X-Git-Url: https://git.giacomolli.uk/onechat/blobdiff_plain/4a605fa39078354654ae648e489c117c7a445197..900ec0e7dad7a913190fcef9016e53495de9bdf8:/compactor.go?ds=sidebyside

diff --git a/compactor.go b/compactor.go
index e899c09..efb183f 100644
--- a/compactor.go
+++ b/compactor.go
@@ -4,77 +4,91 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"slices"
 	"strings"
 	"unicode/utf8"
 )
 
-// compactionTries is the most calls for one node.
-const compactionTries = 5
+// compactorTries is the most calls for one node.
+const compactorTries = 5
 
-// ruler 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 ruler = strings.Repeat("-", nodeBytes)
+// compactorBlock is the number of view lines in one block message.
+const compactorBlock = 4
 
-// 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 comes one
-// user message with the compaction view up to the node and the task.
-//
-// When the line is over nodeBytes, the next try sends the same view and task,
-// plus 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 has the
-// same cached prefix and about the same size. After at most compactionTries
-// 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       *Client
-	model        ModelConfig
-	instructions func() (string, error)
-	tools        []Tool
-}
+// 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 *Client, instructions func() (string, error), model ModelConfig) *Compactor {
+func newCompactor(client *ModelClient, instructions func() (string, error), model ModelConfig) *Compactor {
 	return &Compactor{
 		client:       client,
 		model:        model,
 		instructions: instructions,
-		tools:        agentTools(),
+		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) {
-	view := task.Context + "\n"
-	text, err := compactionTask(task)
+	text, err := compactorTask(task)
 	if err != nil {
 		return "", err
 	}
 
 	instructions, err := c.instructions()
 	if err != nil {
-		return "", fmt.Errorf("compact node (%d, %d): %w", task.Level, task.Index, err)
+		return "", err
+	}
+
+	var first []any
+	for _, message := range compactorTexts(task.Context, text) {
+		first = append(first, modelUserMessage(message))
 	}
 
 	var lines []string
 	note := ""
-	for len(lines) < compactionTries {
-		parts := []string{view, text}
+	for len(lines) < compactorTries {
+		input := first
 		if note != "" {
-			parts = append(parts, note)
+			input = append(slices.Clip(first), modelUserMessage(note))
 		}
 
-		var response Response
-		response, err = c.client.Create(ctx, Request{
+		var response ModelResponse
+		response, err = c.client.Create(ctx, ModelRequest{
 			Model:        c.model.Model,
 			Effort:       c.model.Effort,
 			Instructions: instructions,
 			Tools:        c.tools,
-			Input:        []any{userMessage(parts...)},
-			Node:         &nodeKey{task.Level, task.Index},
+			Input:        input,
+			Node:         &NodeKey{task.Level, task.Index},
 		})
 		if err != nil {
 			break
@@ -91,25 +105,25 @@ func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
 			break
 		}
 
-		note = tooLong(line)
+		note = compactorTooLong(line)
 	}
 
 	if len(lines) == 0 {
-		return "", fmt.Errorf("compact node (%d, %d): %w", task.Level, task.Index, err)
+		return "", err
 	}
 
-	return shortest(lines), nil
+	return compactorShortest(lines), nil
 }
 
-// compactionTask returns the task of a build, with its input.
-func compactionTask(task Task) (string, error) {
+// 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, ruler, task.Input[0]), nil
+</input>`, task.Index, nodeBytes, compactorRuler, task.Input[0]), nil
 	}
 
 	if task.Level > 0 && len(task.Input) == 2 {
@@ -124,24 +138,56 @@ of them from there too.
 <input>
 %s
 %s
-</input>`, first, half, first+half, half, nodeBytes, ruler, first, nodeEnd(task.Level, task.Index)-1, task.Input[0], task.Input[1]), nil
+</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))
 }
 
-// tooLong 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 tooLong(line string) string {
+// 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, cutLine(line))
+%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
 }
 
-// cutLine returns the first nodeBytes bytes of line, without a part of a
-// character at the end.
-func cutLine(line string) string {
+// 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
 	}
@@ -153,15 +199,3 @@ func cutLine(line string) string {
 
 	return cut
 }
-
-// shortest returns the first of the shortest lines.
-func shortest(lines []string) string {
-	best := lines[0]
-	for _, line := range lines[1:] {
-		if len(line) < len(best) {
-			best = line
-		}
-	}
-
-	return best
-}