| 1 | package control |
| 2 | |
| 3 | import ( |
| 4 | "encoding/json" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "log/slog" |
| 8 | "sync/atomic" |
| 9 | |
| 10 | "reasonix/internal/checkpoint" |
| 11 | "reasonix/internal/diff" |
| 12 | "reasonix/internal/event" |
| 13 | "reasonix/internal/provider" |
| 14 | ) |
| 15 | |
| 16 | // ErrRewindCoverageConfirmationRequired is returned by the compatibility |
| 17 | // Rewind path when restoring files from a partially covered checkpoint. New |
| 18 | // callers should preview with PrepareRewind, show the coverage warning, and |
| 19 | // commit only after the user explicitly confirms it. |
| 20 | var ErrRewindCoverageConfirmationRequired = errors.New("partial checkpoint coverage requires explicit confirmation") |
| 21 | |
| 22 | // RewindPlanRequiresConfirmation reports whether a prepared plan can restore |
| 23 | // files but cannot guarantee that every workspace mutation was captured. |
| 24 | func RewindPlanRequiresConfirmation(plan checkpoint.RewindPlan) bool { |
| 25 | wantsFiles := plan.Scope == checkpoint.RewindCode || plan.Scope == checkpoint.RewindBoth |
| 26 | return wantsFiles && plan.CanFiles && (plan.Coverage == checkpoint.CoveragePartial || len(plan.CoverageGaps) > 0) |
| 27 | } |
| 28 | |
| 29 | // conversationApplier bridges checkpoint transactions to controller session state. |
| 30 | type conversationApplier struct { |
| 31 | c *Controller |
| 32 | } |
| 33 | |
| 34 | func (a conversationApplier) ApplyConversationTruncate(boundary int, forward []byte) error { |
| 35 | c := a.c |
| 36 | if c.executor == nil { |
| 37 | return fmt.Errorf("executor unavailable") |
| 38 | } |
| 39 | s := c.executor.Session() |
| 40 | msgs := s.Snapshot() |
| 41 | if boundary > len(msgs) { |
| 42 | return fmt.Errorf("conversation rewind unavailable: the conversation was compacted past this point") |
| 43 | } |
| 44 | if len(forward) == 0 { |
| 45 | var err error |
| 46 | forward, err = json.Marshal(msgs) |
| 47 | if err != nil { |
| 48 | return err |
| 49 | } |
| 50 | } |
| 51 | s.Rewrite(msgs[:boundary], "rewind_truncate") |
| 52 | if err := c.SnapshotRewrite(); err != nil { |
| 53 | _ = a.RestoreConversation(forward) |
| 54 | return fmt.Errorf("persist conversation after rewind: %w", err) |
| 55 | } |
| 56 | return nil |
| 57 | } |
| 58 | |
| 59 | func (a conversationApplier) RestoreConversation(forward []byte) error { |
| 60 | c := a.c |
| 61 | if c.executor == nil { |
| 62 | return fmt.Errorf("executor unavailable") |
| 63 | } |
| 64 | var msgs []provider.Message |
| 65 | if err := json.Unmarshal(forward, &msgs); err != nil { |
| 66 | return err |
| 67 | } |
| 68 | c.executor.Session().Rewrite(msgs, "rewind_restore") |
| 69 | if err := c.SnapshotRewrite(); err != nil { |
| 70 | return fmt.Errorf("restore conversation: %w", err) |
| 71 | } |
| 72 | return nil |
| 73 | } |
| 74 | |
| 75 | func (a conversationApplier) TruncateCheckpoints(fromTurn int) error { |
| 76 | return a.c.checkpoints.truncateFrom(fromTurn) |
| 77 | } |
| 78 | |
| 79 | func (a conversationApplier) RestoreCheckpoints(backup []byte) error { |
| 80 | store := a.c.checkpoints.storeRef() |
| 81 | if store == nil { |
| 82 | return fmt.Errorf("checkpoints unavailable") |
| 83 | } |
| 84 | if err := store.RestoreCheckpointBackupPublic(backup); err != nil { |
| 85 | return err |
| 86 | } |
| 87 | bounds := store.Bounds() |
| 88 | a.c.checkpoints.mu.Lock() |
| 89 | a.c.checkpoints.bound = bounds |
| 90 | a.c.checkpoints.turn = store.NextTurn() |
| 91 | a.c.checkpoints.mu.Unlock() |
| 92 | return nil |
| 93 | } |
| 94 | |
| 95 | // PrepareRewind validates that a rewind can proceed without mutating state. |
| 96 | func (c *Controller) PrepareRewind(turn int, scope RewindScope) (checkpoint.RewindPlan, error) { |
| 97 | if !c.checkpoints.enabled() || c.executor == nil { |
| 98 | return checkpoint.RewindPlan{}, c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 99 | } |
| 100 | if err := c.beginRotation(); err != nil { |
| 101 | if errors.Is(err, errTurnRunningRotation) { |
| 102 | return checkpoint.RewindPlan{}, c.rewindFail(fmt.Errorf("cannot rewind while a turn is running")) |
| 103 | } |
| 104 | return checkpoint.RewindPlan{}, c.rewindFail(err) |
| 105 | } |
| 106 | // Release rotation before file precheck I/O. |
| 107 | c.endRotation() |
| 108 | |
| 109 | boundary, hasBound := c.checkpoints.boundary(turn) |
| 110 | store := c.checkpoints.storeRef() |
| 111 | if store == nil { |
| 112 | return checkpoint.RewindPlan{}, c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 113 | } |
| 114 | if obs := c.mutationObserver; obs != nil { |
| 115 | store.SetActiveWriters(obs.ActiveWriters()) |
| 116 | } |
| 117 | rev := atomic.LoadInt64(&c.sessionRevision) |
| 118 | plan, err := store.PrepareRewind(turn, checkpoint.RewindScope(scope), rev, boundary, hasBound) |
| 119 | if err != nil { |
| 120 | return plan, c.rewindFail(err) |
| 121 | } |
| 122 | if scope == RewindBoth && !plan.CanConversation { |
| 123 | plan.CanFiles = false |
| 124 | if plan.DisabledReason == "" { |
| 125 | plan.DisabledReason = "conversation boundary unavailable" |
| 126 | } |
| 127 | } |
| 128 | return plan, nil |
| 129 | } |
| 130 | |
| 131 | // CommitRewind executes a prepared plan under rotation gate + mutation barrier. |
| 132 | func (c *Controller) CommitRewind(planID string) (checkpoint.RewindResult, error) { |
| 133 | if !c.checkpoints.enabled() || c.executor == nil { |
| 134 | return checkpoint.RewindResult{}, c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 135 | } |
| 136 | if err := c.beginRotation(); err != nil { |
| 137 | if errors.Is(err, errTurnRunningRotation) { |
| 138 | return checkpoint.RewindResult{}, c.rewindFail(fmt.Errorf("cannot rewind while a turn is running")) |
| 139 | } |
| 140 | return checkpoint.RewindResult{}, c.rewindFail(err) |
| 141 | } |
| 142 | defer c.endRotation() |
| 143 | |
| 144 | store := c.checkpoints.storeRef() |
| 145 | if store == nil { |
| 146 | return checkpoint.RewindResult{}, c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 147 | } |
| 148 | if err := store.ValidatePlanSessionRevision(planID, atomic.LoadInt64(&c.sessionRevision)); err != nil { |
| 149 | conflict := checkpoint.RewindConflict{Reason: checkpoint.ConflictStalePlan} |
| 150 | return checkpoint.RewindResult{OK: false, Error: err.Error(), Conflicts: []checkpoint.RewindConflict{conflict}}, c.rewindFail(err) |
| 151 | } |
| 152 | |
| 153 | forward, err := json.Marshal(c.executor.Session().Snapshot()) |
| 154 | if err != nil { |
| 155 | return checkpoint.RewindResult{}, c.rewindFail(err) |
| 156 | } |
| 157 | |
| 158 | result, err := store.CommitRewindWithForward(planID, forward, conversationApplier{c: c}, nil) |
| 159 | if err != nil { |
| 160 | return result, c.rewindFail(err) |
| 161 | } |
| 162 | if result.OK { |
| 163 | if len(result.Written) > 0 || len(result.Deleted) > 0 { |
| 164 | c.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, |
| 165 | Text: fmt.Sprintf("rewound code — %d file(s) restored, %d removed", len(result.Written), len(result.Deleted))}) |
| 166 | } |
| 167 | if result.ConversationOK { |
| 168 | c.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, |
| 169 | Text: "rewound conversation"}) |
| 170 | } |
| 171 | atomic.AddInt64(&c.sessionRevision, 1) |
| 172 | } |
| 173 | return result, nil |
| 174 | } |
| 175 | |
| 176 | // UndoRewind reverses the last committed rewind transaction when still available. |
| 177 | func (c *Controller) UndoRewind(transactionID string) (checkpoint.RewindResult, error) { |
| 178 | if !c.checkpoints.enabled() || c.executor == nil { |
| 179 | return checkpoint.RewindResult{}, c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 180 | } |
| 181 | if err := c.beginRotation(); err != nil { |
| 182 | if errors.Is(err, errTurnRunningRotation) { |
| 183 | return checkpoint.RewindResult{}, c.rewindFail(fmt.Errorf("cannot undo rewind while a turn is running")) |
| 184 | } |
| 185 | return checkpoint.RewindResult{}, c.rewindFail(err) |
| 186 | } |
| 187 | defer c.endRotation() |
| 188 | |
| 189 | store := c.checkpoints.storeRef() |
| 190 | if store == nil { |
| 191 | return checkpoint.RewindResult{}, c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 192 | } |
| 193 | result, err := store.UndoRewind(transactionID, conversationApplier{c: c}) |
| 194 | if err != nil { |
| 195 | return result, c.rewindFail(err) |
| 196 | } |
| 197 | if result.OK { |
| 198 | atomic.AddInt64(&c.sessionRevision, 1) |
| 199 | c.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, Text: "undid last rewind"}) |
| 200 | } |
| 201 | return result, nil |
| 202 | } |
| 203 | |
| 204 | // PrepareFileRevert prepares a single-file restore to the session's first-touch preimage. |
| 205 | func (c *Controller) PrepareFileRevert(path string) (checkpoint.RewindPlan, error) { |
| 206 | if !c.checkpoints.enabled() || c.executor == nil { |
| 207 | return checkpoint.RewindPlan{}, c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 208 | } |
| 209 | store := c.checkpoints.storeRef() |
| 210 | if store == nil { |
| 211 | return checkpoint.RewindPlan{}, c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 212 | } |
| 213 | state, ok := store.FileState(path) |
| 214 | if !ok { |
| 215 | return checkpoint.RewindPlan{ |
| 216 | Path: path, CanFiles: false, DisabledReason: "file is not session-owned", |
| 217 | }, nil |
| 218 | } |
| 219 | _ = state |
| 220 | return store.PrepareFileRevert(path, atomic.LoadInt64(&c.sessionRevision)) |
| 221 | } |
| 222 | |
| 223 | // CommitFileRevert commits a single-file restore with optional conflict resolution. |
| 224 | func (c *Controller) CommitFileRevert(planID string, resolution checkpoint.ConflictResolution) (checkpoint.RewindResult, error) { |
| 225 | if !c.checkpoints.enabled() || c.executor == nil { |
| 226 | return checkpoint.RewindResult{}, c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 227 | } |
| 228 | if err := c.beginRotation(); err != nil { |
| 229 | if errors.Is(err, errTurnRunningRotation) { |
| 230 | return checkpoint.RewindResult{}, c.rewindFail(fmt.Errorf("cannot revert file while a turn is running")) |
| 231 | } |
| 232 | return checkpoint.RewindResult{}, c.rewindFail(err) |
| 233 | } |
| 234 | defer c.endRotation() |
| 235 | |
| 236 | store := c.checkpoints.storeRef() |
| 237 | if store == nil { |
| 238 | return checkpoint.RewindResult{}, c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 239 | } |
| 240 | if err := store.ValidatePlanSessionRevision(planID, atomic.LoadInt64(&c.sessionRevision)); err != nil { |
| 241 | conflict := checkpoint.RewindConflict{Reason: checkpoint.ConflictStalePlan} |
| 242 | return checkpoint.RewindResult{OK: false, Error: err.Error(), Conflicts: []checkpoint.RewindConflict{conflict}}, c.rewindFail(err) |
| 243 | } |
| 244 | result, err := store.CommitFileRevert(planID, resolution) |
| 245 | if err != nil { |
| 246 | return result, c.rewindFail(err) |
| 247 | } |
| 248 | if result.OK { |
| 249 | atomic.AddInt64(&c.sessionRevision, 1) |
| 250 | } |
| 251 | return result, nil |
| 252 | } |
| 253 | |
| 254 | // Rewind is the compatibility wrapper used by CLI and existing desktop paths. |
| 255 | // Conversation failures never leave files half-applied for both-scope: files are |
| 256 | // captured first, restored second, and conversation is persisted last with full |
| 257 | // compensation on failure. |
| 258 | func (c *Controller) Rewind(turn int, scope RewindScope) error { |
| 259 | if !c.checkpoints.enabled() || c.executor == nil { |
| 260 | return c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 261 | } |
| 262 | if err := c.beginRotation(); err != nil { |
| 263 | if errors.Is(err, errTurnRunningRotation) { |
| 264 | return c.rewindFail(fmt.Errorf("cannot rewind while a turn is running")) |
| 265 | } |
| 266 | return c.rewindFail(err) |
| 267 | } |
| 268 | defer c.endRotation() |
| 269 | |
| 270 | boundary, hasBound := c.checkpoints.boundary(turn) |
| 271 | var forward []byte |
| 272 | if scope == RewindConversation || scope == RewindBoth { |
| 273 | if !hasBound { |
| 274 | return c.rewindFail(fmt.Errorf("conversation rewind unavailable for turn %d (resumed session)", turn)) |
| 275 | } |
| 276 | msgs := c.executor.Session().Snapshot() |
| 277 | if boundary > len(msgs) { |
| 278 | return c.rewindFail(fmt.Errorf("conversation rewind unavailable for turn %d: the conversation was compacted past this point", turn)) |
| 279 | } |
| 280 | var err error |
| 281 | forward, err = json.Marshal(msgs) |
| 282 | if err != nil { |
| 283 | return c.rewindFail(err) |
| 284 | } |
| 285 | } |
| 286 | store := c.checkpoints.storeRef() |
| 287 | if store == nil { |
| 288 | return c.rewindFail(fmt.Errorf("checkpoints unavailable")) |
| 289 | } |
| 290 | if obs := c.mutationObserver; obs != nil { |
| 291 | store.SetActiveWriters(obs.ActiveWriters()) |
| 292 | } |
| 293 | rev := atomic.LoadInt64(&c.sessionRevision) |
| 294 | plan, err := store.PrepareRewind(turn, checkpoint.RewindScope(scope), rev, boundary, hasBound) |
| 295 | if err != nil { |
| 296 | return c.rewindFail(err) |
| 297 | } |
| 298 | if (scope == RewindCode || scope == RewindBoth) && !plan.CanFiles { |
| 299 | return c.rewindFail(fmt.Errorf("%s", plan.DisabledReason)) |
| 300 | } |
| 301 | if (scope == RewindConversation || scope == RewindBoth) && !plan.CanConversation { |
| 302 | return c.rewindFail(fmt.Errorf("%s", plan.DisabledReason)) |
| 303 | } |
| 304 | if RewindPlanRequiresConfirmation(plan) { |
| 305 | return c.rewindFail(fmt.Errorf("%w (%d coverage gap(s))", ErrRewindCoverageConfirmationRequired, len(plan.CoverageGaps))) |
| 306 | } |
| 307 | if forward == nil { |
| 308 | forward, err = json.Marshal(c.executor.Session().Snapshot()) |
| 309 | if err != nil { |
| 310 | return c.rewindFail(err) |
| 311 | } |
| 312 | } |
| 313 | result, err := store.CommitRewindWithForward(plan.PlanID, forward, conversationApplier{c: c}, nil) |
| 314 | if err != nil { |
| 315 | return c.rewindFail(err) |
| 316 | } |
| 317 | if len(result.Written) > 0 || len(result.Deleted) > 0 { |
| 318 | c.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, |
| 319 | Text: fmt.Sprintf("rewound code to turn %d — %d file(s) restored, %d removed", turn, len(result.Written), len(result.Deleted))}) |
| 320 | } |
| 321 | if result.ConversationOK { |
| 322 | c.sink.Emit(event.Event{Kind: event.Notice, Level: event.LevelInfo, |
| 323 | Text: fmt.Sprintf("rewound conversation to turn %d", turn)}) |
| 324 | } |
| 325 | atomic.AddInt64(&c.sessionRevision, 1) |
| 326 | return nil |
| 327 | } |
| 328 | |
| 329 | func (c *Controller) recoverCheckpointTransactions() { |
| 330 | store := c.checkpoints.storeRef() |
| 331 | if store == nil || c.executor == nil { |
| 332 | return |
| 333 | } |
| 334 | for _, note := range store.RecoverTransactionsWithApplier(conversationApplier{c: c}) { |
| 335 | slog.Info("controller: checkpoint transaction recovery", "result", note) |
| 336 | } |
| 337 | } |
| 338 | |
| 339 | // wireMutationObserver installs the v2 observer on the executor. |
| 340 | func (c *Controller) wireMutationObserver() { |
| 341 | store := c.checkpoints.storeRef() |
| 342 | if store == nil || c.executor == nil { |
| 343 | return |
| 344 | } |
| 345 | obs := checkpoint.NewMutationObserver(checkpoint.ObserverOptions{ |
| 346 | Store: store, |
| 347 | WriterID: "root", |
| 348 | }) |
| 349 | c.mutationObserver = obs |
| 350 | c.executor.SetMutationObserver(obs) |
| 351 | // Keep legacy pre-edit hook as a secondary path when observer is absent on |
| 352 | // a cloned agent; with observer set, BeforeMutation is preferred. |
| 353 | c.executor.SetPreEditHook(func(ch diff.Change) { |
| 354 | if c.mutationObserver != nil { |
| 355 | c.mutationObserver.BeforeMutationFromChange(ch, "legacy_hook") |
| 356 | return |
| 357 | } |
| 358 | c.checkpoints.snapshot(ch) |
| 359 | }) |
| 360 | } |
| 361 |