Show styled replies only on terminal output

[onechat] / compactor.go

package main

import (
        "context"
        "errors"
        "fmt"
        "slices"
        "strings"
        "unicode/utf8"
)

// compactorTries is the most calls for one node.
const compactorTries = 5

// compactorBlock is the number of view lines in one block message.
const compactorBlock = 4

// compactorRuler shows the model the size limit of a line. Models cannot count
// bytes, and a sample line as the ruler gets its content copied.
var compactorRuler = strings.Repeat("-", nodeBytes)

// newCompactor returns the compactor of an agent, which sends its calls through
// client with model. instructions gives the system prompt of the agent at each
// build, so a build sends the same prompt as the turns of that time.
func newCompactor(client *ModelClient, instructions func() (string, error), model ModelConfig) *Compactor {
        return &Compactor{
                client:       client,
                model:        model,
                instructions: instructions,
                tools:        promptTools(),
        }
}

// Compactor builds the nodes of one agent with a model. A compaction is a call
// like a turn, with the same system prompt and tools, so it reads them from the
// prompt cache of the turns. The tools are never called. After them come the
// compaction view up to the node and the task, as the user messages that
// compactorTexts gives. Each whole block of compactorBlock view lines is its
// own message, and the last message holds the rest of the lines and the task.
// The backend reads a cache entry only at a message end of a request, so a
// compaction whose view begins with the same lines as an earlier one begins
// with the same block messages.
//
// When the line is over nodeBytes, the next try sends the same messages, plus
// one more user message with a note with the size of the newest line and its
// cut at the limit. A try never sends the model's earlier reply or reasoning
// back, so every try begins with the whole earlier try and has about the same
// size. After at most compactorTries tries, the shortest line is the node. A
// few bytes over the limit is fine, because the view measures the real sizes. A
// call that fails before any line fails the build, and the core tries it again
// at the next message.
type Compactor struct {
        client       *ModelClient
        model        ModelConfig
        instructions func() (string, error)
        tools        []ModelTool
}

// Summarize returns the line of the node of task.
func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
        text, err := compactorTask(task)
        if err != nil {
                return "", err
        }

        instructions, err := c.instructions()
        if err != nil {
                return "", err
        }

        var first []any

        for _, message := range compactorTexts(task.Context, text) {
                first = append(first, modelUserMessage(message))
        }

        var lines []string

        note := ""
        for len(lines) < compactorTries {
                input := first
                if note != "" {
                        input = append(slices.Clip(first), modelUserMessage(note))
                }

                var response ModelResponse

                response, err = c.client.Create(ctx, ModelRequest{
                        Model:        c.model.Model,
                        Effort:       c.model.Effort,
                        Instructions: instructions,
                        Tools:        c.tools,
                        Input:        input,
                        Node:         &NodeKey{task.Level, task.Index},
                })
                if err != nil {
                        break
                }

                line := strings.TrimSpace(response.Text)
                if line == "" {
                        err = errors.New("compaction gave no line")
                        break
                }

                lines = append(lines, line)
                if len(line) <= nodeBytes {
                        break
                }

                note = compactorTooLong(line)
        }

        if len(lines) == 0 {
                return "", err
        }

        return compactorShortest(lines), nil
}

// compactorTask returns the task of a build, with its input.
func compactorTask(task Task) (string, error) {
        if task.Level == 0 && len(task.Input) == 1 {
                return fmt.Sprintf(`Compaction: compress message %d into one line of at most %d bytes
(about 70 words), the length of this ruler:
%s
<input>
%s
</input>`, task.Index, nodeBytes, compactorRuler, task.Input[0]), nil
        }

        if task.Level > 0 && len(task.Input) == 2 {
                first := task.Index << task.Level
                half := 1 << (task.Level - 1)

                return fmt.Sprintf(`Compaction: merge lines %d+%d and %d+%d, adjacent, into one line of at most
%d bytes (about 70 words), the length of this ruler:
%s
<chat> may hold their messages, %d to %d, in more detail: take details
of them from there too.
<input>
%s
%s
</input>`, first, half, first+half, half, nodeBytes, compactorRuler, first, nodeEnd(task.Level, task.Index)-1, task.Input[0], task.Input[1]), nil
        }

        return "", fmt.Errorf("node (%d, %d) has %d inputs", task.Level, task.Index, len(task.Input))
}

// compactorTexts returns the texts of the user messages of a first try for the
// compaction view and the task text. Each whole block of compactorBlock lines,
// counted from the first line of the view, is one message. The last message
// holds the rest of the lines, the closing tag, a newline, and the task. The
// first message starts with the opening tag. Joined, the texts are the view, a
// newline, and the task.
func compactorTexts(view, task string) []string {
        body := strings.TrimSuffix(strings.TrimPrefix(view, "<chat>\n"), "</chat>")
        lines := strings.SplitAfter(body, "\n")
        lines = lines[:len(lines)-1]
        whole := len(lines) - len(lines)%compactorBlock

        var texts []string

        for n := 0; n < whole; n += compactorBlock {
                texts = append(texts, viewMessage(n == 0, lines[n:n+compactorBlock]))
        }

        return append(texts, viewMessage(whole == 0, lines[whole:])+"</chat>\n"+task)
}

// compactorTooLong returns the note for a line over nodeBytes: its size and its
// first nodeBytes bytes, which show the model where the limit cuts it.
func compactorTooLong(line string) string {
        return fmt.Sprintf(`Too long: your last line for this <input> was %d bytes,
over the %d-byte limit. Write the whole line again, cutting just
enough of the least valuable items to fit before this cut:
%s| <- LIMIT`, len(line), nodeBytes, compactorCutLine(line))
}

// compactorShortest returns the first of the compactorShortest lines.
func compactorShortest(lines []string) string {
        best := lines[0]
        for _, line := range lines[1:] {
                if len(line) < len(best) {
                        best = line
                }
        }

        return best
}

// compactorCutLine returns the first nodeBytes bytes of line, without a part of
// a character at the end.
func compactorCutLine(line string) string {
        if len(line) <= nodeBytes {
                return line
        }

        cut := line[:nodeBytes]
        for !utf8.ValidString(cut) {
                cut = cut[:len(cut)-1]
        }

        return cut
}