |
| 552 | switch { |
| 553 | case live && !staged: |
| 554 | // The durable marker landed but the first rename did not. No session |
| 555 | // artifact has moved, so discard the empty stage and let a later GC pass |
| 556 | // revalidate coverage before trying again. |
| 557 | return removeEmptyRecoveryTrashStage(itemDir) |
| 558 | case live && staged: |
| 559 | return fmt.Errorf("live and staged recovery transcripts both exist") |
| 560 | case !live && !staged: |
| 561 | return fmt.Errorf("recovery trash stage is missing transcript") |
| 562 | } |
| 563 | |
| 564 | guard, err := TryAcquireSessionRemovalGuard(livePath) |
| 565 | if err != nil { |
| 566 | return err |
| 567 | } |
| 568 | defer guard.Release() |
| 569 | return finishRecoveryTrashStage(dir, livePath, key, itemDir, guard) |
| 570 | } |
| 571 | |
| 572 | func regularRecoveryTrashPath(path string) (bool, error) { |
| 573 | info, err := os.Lstat(path) |
| 574 | if os.IsNotExist(err) { |
| 575 | return false, nil |
| 576 | } |
| 577 | if err != nil { |
| 578 | return false, err |
| 579 | } |
| 580 | if !info.Mode().IsRegular() { |
| 581 | return false, fmt.Errorf("recovery trash path is not a regular file: %s", path) |
| 582 | } |
| 583 | return true, nil |
| 584 | } |
| 585 | |
| 586 | func removeEmptyRecoveryTrashStage(itemDir string) error { |
| 587 | entries, err := os.ReadDir(itemDir) |
| 588 | if err != nil { |
| 589 | return err |
| 590 | } |
| 591 | if len(entries) != 1 || entries[0].Name() != recoveryTrashPendingFile || entries[0].IsDir() { |
| 592 | return fmt.Errorf("recovery trash stage contains artifacts without a transcript") |
| 593 | } |
| 594 | return os.RemoveAll(itemDir) |
| 595 | } |
| 596 | |
| 597 | func reserveRecoveryTrashStage(dir string) (string, error) { |
| 598 | root := filepath.Join(dir, recoveryTrashDir) |
| 599 | if err := os.MkdirAll(root, 0o755); err != nil { |
| 600 | return "", err |
| 601 | } |
| 602 | return os.MkdirTemp(root, recoveryTrashStagingPrefix) |
| 603 | } |
| 604 | |
| 605 | func prepareRecoveryTrashStage(path, key, stageDir string) error { |
| 606 | if err := writeRecoveryTrashPending(stageDir, key); err != nil { |
| 607 | return err |
| 608 | } |
| 609 | return moveRecoveryTrashPath(path, filepath.Join(stageDir, key)) |
| 610 | } |
| 611 | |
| 612 | func finishRecoveryTrashStage(dir, path, key, stageDir string, guard *SessionRemovalGuard) error { |
| 613 | if err := moveRecoveryTrashArtifacts(dir, path, stageDir); err != nil { |
| 614 | return err |
| 615 | } |
| 616 | itemDir, err := publishRecoveryTrashStage(dir, key, stageDir) |
| 617 | if err != nil { |
| 618 | return err |
| 619 | } |
| 620 | if err := writeRecoveryTrashMetaExisting(itemDir, key); err != nil { |
| 621 | if !os.IsNotExist(err) { |
| 622 | return err |
| 623 | } |
| 624 | // A complete entry may be restored or purged as soon as it is published. |
| 625 | // In that case there is nothing left for this producer to finalize. |
| 626 | return guard.RemoveSidecarsAndRelease() |
| 627 | } |
| 628 | if err := clearRecoveryTrashPending(itemDir); err != nil { |
| 629 | return err |
| 630 | } |
| 631 | return guard.RemoveSidecarsAndRelease() |
| 632 | } |
| 633 | |
| 634 | func publishRecoveryTrashStage(dir, key, stageDir string) (string, error) { |
| 635 | root := filepath.Join(dir, recoveryTrashDir) |
| 636 | stem := strings.TrimSuffix(key, filepath.Ext(key)) |
| 637 | stamp := time.Now().UTC().UnixMilli() |
| 638 | for i := range 1000 { |
| 639 | name := key |
| 640 | if i > 0 { |
| 641 | name = fmt.Sprintf("%s-recovery-%d-%d", stem, stamp, i) |
| 642 | } |
| 643 | itemDir := filepath.Join(root, name) |
| 644 | if _, err := os.Lstat(itemDir); err == nil { |
| 645 | continue |
| 646 | } else if !os.IsNotExist(err) { |
| 647 | return "", err |
| 648 | } |
| 649 | if err := os.Rename(stageDir, itemDir); err == nil { |
| 650 | return itemDir, nil |
| 651 | } else if _, statErr := os.Lstat(itemDir); statErr == nil { |
| 652 | continue // another producer won this candidate |
| 653 | } else if !os.IsNotExist(statErr) { |
| 654 | return "", statErr |
| 655 | } else { |
| 656 | return "", err |
| 657 | } |
| 658 | } |
| 659 | return "", fmt.Errorf("could not publish recovery trash target") |
| 660 | } |
| 661 | |
| 662 | func writeRecoveryTrashPending(itemDir, key string) error { |
| 663 | if !validRecoveryTrashKey(key) { |
| 664 | return fmt.Errorf("invalid recovery trash key") |
| 665 | } |
| 666 | b, err := json.MarshalIndent(recoveryTrashPendingMeta{Key: key}, "", " ") |
| 667 | if err != nil { |
| 668 | return err |
| 669 | } |
| 670 | return fileutil.AtomicWriteFileStrict(filepath.Join(itemDir, recoveryTrashPendingFile), b, 0o644) |
| 671 | } |
| 672 | |
| 673 | func readRecoveryTrashPending(itemDir string) (recoveryTrashPendingMeta, error) { |
| 674 | b, err := os.ReadFile(filepath.Join(itemDir, recoveryTrashPendingFile)) |
| 675 | if err != nil { |
| 676 | return recoveryTrashPendingMeta{}, err |
| 677 | } |
| 678 | var meta recoveryTrashPendingMeta |
| 679 | if err := json.Unmarshal(b, &meta); err != nil { |
| 680 | return recoveryTrashPendingMeta{}, err |
| 681 | } |
| 682 | return meta, nil |
| 683 | } |
| 684 | |
| 685 | func clearRecoveryTrashPending(itemDir string) error { |
| 686 | err := os.Remove(filepath.Join(itemDir, recoveryTrashPendingFile)) |
| 687 | if os.IsNotExist(err) { |
| 688 | return nil |
| 689 | } |
| 690 | return err |
| 691 | } |
| 692 | |
| 693 | func validRecoveryTrashKey(key string) bool { |
| 694 | return key != "" && filepath.Base(key) == key && key != "." && key != ".." && |
| 695 | strings.HasSuffix(key, ".jsonl") && !strings.HasSuffix(key, ".events.jsonl") |
| 696 | } |
| 697 | |
| 698 | func reserveRecoveryTrashItemDir(dir, key string) (string, string, error) { |
| 699 | root := filepath.Join(dir, recoveryTrashDir) |
| 700 | if err := os.MkdirAll(root, 0o755); err != nil { |
| 701 | return "", "", err |
| 702 | } |
| 703 | stem := strings.TrimSuffix(key, filepath.Ext(key)) |
| 704 | for i := range 1000 { |
| 705 | name := key |
| 706 | if i > 0 { |
| 707 | name = fmt.Sprintf("%s-recovery-%d-%d", stem, time.Now().UTC().UnixMilli(), i) |
| 708 | } |
| 709 | itemDir := filepath.Join(root, name) |
| 710 | if err := os.Mkdir(itemDir, 0o755); err == nil { |
| 711 | return name, itemDir, nil |
| 712 | } else if !os.IsExist(err) { |
| 713 | return "", "", err |
| 714 | } |
| 715 | } |
| 716 | return "", "", fmt.Errorf("could not reserve recovery trash target") |
| 717 | } |
| 718 | |
| 719 | func finishRecoveryTrashMove(dir, path, key, itemDir string, guard *SessionRemovalGuard) error { |
| 720 | if err := os.MkdirAll(itemDir, 0o755); err != nil { |
| 721 | return err |
| 722 | } |
| 723 | if err := moveRecoveryTrashArtifacts(dir, path, itemDir); err != nil { |
| 724 | return err |
| 725 | } |
| 726 | if err := writeRecoveryTrashMeta(itemDir, key); err != nil { |
| 727 | return err |
| 728 | } |
| 729 | // Keep the branch guard until the trash entry is complete and the hidden |
| 730 | // marker is cleared. No runtime can bind the now-vacant live path in the |
| 731 | // middle and inherit an incomplete cleanup state. |
| 732 | if err := ClearCleanupPending(path); err != nil { |
| 733 | return err |
| 734 | } |
| 735 | return guard.RemoveSidecarsAndRelease() |
| 736 | } |
| 737 | |
| 738 | func moveRecoveryTrashArtifacts(dir, path, itemDir string) error { |
| 739 | for _, src := range recoveryTrashSidecars(path) { |
| 740 | if err := moveRecoveryTrashPath(src, filepath.Join(itemDir, filepath.Base(src))); err != nil { |
| 741 | return err |
| 742 | } |
| 743 | } |
| 744 | return moveRecoverySubagentArtifacts(dir, path, itemDir) |
| 745 | } |
| 746 | |
| 747 | func prepareRecoveryTrashEntry(path, key, itemDir string) error { |
| 748 | if err := writeRecoveryTrashMeta(itemDir, key); err != nil { |
| 749 | return err |
| 750 | } |
| 751 | return moveRecoveryTrashPath(path, filepath.Join(itemDir, key)) |
| 752 | } |
| 753 | |
| 754 | func writeRecoveryTrashMeta(itemDir, key string) error { |
| 755 | meta := recoveryTrashMeta{Key: key, DeletedAt: time.Now().UnixMilli()} |
| 756 | b, err := json.MarshalIndent(meta, "", " ") |
| 757 | if err != nil { |
| 758 | return err |
| 759 | } |
| 760 | if err := os.MkdirAll(itemDir, 0o755); err != nil { |
| 761 | return err |
| 762 | } |
| 763 | return os.WriteFile(filepath.Join(itemDir, recoveryTrashMetaFile), b, 0o644) |
| 764 | } |
| 765 | |
| 766 | func writeRecoveryTrashMetaExisting(itemDir, key string) error { |
| 767 | meta := recoveryTrashMeta{Key: key, DeletedAt: time.Now().UnixMilli()} |
| 768 | b, err := json.MarshalIndent(meta, "", " ") |
| 769 | if err != nil { |
| 770 | return err |
| 771 | } |
| 772 | return os.WriteFile(filepath.Join(itemDir, recoveryTrashMetaFile), b, 0o644) |
| 773 | } |
| 774 | |
| 775 | func recoveryTrashSidecars(path string) []string { |
| 776 | artifacts := append([]string(nil), store.SessionSidecarFiles(path)...) |
| 777 | artifacts = append(artifacts, |
| 778 | path+".telemetry.json", |
| 779 | store.SessionCheckpointDir(path), |
| 780 | store.SessionJobsDir(path), |
| 781 | store.SessionInboxDir(path), |
| 782 | ) |
| 783 | return artifacts |
| 784 | } |
| 785 | |
| 786 | func moveRecoverySubagentArtifacts(dir, path, itemDir string) error { |
| 787 | artifacts, err := ListSubagentsByParent(dir, BranchID(path)) |
| 788 | if err != nil { |
| 789 | return err |
| 790 | } |
| 791 | targetDir := filepath.Join(itemDir, "subagents") |
| 792 | for _, artifact := range artifacts { |
| 793 | paths := []string{artifact.SessionPath, artifact.MetaPath} |
| 794 | paths = append(paths, store.SessionSidecarFiles(artifact.SessionPath)...) |
| 795 | for _, src := range paths { |
| 796 | if err := moveRecoveryTrashPath(src, filepath.Join(targetDir, filepath.Base(src))); err != nil { |
| 797 | return err |
| 798 | } |
| 799 | } |
| 800 | } |
| 801 | return nil |
| 802 | } |
| 803 | |
| 804 | func moveRecoveryTrashPath(src, dst string) error { |
| 805 | if strings.TrimSpace(src) == "" { |
| 806 | return nil |
| 807 | } |
| 808 | if _, err := os.Lstat(src); os.IsNotExist(err) { |
| 809 | return nil |
| 810 | } else if err != nil { |
| 811 | return err |
| 812 | } |
| 813 | if err := os.MkdirAll(filepath.Dir(dst), 0o755); err != nil { |
| 814 | return err |
| 815 | } |
| 816 | return os.Rename(src, dst) |
| 817 | } |
| 818 |