X-Git-Url: http://git.giacomolli.uk/onechat/blobdiff_plain/522d226230a2eae0a8a33ae0fbf32f96c6886b26..1e2df788bbb17653a6df44acd41da2edb109d403:/probe.go

diff --git a/probe.go b/probe.go
index acd4f6b..ba656ff 100644
--- a/probe.go
+++ b/probe.go
@@ -2,6 +2,7 @@ package main
 
 import (
 	"bufio"
+	"cmp"
 	"context"
 	"encoding/json"
 	"errors"
@@ -25,20 +26,27 @@ const (
 	// can grow in a probe run before the run stops.
 	probeQuotaLimit = 15.0
 
+	// probeCallLimit is the most model calls of a probe run.
+	probeCallLimit = 660
+
 	// probeLineBytes is the most bytes of one line of a probe log.
 	probeLineBytes = 1 << 20
 )
 
-// probe runs "onechat probe LOG COUNT". It appends the first COUNT messages of
-// LOG, one JSON object {kind, text} per line, to a new chat in probe/ in the
-// data directory, and builds its tree with the real compactor. After each
-// message it waits for the builds, as a turn does. It stops at the first
-// failed call, or when the used percent of a quota window grows by more than
-// probeQuotaLimit points. At the end it reports the calls, the cached share,
-// the quota, the retries, and the line sizes.
-func probe(ctx context.Context, args []string, out io.Writer) error {
-	if len(args) != 2 {
-		return errors.New("usage: onechat probe LOG COUNT")
+// probe runs "onechat probe LOG COUNT [BUILDS]". It appends the first COUNT
+// messages of LOG, one JSON object {kind, text} per line, to a new chat in
+// probe/ in the data directory, and builds its tree with the real compactor.
+// Before each user message it waits for the builds, as a turn does. It stops at
+// the first failed call, after probeCallLimit calls, or when the used percent
+// of a quota window grows by more than probeQuotaLimit points. At the end it
+// reports the calls, the time per call, the calls at once, the cached share,
+// the quota, the retries, and the line sizes. BUILDS, from 1 to buildMax,
+// limits the builds that call the model at once in this run only. The core
+// still queues up to buildMax builds. The run holds the lock of the token file,
+// so it does not run while the daemon runs.
+func probe(ctx context.Context, args []string, out io.Writer) (err error) {
+	if len(args) != 2 && len(args) != 3 {
+		return errors.New("usage: onechat probe LOG COUNT [BUILDS]")
 	}
 
 	count, err := strconv.Atoi(args[1])
@@ -46,7 +54,15 @@ func probe(ctx context.Context, args []string, out io.Writer) error {
 		return fmt.Errorf("COUNT %q is not a positive number", args[1])
 	}
 
-	entries, err := readProbeLog(args[0], count)
+	builds := buildMax
+	if len(args) == 3 {
+		builds, err = strconv.Atoi(args[2])
+		if err != nil || builds < 1 || builds > buildMax {
+			return fmt.Errorf("BUILDS %q is not a number from 1 to %d", args[2], buildMax)
+		}
+	}
+
+	entries, err := probeReadLog(args[0], count)
 	if err != nil {
 		return err
 	}
@@ -61,6 +77,15 @@ func probe(ctx context.Context, args []string, out io.Writer) error {
 		return err
 	}
 
+	tokens, err := lockTokens(dir)
+	if err != nil {
+		return err
+	}
+
+	defer func() {
+		err = errors.Join(err, tokens.Close())
+	}()
+
 	address, auth, err := modelEndpoint(config, dir)
 	if err != nil {
 		return err
@@ -73,68 +98,161 @@ func probe(ctx context.Context, args []string, out io.Writer) error {
 	}
 
 	usagePath := filepath.Join(root, "usage.jsonl")
-	client := newClient(address, auth, "onechat-"+probeName, usagePath)
-	memory, err := openMemory(ctx, root, time.Now, newCompactor(client, probeName, config.Compactor))
+	client := newModelClient(address, auth, "onechat-"+probeName, usagePath)
+	client.LimitCalls(probeCallLimit)
+	summarizer := ProbeSlotSummarizer{
+		inner: newCompactor(client, probePrompt, config.Compactor),
+		slots: make(chan struct{}, builds),
+	}
+
+	memory, err := openMemory(ctx, root, time.Now, summarizer)
 	if err != nil {
 		return err
 	}
 
-	_, err = fmt.Fprintf(out, "probe: %d messages into %s with %s at %s effort\n", len(entries), root, config.Compactor.Model, config.Compactor.Effort)
+	_, 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)
 	if err != nil {
 		return errors.Join(err, memory.Close())
 	}
 
-	appended, stop := runProbe(ctx, memory, entries, usagePath)
-	sizes, err := modelLines(memory)
+	run := ProbeRun{
+		memory:    memory,
+		client:    client,
+		usagePath: usagePath,
+	}
+
+	appended, stop := run.Run(ctx, entries)
+	sizes, err := probeModelLines(memory)
 	err = errors.Join(err, memory.Close())
 	if err != nil {
 		return err
 	}
 
-	records, err := readUsageFile(usagePath)
+	records, err := probeReadUsage(usagePath)
 	if err != nil {
 		return err
 	}
 
-	return writeProbeReport(out, appended, stop, records, sizes)
+	return probeWriteReport(out, appended, stop, records, sizes)
+}
+
+// ProbeSlotSummarizer lets at most cap(slots) builds call inner at once. A
+// build past the limit waits for a slot.
+type ProbeSlotSummarizer struct {
+	inner Summarizer
+	slots chan struct{}
+}
+
+// Summarize waits for a slot and builds the node with inner.
+func (s ProbeSlotSummarizer) Summarize(ctx context.Context, task Task) (string, error) {
+	select {
+	case s.slots <- struct{}{}:
+	case <-ctx.Done():
+		return "", ctx.Err()
+	}
+
+	defer func() {
+		<-s.slots
+	}()
+
+	return s.inner.Summarize(ctx, task)
+}
+
+// ProbeRun appends the messages of a probe run to its chat and stops the run by
+// the rules of probeStop.
+type ProbeRun struct {
+	memory    *Memory
+	client    *ModelClient
+	usagePath string
+}
+
+// Run appends the entries. It waits for the builds only before a user message,
+// as a turn does, so up to buildMax builds run at once in between, and it waits
+// for all builds at the end. It returns the number of messages appended and why
+// the run stopped early, or "" when it did not. After a stop it waits for the
+// calls in flight, so no sent call is cancelled.
+func (r *ProbeRun) Run(ctx context.Context, entries []ProbeEntry) (int, string) {
+	appended, reason := r.append(ctx, entries)
+	if reason == "" {
+		return appended, ""
+	}
+
+	err := r.memory.Wait(ctx)
+	if err != nil {
+		reason += "; " + err.Error()
+	}
+
+	return appended, reason
 }
 
-// runProbe appends the entries and waits for the builds after each one. It
-// returns the number of messages appended and why the run stopped early, or
-// "" when it did not.
-func runProbe(ctx context.Context, memory *Memory, entries []probeEntry, usagePath string) (int, string) {
+// append appends the entries and stops at the first stop rule.
+func (r *ProbeRun) append(ctx context.Context, entries []ProbeEntry) (int, string) {
 	for n, entry := range entries {
-		_, err := memory.Append(entry.Kind, entry.Text)
-		if err == nil {
-			err = memory.Wait(ctx)
+		if entry.Kind == "user" {
+			reason := r.wait(ctx)
+			if reason != "" {
+				return n, reason
+			}
 		}
 
+		_, err := r.memory.Append(entry.Kind, entry.Text)
 		if err != nil {
 			return n, err.Error()
 		}
 
-		records, err := readUsageFile(usagePath)
-		if err != nil {
-			return n + 1, err.Error()
+		reason := r.check()
+		if reason != "" {
+			return n + 1, reason
 		}
+	}
 
-		reason := probeStop(records)
+	return len(entries), r.wait(ctx)
+}
+
+// wait returns when no build runs, or when a stop rule ends the run. It checks
+// the rules each second, so a run stops soon also during a long wait.
+func (r *ProbeRun) wait(ctx context.Context) string {
+	for {
+		reason := r.check()
 		if reason != "" {
-			return n + 1, reason
+			return reason
+		}
+
+		waitCtx, cancel := context.WithTimeout(ctx, time.Second)
+		err := r.memory.Wait(waitCtx)
+		cancel()
+		if err == nil {
+			return r.check()
 		}
+
+		if ctx.Err() != nil {
+			return ctx.Err().Error()
+		}
+	}
+}
+
+// check returns why the run must stop, or "" when it can go on.
+func (r *ProbeRun) check() string {
+	if r.client.LimitReached() {
+		return fmt.Sprintf("call limit of %d reached", r.client.limit)
+	}
+
+	records, err := probeReadUsage(r.usagePath)
+	if err != nil {
+		return err.Error()
 	}
 
-	return len(entries), ""
+	return probeStop(records)
 }
 
-// probeEntry is one message of a probe log.
-type probeEntry struct {
+// ProbeEntry is one message of a probe log.
+type ProbeEntry struct {
 	Kind string `json:"kind"`
 	Text string `json:"text"`
 }
 
-// readProbeLog reads the first count messages of the probe log at path.
-func readProbeLog(path string, count int) (entries []probeEntry, err error) {
+// probeReadLog reads the first count messages of the probe log at path.
+func probeReadLog(path string, count int) (entries []ProbeEntry, err error) {
 	file, err := os.Open(path)
 	if err != nil {
 		return nil, fmt.Errorf("open probe log: %w", err)
@@ -147,7 +265,7 @@ func readProbeLog(path string, count int) (entries []probeEntry, err error) {
 	scanner := bufio.NewScanner(file)
 	scanner.Buffer(make([]byte, 64*1024), probeLineBytes)
 	for len(entries) < count && scanner.Scan() {
-		var entry probeEntry
+		var entry ProbeEntry
 		err = json.Unmarshal(scanner.Bytes(), &entry)
 		if err != nil {
 			return nil, fmt.Errorf("probe log line %d: %w", len(entries)+1, err)
@@ -168,34 +286,15 @@ func readProbeLog(path string, count int) (entries []probeEntry, err error) {
 	return entries, nil
 }
 
-// probeStop returns why a probe run must stop: a failed call, or a quota window
-// whose used percent grew by more than probeQuotaLimit points since the first
-// call. It returns "" when the run can go on.
-func probeStop(records []usageRecord) string {
-	for _, record := range records {
-		if record.Error != "" {
-			return "call failed: " + record.Error
-		}
-	}
-
-	if len(records) == 0 {
-		return ""
-	}
-
-	first := usedPercents(records[0])
-	for name, used := range usedPercents(records[len(records)-1]) {
-		start, found := first[name]
-		if found && used-start > probeQuotaLimit {
-			return fmt.Sprintf("%s grew from %g to %g", name, start, used)
-		}
-	}
-
-	return ""
+// probePrompt returns the system prompt of a probe run: the prompt of the agent
+// probeName, without instructions.
+func probePrompt() (string, error) {
+	return prompt(probeName, ""), nil
 }
 
-// modelLines returns the sizes of the nodes that the model wrote, in tree
+// probeModelLines returns the sizes of the nodes that the model wrote, in tree
 // order. A node that fits its source with no model call is not one of them.
-func modelLines(memory *Memory) ([]int, error) {
+func probeModelLines(memory *Memory) ([]int, error) {
 	memory.mutex.Lock()
 	defer memory.mutex.Unlock()
 
@@ -211,7 +310,7 @@ func modelLines(memory *Memory) ([]int, error) {
 			return nil, err
 		}
 
-		model, err := builtByModel(memory, key)
+		model, err := probeBuiltByModel(memory, key)
 		if err != nil {
 			return nil, err
 		}
@@ -224,32 +323,10 @@ func modelLines(memory *Memory) ([]int, error) {
 	return sizes, nil
 }
 
-// builtByModel tells if the node key needed a model call: a leaf whose source
-// is over nodeBytes, or a merge whose children do not fit in one node.
-func builtByModel(memory *Memory, key nodeKey) (bool, error) {
-	if key.level == 0 {
-		message, err := memory.log.Message(key.index)
-		if err != nil {
-			return false, err
-		}
-
-		_, free := freeLeaf(message)
-
-		return !free, nil
-	}
-
-	left, right, err := memory.children(key)
-	if err != nil {
-		return false, err
-	}
-
-	_, free := freeMerge(left, right)
-
-	return !free, nil
-}
-
-// readUsageFile reads the records of usage.jsonl at path.
-func readUsageFile(path string) (records []usageRecord, err error) {
+// probeReadUsage reads the records of usage.jsonl at path, in the order of the
+// start of their calls. The file has them in the order of the end of their
+// calls, which differs when calls run at once.
+func probeReadUsage(path string) (records []ModelUsageRecord, err error) {
 	file, err := os.Open(path)
 	if errors.Is(err, os.ErrNotExist) {
 		return nil, nil
@@ -266,7 +343,7 @@ func readUsageFile(path string) (records []usageRecord, err error) {
 	scanner := bufio.NewScanner(file)
 	scanner.Buffer(make([]byte, 64*1024), probeLineBytes)
 	for scanner.Scan() {
-		var record usageRecord
+		var record ModelUsageRecord
 		err = json.Unmarshal(scanner.Bytes(), &record)
 		if err != nil {
 			return nil, fmt.Errorf("decode usage: %w", err)
@@ -280,29 +357,15 @@ func readUsageFile(path string) (records []usageRecord, err error) {
 		return nil, fmt.Errorf("read usage file: %w", err)
 	}
 
-	return records, nil
-}
-
-// usedPercents returns the used percent of each quota window of record, by
-// header name.
-func usedPercents(record usageRecord) map[string]float64 {
-	used := make(map[string]float64)
-	for name, value := range record.Quota {
-		if !strings.HasSuffix(name, "-used-percent") {
-			continue
-		}
-
-		percent, err := strconv.ParseFloat(value, 64)
-		if err == nil {
-			used[name] = percent
-		}
-	}
+	slices.SortStableFunc(records, func(a, b ModelUsageRecord) int {
+		return probeCallStart(a).Compare(probeCallStart(b))
+	})
 
-	return used
+	return records, nil
 }
 
-// writeProbeReport writes the result of a probe run.
-func writeProbeReport(out io.Writer, appended int, stop string, records []usageRecord, sizes []int) error {
+// probeWriteReport writes the result of a probe run.
+func probeWriteReport(out io.Writer, appended int, stop string, records []ModelUsageRecord, sizes []int) error {
 	var input, cached, output, failed int
 	for _, record := range records {
 		input += record.InputTokens
@@ -320,19 +383,23 @@ func writeProbeReport(out io.Writer, appended int, stop string, records []usageR
 
 	lines = append(lines,
 		fmt.Sprintf("messages: %d", appended),
-		fmt.Sprintf("calls: %d (%d failed), %.2f per message", len(records), failed, ratio(len(records), appended)),
-		fmt.Sprintf("model lines: %d, retries: %d", len(sizes), len(records)-failed-len(sizes)),
-		fmt.Sprintf("tokens: %d input, %d cached (%.1f%%), %d output", input, cached, 100*ratio(cached, input), output),
+		fmt.Sprintf("calls: %d (%d failed), %.2f per message", len(records), failed, probeRatio(len(records), appended)),
+		fmt.Sprintf("model lines: %d, retries: %d, %.2f calls per model line", len(sizes), len(records)-failed-len(sizes), probeRatio(len(records), len(sizes))),
+		fmt.Sprintf("tokens: %d input, %d cached (%.1f%%), %d output", input, cached, 100*probeRatio(cached, input), output),
 	)
 
 	if len(sizes) > 0 {
-		lines = append(lines, sizeText(sizes))
+		lines = append(lines, probeSizeText(sizes))
+	}
+
+	if len(records) > 0 {
+		lines = append(lines, probeTimeText(records))
 	}
 
 	if len(records) > 0 {
 		lines = append(lines,
-			"quota at the first call: "+quotaText(records[0]),
-			"quota at the last call: "+quotaText(records[len(records)-1]),
+			"quota at the first call: "+probeQuotaText(records[0]),
+			"quota at the last call: "+probeQuotaText(records[len(records)-1]),
 		)
 	}
 
@@ -341,8 +408,77 @@ func writeProbeReport(out io.Writer, appended int, stop string, records []usageR
 	return err
 }
 
-// sizeText describes the sizes of the model lines.
-func sizeText(sizes []int) string {
+// probeStop returns why a probe run must stop: a failed call, or a quota window
+// whose used percent grew by more than probeQuotaLimit points since the first
+// call. It returns "" when the run can go on.
+func probeStop(records []ModelUsageRecord) string {
+	for _, record := range records {
+		if record.Error != "" {
+			return "call failed: " + record.Error
+		}
+	}
+
+	if len(records) == 0 {
+		return ""
+	}
+
+	first := probeUsedPercents(records[0])
+	for name, used := range probeUsedPercents(records[len(records)-1]) {
+		start, found := first[name]
+		if found && used-start > probeQuotaLimit {
+			return fmt.Sprintf("%s grew from %g to %g", name, start, used)
+		}
+	}
+
+	return ""
+}
+
+// probeBuiltByModel tells if the node key needed a model call: a leaf whose
+// source is over nodeBytes, or a merge whose children do not fit in one node.
+func probeBuiltByModel(memory *Memory, key NodeKey) (bool, error) {
+	if key.level == 0 {
+		message, err := memory.log.Message(key.index)
+		if err != nil {
+			return false, err
+		}
+
+		_, free := freeLeaf(message)
+
+		return !free, nil
+	}
+
+	left, right, err := memory.children(key)
+	if err != nil {
+		return false, err
+	}
+
+	_, free := freeMerge(left, right)
+
+	return !free, nil
+}
+
+// probeCallStart returns the start of the call of record, or the zero time when
+// its date is not valid.
+func probeCallStart(record ModelUsageRecord) time.Time {
+	start, err := time.Parse(time.RFC3339Nano, record.Date)
+	if err != nil {
+		return time.Time{}
+	}
+
+	return start
+}
+
+// probeRatio returns a / b, or 0 when b is 0.
+func probeRatio(a, b int) float64 {
+	if b == 0 {
+		return 0
+	}
+
+	return float64(a) / float64(b)
+}
+
+// probeSizeText describes the sizes of the model lines.
+func probeSizeText(sizes []int) string {
 	sorted := slices.Sorted(slices.Values(sizes))
 	total, over := 0, 0
 	for _, size := range sorted {
@@ -352,11 +488,60 @@ func sizeText(sizes []int) string {
 		}
 	}
 
-	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)
+	return fmt.Sprintf("line bytes: min %d, median %d, mean %.0f, max %d, %d over %d", sorted[0], sorted[len(sorted)/2], probeRatio(total, len(sorted)), sorted[len(sorted)-1], over, nodeBytes)
+}
+
+// probeTimeText describes the time per call and how many calls ran at once. A
+// call runs from its date for its seconds.
+func probeTimeText(records []ModelUsageRecord) string {
+	type event struct {
+		at    time.Time
+		delta int
+	}
+
+	var events []event
+	seconds := make([]float64, 0, len(records))
+	total := 0.0
+	for _, record := range records {
+		start, err := time.Parse(time.RFC3339Nano, record.Date)
+		if err != nil {
+			continue
+		}
+
+		length := time.Duration(record.Seconds * float64(time.Second))
+		events = append(events, event{
+			at:    start,
+			delta: 1,
+		}, event{
+			at:    start.Add(length),
+			delta: -1,
+		})
+		seconds = append(seconds, record.Seconds)
+		total += record.Seconds
+	}
+
+	if len(events) == 0 {
+		return "time: no dated calls"
+	}
+
+	slices.SortFunc(events, func(a, b event) int {
+		return cmp.Or(a.at.Compare(b.at), cmp.Compare(a.delta, b.delta))
+	})
+
+	running, most := 0, 0
+	for _, e := range events {
+		running += e.delta
+		most = max(most, running)
+	}
+
+	wall := events[len(events)-1].at.Sub(events[0].at).Seconds()
+	slices.Sort(seconds)
+
+	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))
 }
 
-// quotaText returns the quota headers of record, sorted by name.
-func quotaText(record usageRecord) string {
+// probeQuotaText returns the quota headers of record, sorted by name.
+func probeQuotaText(record ModelUsageRecord) string {
 	names := slices.Sorted(maps.Keys(record.Quota))
 	parts := make([]string, len(names))
 	for n, name := range names {
@@ -366,11 +551,20 @@ func quotaText(record usageRecord) string {
 	return strings.Join(parts, " ")
 }
 
-// ratio returns a / b, or 0 when b is 0.
-func ratio(a, b int) float64 {
-	if b == 0 {
-		return 0
+// probeUsedPercents returns the used percent of each quota window of record, by
+// header name.
+func probeUsedPercents(record ModelUsageRecord) map[string]float64 {
+	used := make(map[string]float64)
+	for name, value := range record.Quota {
+		if !strings.HasSuffix(name, "-used-percent") {
+			continue
+		}
+
+		percent, err := strconv.ParseFloat(value, 64)
+		if err == nil {
+			used[name] = percent
+		}
 	}
 
-	return float64(a) / float64(b)
+	return used
 }