| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "fmt" |
| 6 | "sync" |
| 7 | ) |
| 8 | |
| 9 | // SubagentSlotStatus is the queue lifecycle shown for background task/fleet |
| 10 | // items that share the session scheduler. |
| 11 | type SubagentSlotStatus string |
| 12 | |
| 13 | const ( |
| 14 | SubagentSlotQueued SubagentSlotStatus = "queued" |
| 15 | SubagentSlotRunning SubagentSlotStatus = "running" |
| 16 | SubagentSlotDone SubagentSlotStatus = "done" |
| 17 | SubagentSlotFailed SubagentSlotStatus = "failed" |
| 18 | ) |
| 19 | |
| 20 | // AcquireRequest describes a sub-agent slot request against the session pool. |
| 21 | type AcquireRequest struct { |
| 22 | // Writer is true for writer-capable runs (task without read_only, profile |
| 23 | // that is not read-only, fleet items that can write). |
| 24 | Writer bool |
| 25 | // WritePaths is the claim held while the slot is active. Empty for |
| 26 | // read-only work. Whole-workspace claims count as writers and serialize |
| 27 | // against every other writer. |
| 28 | WritePaths WritePathSet |
| 29 | // Nested fails immediately when no capacity is free instead of queueing. |
| 30 | // Nested sub-agents must not block waiting for a parent-held slot. |
| 31 | Nested bool |
| 32 | // Label is optional diagnostics text. |
| 33 | Label string |
| 34 | } |
| 35 | |
| 36 | // SubagentScheduler is a session-scoped concurrency controller shared by task, |
| 37 | // fleet, parallel_tasks, profile skills, and nested sub-agents. |
| 38 | type SubagentScheduler struct { |
| 39 | mu sync.Mutex |
| 40 | |
| 41 | maxTotal int |
| 42 | maxWriters int |
| 43 | |
| 44 | activeTotal int |
| 45 | activeWriters int |
| 46 | activeLive []liveClaim |
| 47 | nextClaimID int64 |
| 48 | // parentClaims are write paths held by the parent agent during a write-tool |
| 49 | // Execute. They block overlapping subagent claims without consuming a |
| 50 | // subagent concurrency slot (parent is not a subagent). |
| 51 | parentClaims []WritePathSet |
| 52 | |
| 53 | // waiters are FIFO waiters for non-nested acquires. |
| 54 | waiters []*schedulerWaiter |
| 55 | } |
| 56 | |
| 57 | type schedulerWaiter struct { |
| 58 | req AcquireRequest |
| 59 | ready chan struct{} |
| 60 | failed error |
| 61 | id int64 |
| 62 | } |
| 63 | |
| 64 | // NewSubagentScheduler builds a scheduler with the given limits (normalized). |
| 65 | func NewSubagentScheduler(maxTotal, maxWriters int) *SubagentScheduler { |
| 66 | maxTotal, maxWriters = NormalizeConcurrencyLimits(maxTotal, maxWriters) |
| 67 | return &SubagentScheduler{maxTotal: maxTotal, maxWriters: maxWriters} |
| 68 | } |
| 69 | |
| 70 | // Limits returns the effective total/writer caps. |
| 71 | func (s *SubagentScheduler) Limits() (total, writers int) { |
| 72 | if s == nil { |
| 73 | return DefaultMaxSubagentConcurrency, DefaultMaxParallelWriters |
| 74 | } |
| 75 | return s.maxTotal, s.maxWriters |
| 76 | } |
| 77 | |
| 78 | // Acquire reserves a concurrency slot (and optional write claim). Nested |
| 79 | // requests fail immediately when capacity is exhausted. Non-nested requests |
| 80 | // queue until capacity is free or ctx is cancelled. |
| 81 | // |
| 82 | // The returned release function must be called exactly once when the sub-agent |
| 83 | // finishes. release is safe to call even if Acquire returns an error (no-op). |
| 84 | func (s *SubagentScheduler) Acquire(ctx context.Context, req AcquireRequest) (release func(), err error) { |
| 85 | release, _, err = s.AcquireWithID(ctx, req) |
| 86 | return release, err |
| 87 | } |
| 88 | |
| 89 | // AcquireWithID is Acquire plus the live claim id used by Realize/MarkOpaque. |
| 90 | func (s *SubagentScheduler) AcquireWithID(ctx context.Context, req AcquireRequest) (release func(), claimID int64, err error) { |
| 91 | noop := func() {} |
| 92 | if s == nil { |
| 93 | return noop, 0, nil |
| 94 | } |
| 95 | if ctx == nil { |
| 96 | ctx = context.Background() |
| 97 | } |
| 98 | |
| 99 | s.mu.Lock() |
| 100 | if ok, reason := s.canStartIncomingLocked(req); ok { |
| 101 | id := s.activateLocked(req) |
| 102 | s.mu.Unlock() |
| 103 | return s.makeReleaseID(id), id, nil |
| 104 | } else if req.Nested { |
| 105 | s.mu.Unlock() |
| 106 | return noop, 0, fmt.Errorf("subagent concurrency limit reached (%s); nested subagents fail fast to avoid parent/child slot deadlock", reason) |
| 107 | } |
| 108 | |
| 109 | w := &schedulerWaiter{req: req, ready: make(chan struct{})} |
| 110 | s.waiters = append(s.waiters, w) |
| 111 | s.mu.Unlock() |
| 112 | |
| 113 | select { |
| 114 | case <-w.ready: |
| 115 | if w.failed != nil { |
| 116 | return noop, 0, w.failed |
| 117 | } |
| 118 | return s.makeReleaseID(w.id), w.id, nil |
| 119 | case <-ctx.Done(): |
| 120 | s.mu.Lock() |
| 121 | s.removeWaiterLocked(w) |
| 122 | s.pumpWaitersLocked() |
| 123 | s.mu.Unlock() |
| 124 | select { |
| 125 | case <-w.ready: |
| 126 | if w.failed == nil { |
| 127 | s.makeReleaseID(w.id)() |
| 128 | } |
| 129 | default: |
| 130 | } |
| 131 | return noop, 0, ctx.Err() |
| 132 | } |
| 133 | } |
| 134 | |
| 135 | // TryClaimWritePaths checks whether paths conflict with active claims without |
| 136 | // taking a concurrency slot. Used for diagnostics; prefer ReserveParentWrite |
| 137 | // for parent agent writes so the check is not TOCTOU with subagent Acquire. |
| 138 | func (s *SubagentScheduler) TryClaimWritePaths(paths WritePathSet) error { |
| 139 | if s == nil || paths.Empty() { |
| 140 | return nil |
| 141 | } |
| 142 | s.mu.Lock() |
| 143 | defer s.mu.Unlock() |
| 144 | return s.conflictLocked(paths) |
| 145 | } |
| 146 | |
| 147 | // Realize records path-bound writes against an active claim. Directory and |
| 148 | // whole-workspace declarations shrink to the realized files when no opaque |
| 149 | // mutation has occurred. Same-file realizes from two live writers fail. |
| 150 | func (s *SubagentScheduler) Realize(id int64, paths WritePathSet) error { |
| 151 | if s == nil || id == 0 || paths.Empty() { |
| 152 | return nil |
| 153 | } |
| 154 | s.mu.Lock() |
| 155 | defer s.mu.Unlock() |
| 156 | idx := s.liveIndexLocked(id) |
| 157 | if idx < 0 { |
| 158 | return fmt.Errorf("write path is claimed by a running background subagent; wait for it to finish before writing the same path") |
| 159 | } |
| 160 | claim := s.activeLive[idx] |
| 161 | if claim.opaque { |
| 162 | return nil |
| 163 | } |
| 164 | nextPaths := mergeRealized(claim.realized, paths) |
| 165 | next := fileReservation(claim.declared.WorkspaceRoot, nextPaths) |
| 166 | if err := s.conflictAgainstOthersLocked(id, next); err != nil { |
| 167 | return err |
| 168 | } |
| 169 | claim.realized = nextPaths |
| 170 | s.activeLive[idx] = claim |
| 171 | s.pumpWaitersLocked() |
| 172 | return nil |
| 173 | } |
| 174 | |
| 175 | // MarkOpaque upgrades a live claim to a whole-workspace reservation (bash/MCP). |
| 176 | func (s *SubagentScheduler) MarkOpaque(id int64) error { |
| 177 | if s == nil || id == 0 { |
| 178 | return nil |
| 179 | } |
| 180 | s.mu.Lock() |
| 181 | defer s.mu.Unlock() |
| 182 | idx := s.liveIndexLocked(id) |
| 183 | if idx < 0 { |
| 184 | return fmt.Errorf("write path is claimed by a running background subagent; wait for it to finish before writing the same path") |
| 185 | } |
| 186 | claim := s.activeLive[idx] |
| 187 | if claim.opaque { |
| 188 | return nil |
| 189 | } |
| 190 | next := wholeReservation(claim.declared.WorkspaceRoot) |
| 191 | if err := s.conflictAgainstOthersLocked(id, next); err != nil { |
| 192 | return err |
| 193 | } |
| 194 | claim.opaque = true |
| 195 | s.activeLive[idx] = claim |
| 196 | return nil |
| 197 | } |
| 198 | |
| 199 | // ReserveParentWrite holds paths against overlapping subagent claims for the |
| 200 | // duration of a parent write-tool Execute. It does not consume subagent |
| 201 | // concurrency slots. On conflict it fails immediately (parent cannot queue |
| 202 | // behind background jobs mid-tool-call). release must be called once when the |
| 203 | // write finishes so queued subagents can proceed. |
| 204 | func (s *SubagentScheduler) ReserveParentWrite(paths WritePathSet) (release func(), err error) { |
| 205 | noop := func() {} |
| 206 | if s == nil || paths.Empty() { |
| 207 | return noop, nil |
| 208 | } |
| 209 | s.mu.Lock() |
| 210 | if err := s.conflictLocked(paths); err != nil { |
| 211 | s.mu.Unlock() |
| 212 | return noop, err |
| 213 | } |
| 214 | s.parentClaims = append(s.parentClaims, paths) |
| 215 | s.mu.Unlock() |
| 216 | |
| 217 | var once sync.Once |
| 218 | return func() { |
| 219 | once.Do(func() { |
| 220 | s.mu.Lock() |
| 221 | s.parentClaims = removeClaim(s.parentClaims, paths) |
| 222 | s.pumpWaitersLocked() |
| 223 | s.mu.Unlock() |
| 224 | }) |
| 225 | }, nil |
| 226 | } |
| 227 | |
| 228 | // ActiveWriterClaims returns a snapshot of subagent + parent write claims. |
| 229 | func (s *SubagentScheduler) ActiveWriterClaims() []WritePathSet { |
| 230 | if s == nil { |
| 231 | return nil |
| 232 | } |
| 233 | s.mu.Lock() |
| 234 | defer s.mu.Unlock() |
| 235 | out := make([]WritePathSet, 0, len(s.activeLive)+len(s.parentClaims)) |
| 236 | for _, live := range s.activeLive { |
| 237 | if !live.writer { |
| 238 | continue |
| 239 | } |
| 240 | res := live.reservation() |
| 241 | if res.Empty() { |
| 242 | if live.declared.Empty() { |
| 243 | continue |
| 244 | } |
| 245 | res = live.declared |
| 246 | } |
| 247 | out = append(out, res) |
| 248 | } |
| 249 | out = append(out, s.parentClaims...) |
| 250 | return out |
| 251 | } |
| 252 | |
| 253 | func (s *SubagentScheduler) conflictLocked(paths WritePathSet) error { |
| 254 | if paths.WholeWorkspace { |
| 255 | for _, live := range s.activeLive { |
| 256 | if live.writer { |
| 257 | return fmt.Errorf("write path is claimed by a running background subagent; wait for it to finish before writing the same path") |
| 258 | } |
| 259 | } |
| 260 | } |
| 261 | return s.conflictAgainstOthersLocked(0, paths) |
| 262 | } |
| 263 | |
| 264 | func (s *SubagentScheduler) conflictAgainstOthersLocked(skipID int64, paths WritePathSet) error { |
| 265 | if paths.Empty() { |
| 266 | return nil |
| 267 | } |
| 268 | for _, live := range s.activeLive { |
| 269 | if live.id == skipID { |
| 270 | continue |
| 271 | } |
| 272 | if ScheduleOverlaps(live.reservation(), paths) { |
| 273 | return fmt.Errorf("write path is claimed by a running background subagent; wait for it to finish before writing the same path") |
| 274 | } |
| 275 | } |
| 276 | for _, active := range s.parentClaims { |
| 277 | if ScheduleOverlaps(active, paths) { |
| 278 | return fmt.Errorf("write path is claimed by another parent write in progress") |
| 279 | } |
| 280 | } |
| 281 | return nil |
| 282 | } |
| 283 | |
| 284 | func (s *SubagentScheduler) makeReleaseID(id int64) func() { |
| 285 | var once sync.Once |
| 286 | return func() { |
| 287 | once.Do(func() { |
| 288 | s.mu.Lock() |
| 289 | s.deactivateIDLocked(id) |
| 290 | s.pumpWaitersLocked() |
| 291 | s.mu.Unlock() |
| 292 | }) |
| 293 | } |
| 294 | } |
| 295 | |
| 296 | func (s *SubagentScheduler) liveIndexLocked(id int64) int { |
| 297 | for i, live := range s.activeLive { |
| 298 | if live.id == id { |
| 299 | return i |
| 300 | } |
| 301 | } |
| 302 | return -1 |
| 303 | } |
| 304 | |
| 305 | func (s *SubagentScheduler) canStartLocked(req AcquireRequest) (bool, string) { |
| 306 | if s.activeTotal >= s.maxTotal { |
| 307 | return false, fmt.Sprintf("total concurrency %d/%d", s.activeTotal, s.maxTotal) |
| 308 | } |
| 309 | if !req.Writer { |
| 310 | return true, "" |
| 311 | } |
| 312 | if s.activeWriters >= s.maxWriters { |
| 313 | return false, fmt.Sprintf("writer concurrency %d/%d", s.activeWriters, s.maxWriters) |
| 314 | } |
| 315 | if req.WritePaths.WholeWorkspace { |
| 316 | for _, live := range s.activeLive { |
| 317 | if live.writer { |
| 318 | return false, "whole-workspace claim conflicts with a running writer" |
| 319 | } |
| 320 | } |
| 321 | } |
| 322 | for _, live := range s.activeLive { |
| 323 | if ScheduleOverlaps(req.WritePaths, live.reservation()) { |
| 324 | return false, "write path conflict with a running subagent" |
| 325 | } |
| 326 | } |
| 327 | for _, active := range s.parentClaims { |
| 328 | if ScheduleOverlaps(req.WritePaths, active) { |
| 329 | return false, "write path conflict with a parent write in progress" |
| 330 | } |
| 331 | } |
| 332 | return true, "" |
| 333 | } |
| 334 | |
| 335 | func (s *SubagentScheduler) activateLocked(req AcquireRequest) int64 { |
| 336 | s.activeTotal++ |
| 337 | s.nextClaimID++ |
| 338 | id := s.nextClaimID |
| 339 | if req.Writer { |
| 340 | s.activeWriters++ |
| 341 | } |
| 342 | s.activeLive = append(s.activeLive, liveClaim{id: id, writer: req.Writer, declared: req.WritePaths}) |
| 343 | return id |
| 344 | } |
| 345 | |
| 346 | func (s *SubagentScheduler) deactivateIDLocked(id int64) { |
| 347 | idx := s.liveIndexLocked(id) |
| 348 | if idx < 0 { |
| 349 | return |
| 350 | } |
| 351 | if s.activeTotal > 0 { |
| 352 | s.activeTotal-- |
| 353 | } |
| 354 | if s.activeLive[idx].writer && s.activeWriters > 0 { |
| 355 | s.activeWriters-- |
| 356 | } |
| 357 | s.activeLive = append(s.activeLive[:idx], s.activeLive[idx+1:]...) |
| 358 | } |
| 359 | |
| 360 | func (s *SubagentScheduler) pumpWaitersLocked() { |
| 361 | if len(s.waiters) == 0 { |
| 362 | return |
| 363 | } |
| 364 | remaining := s.waiters[:0] |
| 365 | // A blocked whole-workspace writer is a FIFO barrier for later writers, |
| 366 | // while read-only work may still use otherwise available capacity. |
| 367 | wholeWriterPending := false |
| 368 | for _, w := range s.waiters { |
| 369 | if wholeWriterPending && w.req.Writer { |
| 370 | remaining = append(remaining, w) |
| 371 | continue |
| 372 | } |
| 373 | if ok, _ := s.canStartLocked(w.req); ok { |
| 374 | w.id = s.activateLocked(w.req) |
| 375 | close(w.ready) |
| 376 | continue |
| 377 | } |
| 378 | remaining = append(remaining, w) |
| 379 | if w.req.Writer && w.req.WritePaths.WholeWorkspace { |
| 380 | wholeWriterPending = true |
| 381 | } |
| 382 | } |
| 383 | s.waiters = remaining |
| 384 | } |
| 385 | |
| 386 | func (s *SubagentScheduler) removeWaiterLocked(target *schedulerWaiter) { |
| 387 | if len(s.waiters) == 0 { |
| 388 | return |
| 389 | } |
| 390 | out := s.waiters[:0] |
| 391 | for _, w := range s.waiters { |
| 392 | if w == target { |
| 393 | continue |
| 394 | } |
| 395 | out = append(out, w) |
| 396 | } |
| 397 | s.waiters = out |
| 398 | } |
| 399 | |
| 400 | func removeClaim(claims []WritePathSet, target WritePathSet) []WritePathSet { |
| 401 | for i, c := range claims { |
| 402 | if writeClaimEqual(c, target) { |
| 403 | return append(claims[:i], claims[i+1:]...) |
| 404 | } |
| 405 | } |
| 406 | return claims |
| 407 | } |
| 408 | |
| 409 | func writeClaimEqual(a, b WritePathSet) bool { |
| 410 | if a.WholeWorkspace != b.WholeWorkspace || a.WorkspaceRoot != b.WorkspaceRoot { |
| 411 | return false |
| 412 | } |
| 413 | if len(a.Paths) != len(b.Paths) { |
| 414 | return false |
| 415 | } |
| 416 | for i := range a.Paths { |
| 417 | if a.Paths[i] != b.Paths[i] { |
| 418 | return false |
| 419 | } |
| 420 | } |
| 421 | return true |
| 422 | } |
| 423 |