| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "bytes" |
| 5 | "crypto/sha256" |
| 6 | "encoding/json" |
| 7 | "errors" |
| 8 | "fmt" |
| 9 | "io" |
| 10 | "log/slog" |
| 11 | "os" |
| 12 | "path/filepath" |
| 13 | "time" |
| 14 | |
| 15 | "reasonix/internal/fileutil" |
| 16 | fileencoding "reasonix/internal/fileutil/encoding" |
| 17 | "reasonix/internal/provider" |
| 18 | "reasonix/internal/store" |
| 19 | ) |
| 20 | |
| 21 | const ( |
| 22 | sessionEventSchemaVersion = 1 |
| 23 | sessionEventTypeReplace = "replace" |
| 24 | sessionEventTypeAppend = "append" |
| 25 | // sessionEventReplayMaxBytes caps decoder input before encoding/json can |
| 26 | // allocate an arbitrarily large record. Session logs normally compact far |
| 27 | // below this threshold; the generous ceiling still accommodates histories |
| 28 | // with embedded images while keeping corrupt logs from exhausting the host. |
| 29 | sessionEventReplayMaxBytes = int64(128 << 20) |
| 30 | // A byte limit alone is insufficient: a compact JSON array can expand into |
| 31 | // a much larger graph of messages and event records after decoding. |
| 32 | sessionEventReplayMaxRecords = 100_000 |
| 33 | sessionEventReplayMaxMessages = 100_000 |
| 34 | sessionEventReplayMaxCollectionItems = 100_000 |
| 35 | sessionEventProbeMaxBytes = int64(4 << 10) |
| 36 | // sessionEventLogCompactFloor is the smallest log size that can trigger |
| 37 | // compaction, so short sessions never pay a checkpoint rewrite. |
| 38 | sessionEventLogCompactFloor = int64(256 << 10) |
| 39 | // sessionEventLogCompactFactor bounds the log at this multiple of the live |
| 40 | // transcript's encoded size; past it the log is rewritten to one replace |
| 41 | // event so replace-heavy histories (compaction, rewind) cannot grow the |
| 42 | // file without bound. |
| 43 | sessionEventLogCompactFactor = int64(4) |
| 44 | ) |
| 45 | |
| 46 | // ErrSessionReplayLimitExceeded identifies a session that was left untouched |
| 47 | // because replaying it would exceed the process safety budget. Callers must not |
| 48 | // fall back to an older checkpoint: the event log may contain newer turns. |
| 49 | var ErrSessionReplayLimitExceeded = errors.New("session history exceeds safe replay limits") |
| 50 | |
| 51 | // SessionReplayLimitError carries machine-readable diagnostics while keeping |
| 52 | // Error free of local paths for Desktop surfaces that display startup errors. |
| 53 | type SessionReplayLimitError struct { |
| 54 | Path string |
| 55 | Resource string |
| 56 | Value int64 |
| 57 | Limit int64 |
| 58 | } |
| 59 | |
| 60 | func (e *SessionReplayLimitError) Error() string { |
| 61 | if e == nil { |
| 62 | return ErrSessionReplayLimitExceeded.Error() |
| 63 | } |
| 64 | return fmt.Sprintf("%s: %s=%d, limit=%d; session files were left unchanged", |
| 65 | ErrSessionReplayLimitExceeded, e.Resource, e.Value, e.Limit) |
| 66 | } |
| 67 | |
| 68 | func (e *SessionReplayLimitError) Unwrap() error { |
| 69 | return ErrSessionReplayLimitExceeded |
| 70 | } |
| 71 | |
| 72 | type sessionReplayLimits struct { |
| 73 | maxBytes int64 |
| 74 | maxRecords int |
| 75 | maxMessages int |
| 76 | maxCollectionItems int |
| 77 | } |
| 78 | |
| 79 | var defaultSessionReplayLimits = sessionReplayLimits{ |
| 80 | maxBytes: sessionEventReplayMaxBytes, |
| 81 | maxRecords: sessionEventReplayMaxRecords, |
| 82 | maxMessages: sessionEventReplayMaxMessages, |
| 83 | maxCollectionItems: sessionEventReplayMaxCollectionItems, |
| 84 | } |
| 85 | |
| 86 | func sessionReplayLimitError(path, resource string, value, limit int64) error { |
| 87 | err := &SessionReplayLimitError{Path: path, Resource: resource, Value: value, Limit: limit} |
| 88 | slog.Warn("session: refusing unsafe event-log replay", |
| 89 | "path", path, "resource", resource, "value", value, "limit", limit) |
| 90 | return err |
| 91 | } |
| 92 | |
| 93 | type sessionEventRecord struct { |
| 94 | SchemaVersion int `json:"schema_version"` |
| 95 | Type string `json:"type"` |
| 96 | Revision int64 `json:"revision,omitempty"` |
| 97 | BaseRevision int64 `json:"base_revision,omitempty"` |
| 98 | MessageIndex int `json:"message_index,omitempty"` |
| 99 | Messages []provider.Message `json:"messages,omitempty"` |
| 100 | ContentDigest string `json:"content_digest,omitempty"` |
| 101 | WriterID string `json:"writer_id,omitempty"` |
| 102 | Reason string `json:"reason,omitempty"` |
| 103 | CreatedAt time.Time `json:"created_at"` |
| 104 | } |
| 105 | |
| 106 | // sessionEventWireRecord keeps the messages array encoded until the replay |
| 107 | // budget has been checked. Decoding directly into sessionEventRecord would |
| 108 | // materialize every provider.Message before replay could enforce maxMessages. |
| 109 | type sessionEventWireRecord struct { |
| 110 | SchemaVersion int `json:"schema_version"` |
| 111 | Type string `json:"type"` |
| 112 | Revision int64 `json:"revision,omitempty"` |
| 113 | BaseRevision int64 `json:"base_revision,omitempty"` |
| 114 | MessageIndex int `json:"message_index,omitempty"` |
| 115 | Messages json.RawMessage `json:"messages,omitempty"` |
| 116 | ContentDigest string `json:"content_digest,omitempty"` |
| 117 | WriterID string `json:"writer_id,omitempty"` |
| 118 | Reason string `json:"reason,omitempty"` |
| 119 | CreatedAt time.Time `json:"created_at"` |
| 120 | } |
| 121 | |
| 122 | type sessionEventIndex struct { |
| 123 | SchemaVersion int `json:"schema_version"` |
| 124 | LogSize int64 `json:"log_size"` |
| 125 | MessageCount int `json:"message_count"` |
| 126 | Revision int64 `json:"revision"` |
| 127 | ContentDigest string `json:"content_digest"` |
| 128 | WriterID string `json:"writer_id"` |
| 129 | UpdatedAt time.Time `json:"updated_at"` |
| 130 | } |
| 131 | |
| 132 | func SessionEventLogPath(sessionPath string) string { |
| 133 | return store.SessionEventLog(sessionPath) |
| 134 | } |
| 135 | |
| 136 | func SessionEventIndexPath(sessionPath string) string { |
| 137 | return store.SessionEventIndex(sessionPath) |
| 138 | } |
| 139 | |
| 140 | func sessionEventLogSize(sessionPath string) int64 { |
| 141 | path := store.SessionEventLog(sessionPath) |
| 142 | if path == "" { |
| 143 | return 0 |
| 144 | } |
| 145 | info, err := os.Stat(path) |
| 146 | if err != nil || info.IsDir() { |
| 147 | return 0 |
| 148 | } |
| 149 | return info.Size() |
| 150 | } |
| 151 | |
| 152 | func sessionEventLogOversized(logSize, contentBytes int64) bool { |
| 153 | limit := sessionEventLogCompactFloor |
| 154 | if scaled := contentBytes * sessionEventLogCompactFactor; scaled > limit { |
| 155 | limit = scaled |
| 156 | } |
| 157 | return logSize > limit |
| 158 | } |
| 159 | |
| 160 | // sessionEventReplay is the result of a tolerant event-log replay: the |
| 161 | // transcript up to the last cleanly applied record, plus enough bookkeeping |
| 162 | // for writers to self-heal a torn tail. |
| 163 | type sessionEventReplay struct { |
| 164 | msgs []provider.Message |
| 165 | // collectionItems counts the elements in every JSON array nested below a |
| 166 | // live message. Keeping this alongside msgs bounds slices such as tool calls, |
| 167 | // images, memory citations, and interrupted-turn recovery metadata without |
| 168 | // coupling replay safety to today's provider.Message field list. |
| 169 | collectionItems int |
| 170 | // times mirrors msgs with each message's record CreatedAt. Replace events |
| 171 | // collapse per-turn history, so their messages get the zero time and |
| 172 | // callers fall back to coarser timestamps. |
| 173 | times []time.Time |
| 174 | // records counts cleanly applied events. |
| 175 | records int |
| 176 | // lastGoodEnd is the byte offset just past the last cleanly applied |
| 177 | // record; truncating the log here drops only undecodable bytes. |
| 178 | lastGoodEnd int64 |
| 179 | // size is the log size that was replayed. |
| 180 | size int64 |
| 181 | // damaged is set when replay stopped early on a torn/corrupt record or a |
| 182 | // broken append chain. The prefix in msgs is still a valid historical |
| 183 | // state. |
| 184 | damaged bool |
| 185 | } |
| 186 | |
| 187 | // sessionEventLogProbe classifies whatever sits at the session's event-log |
| 188 | // path. Legacy imports can leave a foreign ".events.jsonl" (e.g. the v0.x |
| 189 | // Claude-style event transcript) at exactly the native log path; writing into |
| 190 | // or over it would corrupt the user's original file, so foreign logs are |
| 191 | // read-ignored and never touched. |
| 192 | type sessionEventLogProbe struct { |
| 193 | size int64 |
| 194 | native bool // missing/empty, or first record is a supported native event |
| 195 | futureSchema bool // first record declares a newer schema than this build |
| 196 | schemaVersion int |
| 197 | } |
| 198 | |
| 199 | // sessionEventSidecarsFit reports whether the event log and index filenames |
| 200 | // stay within the filesystem's name limit. Overlong transcript names (from the |
| 201 | // pre-bounded recovery cascade, until reconcileOverlongSessionFilenames renames |
| 202 | // them) must run checkpoint-only: creating their sidecars would fail with |
| 203 | // ENAMETOOLONG mid-save. |
| 204 | func sessionEventSidecarsFit(sessionPath string) bool { |
| 205 | logName := filepath.Base(store.SessionEventLog(sessionPath)) |
| 206 | indexName := filepath.Base(store.SessionEventIndex(sessionPath)) |
| 207 | return len(logName) <= nameMaxBytes && len(indexName) <= nameMaxBytes |
| 208 | } |
| 209 | |
| 210 | // probeSessionEventLog inspects the first record of the event log to decide |
| 211 | // whether the native persistence layer owns the file. Missing or empty logs |
| 212 | // count as native (we may create/append); an undecodable or foreign first |
| 213 | // record — or a transcript name too long for the sidecars to fit — marks the |
| 214 | // file as not ours. |
| 215 | func probeSessionEventLog(sessionPath string) (sessionEventLogProbe, error) { |
| 216 | return probeSessionEventLogWithLimits(sessionPath, defaultSessionReplayLimits) |
| 217 | } |
| 218 | |
| 219 | func probeSessionEventLogWithLimits(sessionPath string, limits sessionReplayLimits) (sessionEventLogProbe, error) { |
| 220 | path := store.SessionEventLog(sessionPath) |
| 221 | if path == "" { |
| 222 | return sessionEventLogProbe{native: true}, nil |
| 223 | } |
| 224 | if !sessionEventSidecarsFit(sessionPath) { |
| 225 | return sessionEventLogProbe{}, nil |
| 226 | } |
| 227 | info, err := os.Stat(path) |
| 228 | if err != nil { |
| 229 | if os.IsNotExist(err) { |
| 230 | return sessionEventLogProbe{native: true}, nil |
| 231 | } |
| 232 | return sessionEventLogProbe{}, err |
| 233 | } |
| 234 | if info.IsDir() { |
| 235 | return sessionEventLogProbe{}, nil |
| 236 | } |
| 237 | if info.Size() == 0 { |
| 238 | return sessionEventLogProbe{native: true}, nil |
| 239 | } |
| 240 | probe := sessionEventLogProbe{size: info.Size()} |
| 241 | f, err := os.Open(path) |
| 242 | if err != nil { |
| 243 | return sessionEventLogProbe{}, err |
| 244 | } |
| 245 | defer f.Close() |
| 246 | var schemaVersion int |
| 247 | var eventType string |
| 248 | var ok bool |
| 249 | schemaVersion, eventType, ok = probeSessionEventHeader(f) |
| 250 | if !ok && info.Size() <= limits.maxBytes { |
| 251 | // Native writers put both identifying fields in the bounded prefix. For |
| 252 | // other valid in-budget JSON, fall back to a minimal struct decode so |
| 253 | // field order remains a compatibility property rather than a format |
| 254 | // requirement. Unknown fields are not materialized into messages. |
| 255 | if _, err := f.Seek(0, io.SeekStart); err != nil { |
| 256 | return sessionEventLogProbe{}, err |
| 257 | } |
| 258 | var header struct { |
| 259 | SchemaVersion int `json:"schema_version"` |
| 260 | Type string `json:"type"` |
| 261 | } |
| 262 | dec := json.NewDecoder(&io.LimitedReader{R: f, N: limits.maxBytes + 1}) |
| 263 | if err := dec.Decode(&header); err == nil { |
| 264 | schemaVersion, eventType, ok = header.SchemaVersion, header.Type, true |
| 265 | } |
| 266 | } |
| 267 | if !ok { |
| 268 | // Nothing decodable at the head: not a native log this build can own. |
| 269 | return probe, nil |
| 270 | } |
| 271 | probe.schemaVersion = schemaVersion |
| 272 | switch { |
| 273 | case schemaVersion == sessionEventSchemaVersion && |
| 274 | (eventType == sessionEventTypeReplace || eventType == sessionEventTypeAppend): |
| 275 | probe.native = true |
| 276 | case schemaVersion > sessionEventSchemaVersion: |
| 277 | // A newer writer owns this log; ignoring or truncating it would |
| 278 | // silently discard that writer's transcript. |
| 279 | probe.futureSchema = true |
| 280 | } |
| 281 | return probe, nil |
| 282 | } |
| 283 | |
| 284 | // probeSessionEventHeader searches a bounded prefix for the identifying fields. |
| 285 | // Using Decode on a partial struct still buffers the whole JSON value, so native |
| 286 | // writer output must take this fast path before replay's byte budget is checked. |
| 287 | func probeSessionEventHeader(r io.Reader) (schemaVersion int, eventType string, ok bool) { |
| 288 | dec := json.NewDecoder(io.LimitReader(r, sessionEventProbeMaxBytes)) |
| 289 | tok, err := dec.Token() |
| 290 | if err != nil { |
| 291 | return 0, "", false |
| 292 | } |
| 293 | if delim, isDelim := tok.(json.Delim); !isDelim || delim != '{' { |
| 294 | return 0, "", false |
| 295 | } |
| 296 | var haveSchema, haveType bool |
| 297 | for dec.More() { |
| 298 | key, err := dec.Token() |
| 299 | if err != nil { |
| 300 | return 0, "", false |
| 301 | } |
| 302 | name, isString := key.(string) |
| 303 | if !isString { |
| 304 | return 0, "", false |
| 305 | } |
| 306 | switch name { |
| 307 | case "schema_version": |
| 308 | if err := dec.Decode(&schemaVersion); err != nil { |
| 309 | return 0, "", false |
| 310 | } |
| 311 | haveSchema = true |
| 312 | case "type": |
| 313 | if err := dec.Decode(&eventType); err != nil { |
| 314 | return 0, "", false |
| 315 | } |
| 316 | haveType = true |
| 317 | default: |
| 318 | var discard json.RawMessage |
| 319 | if err := dec.Decode(&discard); err != nil { |
| 320 | return 0, "", false |
| 321 | } |
| 322 | } |
| 323 | if haveSchema && haveType { |
| 324 | return schemaVersion, eventType, true |
| 325 | } |
| 326 | } |
| 327 | return 0, "", false |
| 328 | } |
| 329 | |
| 330 | // replaySessionEventLog decodes an event log tolerantly: decoding stops at the |
| 331 | // first record that fails to parse or chain, and the state up to that point is |
| 332 | // returned with damaged=true so writers can self-heal. Unsupported schema |
| 333 | // versions and unknown event types stay hard errors — they mean a newer writer |
| 334 | // owns this log, and truncating it would discard that writer's data. |
| 335 | func replaySessionEventLog(path string) (sessionEventReplay, error) { |
| 336 | return replaySessionEventLogWithLimits(path, defaultSessionReplayLimits) |
| 337 | } |
| 338 | |
| 339 | func replaySessionEventLogWithLimits(path string, limits sessionReplayLimits) (sessionEventReplay, error) { |
| 340 | f, err := os.Open(path) |
| 341 | if err != nil { |
| 342 | return sessionEventReplay{}, err |
| 343 | } |
| 344 | defer f.Close() |
| 345 | info, err := f.Stat() |
| 346 | if err != nil { |
| 347 | return sessionEventReplay{}, err |
| 348 | } |
| 349 | replay := sessionEventReplay{size: info.Size()} |
| 350 | if info.Size() > limits.maxBytes { |
| 351 | return replay, sessionReplayLimitError(path, "encoded_bytes", info.Size(), limits.maxBytes) |
| 352 | } |
| 353 | // Stat and read are not atomic across processes. LimitReader keeps a log |
| 354 | // that grows after Stat inside the same byte budget. |
| 355 | limited := &io.LimitedReader{R: f, N: limits.maxBytes + 1} |
| 356 | dec := json.NewDecoder(limited) |
| 357 | for { |
| 358 | var rec sessionEventWireRecord |
| 359 | if err := dec.Decode(&rec); err != nil { |
| 360 | if limited.N == 0 { |
| 361 | return replay, sessionReplayLimitError(path, "encoded_bytes", limits.maxBytes+1, limits.maxBytes) |
| 362 | } |
| 363 | if errors.Is(err, io.EOF) { |
| 364 | return replay, nil |
| 365 | } |
| 366 | replay.damaged = true |
| 367 | return replay, nil |
| 368 | } |
| 369 | if rec.SchemaVersion != sessionEventSchemaVersion { |
| 370 | return replay, fmt.Errorf("decode session event log %s: unsupported schema version %d", path, rec.SchemaVersion) |
| 371 | } |
| 372 | if replay.records >= limits.maxRecords { |
| 373 | return replay, sessionReplayLimitError(path, "event_records", int64(replay.records+1), int64(limits.maxRecords)) |
| 374 | } |
| 375 | switch rec.Type { |
| 376 | case sessionEventTypeReplace: |
| 377 | msgs, collectionItems, err := decodeSessionEventMessages(path, rec.Messages, 0, 0, limits) |
| 378 | if err != nil { |
| 379 | if errors.Is(err, ErrSessionReplayLimitExceeded) { |
| 380 | return replay, err |
| 381 | } |
| 382 | replay.damaged = true |
| 383 | return replay, nil |
| 384 | } |
| 385 | replay.msgs = msgs |
| 386 | replay.collectionItems = collectionItems |
| 387 | replay.times = make([]time.Time, len(replay.msgs)) |
| 388 | case sessionEventTypeAppend: |
| 389 | if rec.MessageIndex != len(replay.msgs) { |
| 390 | replay.damaged = true |
| 391 | return replay, nil |
| 392 | } |
| 393 | msgs, collectionItems, err := decodeSessionEventMessages( |
| 394 | path, rec.Messages, len(replay.msgs), replay.collectionItems, limits, |
| 395 | ) |
| 396 | if err != nil { |
| 397 | if errors.Is(err, ErrSessionReplayLimitExceeded) { |
| 398 | return replay, err |
| 399 | } |
| 400 | replay.damaged = true |
| 401 | return replay, nil |
| 402 | } |
| 403 | replay.msgs = append(replay.msgs, msgs...) |
| 404 | replay.collectionItems = collectionItems |
| 405 | for range msgs { |
| 406 | replay.times = append(replay.times, rec.CreatedAt) |
| 407 | } |
| 408 | default: |
| 409 | return replay, fmt.Errorf("decode session event log %s: unsupported event type %q", path, rec.Type) |
| 410 | } |
| 411 | replay.records++ |
| 412 | replay.lastGoodEnd = dec.InputOffset() |
| 413 | } |
| 414 | } |
| 415 | |
| 416 | // decodeSessionEventMessages preflights both the top-level message count and |
| 417 | // every nested JSON collection before constructing provider.Message values. |
| 418 | // The token walk is independent of today's provider.Message fields, so future |
| 419 | // slice fields inherit the same aggregate object-graph bound automatically. |
| 420 | func decodeSessionEventMessages( |
| 421 | path string, |
| 422 | raw json.RawMessage, |
| 423 | existingMessages, existingCollectionItems int, |
| 424 | limits sessionReplayLimits, |
| 425 | ) ([]provider.Message, int, error) { |
| 426 | trimmed := bytes.TrimSpace(raw) |
| 427 | if len(trimmed) == 0 || bytes.Equal(trimmed, []byte("null")) { |
| 428 | return nil, existingCollectionItems, nil |
| 429 | } |
| 430 | messageCount, collectionItems, err := preflightSessionEventMessages( |
| 431 | path, trimmed, existingMessages, existingCollectionItems, limits, |
| 432 | ) |
| 433 | if err != nil { |
| 434 | return nil, existingCollectionItems, err |
| 435 | } |
| 436 | dec := json.NewDecoder(bytes.NewReader(trimmed)) |
| 437 | tok, err := dec.Token() |
| 438 | if err != nil { |
| 439 | return nil, existingCollectionItems, err |
| 440 | } |
| 441 | if delim, ok := tok.(json.Delim); !ok || delim != '[' { |
| 442 | return nil, existingCollectionItems, fmt.Errorf("messages must be an array") |
| 443 | } |
| 444 | msgs := make([]provider.Message, 0, messageCount) |
| 445 | for dec.More() { |
| 446 | var msg provider.Message |
| 447 | if err := dec.Decode(&msg); err != nil { |
| 448 | return nil, existingCollectionItems, err |
| 449 | } |
| 450 | msgs = append(msgs, msg) |
| 451 | } |
| 452 | if _, err := dec.Token(); err != nil { |
| 453 | return nil, existingCollectionItems, err |
| 454 | } |
| 455 | return msgs, collectionItems, nil |
| 456 | } |
| 457 | |
| 458 | func preflightSessionEventMessages( |
| 459 | path string, |
| 460 | raw []byte, |
| 461 | existingMessages, existingCollectionItems int, |
| 462 | limits sessionReplayLimits, |
| 463 | ) (messageCount, collectionItems int, err error) { |
| 464 | dec := json.NewDecoder(bytes.NewReader(raw)) |
| 465 | tok, err := dec.Token() |
| 466 | if err != nil { |
| 467 | return 0, existingCollectionItems, err |
| 468 | } |
| 469 | if delim, ok := tok.(json.Delim); !ok || delim != '[' { |
| 470 | return 0, existingCollectionItems, fmt.Errorf("messages must be an array") |
| 471 | } |
| 472 | collectionItems = existingCollectionItems |
| 473 | for dec.More() { |
| 474 | if existingMessages+messageCount >= limits.maxMessages { |
| 475 | return 0, existingCollectionItems, sessionReplayLimitError( |
| 476 | path, "messages", int64(existingMessages+messageCount+1), int64(limits.maxMessages), |
| 477 | ) |
| 478 | } |
| 479 | messageCount++ |
| 480 | if err := preflightSessionEventValue(path, dec, &collectionItems, limits.maxCollectionItems); err != nil { |
| 481 | return 0, existingCollectionItems, err |
| 482 | } |
| 483 | } |
| 484 | if _, err := dec.Token(); err != nil { |
| 485 | return 0, existingCollectionItems, err |
| 486 | } |
| 487 | return messageCount, collectionItems, nil |
| 488 | } |
| 489 | |
| 490 | // preflightSessionEventValue walks one JSON value without materializing maps or |
| 491 | // slices. Each array element is charged before its value is read, so an invalid |
| 492 | // over-limit element cannot allocate a typed provider collection first. |
| 493 | func preflightSessionEventValue(path string, dec *json.Decoder, collectionItems *int, maxCollectionItems int) error { |
| 494 | tok, err := dec.Token() |
| 495 | if err != nil { |
| 496 | return err |
| 497 | } |
| 498 | delim, ok := tok.(json.Delim) |
| 499 | if !ok { |
| 500 | return nil |
| 501 | } |
| 502 | switch delim { |
| 503 | case '{': |
| 504 | for dec.More() { |
| 505 | key, err := dec.Token() |
| 506 | if err != nil { |
| 507 | return err |
| 508 | } |
| 509 | if _, ok := key.(string); !ok { |
| 510 | return fmt.Errorf("object key must be a string") |
| 511 | } |
| 512 | if err := preflightSessionEventValue(path, dec, collectionItems, maxCollectionItems); err != nil { |
| 513 | return err |
| 514 | } |
| 515 | } |
| 516 | end, err := dec.Token() |
| 517 | if err != nil { |
| 518 | return err |
| 519 | } |
| 520 | if end != json.Delim('}') { |
| 521 | return fmt.Errorf("object is not terminated") |
| 522 | } |
| 523 | return nil |
| 524 | case '[': |
| 525 | for dec.More() { |
| 526 | if *collectionItems >= maxCollectionItems { |
| 527 | return sessionReplayLimitError( |
| 528 | path, |
| 529 | "message_collection_items", |
| 530 | int64(*collectionItems+1), |
| 531 | int64(maxCollectionItems), |
| 532 | ) |
| 533 | } |
| 534 | (*collectionItems)++ |
| 535 | if err := preflightSessionEventValue(path, dec, collectionItems, maxCollectionItems); err != nil { |
| 536 | return err |
| 537 | } |
| 538 | } |
| 539 | end, err := dec.Token() |
| 540 | if err != nil { |
| 541 | return err |
| 542 | } |
| 543 | if end != json.Delim(']') { |
| 544 | return fmt.Errorf("array is not terminated") |
| 545 | } |
| 546 | return nil |
| 547 | default: |
| 548 | return fmt.Errorf("unexpected JSON delimiter %q", delim) |
| 549 | } |
| 550 | } |
| 551 | |
| 552 | // loadSessionMessages returns the session transcript, preferring the event log |
| 553 | // when the native layer owns it and it holds at least one decodable record. |
| 554 | // Foreign files squatting the log path (legacy import leftovers) are ignored |
| 555 | // in favor of the .jsonl checkpoint. damaged reports that a native log could |
| 556 | // not be replayed to its end (torn tail or corrupt record); callers that write |
| 557 | // should rewrite-and-compact to heal it. |
| 558 | func loadSessionMessages(sessionPath string) (msgs []provider.Message, fromEvents, damaged bool, err error) { |
| 559 | return loadSessionMessagesWithLimits(sessionPath, defaultSessionReplayLimits) |
| 560 | } |
| 561 | |
| 562 | func loadSessionMessagesWithLimits(sessionPath string, limits sessionReplayLimits) (msgs []provider.Message, fromEvents, damaged bool, err error) { |
| 563 | probe, err := probeSessionEventLogWithLimits(sessionPath, limits) |
| 564 | if err != nil { |
| 565 | return nil, false, false, err |
| 566 | } |
| 567 | if probe.futureSchema { |
| 568 | return nil, true, false, fmt.Errorf("session event log for %s uses schema %d; this build supports up to %d", sessionPath, probe.schemaVersion, sessionEventSchemaVersion) |
| 569 | } |
| 570 | if probe.native && probe.size > 0 { |
| 571 | replay, replayErr := replaySessionEventLogWithLimits(store.SessionEventLog(sessionPath), limits) |
| 572 | if replayErr != nil { |
| 573 | return nil, true, false, replayErr |
| 574 | } |
| 575 | if replay.records > 0 { |
| 576 | return replay.msgs, true, replay.damaged, nil |
| 577 | } |
| 578 | // Defensive: the probe saw a native head but nothing replayed; fall |
| 579 | // back to the checkpoint and let the next save rebuild the log. |
| 580 | msgs, err = loadSessionMessagesFromJSONL(sessionPath) |
| 581 | return msgs, false, true, err |
| 582 | } |
| 583 | msgs, err = loadSessionMessagesFromJSONL(sessionPath) |
| 584 | return msgs, false, false, err |
| 585 | } |
| 586 | |
| 587 | func loadSessionMessagesFromJSONL(path string) ([]provider.Message, error) { |
| 588 | f, err := os.Open(path) |
| 589 | if err != nil { |
| 590 | return nil, err |
| 591 | } |
| 592 | defer f.Close() |
| 593 | |
| 594 | var msgs []provider.Message |
| 595 | dec := json.NewDecoder(f) |
| 596 | for { |
| 597 | var m provider.Message |
| 598 | if err := dec.Decode(&m); err != nil { |
| 599 | if errors.Is(err, io.EOF) { |
| 600 | break |
| 601 | } |
| 602 | return nil, fmt.Errorf("decode %s: %w", path, err) |
| 603 | } |
| 604 | msgs = append(msgs, m) |
| 605 | } |
| 606 | return msgs, nil |
| 607 | } |
| 608 | |
| 609 | // repairSessionEventLogTail truncates undecodable bytes left by a crash or |
| 610 | // disk-full append so the next append cannot bury them mid-log where replay |
| 611 | // would stop forever. Callers must hold the session file lock. The event |
| 612 | // index's LogSize doubles as a cheap intact check so the common case never |
| 613 | // re-reads the log. |
| 614 | func repairSessionEventLogTail(sessionPath string) error { |
| 615 | path := store.SessionEventLog(sessionPath) |
| 616 | if path == "" { |
| 617 | return nil |
| 618 | } |
| 619 | info, err := os.Stat(path) |
| 620 | if err != nil { |
| 621 | if os.IsNotExist(err) { |
| 622 | return nil |
| 623 | } |
| 624 | return err |
| 625 | } |
| 626 | if info.IsDir() || info.Size() == 0 { |
| 627 | return nil |
| 628 | } |
| 629 | if idx, err := readSessionEventIndex(sessionPath); err == nil && idx != nil && idx.LogSize == info.Size() { |
| 630 | return nil |
| 631 | } |
| 632 | replay, err := replaySessionEventLog(path) |
| 633 | if err != nil { |
| 634 | return err |
| 635 | } |
| 636 | if replay.lastGoodEnd >= replay.size { |
| 637 | return nil |
| 638 | } |
| 639 | // Salvage the bytes the truncation below discards. A torn tail is usually |
| 640 | // one partial record, but replay also stops at a buried undecodable or |
| 641 | // out-of-order record (e.g. two runtimes interleaving appends on one log) — |
| 642 | // then everything past it, including intact turns, would be silently and |
| 643 | // permanently lost (#6607). Preservation is best-effort: it must not block |
| 644 | // the repair (the log has to become appendable again either way), and its |
| 645 | // most likely failure — a full disk — is the same condition that tears |
| 646 | // tails in the first place. |
| 647 | if preserveErr := preserveDamagedEventLogTail(sessionPath, path, replay.lastGoodEnd, replay.size); preserveErr != nil { |
| 648 | slog.Warn("session: could not preserve damaged event log tail; truncating anyway", |
| 649 | "path", path, "from", replay.lastGoodEnd, "size", replay.size, "err", preserveErr) |
| 650 | } |
| 651 | if err := os.Truncate(path, replay.lastGoodEnd); err != nil { |
| 652 | return err |
| 653 | } |
| 654 | if replay.lastGoodEnd == 0 { |
| 655 | return nil |
| 656 | } |
| 657 | // The truncation point sits exactly at the end of a JSON value; restore |
| 658 | // the trailing newline so the file stays line-oriented for external tools. |
| 659 | f, err := os.OpenFile(path, os.O_WRONLY|os.O_APPEND, 0o600) |
| 660 | if err != nil { |
| 661 | return err |
| 662 | } |
| 663 | if err := f.Chmod(0o600); err != nil { |
| 664 | _ = f.Close() |
| 665 | return err |
| 666 | } |
| 667 | if _, err := f.Write([]byte{'\n'}); err != nil { |
| 668 | f.Close() |
| 669 | return err |
| 670 | } |
| 671 | return f.Close() |
| 672 | } |
| 673 | |
| 674 | // preserveDamagedEventLogTail appends the about-to-be-truncated byte range of |
| 675 | // the event log to the .damaged salvage sidecar, prefixed with a one-line JSON |
| 676 | // header recording when and where the bytes came from. The sidecar is a |
| 677 | // forensic artifact for recovery, never replayed by the loader, and is removed |
| 678 | // with the session's other sidecars on delete. |
| 679 | func preserveDamagedEventLogTail(sessionPath, logPath string, from, to int64) error { |
| 680 | if to <= from { |
| 681 | return nil |
| 682 | } |
| 683 | src, err := os.Open(logPath) |
| 684 | if err != nil { |
| 685 | return err |
| 686 | } |
| 687 | defer src.Close() |
| 688 | if _, err := src.Seek(from, io.SeekStart); err != nil { |
| 689 | return err |
| 690 | } |
| 691 | dst, err := os.OpenFile(store.SessionEventLogDamaged(sessionPath), os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0o600) |
| 692 | if err != nil { |
| 693 | return err |
| 694 | } |
| 695 | header := fmt.Sprintf("{\"damaged_tail\":true,\"preserved_at\":%q,\"log_offset\":%d,\"bytes\":%d}\n", |
| 696 | time.Now().UTC().Format(time.RFC3339), from, to-from) |
| 697 | if _, err := dst.WriteString(header); err != nil { |
| 698 | dst.Close() |
| 699 | return err |
| 700 | } |
| 701 | if _, err := io.CopyN(dst, src, to-from); err != nil && !errors.Is(err, io.EOF) { |
| 702 | dst.Close() |
| 703 | return err |
| 704 | } |
| 705 | if _, err := dst.WriteString("\n"); err != nil { |
| 706 | dst.Close() |
| 707 | return err |
| 708 | } |
| 709 | return dst.Close() |
| 710 | } |
| 711 | |
| 712 | func appendSessionEvent(sessionPath string, rec sessionEventRecord, sync bool) error { |
| 713 | path := store.SessionEventLog(sessionPath) |
| 714 | if path == "" { |
| 715 | return fmt.Errorf("empty session event log path") |
| 716 | } |
| 717 | if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil { |
| 718 | return err |
| 719 | } |
| 720 | rec.SchemaVersion = sessionEventSchemaVersion |
| 721 | if rec.CreatedAt.IsZero() { |
| 722 | rec.CreatedAt = time.Now().UTC() |
| 723 | } |
| 724 | if rec.WriterID == "" { |
| 725 | rec.WriterID = SessionWriterID() |
| 726 | } |
| 727 | buf, err := json.Marshal(rec) |
| 728 | if err != nil { |
| 729 | return fmt.Errorf("encode session event: %w", err) |
| 730 | } |
| 731 | buf = append(buf, '\n') |
| 732 | f, err := os.OpenFile(path, os.O_CREATE|os.O_WRONLY|os.O_APPEND, 0o600) |
| 733 | if err != nil { |
| 734 | return fmt.Errorf("open session event log: %w", err) |
| 735 | } |
| 736 | // The event log carries the complete transcript. Chmod after opening so |
| 737 | // upgrading a pre-v0.53-boundary 0644 sidecar tightens the existing inode |
| 738 | // before any unredacted message is appended; OpenFile's perm only applies |
| 739 | // when the file is newly created. |
| 740 | if err := f.Chmod(0o600); err != nil { |
| 741 | _ = f.Close() |
| 742 | return fmt.Errorf("protect session event log: %w", err) |
| 743 | } |
| 744 | if _, err := f.Write(buf); err != nil { |
| 745 | _ = f.Close() |
| 746 | return fmt.Errorf("append session event: %w", err) |
| 747 | } |
| 748 | if sync { |
| 749 | if err := f.Sync(); err != nil { |
| 750 | _ = f.Close() |
| 751 | return err |
| 752 | } |
| 753 | } |
| 754 | return f.Close() |
| 755 | } |
| 756 | |
| 757 | func appendSessionReplaceEvent(sessionPath string, msgs []provider.Message, digest [sha256.Size]byte, baseRevision int64, reason string) error { |
| 758 | // Replace events carry the whole transcript and mark intentional history |
| 759 | // rewrites; they are rare and fsynced so a power cut cannot lose one. |
| 760 | return appendSessionEvent(sessionPath, sessionEventRecord{ |
| 761 | Type: sessionEventTypeReplace, |
| 762 | Revision: baseRevision + 1, |
| 763 | BaseRevision: baseRevision, |
| 764 | MessageIndex: 0, |
| 765 | Messages: append([]provider.Message(nil), msgs...), |
| 766 | ContentDigest: digestString(digest), |
| 767 | Reason: reason, |
| 768 | }, true) |
| 769 | } |
| 770 | |
| 771 | func appendSessionAppendEvent(sessionPath string, messageIndex int, msgs []provider.Message, digest [sha256.Size]byte, baseRevision int64) error { |
| 772 | if len(msgs) == 0 { |
| 773 | return nil |
| 774 | } |
| 775 | return appendSessionEvent(sessionPath, sessionEventRecord{ |
| 776 | Type: sessionEventTypeAppend, |
| 777 | Revision: baseRevision + 1, |
| 778 | BaseRevision: baseRevision, |
| 779 | MessageIndex: messageIndex, |
| 780 | Messages: append([]provider.Message(nil), msgs...), |
| 781 | ContentDigest: digestString(digest), |
| 782 | }, true) |
| 783 | } |
| 784 | |
| 785 | // compactSessionEventLog rewrites the log as a single replace event via an |
| 786 | // atomic tmp+fsync+rename, so readers observe either the old log or the |
| 787 | // compacted one and never a partial state. It also heals a damaged log by |
| 788 | // construction. |
| 789 | func compactSessionEventLog(sessionPath string, msgs []provider.Message, digest [sha256.Size]byte, baseRevision int64, reason string) error { |
| 790 | path := store.SessionEventLog(sessionPath) |
| 791 | if path == "" { |
| 792 | return fmt.Errorf("empty session event log path") |
| 793 | } |
| 794 | rec := sessionEventRecord{ |
| 795 | SchemaVersion: sessionEventSchemaVersion, |
| 796 | Type: sessionEventTypeReplace, |
| 797 | Revision: baseRevision + 1, |
| 798 | BaseRevision: baseRevision, |
| 799 | Messages: append([]provider.Message(nil), msgs...), |
| 800 | ContentDigest: digestString(digest), |
| 801 | WriterID: SessionWriterID(), |
| 802 | Reason: reason, |
| 803 | CreatedAt: time.Now().UTC(), |
| 804 | } |
| 805 | buf, err := json.Marshal(rec) |
| 806 | if err != nil { |
| 807 | return fmt.Errorf("encode session event: %w", err) |
| 808 | } |
| 809 | buf = append(buf, '\n') |
| 810 | return fileutil.AtomicWriteFile(path, buf, 0o600) |
| 811 | } |
| 812 | |
| 813 | func readSessionEventIndex(sessionPath string) (*sessionEventIndex, error) { |
| 814 | path := store.SessionEventIndex(sessionPath) |
| 815 | if path == "" { |
| 816 | return nil, nil |
| 817 | } |
| 818 | b, err := fileencoding.ReadFileUTF8(path) |
| 819 | if err != nil { |
| 820 | return nil, err |
| 821 | } |
| 822 | var idx sessionEventIndex |
| 823 | if err := json.Unmarshal(b, &idx); err != nil { |
| 824 | return nil, err |
| 825 | } |
| 826 | if idx.SchemaVersion != sessionEventSchemaVersion { |
| 827 | return nil, fmt.Errorf("unsupported session event index schema %d", idx.SchemaVersion) |
| 828 | } |
| 829 | return &idx, nil |
| 830 | } |
| 831 | |
| 832 | func writeSessionEventIndex(path string, msgs []provider.Message, digest [sha256.Size]byte, revision int64) error { |
| 833 | indexPath := store.SessionEventIndex(path) |
| 834 | if indexPath == "" { |
| 835 | return nil |
| 836 | } |
| 837 | logInfo, err := os.Stat(store.SessionEventLog(path)) |
| 838 | if err != nil { |
| 839 | if os.IsNotExist(err) { |
| 840 | // No log means nothing for the index to describe; drop a stale |
| 841 | // index (e.g. after a force save folded the log away). |
| 842 | if err := os.Remove(indexPath); err != nil && !os.IsNotExist(err) { |
| 843 | return err |
| 844 | } |
| 845 | return nil |
| 846 | } |
| 847 | return err |
| 848 | } |
| 849 | idx := sessionEventIndex{ |
| 850 | SchemaVersion: sessionEventSchemaVersion, |
| 851 | LogSize: logInfo.Size(), |
| 852 | MessageCount: len(msgs), |
| 853 | Revision: revision, |
| 854 | ContentDigest: digestString(digest), |
| 855 | WriterID: SessionWriterID(), |
| 856 | UpdatedAt: time.Now().UTC(), |
| 857 | } |
| 858 | b, err := json.MarshalIndent(idx, "", " ") |
| 859 | if err != nil { |
| 860 | return err |
| 861 | } |
| 862 | b = append(b, '\n') |
| 863 | if err := os.MkdirAll(filepath.Dir(indexPath), 0o755); err != nil { |
| 864 | return err |
| 865 | } |
| 866 | tmp, err := os.CreateTemp(filepath.Dir(indexPath), ".session-event-index.*.tmp") |
| 867 | if err != nil { |
| 868 | return err |
| 869 | } |
| 870 | tmpPath := tmp.Name() |
| 871 | if _, err := tmp.Write(b); err != nil { |
| 872 | tmp.Close() |
| 873 | os.Remove(tmpPath) |
| 874 | return err |
| 875 | } |
| 876 | if err := tmp.Close(); err != nil { |
| 877 | os.Remove(tmpPath) |
| 878 | return err |
| 879 | } |
| 880 | if err := fileutil.ReplaceFile(tmpPath, indexPath); err != nil { |
| 881 | os.Remove(tmpPath) |
| 882 | return err |
| 883 | } |
| 884 | return nil |
| 885 | } |
| 886 |