git.giacomolli.uk - onechat/blob - probe.go

   1 package main
   2
   3 import (
   4         "bufio"
   5         "cmp"
   6         "context"
   7         "encoding/json"
   8         "errors"
   9         "fmt"
  10         "io"
  11         "maps"
  12         "os"
  13         "path/filepath"
  14         "slices"
  15         "strconv"
  16         "strings"
  17         "time"
  18 )
  19
  20 const (
  21         // probeName is the agent name of a probe run. The cache key comes from
  22         // it.
  23         probeName = "Probe"
  24
  25         // probeQuotaLimit is how many points the used percent of a quota window
  26         // can grow in a probe run before the run stops.
  27         probeQuotaLimit = 15.0
  28
  29         // probeCallLimit is the most model calls of a probe run.
  30         probeCallLimit = 660
  31
  32         // probeLineBytes is the most bytes of one line of a probe log.
  33         probeLineBytes = 1 << 20
  34 )
  35
  36 // probe runs "onechat probe LOG COUNT [BUILDS]". It appends the first COUNT
  37 // messages of LOG, one JSON object {kind, text} per line, to a new chat in
  38 // probe/ in the data directory, and builds its tree with the real compactor.
  39 // Before each user message it waits for the builds, as a turn does. It stops at
  40 // the first failed call, after probeCallLimit calls, or when the used percent
  41 // of a quota window grows by more than probeQuotaLimit points. At the end it
  42 // reports the calls, the time per call, the calls at once, the cached share,
  43 // the quota, the retries, and the line sizes. BUILDS, from 1 to maxBuilds,
  44 // limits the builds that call the model at once in this run only. The core
  45 // still queues up to maxBuilds builds. The run holds the lock of the token
  46 // file, so it does not run while the daemon runs.
  47 func probe(ctx context.Context, args []string, out io.Writer) (err error) {
  48         if len(args) != 2 && len(args) != 3 {
  49                 return errors.New("usage: onechat probe LOG COUNT [BUILDS]")
  50         }
  51
  52         count, err := strconv.Atoi(args[1])
  53         if err != nil || count <= 0 {
  54                 return fmt.Errorf("COUNT %q is not a positive number", args[1])
  55         }
  56
  57         builds := maxBuilds
  58         if len(args) == 3 {
  59                 builds, err = strconv.Atoi(args[2])
  60                 if err != nil || builds < 1 || builds > maxBuilds {
  61                         return fmt.Errorf("BUILDS %q is not a number from 1 to %d", args[2], maxBuilds)
  62                 }
  63         }
  64
  65         entries, err := readProbeLog(args[0], count)
  66         if err != nil {
  67                 return err
  68         }
  69
  70         dir, err := dataDir()
  71         if err != nil {
  72                 return err
  73         }
  74
  75         config, err := loadConfig(dir)
  76         if err != nil {
  77                 return err
  78         }
  79
  80         tokens, err := lockTokens(dir)
  81         if err != nil {
  82                 return err
  83         }
  84
  85         defer func() {
  86                 err = errors.Join(err, tokens.Close())
  87         }()
  88
  89         address, auth, err := modelEndpoint(config, dir)
  90         if err != nil {
  91                 return err
  92         }
  93
  94         root := filepath.Join(dir, "probe", time.Now().Format("20060102-150405"))
  95         err = os.MkdirAll(root, 0o700)
  96         if err != nil {
  97                 return fmt.Errorf("create probe directory: %w", err)
  98         }
  99
 100         usagePath := filepath.Join(root, "usage.jsonl")
 101         client := newClient(address, auth, "onechat-"+probeName, usagePath)
 102         client.LimitCalls(probeCallLimit)
 103         summarizer := slotSummarizer{
 104                 inner: newCompactor(client, probePrompt, config.Compactor),
 105                 slots: make(chan struct{}, builds),
 106         }
 107
 108         memory, err := openMemory(ctx, root, time.Now, summarizer)
 109         if err != nil {
 110                 return err
 111         }
 112
 113         _, err = fmt.Fprintf(out, "probe: %d messages into %s with %s at %s effort, %d builds at once\n", len(entries), root, config.Compactor.Model, config.Compactor.Effort, builds)
 114         if err != nil {
 115                 return errors.Join(err, memory.Close())
 116         }
 117
 118         run := probeRun{
 119                 memory:    memory,
 120                 client:    client,
 121                 usagePath: usagePath,
 122         }
 123
 124         appended, stop := run.Run(ctx, entries)
 125         sizes, err := modelLines(memory)
 126         err = errors.Join(err, memory.Close())
 127         if err != nil {
 128                 return err
 129         }
 130
 131         records, err := readUsageFile(usagePath)
 132         if err != nil {
 133                 return err
 134         }
 135
 136         return writeProbeReport(out, appended, stop, records, sizes)
 137 }
 138
 139 // probeRun appends the messages of a probe run to its chat and stops the run
 140 // by the rules of probeStop.
 141 type probeRun struct {
 142         memory    *Memory
 143         client    *Client
 144         usagePath string
 145 }
 146
 147 // Run appends the entries. It waits for the builds only before a user message,
 148 // as a turn does, so up to maxBuilds builds run at once in between, and it
 149 // waits for all builds at the end. It returns the number of messages appended
 150 // and why the run stopped early, or "" when it did not. After a stop it waits
 151 // for the calls in flight, so no sent call is cancelled.
 152 func (r *probeRun) Run(ctx context.Context, entries []probeEntry) (int, string) {
 153         appended, reason := r.append(ctx, entries)
 154         if reason == "" {
 155                 return appended, ""
 156         }
 157
 158         err := r.memory.Wait(ctx)
 159         if err != nil {
 160                 reason += "; " + err.Error()
 161         }
 162
 163         return appended, reason
 164 }
 165
 166 // append appends the entries and stops at the first stop rule.
 167 func (r *probeRun) append(ctx context.Context, entries []probeEntry) (int, string) {
 168         for n, entry := range entries {
 169                 if entry.Kind == "user" {
 170                         reason := r.wait(ctx)
 171                         if reason != "" {
 172                                 return n, reason
 173                         }
 174                 }
 175
 176                 _, err := r.memory.Append(entry.Kind, entry.Text)
 177                 if err != nil {
 178                         return n, err.Error()
 179                 }
 180
 181                 reason := r.check()
 182                 if reason != "" {
 183                         return n + 1, reason
 184                 }
 185         }
 186
 187         return len(entries), r.wait(ctx)
 188 }
 189
 190 // wait returns when no build runs, or when a stop rule ends the run. It checks
 191 // the rules each second, so a run stops soon also during a long wait.
 192 func (r *probeRun) wait(ctx context.Context) string {
 193         for {
 194                 reason := r.check()
 195                 if reason != "" {
 196                         return reason
 197                 }
 198
 199                 waitCtx, cancel := context.WithTimeout(ctx, time.Second)
 200                 err := r.memory.Wait(waitCtx)
 201                 cancel()
 202                 if err == nil {
 203                         return r.check()
 204                 }
 205
 206                 if ctx.Err() != nil {
 207                         return ctx.Err().Error()
 208                 }
 209         }
 210 }
 211
 212 // check returns why the run must stop, or "" when it can go on.
 213 func (r *probeRun) check() string {
 214         if r.client.LimitReached() {
 215                 return fmt.Sprintf("call limit of %d reached", r.client.limit)
 216         }
 217
 218         records, err := readUsageFile(r.usagePath)
 219         if err != nil {
 220                 return err.Error()
 221         }
 222
 223         return probeStop(records)
 224 }
 225
 226 // slotSummarizer lets at most cap(slots) builds call inner at once. A build
 227 // past the limit waits for a slot.
 228 type slotSummarizer struct {
 229         inner Summarizer
 230         slots chan struct{}
 231 }
 232
 233 // Summarize waits for a slot and builds the node with inner.
 234 func (s slotSummarizer) Summarize(ctx context.Context, task Task) (string, error) {
 235         select {
 236         case s.slots <- struct{}{}:
 237         case <-ctx.Done():
 238                 return "", ctx.Err()
 239         }
 240
 241         defer func() {
 242                 <-s.slots
 243         }()
 244
 245         return s.inner.Summarize(ctx, task)
 246 }
 247
 248 // probePrompt returns the system prompt of a probe run: the prompt of the agent
 249 // probeName, without instructions.
 250 func probePrompt() (string, error) {
 251         return prompt(probeName, ""), nil
 252 }
 253
 254 // probeEntry is one message of a probe log.
 255 type probeEntry struct {
 256         Kind string `json:"kind"`
 257         Text string `json:"text"`
 258 }
 259
 260 // readProbeLog reads the first count messages of the probe log at path.
 261 func readProbeLog(path string, count int) (entries []probeEntry, err error) {
 262         file, err := os.Open(path)
 263         if err != nil {
 264                 return nil, fmt.Errorf("open probe log: %w", err)
 265         }
 266
 267         defer func() {
 268                 err = errors.Join(err, file.Close())
 269         }()
 270
 271         scanner := bufio.NewScanner(file)
 272         scanner.Buffer(make([]byte, 64*1024), probeLineBytes)
 273         for len(entries) < count && scanner.Scan() {
 274                 var entry probeEntry
 275                 err = json.Unmarshal(scanner.Bytes(), &entry)
 276                 if err != nil {
 277                         return nil, fmt.Errorf("probe log line %d: %w", len(entries)+1, err)
 278                 }
 279
 280                 entries = append(entries, entry)
 281         }
 282
 283         err = scanner.Err()
 284         if err != nil {
 285                 return nil, fmt.Errorf("read probe log: %w", err)
 286         }
 287
 288         if len(entries) < count {
 289                 return nil, fmt.Errorf("probe log has %d messages, fewer than %d", len(entries), count)
 290         }
 291
 292         return entries, nil
 293 }
 294
 295 // probeStop returns why a probe run must stop: a failed call, or a quota window
 296 // whose used percent grew by more than probeQuotaLimit points since the first
 297 // call. It returns "" when the run can go on.
 298 func probeStop(records []usageRecord) string {
 299         for _, record := range records {
 300                 if record.Error != "" {
 301                         return "call failed: " + record.Error
 302                 }
 303         }
 304
 305         if len(records) == 0 {
 306                 return ""
 307         }
 308
 309         first := usedPercents(records[0])
 310         for name, used := range usedPercents(records[len(records)-1]) {
 311                 start, found := first[name]
 312                 if found && used-start > probeQuotaLimit {
 313                         return fmt.Sprintf("%s grew from %g to %g", name, start, used)
 314                 }
 315         }
 316
 317         return ""
 318 }
 319
 320 // modelLines returns the sizes of the nodes that the model wrote, in tree
 321 // order. A node that fits its source with no model call is not one of them.
 322 func modelLines(memory *Memory) ([]int, error) {
 323         memory.mutex.Lock()
 324         defer memory.mutex.Unlock()
 325
 326         keys, err := memory.tree.Keys()
 327         if err != nil {
 328                 return nil, err
 329         }
 330
 331         var sizes []int
 332         for _, key := range keys {
 333                 node, _, err := memory.tree.Node(key.level, key.index)
 334                 if err != nil {
 335                         return nil, err
 336                 }
 337
 338                 model, err := builtByModel(memory, key)
 339                 if err != nil {
 340                         return nil, err
 341                 }
 342
 343                 if model {
 344                         sizes = append(sizes, node.Size)
 345                 }
 346         }
 347
 348         return sizes, nil
 349 }
 350
 351 // builtByModel tells if the node key needed a model call: a leaf whose source
 352 // is over nodeBytes, or a merge whose children do not fit in one node.
 353 func builtByModel(memory *Memory, key nodeKey) (bool, error) {
 354         if key.level == 0 {
 355                 message, err := memory.log.Message(key.index)
 356                 if err != nil {
 357                         return false, err
 358                 }
 359
 360                 _, free := freeLeaf(message)
 361
 362                 return !free, nil
 363         }
 364
 365         left, right, err := memory.children(key)
 366         if err != nil {
 367                 return false, err
 368         }
 369
 370         _, free := freeMerge(left, right)
 371
 372         return !free, nil
 373 }
 374
 375 // readUsageFile reads the records of usage.jsonl at path, in the order of the
 376 // start of their calls. The file has them in the order of the end of their
 377 // calls, which differs when calls run at once.
 378 func readUsageFile(path string) (records []usageRecord, err error) {
 379         file, err := os.Open(path)
 380         if errors.Is(err, os.ErrNotExist) {
 381                 return nil, nil
 382         }
 383
 384         if err != nil {
 385                 return nil, fmt.Errorf("open usage file: %w", err)
 386         }
 387
 388         defer func() {
 389                 err = errors.Join(err, file.Close())
 390         }()
 391
 392         scanner := bufio.NewScanner(file)
 393         scanner.Buffer(make([]byte, 64*1024), probeLineBytes)
 394         for scanner.Scan() {
 395                 var record usageRecord
 396                 err = json.Unmarshal(scanner.Bytes(), &record)
 397                 if err != nil {
 398                         return nil, fmt.Errorf("decode usage: %w", err)
 399                 }
 400
 401                 records = append(records, record)
 402         }
 403
 404         err = scanner.Err()
 405         if err != nil {
 406                 return nil, fmt.Errorf("read usage file: %w", err)
 407         }
 408
 409         slices.SortStableFunc(records, func(a, b usageRecord) int {
 410                 return callStart(a).Compare(callStart(b))
 411         })
 412
 413         return records, nil
 414 }
 415
 416 // callStart returns the start of the call of record, or the zero time when its
 417 // date is not valid.
 418 func callStart(record usageRecord) time.Time {
 419         start, err := time.Parse(time.RFC3339Nano, record.Date)
 420         if err != nil {
 421                 return time.Time{}
 422         }
 423
 424         return start
 425 }
 426
 427 // usedPercents returns the used percent of each quota window of record, by
 428 // header name.
 429 func usedPercents(record usageRecord) map[string]float64 {
 430         used := make(map[string]float64)
 431         for name, value := range record.Quota {
 432                 if !strings.HasSuffix(name, "-used-percent") {
 433                         continue
 434                 }
 435
 436                 percent, err := strconv.ParseFloat(value, 64)
 437                 if err == nil {
 438                         used[name] = percent
 439                 }
 440         }
 441
 442         return used
 443 }
 444
 445 // writeProbeReport writes the result of a probe run.
 446 func writeProbeReport(out io.Writer, appended int, stop string, records []usageRecord, sizes []int) error {
 447         var input, cached, output, failed int
 448         for _, record := range records {
 449                 input += record.InputTokens
 450                 cached += record.CachedTokens
 451                 output += record.OutputTokens
 452                 if record.Error != "" {
 453                         failed++
 454                 }
 455         }
 456
 457         var lines []string
 458         if stop != "" {
 459                 lines = append(lines, "stopped early: "+stop)
 460         }
 461
 462         lines = append(lines,
 463                 fmt.Sprintf("messages: %d", appended),
 464                 fmt.Sprintf("calls: %d (%d failed), %.2f per message", len(records), failed, ratio(len(records), appended)),
 465                 fmt.Sprintf("model lines: %d, retries: %d, %.2f calls per model line", len(sizes), len(records)-failed-len(sizes), ratio(len(records), len(sizes))),
 466                 fmt.Sprintf("tokens: %d input, %d cached (%.1f%%), %d output", input, cached, 100*ratio(cached, input), output),
 467         )
 468
 469         if len(sizes) > 0 {
 470                 lines = append(lines, sizeText(sizes))
 471         }
 472
 473         if len(records) > 0 {
 474                 lines = append(lines, timeText(records))
 475         }
 476
 477         if len(records) > 0 {
 478                 lines = append(lines,
 479                         "quota at the first call: "+quotaText(records[0]),
 480                         "quota at the last call: "+quotaText(records[len(records)-1]),
 481                 )
 482         }
 483
 484         _, err := io.WriteString(out, strings.Join(lines, "\n")+"\n")
 485
 486         return err
 487 }
 488
 489 // sizeText describes the sizes of the model lines.
 490 func sizeText(sizes []int) string {
 491         sorted := slices.Sorted(slices.Values(sizes))
 492         total, over := 0, 0
 493         for _, size := range sorted {
 494                 total += size
 495                 if size > nodeBytes {
 496                         over++
 497                 }
 498         }
 499
 500         return fmt.Sprintf("line bytes: min %d, median %d, mean %.0f, max %d, %d over %d", sorted[0], sorted[len(sorted)/2], ratio(total, len(sorted)), sorted[len(sorted)-1], over, nodeBytes)
 501 }
 502
 503 // timeText describes the time per call and how many calls ran at once. A call
 504 // runs from its date for its seconds.
 505 func timeText(records []usageRecord) string {
 506         type event struct {
 507                 at    time.Time
 508                 delta int
 509         }
 510
 511         var events []event
 512         seconds := make([]float64, 0, len(records))
 513         total := 0.0
 514         for _, record := range records {
 515                 start, err := time.Parse(time.RFC3339Nano, record.Date)
 516                 if err != nil {
 517                         continue
 518                 }
 519
 520                 length := time.Duration(record.Seconds * float64(time.Second))
 521                 events = append(events, event{
 522                         at:    start,
 523                         delta: 1,
 524                 }, event{
 525                         at:    start.Add(length),
 526                         delta: -1,
 527                 })
 528                 seconds = append(seconds, record.Seconds)
 529                 total += record.Seconds
 530         }
 531
 532         if len(events) == 0 {
 533                 return "time: no dated calls"
 534         }
 535
 536         slices.SortFunc(events, func(a, b event) int {
 537                 return cmp.Or(a.at.Compare(b.at), cmp.Compare(a.delta, b.delta))
 538         })
 539
 540         running, most := 0, 0
 541         for _, e := range events {
 542                 running += e.delta
 543                 most = max(most, running)
 544         }
 545
 546         wall := events[len(events)-1].at.Sub(events[0].at).Seconds()
 547         slices.Sort(seconds)
 548
 549         return fmt.Sprintf("time: %.0f s in all, %.1f calls per minute, %.1f s per call (median %.1f, max %.1f), calls at once: at most %d, %.1f on average", wall, 60*float64(len(seconds))/max(wall, 1e-9), total/float64(len(seconds)), seconds[len(seconds)/2], seconds[len(seconds)-1], most, total/max(wall, 1e-9))
 550 }
 551
 552 // quotaText returns the quota headers of record, sorted by name.
 553 func quotaText(record usageRecord) string {
 554         names := slices.Sorted(maps.Keys(record.Quota))
 555         parts := make([]string, len(names))
 556         for n, name := range names {
 557                 parts[n] = strings.TrimPrefix(name, "x-codex-") + "=" + record.Quota[name]
 558         }
 559
 560         return strings.Join(parts, " ")
 561 }
 562
 563 // ratio returns a / b, or 0 when b is 0.
 564 func ratio(a, b int) float64 {
 565         if b == 0 {
 566                 return 0
 567         }
 568
 569         return float64(a) / float64(b)
 570 }