| 1 | // Package agentrunner runs an optional agent process outside the relay daemon. |
| 2 | package agentrunner |
| 3 | |
| 4 | import ( |
| 5 | "context" |
| 6 | "encoding/json" |
| 7 | "errors" |
| 8 | "fmt" |
| 9 | "os" |
| 10 | "path/filepath" |
| 11 | "sync" |
| 12 | "time" |
| 13 | ) |
| 14 | |
| 15 | var ( |
| 16 | ErrDuplicate = errors.New("agent runner: duplicate event" ) |
| 17 | ErrUnauthorizedRoom = errors.New("agent runner: unauthorized room" ) |
| 18 | ErrQueueFull = errors.New("agent runner: room queue full" ) |
| 19 | ErrStopped = errors.New("agent runner: stopped" ) |
| 20 | ) |
| 21 | |
| 22 | type Mention struct { |
| 23 | Room string `json:"room"` |
| 24 | EventID string `json:"event_id"` |
| 25 | Author string `json:"author"` |
| 26 | Prompt string `json:"prompt"` |
| 27 | Root string `json:"root"` |
| 28 | Kind int `json:"kind"` |
| 29 | CreatedAt int64 `json:"created_at"` |
| 30 | } |
| 31 | type Handler func(context.Context, []Mention) error |
| 32 | type Options struct { |
| 33 | AllowedRooms []string |
| 34 | QueueLimit int |
| 35 | StatePath string |
| 36 | Handle Handler |
| 37 | OnError func(error) |
| 38 | } |
| 39 | type roomQueue struct { |
| 40 | pending, active []Mention |
| 41 | cancel context.CancelFunc |
| 42 | wake chan struct{} |
| 43 | } |
| 44 | type journal struct { |
| 45 | Pending map[string][]Mention `json:"pending"` |
| 46 | Seen []string `json:"seen"` |
| 47 | } |
| 48 | type Runner struct { |
| 49 | mu sync.Mutex |
| 50 | rooms map[string]*roomQueue |
| 51 | seen map[string]bool |
| 52 | order []string |
| 53 | opts Options |
| 54 | ctx context.Context |
| 55 | cancel context.CancelFunc |
| 56 | wg sync.WaitGroup |
| 57 | stopped bool |
| 58 | } |
| 59 | |
| 60 | func New(opts Options) (*Runner, error) { |
| 61 | if opts.Handle == nil || len(opts.AllowedRooms) == 0 || len(opts.AllowedRooms) > 32 { |
| 62 | return nil, errors.New("agent runner: a handler and 1 to 32 rooms are required" ) |
| 63 | } |
| 64 | if opts.QueueLimit <= 0 { |
| 65 | opts.QueueLimit = 32 |
| 66 | } |
| 67 | if opts.QueueLimit > 256 { |
| 68 | return nil, errors.New("agent runner: queue limit cannot exceed 256" ) |
| 69 | } |
| 70 | r := &Runner{opts: opts, rooms: map[string]*roomQueue{}, seen: map[string]bool{}} |
| 71 | for _, room := range opts.AllowedRooms { |
| 72 | if room == "" { |
| 73 | return nil, ErrUnauthorizedRoom |
| 74 | } |
| 75 | r.rooms[room] = &roomQueue{wake: make(chan struct{}, 1)} |
| 76 | } |
| 77 | if err := r.restore(); err != nil { |
| 78 | return nil, err |
| 79 | } |
| 80 | return r, nil |
| 81 | } |
| 82 | func (r *Runner) Start(ctx context.Context) { |
| 83 | r.mu.Lock() |
| 84 | defer r.mu.Unlock() |
| 85 | if r.ctx != nil || r.stopped { |
| 86 | return |
| 87 | } |
| 88 | r.ctx, r.cancel = context.WithCancel(ctx) |
| 89 | for room, q := range r.rooms { |
| 90 | r.wg.Add(1) |
| 91 | go r.work(room, q) |
| 92 | } |
| 93 | } |
| 94 | func (r *Runner) Enqueue(m Mention) error { |
| 95 | if m.EventID == "" || m.Author == "" || len(m.Prompt) > 64*1024 { |
| 96 | return errors.New("agent runner: invalid mention" ) |
| 97 | } |
| 98 | r.mu.Lock() |
| 99 | defer r.mu.Unlock() |
| 100 | if r.stopped { |
| 101 | return ErrStopped |
| 102 | } |
| 103 | q, ok := r.rooms[m.Room] |
| 104 | if !ok { |
| 105 | return ErrUnauthorizedRoom |
| 106 | } |
| 107 | if r.seen[m.EventID] { |
| 108 | return ErrDuplicate |
| 109 | } |
| 110 | if len(q.pending) >= r.opts.QueueLimit { |
| 111 | return ErrQueueFull |
| 112 | } |
| 113 | q.pending = append(q.pending, m) |
| 114 | r.seen[m.EventID] = true |
| 115 | r.order = append(r.order, m.EventID) |
| 116 | if err := r.save(); err != nil { |
| 117 | q.pending = q.pending[:len(q.pending)-1] |
| 118 | delete(r.seen, m.EventID) |
| 119 | r.order = r.order[:len(r.order)-1] |
| 120 | return err |
| 121 | } |
| 122 | select { |
| 123 | case q.wake <- struct{}{}: |
| 124 | default: |
| 125 | } |
| 126 | return nil |
| 127 | } |
| 128 | func (r *Runner) CancelFrom(room, author string) bool { |
| 129 | r.mu.Lock() |
| 130 | defer r.mu.Unlock() |
| 131 | q := r.rooms[room] |
| 132 | if q == nil { |
| 133 | return false |
| 134 | } |
| 135 | allowed := false |
| 136 | for _, m := range append(append([]Mention{}, q.active...), q.pending...) { |
| 137 | if author == "" || m.Author == author { |
| 138 | allowed = true |
| 139 | break |
| 140 | } |
| 141 | } |
| 142 | if !allowed { |
| 143 | return false |
| 144 | } |
| 145 | filtered := q.pending[:0] |
| 146 | for _, m := range q.pending { |
| 147 | if author == "" || m.Author == author { |
| 148 | continue |
| 149 | } |
| 150 | filtered = append(filtered, m) |
| 151 | } |
| 152 | q.pending = filtered |
| 153 | if q.cancel != nil && anyAuthor(q.active, author) { |
| 154 | q.cancel() |
| 155 | } |
| 156 | if err := r.save(); err != nil { |
| 157 | r.report(err) |
| 158 | } |
| 159 | return true |
| 160 | } |
| 161 | func (r *Runner) Cancel(room string) error { r.CancelFrom(room, "" ); return nil } |
| 162 | func (r *Runner) Stop() { |
| 163 | r.mu.Lock() |
| 164 | r.stopped = true |
| 165 | if r.cancel != nil { |
| 166 | r.cancel() |
| 167 | } |
| 168 | r.mu.Unlock() |
| 169 | r.wg.Wait() |
| 170 | } |
| 171 | func (r *Runner) work(room string, q *roomQueue) { |
| 172 | defer r.wg.Done() |
| 173 | for { |
| 174 | select { |
| 175 | case <-r.ctx.Done(): |
| 176 | return |
| 177 | default: |
| 178 | } |
| 179 | r.mu.Lock() |
| 180 | if len(q.pending) == 0 { |
| 181 | r.mu.Unlock() |
| 182 | select { |
| 183 | case <-r.ctx.Done(): |
| 184 | return |
| 185 | case <-q.wake: |
| 186 | } |
| 187 | continue |
| 188 | } |
| 189 | batch := takeBatch(q.pending) |
| 190 | q.pending = q.pending[len(batch):] |
| 191 | q.active = batch |
| 192 | ctx, cancel := context.WithCancel(r.ctx) |
| 193 | q.cancel = cancel |
| 194 | if err := r.save(); err != nil { |
| 195 | r.report(err) |
| 196 | cancel() |
| 197 | q.pending = append(batch, q.pending...) |
| 198 | q.active = nil |
| 199 | q.cancel = nil |
| 200 | r.mu.Unlock() |
| 201 | return |
| 202 | } |
| 203 | r.mu.Unlock() |
| 204 | err := r.opts.Handle(ctx, batch) |
| 205 | cancel() |
| 206 | r.mu.Lock() |
| 207 | if r.ctx.Err() != nil { |
| 208 | q.pending = append(batch, q.pending...) |
| 209 | } |
| 210 | q.active = nil |
| 211 | q.cancel = nil |
| 212 | if saveErr := r.save(); saveErr != nil { |
| 213 | r.report(saveErr) |
| 214 | } |
| 215 | r.mu.Unlock() |
| 216 | if err != nil && !errors.Is(err, context.Canceled) { |
| 217 | r.report(fmt.Errorf("agent room %s: %w" , room, err)) |
| 218 | } |
| 219 | } |
| 220 | } |
| 221 | |
| 222 | func takeBatch(pending []Mention) []Mention { |
| 223 | if len(pending) == 0 { |
| 224 | return nil |
| 225 | } |
| 226 | root := pending[0].Root |
| 227 | n := 1 |
| 228 | for n < len(pending) && pending[n].Root == root { |
| 229 | n++ |
| 230 | } |
| 231 | return append([]Mention(nil), pending[:n]...) |
| 232 | } |
| 233 | |
| 234 | func anyAuthor(events []Mention, author string) bool { |
| 235 | if author == "" { |
| 236 | return len(events) > 0 |
| 237 | } |
| 238 | for _, event := range events { |
| 239 | if event.Author == author { |
| 240 | return true |
| 241 | } |
| 242 | } |
| 243 | return false |
| 244 | } |
| 245 | func (r *Runner) report(err error) { |
| 246 | if r.opts.OnError != nil { |
| 247 | r.opts.OnError(err) |
| 248 | } |
| 249 | } |
| 250 | func (r *Runner) restore() error { |
| 251 | if r.opts.StatePath == "" { |
| 252 | return nil |
| 253 | } |
| 254 | raw, err := os.ReadFile(r.opts.StatePath) |
| 255 | if errors.Is(err, os.ErrNotExist) { |
| 256 | return nil |
| 257 | } |
| 258 | if err != nil { |
| 259 | return fmt.Errorf("read agent state: %w" , err) |
| 260 | } |
| 261 | if len(raw) > 32<<20 { |
| 262 | return errors.New("agent state exceeds 32 MiB" ) |
| 263 | } |
| 264 | var state journal |
| 265 | if err := json.Unmarshal(raw, &state); err != nil { |
| 266 | return fmt.Errorf("decode agent state: %w" , err) |
| 267 | } |
| 268 | for _, id := range state.Seen { |
| 269 | if !r.seen[id] { |
| 270 | r.seen[id] = true |
| 271 | r.order = append(r.order, id) |
| 272 | } |
| 273 | } |
| 274 | for room, queue := range state.Pending { |
| 275 | q := r.rooms[room] |
| 276 | if q == nil { |
| 277 | return errors.New("agent state contains an unconfigured room" ) |
| 278 | } |
| 279 | if len(queue) > r.opts.QueueLimit*2 { |
| 280 | return ErrQueueFull |
| 281 | } |
| 282 | q.pending = queue |
| 283 | } |
| 284 | return nil |
| 285 | } |
| 286 | func (r *Runner) save() error { |
| 287 | if r.opts.StatePath == "" { |
| 288 | return nil |
| 289 | } |
| 290 | // Keep a bounded replay history while retaining every queued event. |
| 291 | active := map[string]bool{} |
| 292 | pending := map[string][]Mention{} |
| 293 | for room, q := range r.rooms { |
| 294 | pending[room] = append(append([]Mention{}, q.active...), q.pending...) |
| 295 | for _, m := range pending[room] { |
| 296 | active[m.EventID] = true |
| 297 | } |
| 298 | } |
| 299 | if len(r.order) > 10000 { |
| 300 | keep := make(map[string]bool, len(active)+10000) |
| 301 | for id := range active { |
| 302 | keep[id] = true |
| 303 | } |
| 304 | remaining := 10000 - len(active) |
| 305 | if remaining < 0 { |
| 306 | remaining = 0 |
| 307 | } |
| 308 | for i := len(r.order) - 1; i >= 0 && remaining > 0; i-- { |
| 309 | if !keep[r.order[i]] { |
| 310 | keep[r.order[i]] = true |
| 311 | remaining-- |
| 312 | } |
| 313 | } |
| 314 | kept := make([]string, 0, len(keep)) |
| 315 | for _, id := range r.order { |
| 316 | if keep[id] { |
| 317 | kept = append(kept, id) |
| 318 | } |
| 319 | } |
| 320 | for id := range r.seen { |
| 321 | if !keep[id] { |
| 322 | delete(r.seen, id) |
| 323 | } |
| 324 | } |
| 325 | r.order = kept |
| 326 | } |
| 327 | raw, err := json.Marshal(journal{Pending: pending, Seen: r.order}) |
| 328 | if err != nil { |
| 329 | return err |
| 330 | } |
| 331 | dir := filepath.Dir(r.opts.StatePath) |
| 332 | if err := os.MkdirAll(dir, 0700); err != nil { |
| 333 | return err |
| 334 | } |
| 335 | file, err := os.CreateTemp(dir, ".tiny-agent-*" ) |
| 336 | if err != nil { |
| 337 | return err |
| 338 | } |
| 339 | name := file.Name() |
| 340 | defer os.Remove(name) |
| 341 | if _, err = file.Write(raw); err != nil { |
| 342 | file.Close() |
| 343 | return err |
| 344 | } |
| 345 | if err = file.Sync(); err != nil { |
| 346 | file.Close() |
| 347 | return err |
| 348 | } |
| 349 | if err = file.Close(); err != nil { |
| 350 | return err |
| 351 | } |
| 352 | return os.Rename(name, r.opts.StatePath) |
| 353 | } |
| 354 | |
| 355 | // Drain waits until queued work has finished, or the caller cancels. |
| 356 | func (r *Runner) Drain(ctx context.Context) error { |
| 357 | ticker := time.NewTicker(10 * time.Millisecond) |
| 358 | defer ticker.Stop() |
| 359 | for { |
| 360 | r.mu.Lock() |
| 361 | empty := true |
| 362 | for _, q := range r.rooms { |
| 363 | empty = empty && len(q.pending) == 0 && len(q.active) == 0 |
| 364 | } |
| 365 | r.mu.Unlock() |
| 366 | if empty { |
| 367 | return nil |
| 368 | } |
| 369 | select { |
| 370 | case <-ctx.Done(): |
| 371 | return ctx.Err() |
| 372 | case <-ticker.C: |
| 373 | } |
| 374 | } |
| 375 | } |