X-Git-Url: http://git.giacomolli.uk/onechat/blobdiff_plain/baf7a940be8927278e3b4b87e7366cebfc16532f..0a98ccb67c3df55c0fe6d76b45afcbaa58c2eda1:/compactor.go

diff --git a/compactor.go b/compactor.go
index 5e73cfa..e7a6866 100644
--- a/compactor.go
+++ b/compactor.go
@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"slices"
 	"strings"
 	"unicode/utf8"
 )
@@ -11,6 +12,9 @@ import (
 // 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)
@@ -29,16 +33,22 @@ func newCompactor(client *ModelClient, instructions func() (string, error), mode
 
 // 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.
+// 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 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 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.
+// 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
@@ -48,7 +58,6 @@ type Compactor struct {
 
 // 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 := compactorTask(task)
 	if err != nil {
 		return "", err
@@ -59,12 +68,17 @@ func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
 		return "", fmt.Errorf("compact node (%d, %d): %w", task.Level, task.Index, err)
 	}
 
+	var first []any
+	for _, message := range compactorTexts(task.Context, text) {
+		first = append(first, modelUserMessage(message))
+	}
+
 	var lines []string
 	note := ""
 	for len(lines) < compactorTries {
-		parts := []string{view, text}
+		input := first
 		if note != "" {
-			parts = append(parts, note)
+			input = append(slices.Clip(first), modelUserMessage(note))
 		}
 
 		var response ModelResponse
@@ -73,7 +87,7 @@ func (c *Compactor) Summarize(ctx context.Context, task Task) (string, error) {
 			Effort:       c.model.Effort,
 			Instructions: instructions,
 			Tools:        c.tools,
-			Input:        []any{modelUserMessage(parts...)},
+			Input:        input,
 			Node:         &NodeKey{task.Level, task.Index},
 		})
 		if err != nil {
@@ -130,6 +144,26 @@ of them from there too.
 	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 {