package main

import (
	"bufio"
	"cmp"
	"context"
	"encoding/json"
	"errors"
	"fmt"
	"io"
	"maps"
	"os"
	"slices"
	"strconv"
	"strings"
	"time"
)

// buildLineBytes is the most bytes of one line in a build input or usage log.
const buildLineBytes = 1 << 20

// buildQuotaLimit is the largest increase in used quota points of a run.
const buildQuotaLimit = 15.0

// BuildRun appends messages, builds their tree, and stops the run by the rules
// of buildStop.
type BuildRun 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 *BuildRun) Run(ctx context.Context, entries []DatedMessage) (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
}

// append appends the entries and stops at the first stop rule.
func (r *BuildRun) append(ctx context.Context, entries []DatedMessage) (int, string) {
	for n, entry := range entries {
		if entry.Kind == "user" {
			reason := r.wait(ctx)
			if reason != "" {
				return n, reason
			}
		}

		if ctx.Err() != nil {
			return n, ctx.Err().Error()
		}

		_, err := r.memory.AppendDated(entry.Kind, entry.Text, entry.Date)
		if err != nil {
			return n, err.Error()
		}

		reason := r.check()
		if reason != "" {
			return n + 1, reason
		}
	}

	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 *BuildRun) wait(ctx context.Context) string {
	for {
		reason := r.check()
		if reason != "" {
			return reason
		}

		err := r.waitBuilds(ctx)
		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 *BuildRun) check() string {
	if r.client.LimitReached() {
		return fmt.Sprintf("call limit of %d reached", r.client.CallLimit())
	}

	records, err := buildReadUsage(r.usagePath)
	if err != nil {
		return err.Error()
	}

	return buildStop(records)
}

// waitBuilds waits up to one second without leaving a child context open.
func (r *BuildRun) waitBuilds(ctx context.Context) error {
	ctx, cancel := context.WithTimeout(ctx, time.Second)
	defer cancel()

	return r.memory.Wait(ctx)
}

// DatedMessage is one message with an optional RFC 3339 date.
type DatedMessage struct {
	Kind string `json:"kind"`
	Text string `json:"text"`
	Date string `json:"date,omitempty"`
}

// buildReadUsage 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 buildReadUsage(path string) (records []ModelUsageRecord, err error) {
	file, err := os.Open(path)
	if errors.Is(err, os.ErrNotExist) {
		return nil, nil
	}

	if err != nil {
		return nil, fmt.Errorf("open usage file: %w", err)
	}

	defer func() {
		closeErr := file.Close()
		if closeErr != nil {
			err = errors.Join(err, fmt.Errorf("close build usage: %w", closeErr))
		}
	}()

	scanner := bufio.NewScanner(file)
	scanner.Buffer(make([]byte, 64*1024), buildLineBytes)

	for scanner.Scan() {
		var record ModelUsageRecord

		err = json.Unmarshal(scanner.Bytes(), &record)
		if err != nil {
			return nil, fmt.Errorf("decode usage: %w", err)
		}

		records = append(records, record)
	}

	err = scanner.Err()
	if err != nil {
		return nil, fmt.Errorf("read usage file: %w", err)
	}

	slices.SortStableFunc(records, func(a, b ModelUsageRecord) int {
		return buildCallStart(a).Compare(buildCallStart(b))
	})

	return records, nil
}

// buildWriteReport writes the result of a build run.
func buildWriteReport(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
		cached += record.CachedTokens
		output += record.OutputTokens
		if record.Error != "" {
			failed++
		}
	}

	var lines []string

	if stop != "" {
		lines = append(lines, "stopped early: "+stop)
	}

	lines = append(lines,
		fmt.Sprintf("messages: %d", appended),
		fmt.Sprintf("calls: %d (%d failed), %.2f per message", len(records), failed, buildRatio(len(records), appended)),
		fmt.Sprintf("model lines: %d, retries: %d, %.2f calls per model line", len(sizes), len(records)-failed-len(sizes), buildRatio(len(records), len(sizes))),
		fmt.Sprintf("tokens: %d input, %d cached (%.1f%%), %d output", input, cached, 100*buildRatio(cached, input), output),
	)

	if len(sizes) > 0 {
		lines = append(lines, buildSizeText(sizes))
	}

	if len(records) > 0 {
		lines = append(lines, buildTimeText(records))
	}

	if len(records) > 0 {
		lines = append(lines,
			"quota at the first call: "+buildQuotaText(records[0]),
			"quota at the last call: "+buildQuotaText(records[len(records)-1]),
		)
	}

	_, err := io.WriteString(out, strings.Join(lines, "\n")+"\n")
	if err != nil {
		return fmt.Errorf("write build report: %w", err)
	}

	return nil
}

// buildStop returns why a build run must stop: a failed call, or a quota window
// whose used percent grew by more than buildQuotaLimit points since the first
// call. It returns "" when the run can go on.
func buildStop(records []ModelUsageRecord) string {
	for _, record := range records {
		if record.Error != "" {
			return "call failed: " + record.Error
		}
	}

	if len(records) == 0 {
		return ""
	}

	first := buildUsedPercents(records[0])
	for name, used := range buildUsedPercents(records[len(records)-1]) {
		start, found := first[name]
		if found && used-start > buildQuotaLimit {
			return fmt.Sprintf("%s grew from %g to %g", name, start, used)
		}
	}

	return ""
}

// buildCallStart returns the start of the call of record, or the zero time when
// its date is not valid.
func buildCallStart(record ModelUsageRecord) time.Time {
	start, err := time.Parse(time.RFC3339Nano, record.Date)
	if err != nil {
		return time.Time{}
	}

	return start
}

// buildRatio returns a / b, or 0 when b is 0.
func buildRatio(a, b int) float64 {
	if b == 0 {
		return 0
	}

	return float64(a) / float64(b)
}

// buildSizeText describes the sizes of the model lines.
func buildSizeText(sizes []int) string {
	sorted := slices.Sorted(slices.Values(sizes))
	total, over := 0, 0
	for _, size := range sorted {
		total += size
		if size > nodeBytes {
			over++
		}
	}

	return fmt.Sprintf("line bytes: min %d, median %d, mean %.0f, max %d, %d over %d", sorted[0], sorted[len(sorted)/2], buildRatio(total, len(sorted)), sorted[len(sorted)-1], over, nodeBytes)
}

// buildTimeText describes the time per call and how many calls ran at once. A
// call runs from its date for its seconds.
func buildTimeText(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))
}

// buildQuotaText returns the quota headers of record, sorted by name.
func buildQuotaText(record ModelUsageRecord) string {
	names := slices.Sorted(maps.Keys(record.Quota))
	parts := make([]string, len(names))
	for n, name := range names {
		parts[n] = strings.TrimPrefix(name, "x-codex-") + "=" + record.Quota[name]
	}

	return strings.Join(parts, " ")
}

// buildUsedPercents returns the used percent of each quota window of record, by
// header name.
func buildUsedPercents(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 used
}